H.l (lateral hypothalamus) is a key node in the foraging system and has an interesting capability of distinguishing a food zone from a non-food zone [Jennings et al 2015]. In a sense foraging is searching for a food zone and then eating.
Foraging as a state machine.
The above diagram shows the foraging phases that I’ve already covered in earlier essays. Importantly, each phase is an independent action path as part of a distributed system, not merely a state in a state machine. To force the separate action paths to act like a state machine, each transition needs to suppress the preceding and following state. In particular the eating phase needs to inhibit the seeking system. This lateral inhibition is important because circuitry is required to force activation of only a single system at a time.
The food zone is particularly interesting for filter feeding, which is naturally area based and long term, as opposed to snapping up a morsel of food. Non-vertebrate chordates are filter feeders, lamprey larvae are filter feeders, and early jawless vertebrates were also likely filter feeders [D’Aniello et al 2023]. Tunicate ascidians, the closest non-vertebrate chordates, have an extreme version of this foraging loop, where the tadpoles find a feeding place after swimming for 12 hours and then settling in place for their adult life [Anselmi et al 2024]. The ascidian foraging state marine is a straight line that ends in the eating phase in the food zone, not continuing in a loop. The ascidian search and settle might give a hint how the vertebrate foraging circuitry is organized.
Ascidians
As covered in essay 30, the ascidian larva nervous system has several seeking (taxis) systems: geotaxis (gravity avoidance – moving up), phototaxis (light avoidance), and dimming for predator and obstacle avoidance. Ascidian navigation disperses the larva from its parent and prefers to settle on the underside of ledges by avoiding gravity while avoiding light. Its settling sensors also avoid toxic or irritating areas and may try to find food-friendly areas, although the specific sensor capabilities aren’t well known. When the larva finds an appropriate place, around 14 hours after hatching, it settles for life [Hoyer et al 2024].
Functional organization of the ascidian larva navigation and settling circuit.
The above diagram is a functional representation of the ascidian larva navigation brain. For this essay the important part is the palp and food-zone sensor and the settling neurons that inhibit motor neurons. The palms are three tentacle-like protrusions from the larva head, which attach the ascidian to a rock with cement glands [Johnson et al 2024]. They contain chemosensory and mechanosensors that distinguish the settling zone from non-settling zones [Hoyer et al 2024]. Interestingly, the genetic markers for the palp neurons are similar to markers for the vertebrate forebrain.
Head cement glands still exist in some fish larvae [Pottin et al 2010] and most frog tadpoles [Nokhbatolfoghahai and Downie 2005], [Rétaux and Pottin 2011], [Sive and Bradley 1996]. Frog tadpoles will swim up and attach to the underside of leaves or to the water surface. This cement gland and settling system may have existed in the pre-vertebrate ancestor and shared for tunicates and vertebrates. Unlike the ascidians the pre-vertebrates likely did not permanently settle. For the sake of this essay, let’s assume they temporarily settled to filter feed in a location and only moved on if filter feeding was unsuccessful or if forced to move by predators, competitors, or environmental hazards.
Ascidian larva navigation and palp settling circuit with the settling circuit highlighted. Each of the boxes represents a single neuron or a small (5-10) group of neurons. Labels are neuron names.
The above diagram shows specific neurons in the ascidian larva brain. The importance here is the glutamate pnIN (palp interneuron) to GABA pnRN (palp relay neuron), which inhibits all motor neurons and interneurons. Comparing vertebrate and ascidian neural systems is sketchy and probably should be avoided because both have diverged [Holland 2016]. For this essay, I’ll ignore that sound advice to try to motivate part of the vertebrate nervous system.
H.stn as an analogous node to the settling neurons. H.stn (subthalamic nucleus), MLR (midbrain locomotor region), Ob (olfactory bulb), OT (optic tectum), P.v (ventral pallidum), R.rs (reticulospinal motor command), S.v (ventral striatum), S.nr (substantia nigra pars reticulata), V.pt (posterior tubuculum)
The above diagram shows the H.stn (subthalamic nucleus) as fulfilling a similar role as the pnIN from ascidian Ciona, suppressing seek in preparation for eating. Part of P.v (ventral pallidum) suppresses S.v (ventral striatum) during eating [Vachez et al 2021]. This P.v “arkypallidal” subset is named after similar neurons in P.ge (globus pallidus) that suppresses S.d (dorsal striatum). Although the driver of this eating suppression isn’t known, the timing of the arkypallidal activation closely matches V.dr serotonin food activation [Spring and Nautiyal 2024], ramping at the end of seek and peaking after eating. Also, H.stn and P.ge form an oscillating pair, evident in Parkinson’s disease. So, it’s plausible that H.stn drives persistent suppression of the seek path in S.v through its projection to P.v, possibly influenced or driven by V.dr (dorsal raphe, serotonin). This specific path is speculation but seems compatible with experiments. The second suppression path is the well-known H.stn to S.nr (substantia nigra pars reticulata) that suppresses motor activity. Snr has a widespread suppression or MLR (midbrain locomotor region), R.rs (reticulospinal motor command), and Snr suppresses Snc (substantia nigra parsa compacta dopamine). Note that the medial H.stn, the area connected with P.v, merges with H.l with minimal boundary [Haynes and Haber 2013].
Food zone
Let’s return to the H.l food zone in [Jennings et al 2015] and consider where the food zone information might come from. Following [Jacobs 2012], let’s treat olfaction as the central sense for navigation, which is particularly compelling for food zones.
The diagram below shows the H.l main connectivity. Not displayed is the H.l internal sensing of nutrient information peptides like glucose sensing and leptin fat sensing. H.l doesn’t receive direct sensory input with the exception of R.pb (parabrachial nucleus), which sends nociceptive information like itch or pain. Because an itchy or painful place is a poor choice for filter feeding, this R.pb input is negative place information for a filter-feeding zone, but R.pb doesn’t give positive reasons to stay like food odors.
H.l connectivity encompasses much of the limbic system, driven by olfactory information. A.bl (basolateral amygdala), E.hc (hippocampus), F.pfc (prefrontal cortex), H.arc (hypothalamus arcuate), H.l (lateral hypothalamus), H.pv (paraventricular hypothalamus), H.stn (subthalamic nucleus), Hb.l (lateral habenula), M.pag (periaqueductal gray), Ob (olfactory bulb), O.pir (piriform cortex), P.bst (bed nucleus of the stria terminalis), P.v (ventral pallidum), R.pb (parabrachial), S.a (central amygdala), S.ls (lateral striatum), S.v (ventral striatum), V.dr (dorsal raphe – serotonin), Vta (ventral tegmental area – dopamine)
As the diagram suggests, the information H.l receives about food sources is very abstract. It receives cue information from A.bl (basolateral amygdala), place information from E.hc (hippocampal complex), value-like information from F.ofc (orbitofrontal cortex) and task-like information from F.vm (ventromedial prefrontal cortex). All of those areas are strongly connected with the olfactory system. While H.l doesn’t receive odor place information directly from sensors, it receives multiple organizational perspectives on odor information. P.bst (bed nucleus of the stria terminalis) receives very similar olfactory input as H.l, and it also receives negative information from R.pb. However, R.pb sends different nociceptive information to the S.a (central amygdala)/P.bst extended amygdala than it sends to H.l [Arthurs et al 2023]. The R.pb projections to H.l compared to S.a/P.bst are not redundant.
Not only are the H.l inputs abstract, but the outputs are also abstract, in contrast to direct action paths. This abstraction might be a later evolutionary development, similar to V.pt (posterior tuberculum) in zebrafish. V.pt is roughly homologous to Vta (ventral tegmental area) in mammals, but V.pt has more direct locomotor output to MLR (midbrain locomotor region), while most of Vta’s output is generally abstract.
As a note, the diagram does not include H.l ox (orexin) or H.l mch (melanin-concentrating hormone), partially for simplicity and partially because the zebrafish H.l is distinct from the ox and mch populations, suggesting that the mammalian ox and mch areas of H.l can be separated from the rest of H.l function. The diagram also omits some other connections like Ppt (pedunculopontine nucleus).
Food and serotonin
Returning to the foraging state diagram, it’s important that each “state” is a large, distributed, complex system, not a state in a state machine. The seek state includes areas like S.v, Vta, H.l, E.hc, F.pfc, and the motor regions MLR and R.rs (reticulospinal motor command) with the help of cortical areas and can include OT (optic tectum). Although the eating state is small, it is still comprised of many areas, including V.dr (dorsal raphe), OT.d, R.my.irt (medulla eating), H.l, H.pstn (parasubthalamic nucleus), R.pb and possibly some Vta and S.v subareas. Although the system is not a state machine, each “state” needs to laterally suppress the other systems to prevent multiple action paths from colliding.
Foraging state machine with dopamine and serotonin modulation. DA (dopamine), V.dr (dorsal raphe), Vta ventral tegmental area, 5HT (serotonin).
The split between eat and seek is important, because many studies merge the behavior into a general category “feeding.” Because some experiments only measure total feeding, it can be difficult to distinguish whether the experiment is measuring a seek effect or an eating effect. For example, eating needs to suppress seek to keep the animal from wandering away from the food. If an experiment stimulates eat but inhibits seek, the animal might not search for food even if it’s ready to eat. If it doesn’t seek food, it doesn’t find food.
This distinction between eating and seeking is exhibited by the question of serotonin, which is a heterogeneous system that has a role in feeding. The serotonin from V.dr is a heterogenous system with V.dr having at least 14 different genetic clusters [Okaty et al 2020] with at least 11 different projection patterns [Ren et al 2014]. Earlier studies noted that 30% of V.dr were active during eating [Fornal et al 1996], and many others have noted V.dr being active for “reward” (eating).
Suppose one component of V.dr serotonin encourages eating while discouraging seeking. If an experiment floods the brain with serotonin, it might see total feeding drop because serotonin suppresses seeking food, even if it encourages long meals when it finds food. The confusion becomes greater for studies looking for the even more abstract “reward” as opposed to concrete eating. The point being that serotonin in particular is a complicated system, not reducible to a single value or function.
Eating related effects of serotonin. DA (dopamine), H.arc (hypothalamus arcuate), H.stn (subthalamic nucleus), P.v (ventral pallidum), S.nr (substantia nigra pars reticulata), V.dr (dorsal raphe), Vta (ventral tegmental area), Vta.g (GABA neurons of Vta), 5HT (serotonin)
The above diagram shows some of the eating-related projections. Only a few of the 14 V.dr subtypes are know. The V.dr to Vta connection is one of the known projections and drives the seek system [Courtiol et al 2021], [Wang HL et al 2019]. Unfortunately, the other projections are not known, in particular the 30% of V.dr that is active while eating [Bromberg-Martin et al 2010].
V.dr enhances satiety with 5HT2c.q (serotonin G-q stimulating receptor) in H.arc POME satiety neurons, which suppresses the AgRP hunger peptide. Note that AgRP drops just before eating, suggesting that it’s a seek-promoting system, but an eating-promoting system [Bhave and Nettow 2021]. The prediction suppression only occurs after training and V.dr serotonin shows inverse behavior, possibly suggesting V.dr as suppressing H.arc. Untrained V.dr serotonin only responds after tasting [Li et al 2016], but trained V.dr serotonin responds about 2 seconds before eating [Zhong et al 2016].
Filter feeding and foraging theory
Let’s the consider filter feeding using foraging theory. Foraging theory studies how animals browse patches of food, such as a cluster of flowers for a bee or worms in pine cones for birds [Krebs et al 1974] or a hunting spot for a predator. In particular, foraging theory considers how long the animal should stay at a particular patch before deciding to move on: measuring the give up time. A filter-feeding proto-vertebrate needs to decide if the current food rate is good enough to stay at the current food zone.
The MVT (marginal value theorem) suggests that an animal should move on if the current patch has less food than the environment average [Charnov 1976]. MVT has simplifying assumptions that are challenged by the complexity in the world [Pyke 1984], [Wajnberg et al 2006]. MVT assumptions include omniscience, immortality, determinism, no competition, no predation, and no hunger. Some of those complexities are important to the essay, particularly the omniscience. In MVT the animal knows the average environment food value, but this omniscience isn’t plausible for simple animals [Tenhumberg et al 2001], and the essay animal has almost no learning at all. Realistic search is stochastic and can fail, such as a predator hunting, which is particularly important if the animal is starving. Starvation and satiation are also not covered by the MVT. If the animal is starving, it might stick with a non-optimal, low quality food source below the environment average because not finding a better patch is too risky. Simple organisms use rules of thumb instead of complex strategy, and even birds seem to use a constant give up time [Krebs et al 1974].
As a side note, the foraging terms for eating (“exploiting”) and searching for a new patch (“exploring”) have been appropriated by RL (reinforcement learning) [Sutton and Barto 2018] with some differences in meaning. Reinforcement learning use an n-armed bandit (gambling slot machine) model, where exploring means finding the reward rates of the other arms before deciding on the best arm to exploit. The RL focus is on gather information, generally in a finite and persistent system. In contrast, this essay uses the original foraging terminology.
Covered in essay 36, vertebrate food motivation divides into hunger-driven (“homeostatic”) and opportunistic (“hedonic”) foraging. These form two levels of search and involve different circuits with some overlap. When no longer hungry, mice will not eat plain food but will still eat rich food. In terms of foraging theory, hungry mice will stay longer at poor patches, while sated mice will leave more quickly.
Simulation complexity
After starting to implement the simulation, the issue of complication became overwhelming. Specifically, adding the striatum is too complicated. Consider the issue of distinguishing the eating function of dopamine vs serotonin, when both are responsive to eating food. That similarity makes it difficult to find the system function. The system must have developed from a simpler system because the ascidian feeding or amphioxus feeding is not overly complicated. For the sake of the simulation, I’m backing off and considering only the hindbrain and hypothalamus systems, treating the striatum as a later enhancement.
Hypothalamus and raphe nuclei
The core of the simulation is the pair of H.l and V.dr. As mentioned above, H.l is driven by food zone indicators and can drive both seeking and eating. V.dr is responsive to eating and as part of the hindbrain (it derives from r1) it is a good candidate for primitive, tunicate-like filter feeding circuitry.
Simulation eating model. Ob and H.l form the forebrain food zone system, while V.dr and R.nts form the hindbrain eating system. H.l (lateral hypothalamus), Ob (olfactory bulb), R.nts (nucleus of the solitary tract), V.dr (dorsal raphe).
The diagram above is a simplification, where the Ob to H.l connection represents an ancient version of the food zone system. The V.dr to R.nts (nucleus of the solitary tract) connection includes more hindbrain structures such as medulla eating circuits. The simplification has H.l as a food zone controller and V.dr as an eating sustaining manager.
Although V.dr is a serotonin system, not V.dr neurons are non-serotonin, both glutamate and GABA. As mentioned above the V.dr and V.mr (median raphe) serotonin neurons have at least 11-14 distinct neuron types and projection types. For the essay I’m assuming at least one serotonin neuron type is a measure of eating food. In the simulation successful filter feeding increases the serotonin for eating.
Start and sustain
Let’s return to foraging, where the central decision is when to stop exploiting a patch if it’s not effective. Consider a simple where the animal gives up on a patch if the feeding rate drops below a fixed threshold. Filter feeding naturally has delays between starting filter feeding, trapping some prey, and later receiving nutrients in the gut. This raises a problem: the feeding rate is zero until some food is digested, which implies the animal should give up immediately.
Foraging give-up occurs when the combination of a start signal and sustain signal drop below a threshold.
One solution is to prime the system with a start signal. While the start signal exists, the animal won’t leave even if it hasn’t digested any nutrients. In the simulation H.l is responsible for the start signal and V.dr is responsible for both the sustain and for integrating the two systems. The H.l start signal comes from the food zone detection.
However, the start signal raises a new issue because the start signal must stop to allow sustain to act as the primary decision variable. If H.l always sends the food zone signal to V.dr, it will remain active as long as the animal is in the food zone, preventing the animal from leaving the zone. So, H.l itself needs a timeout. The simulation uses a striatum timeout to disable the H.l food zone signal. The striatum connection can either represent the striatum layer between the olfactory and cortical layers and H.l, or it can represent H.l reciprocal input to the striatum.
The start timeout has the same issues as other striatum systems. Specifically, it needs to remain timed out until the animal leaves the food zone.
Simulation
The screenshot below shows the animal feeding from a low-quality food zone. The grey star is a food zone (grey represents poor food). The nearby purple checkerboard is an avoidance zone, representing an aversive area such as itch or high carbon dioxide.
Simulation of the animal filter feeding at a poor food zone just before giving up.
In the screenshot the startup signal from H.l is temporarily sustaining feeding. It will soon timeout and the animal will abandon the food zone.
Avoidance response and search
The simulation adds two other serotonin-based systems: one for avoiding toxic areas and one for search. Avoidance is one of the V.mr functions. The search serotonin represents the V.dr to Vta connection, despite the current essay disabling the seek function. These two functions may not be serotonin functions because V.mr avoidance is largely non-serotonin, and the V.dr to Vta connection is primarily glutamate. Because the avoidance and search are not the primary focus of the essay, I’m putting off the question of accuracy to a later essay.
Discussion
The essay’s big questionable decision is the omission of the striatum, particularly because I’ve already used the striatum for give-up timing. For eating as opposed to seeking, one possible area appears to be S.dl.vl, which is the orobranchial, mouth area [Foster et al 2021]. Because S.dl receives late dopamine from food in the gut, it might be a good candidate for filter feeding sustain.
Map of the striatum. dl (dorsal lateral striatum), dm (dorsal medial striatum), lsh (lateral shell), msh.d (dorsal medial shell), msh.v (ventral medial shell), ot (olfactory tubercle)
A second area is S.msh.d (dorsal medial shell) which responds to hedonic “liking” and drives strong eating [Castro et al 2016], [Richard and Berridge 2011], [Richard et al 2013]. S.msh.d drives H.l, which is central to the essay. In addition S.msh has longer, sustained dopamine (5-10s) contrasted with shorter dopamine in S.dl (100ms) [de Jong et al 2022].
From a motivational perspective, S.dl.vm and S.msh.d are strong candidates, but they lack the lateral inhibition of seek that’s necessary for the state machine to work. S.dl.vl also works through OT.d.l (optic tectum deep motor areas), which would add more complexity to this essay. In contrast the V.dr serotonin is already part of the hindbrain motor areas, and serotonin is already inhibitory toward seek. V.dr requires fewer additional systems to work. For future work, the two striatum areas are strong areas to research.
Foster NN, Barry J, Korobkova L, Garcia L, Gao L, Becerra M, Sherafat Y, Peng B, Li X, Choi JH, Gou L, Zingg B, Azam S, Lo D, Khanjani N, Zhang B, Stanis J, Bowman I, Cotter K, Cao C, Yamashita S, Tugangui A, Li A, Jiang T, Jia X, Feng Z, Aquino S, Mun HS, Zhu M, Santarelli A, Benavidez NL, Song M, Dan G, Fayzullina M, Ustrell S, Boesen T, Johnson DL, Xu H, Bienkowski MS, Yang XW, Gong H, Levine MS, Wickersham I, Luo Q, Hahn JD, Lim BK, Zhang LI, Cepeda C, Hintiryan H, Dong HW. The mouse cortico-basal ganglia-thalamic network. Nature. 2021 Oct;598(7879):188-194.
Let’s revisit the striatum timeout from essay 31: striatum LTD, where food seeking used the striatum as a timeout to avoid perseveration. Without the timeout, the animal continues to seek toward the odor source even if the food was missing. This essay adds to the timeout by adding an odor context as a cached set of locations to avoid until the timeout, as opposed to avoiding all locations until the timeout. This odor neighborhood resembles the olfactory spatial hypothesis [Jacobs 2012], which considers olfaction as primarily a navigation sense. The added specificity to failed-seek avoidance improves search for other nearby food sources.
Recap of essay 31: striatum LTD
The food seek logic from essay 31 has two search states: a general roaming search and an odor seek. If the odor seek times out, the animal avoids the current area to prevent perseveration. Essay 35 on hippocampal sequences explored using a sequence to specify the avoidance timeout.
Foraging state machine with two search modes: a general roaming search and a target-specific seek.
For the timeout, the essay uses the Sv (ventral striatum aka nucleus accumbens) to suppress failed food seeking [Lafferty et al 2020]. Without the S.v timeout, the animal perseverates at the seek task and gets stuck in the center of the odor plume.
Circuit of S.v for timing out a failed food seek. Adenosine drives a ramping timeout signal that reduces motivation by switching from the seek path via V.pt to the avoidance path via Hb.l. Ad (adenosine), Hb (habenula), Ob (olfactory bulb), Pv (ventral pallidum), S.d1 (striatum projection neuron with D1 dopamine receptor), S.d2 (striatum projection neuron with D2 projection receptor), V.pt (posterior tuberculum – Vta/Snc)
The above diagram shows the essay 31 circuit, largely based on the lamprey. V.pt (posterior tuberculum) is a locomotion hub that receives a direct signal from Ob.m (medial olfactory bulb) and drives downstream motor areas [Derjean et al 2012]. Hb.l (lateral habenula) drives place avoidance. S.v (ventral striatum) drives the timeout selection in Pv (ventral pallidum). Ad (adenosine) is the timeout variable, which increases as neural activity in S.d1 (striatum D1 projection neuron) and S.d2 (striatum D2 projection neuron) continues. Adenosine is a byproduct of ATP (adenosine triphosphate) energy production, and is also a gliotransmitter from astrocytes that monitor synapse activity. Essentially, the Sv subcircuit in red acts as a timeout for the main seek circuit.
Importantly, because the essay 31 timeout only uses the seek odor itself as a key, it can’t distinguish spatially distinct odors, such as different flowers for a honeybee.
Neighborhood odor as context
Because essay 31 only used the Ob seek odor as a signal, a timeout of that odor locks out all food search for that odor. That lockout may be long because the S.d2 LTD (long term depression) recovery time is on the order of 20 to 60 minutes. Consider an analogy to a bee searching a field of flowers for nectar. If one flower is missing nectar, the bee should give up on that flower, but it shouldn’t abandon the entire task until a 60 minute timer expires.
Odor neighborhoods with food odor plumes. Each colored area is an odor neighborhood and each cloud is an odor plume. Only the starred areas contain food.
In the above diagram, the stars represent food locations and the clouds represent food odor plumes. Odor plumes without food are false odors. The colors of the regions represent odor neighborhoods, where non-food odors distinguish the areas. Suppose the animal first searches in the dark orange area and fails to find food. If it next reaches the green area with the star, the timeout from the failed orange search will block the search unless the timeout is specific to the orange neighborhood.
Olfactory spatial hypothesis
The olfactory spatial hypothesis argues that a primary function for olfaction is navigation, as opposed to simply proving identification [Jacobs 2012]. This navigation-centric idea is fleshed out in the parallel map theory, which argues that the hippocampus is primarily organized around two maps: a bearing map using gradients to distant odor landmarks, and a sketch map with local landmark cues [Jacobs and Schenk 2003]. The parallel map theory associates the distant bearing map with E.dg (dentate gyrus of the hippocampus) and the local sketch map with E.ca1 (CA1 region of the hippocampus).
The current essay uses the broad idea of the olfactory spatial hypothesis and the idea of a local olfactory neighborhood. The olfactory neighborhood provides a context to restrict the striatum timeout. Functionally it resembles the local sketch map, but it’s not strictly speaking a map, only a cache of failed locations.
Lamprey dual odor path
The lamprey is a useful animal model because it represents the older jawless vertebrates that preceded the development of the jaw and the majority of more complex vertebrates and because it has a simpler brain. In the lamprey, Ob.m directly drives locomotion via V.pt (posterior tuberculum), which is homologous to the mammalian midbrain dopamine areas Vta (ventral tegmental area) and Snc (substantia nigra pars compacta). Unlike the mammalian dopamine areas, the lamprey V.pt drives locomotion directly to MLR (midbrain locomotor region) and R.rs (reticulospinal motor neurons) [Beauséjour et al 2020].
The rest of the lamprey Ob drives the pallium (cortex) and subpallium (basal ganglia). Unlike the mammalian Ob which only drives specific olfactory cortical areas, the lamprey Ob broadly connects to the entire pallium [Derjean et al 2010], [Suryanarayana et al 2021]. Note that the lamprey pallium is smaller than the Ob [Pombal and Megías 2019].
Dual olfactory projections: direct to locomotor via V.pt and indirectly through the S.ot/P.v (basal ganglia). Hb.l (lateral habenula), MLR (midbrain locomotor region), Ob.l (lateral olfactory bulb), Ob.m (medial olfactory bulb), Pv (ventral pallidum), R.rs (reticulospinal motor command), S.ot (olfactory tubercle), V.pt (posterior tuberculum)
The above diagram illustrates the dual olfactory projection. The main action path is Ob.m to the V.pt to the MLR locomotion [Derjean et al 2010], [Beauséjour et al 2020], [Beauséjour et al 2024]. Not shown is the Ob.m projection to the Hb.m (medial habenula) – R.ip (interpeduncular area) for chemotaxis. The previous essay included the Ob.m to S.ot path for the timeout, which suppressed chemotaxis to avoid perseveration. Ob.l is the new addition, providing distinguishing context to the S.ot circuit.
Striatal discrimination
To represent distinct timeouts, different context or olfactory neighborhoods need distinct neurons or at least different dendrite spines. The striatum architecture is well-suited for this task because of the very large number of S.pn (striatal projection neurons aka medium spiny neurons). Each S.pn can represent a distinct combination of signal and context.
Striatum architecture to represent multiple timeouts, each with a unique context key built from unique distinguishing combination of inputs. cxt-1 (context input), Ob.m (medial olfactory bulb), Pv (ventral pallidum), S.pn (striatum projection neuron).
The above diagram shows the context-keyed timeout architecture. Each S.pn is associated with a distinguishing context, but all of these use the same primary signal. Because S.pn stores the timeout in the LTP (long term potentiatiation) / LTD in its dendrite spines, the multiple S.pn neurons allow for distinct persistent timeout variables. Furthermore, a single S.pn can support multiple contexts because each S.pn has several dendrites, on the order of 8-12, each of which can respond to a distinct input combination.
Note the similarity of this fan-out to granule cells in the hippocampus and cerebellum, and the Kenyon cells in Drosophila fruit fly. This expansion of the coding dimensionality allows for a large space to place odors while reducing overlaps [Laurent 2002].
Striatum UP states
In mammals S.pn are only active with sustained input from multiple distributed cortical sources [Shipp 2017]. This sustain input the S.pn into an UP state, which allows a primary signal to drive the neuron, but doesn’t drive an AP (action potential) directly. Typically the context UP state inputs drive distal dendrites and spines, and the primary signal drives the proximal dendrite. S.pn are hyper polarized at rest, making it difficult for a signal to drive an AP directly. The UP state depolarizes the S.pn, allowing the signal to drive an AP. Essentially this means the context neurons are required gates for the signal.
Combinations of context neurons drive a dendrite UP state, which allows the signal to drive the projection neuron. CN (context neuron), S.nr (substantia nigra pars reticulata), S.pn (striatum projection neuron).
The above diagram shows how each S.pn has an associated context made from a conjunction of several context neurons. Each S.pn has a different combination of context neurons, each differing greatly from its neighbor [Bolam and Bevan 2006]. Multiple simultaneous context neurons are necessary for an UP state.
Broad circuit
Taking an overview of this system, let’s see how addition of this context information affects the seek and timeout circuit affects the earlier circuit.
Olfactory timeout circuit with Ob.l added as a context input to S.v. Ad (adenosine), Hb (habenula), Ob (olfactory bulb), Pv (ventral pallium), S.d1 (striatum D1 projection neuron), S.d2 (striatum D2 projection neuron), V.pt (posterior tubuculum)
The above diagram shows the addition of Ob.l to S.ot was the only change necessary, along with the dimension expansion of the S.pn.
Cache-like model for simulation
The striatum architecture poses a scaling problem for the simulation. The striatum has a large number of neurons, each with a large number of essentially random inputs. This architecture works because the possible combinations are predefined. Each odor neighborhood is a conjunction of odor features, each corresponding to an Ob glomeruli and O.mc (olfactory mitral cells). The many predefined conjunctions are likely to match any new odor combination. However, a simulation model using this architecture would be overly large.
Because the essay model is a toy model, it can use a much simplified system. A cache-like architecture can work because only a few odor locations are active at any time. The cache only holds the recent odor locations, and the cache entry for an odor location is removed when the timeout expires. The simulation cache only needs to store the active locations, unlike the striatum, which holds the much larger number of possible distinct locations.
Simulation
The simulation adds a simplification of odor neighborhoods. Instead of simulating accurate odor plumes, each location has a place code, which then produces an odor code. In the screenshot below, the hexagonal colors represent these place codes that produce odor neighborhoods.
Simulation screenshot of the animal reaching food in a different neighborhood than the previously avoided neighborhood.
The above diagram shows two different odor neighborhoods (teal vs red). The animal avoids the red neighborhood after failing to find food, but seeks in the teal neighborhood to find the food. If the animal had first searched the teal neighborhood without food, it would have avoided have avoided the teal neighborhood with food.
Discussion
A major simplification in the simulation is consistency and precision in odor cues. In an actual environment, odors are not reliable. For now I’m not adding that complexity, but it might explain the need for cortical circuits in O.pir (piriform olfactory cortex) and E.hc (hippocampus). If an odor is irregular, some circuit needs to maintain a consistent odor neighborhood for the timeout circuit to work. In the simulation because the Ob perfectly represents the odor neighborhood and food plume, the downstream circuits can use the Ob signal directly. If the odor varies slightly within a neighborhood, or is lost intermittently, the S.ot timeout circuit could shift to a different S.pn timeout, breaking the logic of the circuit. A later essay might explore how cortical areas like O.pir might be necessary to create a stable neighborhood.
Bolam, J. P., & Bevan, M. D. (2006). Microcircuits of the striatum. In Basal Ganglia and Thalamus in Health and Movement Disorders (pp. 29-39). Boston, MA: Springer US.
Previous essay 23 (feeding and neuropeptide core) and essay 27 (feeding state machine) covered feeding, but were not based on the ascidian tunicate brain. After studying tunicates in essay 30 (RTPA – real time place avoidance), I think using the adult ascidian brain is a better foundation. The earlier essays chose the center of feeding at H.l (lateral hypothalamus), which is a foraging locomotion center, but this essay centers feeding around eating in R.pb (parabrachial). Where R.pb is centered around eating, tasting, and digestion, H.l is centered of seeking and locomotion. Because adult ascidians are sessile, they have no locomotion, but they do eat by filter feeding, suggesting that the eating functions in R.pb are more fundamental to the feeding process than the locomotive foraging.
Adult ascidian
The adult ascidian tunicate (specifically Ciona “sea squirt”) is a sessile filter feeder. Like other chordates, ascidian filter feeding uses pharyngeal slits, which developed into gills for vertebrates, and later into the jaw. The ascidian larva phase is only 24 hours and can swim like vertebrates, using phototaxis and geotaxis to find an appropriate permanent settling place. The adult ascidian brain dissolves the larva navigation areas and expends the “neck” of the larva brain, between the sensory ganglia and the motor ganglia [Gigante et al 2023].
Rough description of the adult ascidian Ciona brain.
The adult ascidian brain is marked by the Phox2 genetic transcription factor [Gigante et al 2023], which corresponds to specific vertebrate hindbrain areas. In vertebrates Phox2b corresponds to much of the hindbrain, including N5 (trigeminal), N7 (facial), N10 (vagus), N11 (accessory) nerves, R.nts (nucleus of the solitary tract), and R.na (nucleus ambiguous). The vertebrate Phox2a corresponds to V.lc (locus coeruleus – norepinephrine), N3 (oculomotor), and N4 (trochlear) nerves [Dufour et al 2006]. To put it another way, the branchial (pharyngeal) motor neurons are a distinct group consisting of N5, N7, N9 (glossopharyngeal), N10, and N11 [Fritzsch et al 2017].
The Phox2b nuclei essentially match the parasympathetic system:
N5 – trigeminal – jaw, chewing, face and mouth
N7 – facial – facial expressions and taste in tongue
N9 – glossopharyngeal – taste in tongue, swallowing, blood pressure
N10 – vagus – heart, breathing, digestion, etc.
N11 – accessory – shoulder and neck
R.nts – nucleus of the solitary tract – visceral and taste sensorimotor
Note that the jaw develops from the first pharyngeal arch, and the pectoral girdle (shoulder and neck) may develop from the last pharyngeal arch [Brazeau et al 2023].
The Phox2a nuclei are:
N3 – oculomotor – most eye movement
N4 – trochlear – additional eye movement
V.lc – locus coeruleus – main source of norepinephrine
Because the ascidian adult brain is necessarily self-contained, corresponding to a restricted set of hindbrain ganglia, it’s a good center module for vertebrates, as a thought experiment. However, it’s important to remember that tunicates are highly divergent from the original vertebrate/tunicate ancestor and trying to derive that ancestor from ascidians is extremely suspect [Holland 2015].
R.pb parabrachial nucleus
For the sake of the simulation, I’m combining R.pb with associated nuclei like R.nts, R.na, and R.plc (pre locus coeruleus). Science research needs to distinguish the nuclei, of course, but complicating the essay’s model wouldn’t add much verisimilitude.
R.pb is a highly heterogenous area [Pauli et al 2022] with at least 36 gene expression clusters [Nardone et al 2024]. Its neurons derive from two distinct progenitor populations marked by transcription factors lmx1b and foxp2 [Karthik et al 2022]. The lmx1b population projects to the Phox2b hindbrain population, S.a (central amygdala), P.bst (bed nucleus of the stria terminalis), H.pstn (presubthalamic nucleus), T.vpm (thalamus gustatory), and C.i (insular cortex). The foxp2 population projects to hypothalamic area Poa (preoptic area), H.l (lateral hypothalamus), and H.vm (ventromedial hypothalamus).
R.pb parabrachial descending projections are to Phox2b hindbrain nuclei. N5 (trigeminal), R.my (medulla), R.nts (nucleus of the solitary tract), R.pb (parabrachial nucleus).
Many R.pb populations transform short stimuli (10ms) into long effects, often to suppress feeding. For example an experimental 10s tail shock suppresses feeding for 30s to 60s using intracellular mechanisms including cAMP and PKA, which suppressed licking but not suppress locomotion [Singh Alvarado et al 2024]. In R.pb, cAMP half life was 33s. Other areas like Po.m (medial preoptic area) with extended firing have longer decay times.
Sleep, wake, and breathing
The R.pb area, particularly around R.pb.m (medial R.pb) and including the neighboring V.lc, is one of the most critical wake areas. If the area is lesioned, the animal can become comatose [Fuller et al 2011]. R.pb astrocytes promote wakefulness [Liu PC et al 2023], and R.pb stimulation wakes from anesthesia [Luo et al 2018]. R.pb to H.l strongly correlates with H.l orexin neurons, which are wake promoting [Huang et al 2021]. R.pb is strongly connected with P.bf (basal forebrain), which manages cortical arousal [McKenna et al 2021]. R.pb’s neighboring R.sld (sublaterodorsal nucleus) is responsible for REM atonia [Peever and Fuller 2016]. The nearby R.pz (parafacial zone) suppresses R.pb to promote sleep [Anaclet et al 2014].
R.pb is also strongly associated with breathing, and the Phox2b region of the hindbrain is required for breathing [Dutchemann and Dick 2012]. Because vertebrate gills and lungs developed from the earlier filter-feeding pharyngeal arches, breathing is a natural extension of the earlier feeding function.
Malaise: LPS and CGRP
Proto-vertebrate filter feeding includes feeding on bacteria, but some bacteria are toxic. LPS (lipopolysaccharide) is a marker for bacterial inflammation [Essner et al 2017], immediately halts feeding in mice, and is often used in behaviorist training for CTA (conditioned taste aversion) [Palmiter 2018], [Parker 2003]. The LPS sensors are in the gut, travel through N10 to R.nts and to R.pb using the CGRP peptide marker [Campos et al 2016]. The R.pb.cgrp (CGRP R.pb neurons) halt eating [Carter et al 2013] and produce CTA (conditioned tasted aversion) [Carter et al 2015].
Sickness inhibition of eating. R.nts.r (rostral nucleus of the solitary tract), R.pb.el (external lateral parabrachial nucleus).
Importantly, note that R.pb neurons are almost entirely glutamate. Although the limitation of lacking inhibition doesn’t matter yet, it will become important soon and motivate the H.arc (arcuate hypothalamus) and S.am (central medial amygdala). So, even though the effect of R.pb.cgrp is to stop eating, R.pb is an active command to stop eating, not an inhibition gating an eating action. I’m assuming eating defaults to enabled, which may make sense for a filter feeder. Alternatively, the eating motor neurons in the Phox2b area may have additional requirements such as food touching the lips via N5 (trigeminal). In the above diagram R.nts.lps is only active when the gut sensors detect LPS, when then triggers R.pb.cgrp, which stops eating.
In the case of LPS, which is essentially food poisoning, the animal might trigger vomiting to remove the toxins from the but, may want to avoid the area to stop filter feeding the disease, and when learning is available to learn some signs to avoid getting sick again.
On review of this section, although LPS is a trigger for R.pb CGRP, much of this discussion applies to LiCl, which is a distinct gut warning peptide. A later essay should clarify when LPS and LiCl can be treated as equivalent and when they need to be distinguished.
Gut satiation: CCK and oxytocin
Satiation also suppresses eating. Like LPS, satiation stops eating or drinking, but unlike LPS, it’s not a negative effect. In fact if the animal is pleasantly full, it might remember the area. CCK is a gut peptide that marks the gut as being full and can also be a nutrient sensor [Palmiter 2018]. In mammals, oxytocin signals a thirst satiation. If the animal it’s thirsty it doesn’t drink.
Satiation suppresses eating and drinking via R.pb. H.pv (periventricular hypothalamus), OXT (oxytocin), R.nts (nucleus of the solitary tract), R.pb (parabrachial)
H.pv (paraventricular hypothalamus) also produces satiety signals. H.pv MC4 and H.pv dyn neurons each produce 50% of H.pv satiety [Li MM et al 2019]. The H.pv targets are distinct from R.pb CGRP.
Bitter tastes
For safety animals taste food before eating because it’s inefficient and dangerous to detect bad food only after developing food poisoning. In mice bitter tastes stimulate R.nts.r, which drives orofactial expressions like “gaping” by the N9 (glossopharyngeal) motor complex [Kinzeler and Travers 2008]. Because mice can’t vomit, rejecting dangerous-tasting food is particularly important. R.nts.r also detects sweet tastes and produces positive orofacial expressions like licking [Roussin et al 2012]. At a higher level bitter taste activates R.pb.el CGRP neurons, which feeds into the eating termination circuitry used by LPS detection.
Bitter response circuitry in R.pb. R.nts bitter tastes excite R.pb CGRP neurons which stop eating. H.arc AgRP hunger can suppress bitter or neophobia. H.arc (arcuate hypothalamus), N9 (glossopharyngeal nerve), R.nts (nucleus of the solitary tract), R.pb (parabrachial), AgRP (hunger peptide), CGRP (alarm peptide).
Interestingly, these two orofacial expressions — gaping and licking — are used as signs of hedonic pleasure in studies of the basal ganglia, which distinguish motivation from pleasure [Berridge 2019].
Hunger: “homeostatic feeding”
Until now I’ve assumed eating as a default activation with R.pb as a brake to stop eating, whether from toxins or satiation, but vertebrates are positively motivated by hunger, such as the AgRP hunger peptide produced by H.arc (arcuate hypothalamus). If the H.arc AgRP neurons are inhibited, mice will starve [Roman et al 2016]. This anorexia can be reversed if the R.nts CCK satiety neurons are also disabled [Roman et al 2016].
Hunger from H.arc suppresses neophobia and bitter to disinhibit eating. AgRP (hunger peptide), H.arc (arcuate hypothalamus), CCK (gut satiation peptide), CGRP (alarm peptide), R.nts (nucleus of the solitary tract), R.pb (parabrachial)
The above diagram shows the system of hunger disinhibiting eating by suppressing neophobia, gut satiation and bitterness. By default R.nts CCK neurons tonically activate R.pb.el CGRP neurons, which disables eating. When the animal is hungry, H.arc AgRP neurons suppress the tonic food inhibition from R.pb CGRP neurons, letting the animal eat. This system is also used for food neophobia, where the animal won’t eat a new food until it’s known to be safe [Campos et al 2018], [Palmiter 2018].
The hunger signal can be circadian; it’s not necessarily triggered by low energy stores. For example rodents typically forage as soon as they’re awake [Blum et al 2014]. The system doesn’t need to wait until the animal is starving with low blood sugar before eating. Along with circadian input, the H.arc neurons are modulated by several factors including signals for fat availability (leptin) [Andermann and Lowell 2017].
Note that this system allows for the combination of “not hungry” with “not sated.” If the animal doesn’t have AgRP (not hungry) but also not CCK (not sated), it will still not eat because of the default R.pb inhibition. This “homeostatic feeding” allows for a high quality “hedonic feeding,” where the animal will eat rich food (typically sugar or fat) if it finds some, but won’t fill itself with low quality food beyond what is necessary to stop hunger [Tang et al 2022].
Sweetness: hedonic feeding
After the animal has satisfied its base hunger, it might still eat if it can find some rich food (sweet or fatty). There’s always room for dessert. If high quality rich food is unavailable, circadian or hunger-driver feeding will keep the animal from starving. Once the animal avoids minimal starvation, it can afford to search for better food sources. After nibbling the food, the animal will only eat if it’s sweet or savory.
Sweetness as disinhibiting satiation or bitterness. CGRP (alarm peptide), M.dp (deep midbrain), N9 (glossopharyngeal nerve), R.nts (nucleus of the solitary tract), R.pb (parabrachial), S.am (central medial amygdala), SST (somatostatin peptide)
In the diagram above, sweet tastes inhibit the R.pb CGRP neurons, allowing the animal to eat when not hungry but also not sated. R.pb sweet taste neurons are marked by satb2 [Fu et al 2019]. R.pb satb2 neurons project to S.am (central medial amygdala) [Jaramillo et al 2021], [Jarvie et al 2021]. S.am (somatostatin) neurons also directly project to M.dp (deep midbrain reticular) to lick after tasting sweet [Zheng D et al 2022]. Ethanol can act like sweet tastes to increase drinking via S.am nts (neurotensin) to R.pb [Torruella-Suárez et al 2020].
Note that this diagram doesn’t show the only source of sweet motivation. The adjacent R.plc (pre locus coeruleus) nucleus can also drive eating for sweet tastes. Inhibition of R.plc glutamate neurons will eat sweet [Gong et al 2020]. Like R.pb, R.plc has a long duration effect. Inhibiting R.plc extends eating for approximately 15s.
An animal needs to interrupt its eating when it’s alarmed, whether a problem from an environmental threat such as high heat, CO2, itch, injury, or pain from eating, such as capsicum from a pepper. R.pb CGRP neurons serve as a general alarm, which stops eating [Campos et al 2018], [Jaramillo et al 2019].
S.a threats as suppressing S.a sweetness to stop eating. CGRP (alarm peptide), LPS (sickness peptide), M.dp (deep reticular midbrain), N.sp.l1 (layer 1 spinal cord), N9 (glossopharyngeal), pkcδ (gene marker for S.al halting), R.nts (nucleus of the solitary tract), R.pb (parabrachial)
The above diagram adds pain and alarm eating suppression to the top of the previous S.am diagram. This diagram only shows two R.pb.el CGRP inputs, but other inputs also feed into the R.pb.el CGRP neurons, including N5 (trigeminal) pain sensors for the jaw and mouth [Campos et al 2018], [Carter et al 2013], [Palmiter 2018].
Note the similarity between the S.al and S.am circuit and the Sv.d2 (ventral striatum D2 neurons) and Sv.d1 (ventral striatum D1 neurons) circuit. Some definitions of the external amygdala include S.msh (medial shell of the ventral striatum) along with P.bst (bed nucleus of the stria terminalis.)
Consensus loops
Before the circuits become too complicated, let’s step back to explore how to manage the complexity. The problem of complexity will quickly multiply in this essay as it covers more modules that each modulate eating. Essay 30 RTPA (real time place avoidance) ran into the same problem and added a notion of a consensus loop, where multiple independent, distributed drives voted continually to decide on an action.
Consensus voting model to select a drive to motor action.
The above diagram shows the essay 30 consensus voting model. Instead of a single action path from stimulus to motor, a set of drives and actions consult each other in a mutually-inhibited winner-take-all system to choose an action. In this essay’s motivation for feeding, it may be simpler to consider the consensus system as managing a number of shared peptides, instead of drawing ever more complicated diagrams. This model resembles the old notion of an “isodendritic core,” which included areas like R.pb, H.l and similar reticular areas [Ramón-Moliner and Nauta 1966], [Agnati et al 2010], where broadcast transmission of peptides is more important than axon synaptic connectivity.
Another model for managing action selection uses diffusion models and Langevin dynamics to describe complex distributed choice in the brainstem [Richman et al 2023]. In their experiment where mice are both hungry and thirsty, the mice choose water or food rewards in a sticky, stochastic manner, choosing water rewards several times in a row before switching to food and then switching back. The brainstem regions in that study’s choice are highly distributed with over 10 regions providing significant management, without any one region serving as a central decision hub.
Weak attractor, diffusion model showing the stochastic, sticky action selection. The choice sticks to a shallow attractor basin, which is weak enough to allow for switching to the other basin..
The above diagram shows the eat vs drink decision as a weak stochastic attractor model. The choice switches between two shallow attractor basins. Because the forces on the choice are stochastic, the choice switches between basins, but stays in a particular basin for a short time. Other models might stick to one choice in a winner-take-all model or switch consistently in an oscillator model.
Going forward in this essay, consider each new module as contributing to a shared consensus decision as opposed to considering the model as a strict connection circuit.
H.pstn hunger and eating suppression
H.pstn (presubthalamic nucleus) is part a part of H.l directly focused on eating, in contrast to the H.l focus on waking, arousal, exploration, and seeking. H.pstn is more responsive to rich food than plain food, part of the hedonic eating as opposed to homeostatic hunger [Chometton et al 2016]. H.pstn is strongly connected with S.a and P.bst and is the main H.l target of S.a [Shah et al 2022].
H.pstn rich food circuit, as a parallel and connected circuit to S.am. H.pstn (presubthalamic nucleus), pkcδ (central amygdala aversive marker), R.nts (nucleus of the solitary tract), R.plc (pre locus coeruleus), R.pb (parabrachial), S.al (lateral central amygdala), S.am (medial central amygdala), SST (S.a attractive marker)
The above diagram shows some of the H.pstn connectivity, focusing on its connections to R.pb and S.a. R.plc was mentioned previously for rich food. H.pstn connections include P.ipac (interstitial nucleus of the posterior limb of the anterior commissure), P.si (substantia innominata aka ventral pallidum), P.bst, T.pv (paraventricular thalamus), R.m5 (trigeminal motor nucleus), R.pz (parafacial zone), R.nts, R.dmv (dorsal motor of vagus nerve). H.pstn also receives input from F.ai (granular insular cortex), H.pv, M.pag.v (ventral periaqueductal gray), R.plc [Shah et al 2022]. In other words H.pstn has many connections and can also directly drive hindbrain eating motor areas.
Because S.a is a GABA nucleus and H.pstn is primarily glutamate, the two areas may work synergistically [Shah et al 2022]. Because S.a, P.bst, and H.l GABA neurons cluster into a genetically related group [Yao et al 2023], an ancestral vertebrate may have had a combined S.am and H.l GABA with H.pstn and H.l glutamate. This H.pstn connection is a reason to consider the consensus loop model because trying to model it more precisely is getting into the weeds.
S.v and taste
Taste in S.v (ventral striatum) can drive eating is parallel with the S.a (central amygdala) system. S.v is a heterogenous system, not only with core, lateral shell and medial shell, but S.msh (medial shell) is itself highly heterogenous [Castro et al 2015], [Chen R et al 2021]. The S.v taste connectivity is more specialized and uses different connectivity than the canonical basal ganglia connectivity.
S.v taste and eating extension. DA (dopamine), GLP-1 (sickness peptide), R.ap (area postrema), R.my (medulla), R.nts (nucleus of the solitary tract), R.pb (parabrachial), Sv (ventral striatum), Vta (ventral tegmental area)
The above diagram shows some Sv connectivity to R.pb, R.nts, R.my, and R.ap (area postrema). A sweet taste drives the Sv.d1 (Sv projection neuron with D1 receptor) neuron to extend eating. The companion Sv.d2 (Sv projection neuron with D2 receptor) suppresses eating directly to R.nts, and suppresses the Sv.d1 eating extension. [Sandoval-Rodríguez et al 2023]. As with S.v, adenosine quickly times out the Sv.d1 path and enables the Sv.d2 path. This might allow for a hesitant lick to taste, preventing full eating. If the taste is sweet, then DA (dopamine) is activated via Vta (ventral tegmental area), and the DA potentiates the Sv.d1 path and suppresses the Sv.d2 path, which will extend eating.
Ignoring dopamine for a moment, the above Sv functionality is similar to Sa functionality, where Sv.d1 promotes eating with sweet taste, like S.am extends eating for sweet taste. Similarly, Sv.d2 suppresses eating for GLP-1 (LPS peptide marker), like S.al suppresses eating for R.pb CGRP alarms. As essay 31 explored, adenosine in S.v can timeout Sv-enabled actions. Without extra domaine, the sweet taste only allows for a minimal extension of eating.
R.pb directly drives Vta for DA to S.v, either enhancing DA for sweet taste or inhibiting for food shocks [Coizet et al 2010], [Han W et al 2018], [Nagashima et al 2023], [Tsou et al 2023]. Looking at the diagram above, enhancing DA extends eating by encouraging the Sv.d1 path, while suppressing DA curtails eating by encouraging the Sv.d2 path.
R.pb S.msh
Let’s return to the S.v modulation of taste and eating. Earlier, sweet tests drove DA to extend eating by driving the S.v.d1 path that extends eating, but an alarm while eating the sweet food should curtail eating.
Alarms suppress dopamine to curtail eating. A CGRP alarm drives RMTg to suppress Vta dopamine, forcing a stratum timeout. CGRP (alarm peptide) DA (dopamine), R.ap (area postrema), R.nts (nucleus of the solitary tract), R.pb (parabrachial), RMTg (rostromedial tegmental nucleus), Sv (ventral striatum), Vta (ventral tegmental area).
The above diagram extends the previous circuit by suppressing dopamine for an alarm, such as sickness to suppress sweet taste. An alarm drives R.pb.el, which drives Vta.rmtg (rostromedial tegmental nucleus) to suppress Vta [Coizet et al 2010], which suppresses dopamine, which forces a timeout of the striatum circuit, driving the Sv.d2 path to suppress eating.
The Vta.rmtg path has another interesting input from the R.pb sweet neurons. R.pb sweet drives P.ms.g SST (median septum somatostatin, which inhibits Hb.l (lateral habenula) to inhibit the tonic suppression of Vta.rmtg [Shen et al 2022]. In short, R.pb sweet disinhibits Vta DA through the P.ms.g to Hb.l path.
Median septum sweet-promoting path. A sweet taste drives P.ms.g SST neurons, which disinhibits Vta through the Hb.l to Vta.rmtg path. DA (dopamine), Hb.l (lateral habenula), P.ms (median septum), R.pb (parabrachial), RMTg (rostromedial tegmental nucleus), S.core (ventral striatum core), Vta (ventral tegmental area)
Because Hb.l glutamate neurons are tonically active, providing a midpoint suppress of Vta DA, Hb.l can be driven in either direction to increase or decrease DA.
Poa temperature
The Poa (preoptic area) has many functions, but the function most directly connected to R.pb is temperature defense. R.pb sends both too hot and too cold signals to Poa, which responds to temperature for multiple coping methods including vasoconstriction and vasodilation, and thermotaxis to leave the too hot or too cold area [Norris et al 2021], [Yahiro et al 2017].
Temperature modulation in the preoptic area driven by parabrachial signals. Hb.m (medial habenula), Poa (preoptic area), R.ip (interpeduncular area), R.pb parabrachial.
Pain
R.pb is a center of sustained pain, as opposed to sharp, acute pain. To keep this essay from growing out of control, I’ll not dive too deeply into the details. For simplicity I’ll treat pain as just another input to the R.pb.l.cgrp input. This pain R.pb.l CGRP signal then activates S.al and P.bst, which curtails eating.
R.pb to hypothalamus
There seems to be three distinct R.pb to hypothalamus connections: eating, pain avoidance, and waking / seeking. As mentioned above H.pstn is highly concerned with eating, particularly rich food and is highly connected with S.a, which has similar connectivity and activation [Chometton et al 2016]. Because S.a is primarily GABA and H.pstn is primarily glutamate, their connectivity may be synergistic. R.pb has a strong connectivity with the H.l orexin neurons for wake and seeking, complementary with the R.pb role in sleep and wake.
Essay model
I’ve chosen to group eating-related areas functionally as opposed to anatomically. BodyEat represents physical body simulation, such as a short digestion delay in the gut and accumulating glucose as the food digests. HindEat includes the Phox2b hindbrain areas including R.nts, R.ap and the branchial derived motor nerves N5 (trigeminal), N7 (facial), N9, N10 (vagus), and N11. MotiveEat includes R.pb eating-related areas, S.am eating, S.al bitter / sickness response, H.pstn rich food, and also includes S.v taste response.
A different model might make each anatomical area into its own module, such as a module for R.pb or split into finer R.pb.l and R.pb.m. However, that approach would run quickly into the issue of module boundaries, particularly because R.pb is more organized by genetic clusters than anatomical boundaries. As important, spliggin modules between R.pb, S.a, H.arc, H.pv, and H.pstn would obscure the shared functionality.
Essay simulation modules
HindMove includes reflexive orofacial movements such as eating (licking), vomiting, and gaping, which is a rodent facial expression functionally equivalent to spitting food out. These behavior are equivalent to R.nts triggering of orofacial movement.
MotiveEat does not include foraging locomotive decisions, such as seeking toward an odor plume, roaming search, and avoidance. When the animal tries to eat a bitter food, it not only stops eating and initiates gaping, but it needs to reverse foraging and avoid the current area to prevent a return to the same bad food patch.
Simulation
The essay simulation expands food items into multiple kinds: plain, bitter, sweet, and sickness. The diagram below shows the setup. The different star colors show the food kinds: red for sweet, orange for plain, and yellow for bitter. Because odors are not distinct, the animal can’t predict the kind of food until eating it.
Simulation screenshot showing the animal avoiding a bitter food area.
An immediate problem in the simulation was bitter food causing a freeze, because the initial implementation stopped eating, and produce gaping, but didn’t avoid the area. Because avoidance never triggered, the animal never left the bitter area. To fix that problem, a bitter taste triggers a general alarm to avoid the current area.
Because sweetness and sampling before eating when the animal is no longer hungry turns out to be relatively complicated, I’ve postponed the implementation because it’s not clear how sampling works. For example it may require Sv or Sd to quickly halt eating, possibly using adenosine as in essay 31. But adding that striatum complexity would derail the current R.pb focus.
Discussion
The main idea of the essay is to pub R.pb in the center of feeding and even as a seed at the center of vertebrate behavior. That center is a seed to build more complicated behavior around. The justification for this idea is an analogy with the ascidian adult brain, which only controls filter feeding and visceral activity such as digestion, respiration, and heart rate. The common vertebrate and tunicate ancestor probably didn’t have the ascidian degenerate brain, but high competition in the Cambrian may have forced a bottleneck with a smaller, focused brain.
This approach has a large advantage of organizing many distributed systems. S.am and S.al make sense as managing R.pb taste and eating with inhibition. H.pstn and S.am manage eating rich food beyond foraging for plain food. H.arc and H.pv manage eating by modulating R.pb CGRP inhibition, driven in part of circadian drives. P.bst, H.l, and Hb.l manage the avoidance of food, driven by R.pb bitterness or sickness signals.
Chen R, Blosser TR, Djekidel MN, Hao J, Bhattacherjee A, Chen W, Tuesta LM, Zhuang X, Zhang Y. Decoding molecular and cellular heterogeneity of mouse nucleus accumbens. Nat Neurosci. 2021 Dec;24(12):1757-1771. Chiang, M. C., Bowen, A., Schier, L. A., Tupone, D., Uddin, O., & Heinricher, M. M. (2019). Parabrachial Complex: A Hub for Pain and Aversion. The Journal of neuroscience : the official journal of the Society for Neuroscience, 39(42), 8225-8230.
Dufour HD, Chettouh Z, Deyts C, de Rosa R, Goridis C, Joly JS, Brunet JF. Precraniate origin of cranial motoneurons. Proc Natl Acad Sci U S A. 2006 Jun 6;103(23):8727-32.
Yao Z, van Velthoven CTJ, Kunst M, Zhang M, McMillen D, Lee C, Jung W, Goldy J, Abdelhak A, Baker P, Barkan E, Bertagnolli D, Campos J, Carey D, Casper T, Chakka AB, Chakrabarty R, Chavan S, Chen M, Clark M, Close J, Crichton K, Daniel S, Dolbeare T, Ellingwood L, Gee J, Glandon A, Gloe J, Gould J, Gray J, Guilford N, Guzman J, Hirschstein D, Ho W, Jin K, Kroll M, Lathia K, Leon A, Long B, Maltzer Z, Martin N, McCue R, Meyerdierks E, Nguyen TN, Pham T, Rimorin C, Ruiz A, Shapovalova N, Slaughterbeck C, Sulc J, Tieu M, Torkelson A, Tung H, Cuevas NV, Wadhwani K, Ward K, Levi B, Farrell C, Thompson CL, Mufti S, Pagan CM, Kruse L, Dee N, Sunkin SM, Esposito L, Hawrylycz MJ, Waters J, Ng L, Smith KA, Tasic B, Zhuang X, Zeng H. A high-resolution transcriptomic and spatial atlas of cell types in the whole mouse brain. bioRxiv [Preprint]. 2023 Mar 6:2023.03.06.531121.
This essay explores using the hippocampus as a sequence generator [Buzsáki and Tingley 2018] to precisely time the avoidance action after a failed seek. In essay 31, the animal started with a roaming random search, which turned into a directed seek when it smelled a food odor. If the animal failed to find food after a timeout, it would avoid the area. This essay expands on that model by improving the avoidance action. Previously, the avoidance time was modeled on a cellular timeout, which is imprecise. Instead, we can use an hippocampus sequence to time the avoidance.
State diagram for the foraging task. A random walk roaming search switched to a directed seek when the animal smells a food odor. It switches to avoidance if the seek times out.
This foraging search resembles a Levy walk, which combines a area-restricted brownian walk with longer movements to avoid repeated searches in an area.
In the previous essay, the seek action timed out using astrocyte-managed adenosine in the striatum, but the avoid timeout wasn’t specified, presumably piggybacking on the astrocyte adenosine. Improving the avoidance time and distance can use a sequence generated by the hippocampus. In mice, this kind of distance measurement and timeout is seen in C.pp (posterior parietal cortex) with neurons tiling the delay period, forming a sequence [Harvey et al 2012], [Rajan et al 2016] and in E.ca1 (hippocampus CA1 area) [Pezzulo et al 2017].
Because this essay remains as primitive as possible, and we haven’t added cortex regions yet, we can only add one simplified cortical area. One option is to consider C.pp as a primitive cortex region that can self-generate the necessary sequences. Another option is to consider E.hc (hippocampus) as the main sequence generator and treat C.pp as a later specialization for sophisticated vertebrates like mammals.
The hippocampus can be seen as a sequence generator [Buzsáki and Tingley 2018]. E.hc (hippocampus) sequences are approximately 7s from first place cell neuron to the last neuron in the sequence [Pezzulo et al 2017], and E.hc delay timing is approximately 8s [Abela et al 2015]. Similarly F.pl (prelimbic prefrontal cortex) sequence neurons tile choice encoding for 7s across an experimental trial. C.pp neurons can tile the distance from a start position in mice [Harvey et al 2012], [Rajan et al 2016]. In mice E.ca1 path integration has a maximum of 2m unless extended by landmarks [Fischler-Ruiz et al 2021].
Additionally, the timeout sequence needs a mechanism to trigger it and to drive the avoidance action. I’ll cover a possible trigger from H.sum (supramammillary) and an output from E.hc either directly to H.l (lateral hypothalamus), H.sum, or Poa (preoptic area), or using S.ls (lateral septum as an intermediary.
E.hc place cell for delay
A study in [Fischler-Ruiz et al 2021] studied E.ca1 (CA1 region of hippocampus) with mice traveling a virtual maze, where the liquid reward was 4m from the initial starting point. Instating of waiting for the full 4m, mice expected reward at 2m unless an odor landmark near 2m extended the range to 4m. Neurons in E.ca1 tracked the distance traveled with short time neuron fragment tiling the delay period. A sparse number of neurons active for a short period and as one neuron ends another, new neuron takes its place, like a long thread is made of shorter fibers. This thread frays at around 2m when mice lose track of the task.
Diagram of neurons tiling a delay period. After a simultaneous initial spike of many neurons at top left, a sequence of neurons fire and replace each other.
Mouse C.pp neurons in virtual maze tile the delay period [Kamiński and Rutishauser 2020]. The above diagram shows the neural tiling. Each row is a single neuron, and time is on the x axis, and the neurons are sorted by their activity. At top is an initial simultaneous burst of many neurons that respond to some event. This large burst is followed by smaller sets of neurons that persist for a small time. When one neuron stops, another neuron takes its place, until the sequence frays and ends.
Roam vs Seek
H.sum is associated with active exploration and is quiet while eating [Kesner et al 2021]. H.sum is activated by food anticipation, food restriction, or gherkin, a hunger peptide [Le May et al 2019].
[Wee et al 2019] H.l (lateral hypothalamus) and Hc (caudal hypothalamus) are both feeding related but are anti-correlated. Hc activates when the zebrafish is hungry and drives the roaming search for food, but when a specific target (paramecium for zebrafish) is detected, H.l activates and Hc deactivates for specific hunting and for eating [Wee et al 2019].
In zebrafish, Hc is comparable to the mammal posterior hypothalamus, which includes H.sum and H.mb (mammillary body), H.tu (tuberal hypothalamus), and H.arc (arcuate nucleus). Zebrafish Hc includes all these areas [Schredelseker and Driver 2020], which are genetically and functionally distinct. Unfortunately this makes it unclear which area is driving the roaming search. H.arc for example is associated with hunger and anti-correlated with eating, using hunger neuropeptides AgRP and NPY in H.arc [Berrios et al 2021]. H.sum is also activated by food anticipation or hunger [Le May et al 2019].
AgRP drops quickly on food cues, drops more slowly for gut nutrient detection, and is slow or permanent in energy balance like blood glucose [Berrios et al 2021]. The specific path for seek dropping AgRP is H.l glutamate to H.dm (dorsal medial hypothalamus) GABA to H.arc AgRP. The H.dm effect disappears with sleep time. RTPP (real-time place preference) is 80% if AgRP is inhibited.
In addition, zebrafish H.l is not identical with the mammal H.l. For example, the orexin and MCH neuropeptide neurons in mammals are part of H.l, while they are in distinct non-H.l areas in zebrafish [Schredelseker and Driver 2020].
In other words, while this essay is taking H.sum as embodying the roam as opposed to H.l embodying the seek, the exact correlations of roam and seek are not yet known.
E.hc input and output
If E.hc drives the seek timeout’s avoid action sustain, how does the seek timeout trigger the E.hc sequence timer to start, and how does E.hc sustain the avoidance? E.hc has three major avoidance action paths it could influence: the obstacle avoidance path through OT (optic tectum) from essay 34, the temporal gradient taxis path through Hb.m – R.ip (medial habenula to interpeduncular nucleus) from essay 33, and the related motivational path through H.l (lateral hypothalamus) and Hb.l (lateral habenula).
Action paths for avoidance to the hippocampus. The left panel shows obstacle avoidance from the optic tectum via C.pp. The right panel shows taxis avoidance from the Hb.m to R.ip path. C.pp (posterior parietal cortex), E.hc (hippocampus), E.mec (medial entorhinal cortex), E.por (postrhinal cortex), H.sum (supramammillary), Hb.m (medial habenula), OT (optic tectum), P.ldt (laterodorsal nucleus), P.ms (median septum), R.in (nucleus incepts), R.ip (interpeduncular nucleus), V.mr (median raphe)
The above diagram shows possible input paths to E.hc. On the left, OT (optic tectum) sends visual threat and obstacle information to C.pp and E.por (postrhinal / parahippocampal cortex) via T.lp (lateral posterior / pulvinar thalamus), then to E.mec (medial entorhinal cortex), and finally to E.hc. In the right panel, Hb.m-R.ip sends feedback to E.hc through multiple paths, including H.sum, V.mr (median raphe), R.in (nucleus incepts) and through P.ms (median septum). These two paths differ not only on the information they convey, but also their circuit effects on E.hc. While the left OT path is a data-driven path that excites glutamate neurons, the right Hb.m-R.ip path is a control path that drives E.hc interneuron control circuitry using GABA, ACh (acetylcholine), and 5HT (serotonin) to modulate the action and timing of E.hc.
For this essay, the control path is more interesting, is part because it’s more closely tied to the seek-roam control, and in part because the control circuitry is more fundamental to E.hc operation. The interneuron circuit for the cortex and the hippocampus are highly conserved for all vertebrates, but at some point in evolution they must have been new. Because this essay is exploring adding the first cortical-like area, it needs to address the interneuron controls and not merely assume their existence.
Motivational avoid feedback to the hippocampus. A motivational avoid signal from Hb.l drives E.c through V.mr. E.hc (hippocampus), H.sum (supramammillary), Hb.l (lateral habenula), P.ms (median septum), V.mr (median raphe)l
A different, motivational path using V.mr arrives from Hb.l (lateral habenula) instead of the Hb.m – R.ip taxis path. Hb.l is more motivational and Hb.m it more of a physical taxis function. As I’ll be using later in the essay, the full avoid path first travels through the motivational Hb.l, then to E.hc for the avoid time/distance sequence, and finally to Hb.m – R.ip for physical avoidance.
Output paths for timed avoidance. The left panel shows a possible path to OT via C.pp. The right panel shows a path to Poa and H.sum via S.ls. C.pp (posterior parietal cortex, E.ca1 (hippocampus CA1 area), E.sub (hippocampus subiculum), H.l (lateral hypothalamus), H.sum (supramammillary), OT (optic tectum), Poa (preoptic area), S.ls (lateral septum).
The avoidance output path from E.hc could use one of three action paths. E.hc can drive obstacle-like avoidance through C.pp to OT. Secondly it can drive physical gradient taxis avoidance through Hb.m via P.ts (triangular septum) and P.bac (bed nucleus of the anterior commissure) [Yamaguchi et al 2013]. Finally it can drive motivational avoidance to H.l, H.sum, and Poa (preoptic area) through S.ls (lateral septum). The above diagram shows the obstacle avoidance path through C.pp to OT and the motivational avoidance path through H.l, H.sum, and Poa. (Note: Poa to M.pag to MLR is more direct avoidance than motivational.) The Poa path drives movement directly through M.pag (periaqueductal gray) and MLR (midbrain locomotor region), which the diagram omits for brevity.
Another E.hc motivational output uses S.v (ventral striatum / nucleus accumbens) and P.v (ventral pallidum / endopeduncular nucleus) to H.l and Vta (ventral tegmental area) for seeking and to Hb. for avoidance. Although this path is similar to the S.ls motivational path above, in mammals at least it’s a distinct circuit. Unlike the S.ls path, the S.v motivational path does not target H.sum or Poa but does strongly target Hb. Note that S.v and S.ls are similar structures both derived from the same progenitor domain LGE (lateral ganglionic eminence) and is neighbors with S.v, also known as nucleus accumbens septi, Latin for nucleus adjacent to the septum.
Output path from hippocampus to medial habenula via P.bac and P.ts. E.hc (hippocampus), Hb.m (medial habenula), P.bac (bed nucleus of the anterior commissure), P.ts (triangular septum).
Mammals also have a direct E.hc output to Hb.m-R.ip via P.ts (triangular septum) and P.bac (bed nucleus of the anterior commissure, which is sometimes identified with part of P.bst bed nucleus of the stria terminalis) [Proulx et al 2014]. As described in previous essays, Hb.m has a direct function for phototaxis [Chen and Engert 2014], chemotaxis [Beretta et al 2012], thermotaxis, and some social conflict [Agetsuma et al 2010]. This mammal P.ts/P.bac is understudied, or I haven’t found any study describing its function, but by its connection E.hc could drive taxis via the Hb.m-R.ip connection. This connection does not exist in lamprey [Stephenson-Jones et al 2012], but I haven’t read any studies on this connectivity for other non-mammal animals.
Ancestral vertebrate forebrain
Stepping back to cover the ancestral vertebrate forebrain (cortex and basal ganglia) to understand likely primitive areas, particularly the E.hc equivalent. Ancestral models generally use Pa (pallium) to describe cortical-like areas and divide pallial areas into multiple regions, generally from four to six, depending on the theory. These areas are named by their location, where MPa (medial pallium) is in the middle, DPa (dorsal pallium) is on top, and LPa (lateral pallium) on the side.
The above diagram shows the quadripartite model of the ancestral vertebrate forebrain [Hegarty et al 2024], [Pessoa et al 2019]. The dorsal areas are cortical-like and the ventral (basal) areas are striatal and pallidal (basal ganglia). The four areas match most vertebrate structure. MPa (medial pallium) is hippocampal, DPa (dorsal pallium) is sensory neocortex, LPa (lateral pallium) is the insula cortex (eating and tasting), E.lec (lateral entorhinal) and F.ofc (orbitofrontal), and VPa (ventral pallium) is the olfactory cortex and amygdala. In the diagram, the notch at the top is deliberate because the neural tube is formed by curling up a neural sheet until the two ends nearly match at the top. In most vertebrates the two ends at MPa curl more until they reach the base, forming two hemispheres, but in teleost fish the two ends curl out, putting MPa on lateral outside (Dl in teleosts) and VPa in the middle (Dm in teleosts) [Hegarty et al 2024], [Roth and Dicke 2013], [Porter and Mueller 2020].
Although this model is useful, it does minimize differences between vertebrates. For example, the mammalian DPa, the “neocortex,” doesn’t nicely match up with reptiles, instead the amygdala-like areas (DVR) take a larger role.
Unrolled hippocampus, showing the consistent order among amniotes. C (isocortex/neocortex), DPa (dorsal pallium), E.ca1, E.ca2, E.ca3 (hippocampus CA1, CA2, CA3 areas), E.dg (hippocampus dentate gyrus), E.ec (entorhinal cortex), E.sub (hippocampus subiculum).
In amniotes (lizards, birds, and mammals) E.hc has a similar structure, including E.dg (dentate gyrus), E.ca1, E.ca3, E.sub (subiculum), E.lec (lateral entorhinal cortex), and E.mec (medial entorhinal cortex) [Medina et al 2017]. As mentioned previously, although the fish Dl is likely an E.hc-like structure, its internals differ from the amniotes. If the fish Dl was a subset of the mammal E.hc, it would be useful to know which areas are more primitive.
The lamprey is interesting because it may not have an E.hc / MPa equivalent at all. A recent genetic cell analysis [Lamanna et al 2023], [Hervas-Sotomayor 2023] suggests that MPa is not ancestral because the medial pallial area in lampreys matches H.em (prethalamic eminence) instead, which is more closely associated with the hypothalamus and habenula, not the forebrain and specifically not matching hippocampal markers. In addition, the lamprey extended amygdala is distinct, well-defined, and separate from the lateral pallium, suggesting it may be more helpful to treat the extended amygdala as a distinct area instead of as part of the pallium. In the case of mammals, that distinction may not apply to A.bl (basolateral amygdala), which resembles the cortical DPa more than other parts of the amygdala [Moreno and González 2007]. [Lammana et al 2023] suggest that the quadripartite model may only apply to later vertebrates, and lampreys having an undifferentiated pallium, an amygdala, and a H.em with unknown purpose. These genetic results reinforce earlier suggestions that the lamprey “MPa” is actually an expended H.em.
For this study, which only uses the hippocampus / MPa, the lack of a lamprey MPa suggests that the sequential hippocampal model is a later vertebrate development. Even if accurate, it would be unknown if it developed specifically with jawed vertebrates (gnathostomes) or if it was an innovation ozone of the many preceding jawless vertebrates.
C.pp navigation delay
In mice, C.pp (posterior parietal cortex) tracks time and distance from a start position with neurons that tile the delay period, as shown in the earlier neuron firing diagram. Each neuron is only active for a fragment of the task, and as one neuron leaves the ensemble, another starts [Harvey et al 2012]. In their task, C.pp was required for memory related success, not needed for cue seeking.
C.pp is “vision for action” [Goodale and Milner 1992]. C.pp tuned to referred self motion and acceleration [Whitlock et al 2012]. In this context, the timing we need it motion timing, so self motion is critical. Some C.pp neurons fire up to 500ms before action and drop immediately on action.
Although these locomotion delays match what the essay needs, Cpp has a number of problems for this purpose. First, C.pp is part of the neocortex, which is specific to mammals. Birds, reptiles, and fish have a cortex (called pallium), but the area around C.pp is organized differently with a larger amygdala area (DVR) than cortical (dorsal pallium) [Aboitiz et al 2003]. Mammalian E.hc as an odor-motor region with 1D maps [Aboitiz and Montiel 2015].
Second, the most direct subcortical C.pp output is to OT (optic tectum), but otherwise C.pp requires C.mo (motor cortex) for any action. But in amphibians, the OT does not receive input from a cortical equivalent, but connects directly with the striatum [Pessoa et al 2019]. In the context of the essay, the OT doesn’t seem related to the place avoidance needed.
Third, the C.pp inputs are primarily indirect through other cortical areas, although using T.lp (lateral posterior thalamus) from OT. So, if C.pp is a primitive cortical area, it seems much more likely to be an OT-focused area, not a foraging area.
As a counterpoint, the electric fish Dc also has direct output to OT, which suggests a C.pp-like role. Dc is driven by the Dl (lateral cortex, possible E.dg) through DD (dorsal cortex, as possible E.ca3) [Fotowat et al 2019]. That study suggested the entire Pa (cortex) area is silent unless actively sensing and sequences were not studied. Other research suggests that only Dl.v (ventral Dl) is equivalent to E.hc, particularly to E.sub [Hegarty et al 2024]. [Rodríguez-Expósito et al 2017], which makes the comparison more tenuous, unless DD is something like C.rs (retrospenial cortex), which in mammals is between E.hc and C.pp.
Similarly, the lamprey LPA (lateral pallium / cortex) has a direct OT projection [Suryanarayana et al 2022], giving it a possible C.pp-like role, but the lamprey does not appear to have a hippocampus equivalent [Lamanna et al 2023]. It does have an expanded H.em (prethalamic eminence) which is physically located where the MPa (medial pallium / hippocampus) would be. Unfortunately, no current studies have covered H.em function in the lamprey, but it seems likely to be significantly different from the hippocampus.
In favor of C.pp as relevant here, a study of electric fish shows a path from Dl (E.dg – dentate gyrus equivalent) to Ddi (E.ca3 equivalent) to Dc (C.rs.5/C.pp.5 equivalent) to OT [Fotowat et al 2019]. However, in that study the entire Pa is silent except when swimming or active sensing (electro-sensing in this case), and it’s unclear if there are any sequences as in the mouse C.pp.
A different description of Dl points out that it primarily projects to Poa and Hc (caudal hypothalamus) [Northcutt 2006]. The fish Hc .sum set of efferents is that Dl projects to the area that includes H.arc, H.mb, H.sum, H.pm, and H.tu, which are developmentally conserved between fish and mammals [Wullimann 2022]. Because these areas have different functions, it would be useful to known exactly which specific areas Dl projects to. In mammals, Poa and H.sum are interconnected and are both associated with exploration and roaming [Escobedo et al 2023], [Ryoo et al 2021].
H.sum
As mentioned previously, H.sum is part of the basal hypothalamus and is highly conserved, existing in sharks [Santos-Durán et al 2022], teleost fish [Wullimann 2022], amphibians and reptiles [Domínguez et al 2016], birds [Kim DW et al 2022], and mammals [Bedont et al 2015], [Croizier et al 2015], [Ferran et al 2015]. I haven’t found an equivalent lamprey hypothalamus study, which would be especially interesting because of the lamprey’s lack of MPa / E.hc equivalent. If H.sum does exist in lampreys, its function and connectivity might show a simpler, more primitive function before adding MPa / E.hc capabilities.
H.sum has several distinct circuits, some with overlapping behavior. H.sum tac1 (Substance-P marker) is highly associated with voluntary locomotion, but not with E.hc theta [Farrell et al 2021]. H.sum is associated with active exploration and produces RTPP (real-time place preference) but is quiet while eating [Kesner et al 2021]. H.sum to Poa (preoptic area) projections are associated with avoidance to threads and shows a strong RTPA (real-time place avoidance) but not anxiety or CPA (conditioned place avoidance) [Escobedo et al 2023]. The H.sum to Poa connection has collaterals to E.ca2 but not to Hb.m or E.dg (hippocampus dentate gyrus). H.sum stimulation inhibits eating and its inhibition is required for eating. The projections to E.dg are related to object novelty and open field exploration [Pan et al 2004], [Chen S et al 2020], while the projection to E.ca2 is related to social novelty and temporal memory [Chen Z et al 2022], [Thirtamara et al 2024]. H.sum is also part of the food reward circuit and is reinforcing [Ikemoto 2010].
E.hc – hippocampus as sequence generator
One model of E.hc (hippocampus) is as a general sequence generator [Buzsáki and Tingley 2018]. The idea notes that E.hc is blind to its inputs, whether olfactory, visual, or vestibular, but what is consistent it its ability to tile gaps between events, and its internal timing structures such as fitting gamma (40-100Hz) bursts inside theta cycles (~8Hz), such as a consistent seven-ish game cycle bursts within a single theta cycle.
E.ca2 – hippocampal CA2
E.hc is a complicated structure with many parts, but must have evolved from an initial core area. Studies with other vertebrates show comparable areas to the hippocampus E.dg (dentate gyrus), E.ca1 (CA1 area), E.ca3 (CA3 area), and E.sub (subiculum area).
Unrolled hippocampus
For this essay, I’m considering E.ca2 as the most primitive because it’s strongly tied to sequence generation [Bhasin and Nair 2022], [He et al 2021], [Lehr et al 2021], [MacDonald and Tonegawa 2021], [Stöber et al 2020] and is reciprocally connected with H.sum, which I’m already using for its connection with RTPA and exploration. E.ca2 is reciprocally connected with H.sum in the early embryonic state, earlier than E.dg and E.ca3 connections [Diethorn and Gould 2023].
In mice, E.ca2 sustains internal sequence memory during delay [Lehr and Stöber 2021]. When E.ca2 is disabled, sequences are destabilized in E.ca1 [Lehr et al 2021]. Interestingly, during sleep E.ca2 appears to fire consistently to remember the animal’s current location [Kay et al 2016].
E.ca2 is distinguished from E.ca3 by its lack of E.dg input [Insausti et al 2023], making it an interesting candidate for an early area because it has fewer dependencies and requirements.
Neural models for sequences and persistence
Neural models for sequences are much less studied than models for persistence and memory. I’m using sequences for a delay period and delay periods are well studied. In behaviorist experiments, “trace conditioning” has a short delay between the cue to be remembered and the food reward or shock punishment.
Before approaching sequences, it seems best to consider persistence, which is better studied, such as persistent working memory. [Zylberberg and Strowbridge 2017] reviews multiple models of persistence. The main difference I’ll highlight is between models focused on intrinsic cellular responses such as bursting, in contrast with models that focus on recurrent connectivity, such as Hopfield networks [Hopfield 1982]. Experimentally, the evidence between cell-autonomous vs recurrent network is an open question [Kamiński and Rutishauser 2020].
Persistent bursts are 3-6 spikes [Zylberberg and Strowbridge 2017]. Epilepsy is related to these bursts and may be a failure in inhibitory circuits to contain expansion of these bursts. These bursts are enabled by metabotropic receptors such as mACh (acetylcholine metabotropic receptor). The hindbrain includes bursting neurons, such as VOR (vestibuloocular reflex), long plateau potentials in spinal motoneurons with Ca2+ (calcium) channels and with K+ (potassium) channels. These bursting mechanisms can rebound after inhibition.
Recurrent models of persistence
Recurrent models of persistence use feedback, recurrent connections of neurons back to the same area. If properly calibrated, these loops produce produce attractors that can remember data [Hopfield 1982]. A difficulty in recurrent models is the issue of fast neural circuits (10-20ms) extending to seconds, and how to maintain stability when 100-fold timescales are required [Zylberberg and Strowbridge 2017]. One common response is to use longer time circuits such as NMDA receptors with 100ms time constants. The longer basal constants reduce the stability issues. For example, one of the models for recurrent storage uses NMDA in a model of the VOR (vestibular-ocular reflex) [Seung 1996], where the eye target is maintained in a line attractor.
Criticisms of recurrence and attractors note the requirement for fine tuning [Zylberberg and Strowbridge 2017]. Continuous attractors require fine turning. For example, line attractors require precise recurrence, while noise and modulation affects the tractors. In addition, the timescale of excitation vs inhibition is important to the stability. In addition neurons tend to switch abruptly between discrete states as opposed to smooth, continuous variation. To reduce the issue, finely spaced concrete attractors are more tolerant than continuous attractors.
Continuous attractors use a smooth neuron rate encoded values, but firing rate models difficult to match actual neuron behavior [Compte et al 2000]. Spontaneous activity at 3.5Hz does not trigger attractor. [Lundqvist et al 2016] notes that individual neurons bridging a multi-second memory delay is rare, but a simple attractor model would have multiple neurons bridging the entire gap. [Cui and Strowbridge 2018] also note there is little experimental support for recurrent synaptic reverberation.
Intrinsic neural persistence
Intrinsic neural persistence models focuses more on cellular mechanisms for persistence, where an individual neuron fires for longer times, such as 2s without requiring external prompting like recurrence. Since the Lundqvist criticism [Lundqvist et al 2016] also applies here, that single neurons do not cover the entire gap, intrinsic cellular persistence alone is insufficient. However, extending the time constant from the 100ms of NMDA to 500ms or even multiple seconds reduces the need for precise calibration.
Persistent firing exists in C (cortex) and Ob (olfactory bulb) using cellular mechanisms, initiates by a burst of AP (action potentials) [Cui and Strowbridge 2018]. Many cellular persistent firings are enabled by m1.q (acetylcholine metabotropic receptor tied to G-q/11 proteins), and additionally require cellular Ca2+ (calcium ions) for bursting, triggered by the input bursts. ACh enhances excitability by reducing ERG (ether-a-go-go) K+ (potassium currents). The ERG K+ currents typically suppress bursting and persistence. By suppressing ERG K+, the ACh to m1.q path disinhibits bursting and persistence. ERG is highly expressed in deep C (layer 5/6 cortex), Snc DA (substantia nigra compacta dopamine), hindbrain, and E.ca1. ERG blockers abolish persistent firing in C.tea (temporal association area) and F.pfc (prefrontal cortex).
In the Cui and Strowbridge study [Cui and Strowbridge 2018], ACh enables persistence for at least 5s at 4.3Hz to 6.4Hz. With normal, non-ACh ERG behavior, input bursts produce a short AP burst followed by suppressed output. When ACh blocks ERG currents, input bursts produce an AP burst followed by persistent firing that requires Ca2+. In short, ACh enables persistent firing in the range of 500ms to 2ms for a single neuron without needing external recurrence.
Other studies show similar modulated persistence in other areas. In E.sub, ACh stimulation of m1.q enables extended plateau potentials [Kawasaki et al 1999], allowing sustained tonic firing. In layer 5 cortex, when ACh enables persistent firing, even the smallest apical depolarization produces L5 repetitive bursts [Schwindt and Crill 1999]. The plateau potential and prolonged response is over 400ms. Similar effects for ACh and m1.q promote sustained persistence in CB (cerebellum) and O.mt (olfactory bulb mitral cell) [Bauer and Schwarz 2018]
Tiling and threading
Combining longer persistent intrinsic cellular activity with recurrence provides many options for persistence and sequence generation. As mentioned above, mouse C.pp neurons in a virtual maze tiles a delay period [Kamiński and Rutishauser 2020]. Each neuron has an intrinsic persistent between 500ms and 1500ms, and only a sparse number like 10% are active at a time [Pastalkova et al 2008]. As one neuron ends its extended 500ms activity, another neuron takes its place. When the animal acts, the sequence is extinguished [Zylberberg and Strowbridge 2017].
One component of the sequence control is the E/I (excitation / inhibition) balance, managed by cortical interneurons. The E / I balance prevents runaway excitation, as in epilepsy, and instead restricts activity to a small subset of neurons. This balance is often oscillatory and combined with winner-take-all systems, where the first neurons to fire suppress slower neurons. In the case of sequences, when one neuron stops firing when its internal persistence ends, the E / I balance shifts towards allowing a new neuron to take its place.
A synfire chain, where each neuron triggers the next neuron in the sequence.
An older model of sequences is the synfire chain, where each neuron triggers the next neuron in the sequence. An argument against the strict synfire chain its developmental plausibility, because it requires precise connectivity.
Using excitation / inhibition balance with semi-random connectivity to allow dynamic sequences. The blue neuron is currently active, and the faded blue neurons are possible next steps in the sequence.
Instead of a precise synfire chain or the precise dynamics required for recurrent systems like [Hopfield 1982], [Rajan et al 2016] proposes a developmental model for sequences. Consider a set of neurons randomly connected, and modify only some of the connections to produce sequences. Neuron development includes calibration and pruning during early development, including intrinsic firing that would allow for sequence selection. Once the network is formed, sequences propagate by cooperation between recurrence and the input.
Adding sequence projection to a readout layer for sequence output
Because sequences are dynamic, their output may require a projection layer to a more meaningful subspace [Kamiński and Rutishauser 2020]. For example, this essay only needs a single active signal to continually drive the avoidance until the sequence ends.
Illustration of a sequence managed by the E / I balance of interneurons and read-out by a projection layer.
The diagram above shows how these ideas fit together. The pyramidal neurons in the middle form the backbone of the sequence. Each neuron has an intrinsic persistent firing for ~500ms. The PV (parvalbumin) interneurons prevent more than one neuron firing. Because the current neuron is already firing, it blocks the next neuron in the chain until the first neuron completes. The sequence is decoded by a separate output layer, such as S.ls (lateral septum) or E.sub (subiculum) before sending to the destination.
Simulation choices
Taking the above research into consideration, the simulation needs to make some choices of a plausible action path. Because this essay builds on the previous essay 31, which used S.v to Hb.l as a seek timeout, it makes sense to continue that path from Hb.l to P.ms and E.hc via V.mr and H.sum.
Action path for the avoid timeout used by the simulation. E.hc (hippocampus), H.sum (supramammillary), Hb.l (lateral habenula), P.ms (median septum), Poa (preoptic area), S.ls (lateral septum), S.v (ventral striatum / nucleus accumbens), V.mr (median raphe)
The output side has two tempting directions: E.hc to Poa and H.sum via S.ls, or E.hc to Hb.m-R.ip via S.ts and S.bac. The E.hc to Hb.m-R.ip has an advantage of exploring differences between Hb.l and Hb.m. One model is Hb.l as the motivational avoid and Hb.m as the physical taxis avoid, where evolution secondarily adds motivational Hb.l avoid as a more sophisticated indirection to the older Hb.m direct avoid taxis. The other option using S.ls to Poa and H.sum has the advantage of possibly being more primitive, because the P.ts / P.bac to Hb.m path may be restricted to mammals and therefore newer. The S.ls path also has the advantage of introducing the strong E.hc to hypothalamus connection via S.ls.
Sequence design
Because the simulation is a toy model where understandability is important, the simulation doesn’t directly implement individual neurons. One advantage is that neuron firing is generally sparse. For example, only about 10% of O.pir (olfactory cortex) neurons fire for any particular odor. Taking an idea from high-dimensional computing [Laiho et al 2015], an engram (neural assembly) can be represented by a vector of digits. Each digit is essentially one sparse neuron selected from a larger set. Base-16 is one neuron selected from 16. This digital representation can be displayed as a normal hex number such as 78af. For display purposes, base-64 is the largest feasible base, because its numerical digits plus lower and upper letters produce 62 with two special digits. Zero is reserved to represent no neurons firing in the digit. So 4305 represents 3 neurons firing, one from each of three active digits with the last digit silent.
In natural sequences, after one neuron falls out, a new random neuron takes its place. In simulation, this randomness adds complexity because to avoid repetition it would need to keep track of earlier values. For neurons this repetition avoidance is built-in because neurons can have a refractory period where they’re less responsive, giving fresh neurons a better chance to be chosen. Instead, the simulation reserves the lowest bits of each digit for sequences, leaving the highest bits for the random engram. The sequence also updates each digit sequentially instead of randomly. Zero represents the end of the sequence when all neurons stop firing. A sequence with a single base-16 digit might look like “8, 9, a, b, 0” or “4, 5, 6, 7, 0” if two bits are sequence bits, and a sequence with two digits might look like “84, 85, 95, 96, a6, a7, ....“
Simulation results
The following diagram shows a sample trajectory where the animal first seeks to the center of the odor plume and spins at the center until the S.v striatum adenosine times out as in essay 31. The timeout triggers a motivational avoid, representing V.mr and H.sum via Hb.l. The motivational avoid starts the E.hc sequence via P.ms, which activates the physical avoid representing either S.ls to Poa or Hb.m to R.ip depending on interpretation. When the sequence ends, E.hc stops driving the physical avoid, and the animal resumes its roaming search.
Simulation of the sequence avoid. The animal has just timed out of a failed search and is currently moving forward, avoiding the previous area.
Discussion
This essay raises a number of issues, mainly because the situation is understudied, and because it involves many regions that are generally studied independently. As a reminder, this essay is at best a thought experiment.
The lamprey’s lack of a MPa / E.hc (medial pallium / hippocampus) is the first interesting issue. H.em (prethalamic eminence) in the expected MPa location is also interesting. What does H.em do? In mammals it’s a progenitor area that disappears, and it produces neurons for P.epn (endopeduncular nucleus) and Cajal-Retzius neurons that populate neocortical layer 1 and organize the neocortical layers. But in lampreys the H.em functionality may be entirely different. It its function an any way similar to a primitive E.hc? Does the lamprey lack all internally generated sequences?
In addition, what happens in lampreys to all the areas connected with MPa / E.hc? Does H.sum and H.mb (mammillary body) exist in lampreys, and if it exists, what does it do? Similar questions for V.mr R.nin (nucleus incepts) and S.ls and P.ms. If E.hc is missing in lampreys, studying those areas might give more information about their function.
More Zebrafish studies about the Dl / E.hc would also be very interesting, because it seems to have a different structure from the common amniote E.hc. Consider the hunger study with the reciprocal activation of Hc (caudal hypothalamus) and H.l (lateral hypothalamus) when the animal detects a food cue [Wee et al 2019]. Because Hc is composed of H.sum, H.mb, H.tu and H.arc, which have distinct functions, learning exactly which areas are activated by hunger and suppressed during seek would be interesting. H.sum and H.mb are also understudied in zebrafish. Do they have similar connectivity to Dl (MPa / E.hc) and function as in mammals? Does Dl have similar sequential functionality? It does have SWF (sharp wave ripples) [Blanco et al 2024], which suggests zebrafish has Dl sequences, but does it have the same threading / tiling of delays?
Although this essay has used E.ca2 as the center of E.hc sequences, but E.ca2 has only been identified in mammals. Genetic transcription comparisons with other animals suggests E.ca3 as better conserved. If E.ca2 is a mammalian innovation, how does that affect how E.hc works in non-mammals?
Chen S, He L, Huang AJY, Boehringer R, Robert V, Wintzer ME, Polygalov D, Weitemier AZ, Tao Y, Gu M, Middleton SJ, Namiki K, Hama H, Therreau L, Chevaleyre V, Hioki H, Miyawaki A, Piskorowski RA, McHugh TJ. A hypothalamic novelty signal modulates hippocampal memory. Nature. 2020 Oct;586(7828):270-274.
Farrell JS, Lovett-Barron M, Klein PM, Sparks FT, Gschwind T, Ortiz AL, Ahanonu B, Bradbury S, Terada S, Oijala M, Hwaun E, Dudok B, Szabo G, Schnitzer MJ, Deisseroth K, Losonczy A, Soltesz I. Supramammillary regulation of locomotion and hippocampal activity. Science. 2021 Dec 17;374(6574):1492-1496.
Roth, G., Dicke, U. (2013). Evolution of Nervous Systems and Brains. In: Galizia, C., Lledo, PM. (eds) Neurosciences – From Molecule to Behavior: a university textbook. Springer Spektrum, Berlin, Heidelberg.
In an earlier essay that covered tunicates, the tunicate larva has two distinction visual action paths, one for phototaxis and one for looming. The two paths use different photoreceptors. Phototaxis photoreceptors are directional with pigment cells blocking light from one direction, while dimming photoreceptors are unidirectional with no shadow from pigment cells.
Looming and dimming are signals of both predators above the animal that block light from the sky, and of obstacles, which also blocks out light from the sky as the animal nears the barrier. In this essay I’ll be focusing on obstacle avoidance using a similar simulation approach as [Zhao et al 2023]. In general, the sky is the brightest, the ground is also light, such as sand, and obstacles are darker. So, if the eye is next to a barrier the average light is dim, while if it’s far from the wall the light is bright because the sky above and the lighter ground below are unobstructed.
Views from the left and right eye for a barrier to the left of an animal. The left is darker because it’s close to the wall, and the right is brighter because it has a clear horizon.
The above screenshot shows a fish-like create with a wall to its left. The left eye is next to a wall, and the right eye views the open field. If the image is reduced to a single average value, the left eye is dimmer, while the right is almost as bright as a full open field. As the first approaches the wall, the image dims rapidly.
Tunicate ascidian dimming
The tunicates (ascidian sea squirts) are the closest non-vertebrate chordates, although evolution has optimized them by removing features, making it difficult to draw direct comparisons to vertebrates [Holland 2016]. The ascidians have a simple larva form that swims for less than 24 hours before settling and becoming a sessile filter feeder.
The ascidian larva has a single ocellus (simple, non-image-forming photoreceptor area) has two distinct photoreceptor types and corresponding action paths, one that produces phototaxis and another that responds to rapid dimming [Ryan et al 2016].
Ascidian nervous system for both phototaxis action path and dimming escape. PR-1 is the directional photoreceptor for phototaxis. PR-2 is the unidirectional photoreceptor for looming escape. mgIN-L and MN-L are motor neurons. AMG is ascending motor feedback.
The above diagram of part of the ascidian larva nervous system, the PR-1 photoreceptors are directional for the top phototaxis path, while PR-2 non-directional photoreceptors produce dimming. The boxes above represent individual neurons, not larger functional groups. MgIN and MN are motor control and motor neurons [Ryan et al 2016].
Larval lamprey primitive eye
Lampreys and hagfish are the only remaining agnathans (non-jawed vertebrates), representing a much larger agnathan vertebrate group that preceded the jawed vertebrates, most of which were filter feeders or sediment feeders [Mallat 2023]. Lamprey larvae are unique among vertebrates in having a primitive non image-forming eye, more like the ascidian ocellus [Bayramov et al 2022]. The adult image-forming retina expands in rings around the more primitive center [Barandela et al 2023].
This central primitive eye is responsive to dimming, and it projects to an equivalent M.pot (pretectum), which handles several optical action paths in zebrafish, including dimming responses, phototaxis, OMR (optomotor reflexes), OKR (optokinetic reflex), and hunting.
Zebrafish retina arborization fields
The zebrafish N.rgc (retina ganglion cell) projects to ten distinct AFs (arborization fields), each with a distinct purpose, from AF1 to H.scn (suprachiasmatic nucleus) for circadian timing to AF10 to OT (optic tectum) [Baier and Wullimann 2021]. In most cases distinct N.rgc neuron types project to distinct arborization fields. Even with the largest field AF10 for OT, individual N.rgc neurons project to distinct OT layers. The temporal phototaxis of a previous essay used the projection to AF4 to thalamus to Hb.m (medial habenula, dorsal Hb in zebrafish) [Cheng et al 2017], [Chen and Engert 2014].
The diagram above shows the zebrafish arborization fields and their targets, although the function of many of the targets is not fully known. Dimming fields include AF6, AF8 and OT [Baier and Wullimann 2021], [Temizier et al 2015]. It seems likely that AF6 and AF8 have distinct functionality, although the distinction is not yet known. In lamprey the central ocellus-like photoreceptors project to M.pot, while the outer, lateral image areas project to OT [Cornide-Petronio et al 2011]. The OT dimming response is directional, dimming in one side produces turning [do Carmo et al 2018].
Optic tectum
OT (optic tectum) has the largest arborization and it is the largest nucleus in the midbrain, larger than the entire zebrafish forebrain (cortex / basal ganglia). The optic tectum responds to looming objects and in zebrafish is used for visual hunting [Liu et al 2022]. Since the early vertebrates were filter feeders, the hunting functionality would be unnecessary, leaving obstacle and predator avoidance.
Optic tectum comparison between zebrafish and mammals. C (cortex/pallium), M.pag (periaqueductal grey), M.tl (torus longitudinal), N.rgc (retina ganglion cells), OT (optic tectum)
The optic tectum is layered with retina information arriving in the superficial layers, integrative information from other senses in intermediate layers, and motor actions from the intermediate and deep layers.
The above diagram shows a rough correspondence between zebrafish and mammal optic tectum layers. For simplicity, I’ll use the mammalian names. It’s not clear to me if the PV layer (periventricular layer) of zebrafish is equivalent to M.pag.d (dorsal periaqueductal grey), but I haven’t read any study addressing the physical similarity either as homologous or non-homologous, so it’s probably best to assume the location similarity is merely coincidental.
In zebrafish, the OT.s (superficial grey layer, SFGS in zebrafish) itself is layered with each layer receiving distinct N.rgc input [Liu et al 2022]. Dimming input goes to the deepest layer of OT.s [Temizier et al 2015], which is used by OT.d for looming responses [Heap et al 2018]. In mammals, OT.s receives retina input, OT.i produces turn actions, OT.d produces seek and avoid actions, and M.pag.dm produces fast escape from predators.
In vertebrates, OT uses visual expansion (looming) in combination with dimming [Nakagawa and Hongjian 2010], and dimming by itself does not trigger escape [Dunn et al 2016]. However, in the context of the essay’s simulation of a more primitive animal, expansion requires more sophisticated processing from an image-forming eye, which is only available for later vertebrates and not available to even the larval lamprey.
Torus longitudinus
In teleosts (most bony fish), M.tl (torus longitudinal) is a unique nucleus between the left and right OT. M.tl averages the dimming value between the right and left [Folgueira et al 2020]. It also has a sustain role, maintaining behavior after an initial signal.
Torus longitudinus between both OT. M.tl (torus longitudinal), OT (optic tectum).
Interestingly, M.tl is a CB-like (cerebellum-like) structure [Folgueira et al 2020]. Other CB-like ares such as MON (CB-like for LL) and DON (CB-like for electro sensation) act like adaptive filters for the lateral line to cancel out self-motion effects from sensors [Bell et al 2008], [Montgomery et al 2012].
Note that M.pot also communicates with its opposite side through the posterior commissure [Suzuki et al 2015], which could resemble an ancestral visual system. So, although M.tl is directly relevant to the looming response in zebrafish, it may be a specific teleost system, not an indication of an ancestral architecture.
nMLF optical motor output
The zebrafish reticulospinal motor control neurons are divided into several groups with distinct action paths. Optical motor output uses M.nmlf (nucleus of the medial longitudinal fasciculus), a midbrain reticulospinal group composed of 20 neurons on each side [Severi et al 2014]. M.nmlf avoidance is distinct from the Mauthner cell startle circuit in r4 in R.mrs. Although the OT looming / dimming can trigger the startle response [Temizer et al 2015], it generally uses the lower-priority M.nmlf [Bhattacharyya et al 2017].
nMLF as the output of the dimming/looming response. AF (retina arborization field), M.pot (pretectum), N.rgc (retina ganglion cell), N.sp (spinal cord), OT (optic tectum).
This direct OT to M.nmlf projection applies to early zebrafish larva. As the fish ages, OT adds projections to R.mrs (middle reticulospinal) in r4-r6 of the hindbrain [Barandela et al 2023], including turning neurons marked by chx10 [Cregg et al 2020]. For this essay, I’m using the simpler early projection to M.nmlf.
Looming can produce zebrafish O-bends (u-turns) as well as directional turns [Portugues and Engert 2009], [Marques et al 2018]. For this essay, I’m assuming that M.pot produces a base O-bend command that the OT can modify by choosing a turn direction. This split between motivation and turning also occurs in R.mrs, where MLR (midbrain locomotive region) produces a non-directional forward movement, while chx10 neurons in R.mrs receive OT turning commands for looming [do Carmo et al 2018], [Cregg et al 2020].
Simulation
This essay’s simulation uses dimming as an obstacle avoidance system, similar to the simulation in [Zhao et al 2023], but with a minimal dimming input. The essay’s simulation condenses the input to the simplest dimming structure, where each eye has only a single averaged luminance value. The retina also calculates a dimming value as the difference between the current luminance and the previous value. Although the vertebrate retina uses distinct unsigned ON and OFF channels, the simulation uses a single signed value.
The looming module triggers a looming response when the dimming value passes a threshold as a proportion of the current luminance. This part of the model represents M.pot (pretectum). If no further information is available, the looming triggers a u-turn (O-bend in zebrafish) using M.nmlf.
If the left and right eyes have a difference in brightness, the model converts the u-turn into a left turn or right turn. This part of the model represents the OT’s dimming response. Like the M.pot output, this OT turn signal uses M.nmlf, as in the early zebrafish larva.
Screenshot showing the animal avoiding a wall to its left. The left and right retina displays are for human viewing.
The above screenshot shows the animal avoiding an obstacle to its left. The two low-resolution images at the lower right are for human viewing and are higher resolution than the animal uses. The animal itself only uses a single averaged value for each eye. This view from the left eye is dominated by the wall, which blocks the light. The right eye mostly sees a clear view to the horizon.
Discussion
Qualitatively, the system works surprisingly well despite its simplicity. In some of the narrow corridors the u-turn behavior will reverse out of the corridor, and the entrance to the corridors is something of a barrier because only the center of the corridor will avoid triggering avoidance.
The model doesn’t adjust speed, which is an interesting potential improvement. If the animal slowed near obstacles, raised the threshold for obstacle avoidance, and reduced the turn angles, it might more easily navigate corridors. Since searching already has a roam vs dwell mode for ARS (area restricted search), triggered by serotonin, a slow-moving obstacle avoidance mode could use the same mechanism. V.dr (dorsal raphe serotonin) does reduce looming defense [Huang et al 2017]. Alternatively, since OT.d looming does habituate [Lee et al 2020], that habituation could reduce the excessive u-turning of the model. H.lgn.v (ventral lateral geniculate nucleus), which responds to overall light levels, can also inhibit the looming response [Fratzl et al 2021].
When seeking an odor, vertebrate swimming undulates left and right, naturally moving the nose perpendicular to the body motion. This lateral motion can help navigation if odor sampling can be coordinated with the movement, enabling a spatiotemporal gradient calculation along the path of the nose movement. This lateral sampling over time is called klinotaxis (“leaning navigation”) or weathervaning.
Essay 24 and essay 25 explored head-direction navigation as inspired by the fruit fly Drosophila fan-shaped body and ellipsoid body. The idea was to use head direction to translate egocentric movement into an allocentric memory of past samples, independent of the current body direction. In contrast, klinotaxis uses an egocentric system, where the lateral motion is relative to the current direction, not an independent, compass or map-like system.
Klinotaxis in Drosophila larva and C. elegans
Klinotaxis has been largely studied in the fruit fly Drosophila larva and the roundworm C. elegans. Drosophila larva have a distinct “cast” movement, where they pause and wave their heads side to side, either a single time (1-cast) or multiple times (n-cast) [Zhao et al 2017]. Larva movements break down into five major types [Gomez-Marin and Louis 2014]:
Forward
Backward
Stop
Turn
Cast
C. elegans has two major seek movements: pirouettes and weathervaning [Lockery 2011]. Pirouettes are a u-turn when the animal is moving away from the odor. Weathervaning is a side-to-side head movement that manages turning.
Both systems are temporal gradient systems, requiring measurements at different times and a memory of the older measurement [Chen X and Engert 2014]. Klinotaxis requires a basic form of memory [Karpenko et al 2020], but the comparison can be a simple ON or OFF result [Lockery 2011]. Pirouetts use a gradient parallel to body motion and reverse direction when the animal is moving away from the odor [Iino and Yoshida 2009]. Weathervaning uses a gradient perpendicular to body motion, measured with a lateral head movement [Lockery 2011].
This klinotaxis contrasts with a bilateral spatial navigation that compares two lateral sensors [Chen X and Engert 2014], such as bilateral eyes, ears, or nostrils. In Drosophila larva, odor turning is proportional to the lateral gradient more than the parallel gradient [Martinez 2014]. The odor navigation is not simply bilateral because disabling one side of O.sn (olfactory sensory neuron) only minimally impairs navigation [Gomez-Marin and Louis 2014].
As a slight digression, let’s return to the adult Drosophila navigation, because the structure can be a useful analogy for understanding vertebrate klinotaxis navigation, despite using a different allocentric system.
Adult Drosophila FSB
Below is a rough sketch of the Drosophila navigation circuit, focused on the fan-shaped body [Hulse et al 2021]. The ellipsoid body (EB) and protocerebral bridge (PB) calculate head direction and sort it into 18 columns. This head direction is allocentric, independent of the animal’s current direction, like a compass direction or a map. Input from odor areas like the mushroom body (MB) and lateral horn (LN) are organized into 9 rows. The fan-shaped body combines these 18 head direction columns and 9 sense data rows into a memory table.
Drosophila navigation, focusing on head direction from PB, odor data from MB and LH, and allocentric table of FB. EB ellipsoid body, FB fan-shaped body, LH lateral horn, MB mushroom body.
Motor navigation reads out from the fan-shaped-body table. These motor commands include left and right, but also include a separate u-turn command [Westeinde et al. 2022]. Although this allocentric navigation system differs from egocentric klinotaxis, its motor output includes both the left vs right from weathervaning and the u-turn from pirouette.
The previous essay 24 and essay 25 attempts followed this model. As the animal moves in space, the model saved the forward odor gradient according to the current head direction. By comparing stored values for other head directions, the animal would improve its heading toward the direction with the strongest odor.
The fan-shaped body then becomes a record of samples of all the older directions that the animal had measured. Output is then calculated for left (PFL3L), right (PFL3R), and u-turn (PFL2) signals. [Westeinde et al 2024]. The current head direction is represented as a sinusoidal neural pattern and combined with the stored values to produce an output.
This system was only partially successful for the essay. Although it was an improvement over no memory, because the animal was continually moving in space, the table was always obsolete. Even when the table memory times out to represent loss in accuracy as the animal moves, the rapid obsolescence made navigation difficult, particularly as the animal neared the target.
So, this essay simplifies the circuit and lowers the ambition. Instead of trying to record every direction and keeping perfect allocentric compass direction, the animal could simple save its left and right oscillation as it swims naturally.
Vertebrate Hb.m and R.ip
The vertebrate Hb.m (medial habenula) to R.ip (interpeduncular nucleus) is used for phototaxis [Chen X and Engert 2014], Chemotaxis [Chen WY et al 2019] and thermotaxis [Palieri et al 2024]. In a clever experiment creating a virtual light circle, Chen and Engert shows that the zebrafish phototaxis is not simply comparing light between the eyes for a spatial gradient (tropotaxis) but is a temporally-based gradient (klinotaxis), relying on a short term memory of the previous light. This phototaxis uses the Hb.m to R.ip circuit [Chen X and Engert 2014].
Head direction from R.dgt (dorsal tegmental nucleus) tiles R.ip vertically [Petrucco et al 2023], while olfactory and light input is organized horizontally [Chen WY et al 2019], [Zaupa et al 2021]. After combining the odor with the head direction and comparing with the stored values, it sends motor commands to R.rs (reticulospinal) using P.ldt (laterodorsal tegmental nucleus) and V.mr (median raphe). The vertebrate R.ip has 6 columns of head direction input from R.dtg, resembling the Drosophila fan-shaped body, but instead of 18 columns for the fan-shaped body, R.ip only has 6, three to a side [Petrucco et al 2023].
Essay 25 explored a model which used the Drosophila fan-shaped body allocentric navigation in R.ip with some limited but not overwhelming success. Instead, this essay will try a different interpretation, where R.ip is only storing side to side weathervaning of the head while swimming, instead of a full 360 degree table like Drosophila.
Vertebrate klinotaxis
As a different approach, suppose the head direction to R.ip is not an allocentric map-making coordinator as in the adult Drosophila, but a simpler egocentric weathervaning or casting coordinator, storing only the lateral gradient from head direction changes from natural swimming, or possibly deliberate larger turns like casting to gather wider lateral gradient information.
Klinotaxis simplifies the need for precise head direction. Instead of the Drosophila 18 head direction columns calibrated to the outside world, we use only three, two lateral and one central, that only require motor efference copies of left and right muscle turns. Studies from the zebrafish R.ip suggest three columns to a side, which isn’t connected to the vestibular system [Petrucco et al 2023]. To me, this suggests to me that the head direction might not be an allocentric signal that requires precise direction, but a simple egocentric lateral measurement, which doesn’t need vestibular information.
The above diagram illustrates the system. Olfactory samples arrive through Hb.mand head direction arrives from R.dtg. Like the Drosophila fan-shaped body, R.ip combines odor samples with lateral head movement into a simple memory table, and it reads out left and right motor commands. A similar system can save odor measurements parallel to body movement, using velocity instead of head direction, to trigger a u-turn when the animal is moving away from the odor.
Discussion
Compared to the parallel-only gradient, allocentric system of essay 25, this lateral navigation is far simpler and more effective. Even with only three bins compared to the 8 bins in essay 25, the lateral weathervaning turned out to be more effective and less brittle. If R.ip does implement a lateral klinotaxis system like this essay, it’s plausible that the 6 directions reported by [Westeinde et al 2024] are sufficient for accurate seek navigation. In contract, those 6 directions seem insufficient for an allocentric navigation compared to the Drosophila 18 directions.
Interestingly, the pirouette also highly effective, even without lateral klinotaxis. In the simulation, when the animal moved away from the odor source, it makes a u-turn. This system served to ratchet the animal closer and closer to the target. Even when most of the movement was random, the pirouette locks in any improvement. Pirouette itself is also simple, only requiring two averages: a short average and a long average, where a short average tracks the odor across a single swim cycle and a long average uses two swim cycles. When the short average has a stronger odor value than the long average, the animal is moving toward the odor.
In both cases, the simulation used a binary OFF for the motor command instead of attempting finer precision from the gradient. This simple OFF strategy was sufficient for the simulation. A C. elegans study suggested that ON-OFF coding was energy efficient, and the worm rarely orients perfectly to the gradient [Lockery 2011].
The ascidian circuit in essay 30 had an interesting dopamine subcircuit that looks like an indirect search, where the ascidian coronet cells modulate the underlying phototaxis and geotaxis circuits. While the function of the coronet cells is unknown, if these cells are another seeking system like following an odor, then the coronet sub circuit follows odor by modulating different seek circuits: phototaxis and geotaxis.
Ascidian analogy
Tunicates are the closest non-vertebrate chordates evolutionarily, but they have developed in vastly different directions from the vertebrates, and likely very differently from the shared common ancestor [Holland 2015]. The ascidian tunicates, which are the most studied tunicates, live their adult life as sessile filter feeders like sponges. Their eggs hatch in only 20 hours and their brief tadpole form lasts only for a few hours, just enough to swim and disperse to find a likely permanent settlement place. Their locomotive strategy is to swim up using geotaxis in the morning and swim down using phototaxis in the afternoon. If they’re lucky enough to find a ledge, they swim up into the ledge’s shadow to settle because hanging like a bat from a ledge offers more protection from some predators than resting on the ocean floor [Zega et al 2006].
As would be expected from a 20-hour brain, the navigation circuit is fairly simple. There are two distinct action paths, one for geotaxis using a heavy pigment cell and one for phototaxis using photoreceptors and another pigment cell as a shadow to provide photo-directionality. The two action paths are connected, where dimming produces upward swimming [Bostwick et al 2020].
Ascidian tadpole sub circuit for geotaxis and phototaxis. The horizontal neurons are the main action paths. The coronet DA cells modulate the action paths.
In the above diagram, the geotaxis action path starts from the otolith (“ear stone”) receptor ant2, which is functionally similar to the vestibular system (but not related), passes input to antenna relay neurons (antRN) and then to the right side motor neurons (mgIN-R and MN-r) [Ryan et al 2016]. Similarly, the phototaxis action path starts from the ocellus (eyespot) to the phototaxis relay (prRN) and to the left motor neurons, providing an opposing direction from geotaxis. Importantly for the following discussion, each path has a weak connection to the opposite direction, possibly to add some stochasticity to the movement to improve dispersion of the many tadpoles.
The function of the coronet cells is unknown, although they have some genetic connection the palp sensory cells [Cao et al 2019]. Other papers compare the corona cells to dopamine cells in the hypothalamus and Ob (olfactory bulb) [Horie et al 2018] or ancestral photo-hypothalamus and retina [Sharma et al 2019], possibly related to the fish saccus vasculosus area of the hypothalamus, responsible for some circadian behavior. However, the ascidian tadpole has lost circadian clock genes, which argues against circadian timing [Chung et al 2023]. The coronet cells can accumulate serotonin and the DA might promote onset of metamorphosis [Razy-Kraika et al 2012]. So, the coronet may be involved in triggering metamorphic changes at twilight, which causes the tadpole to dive to deeper waters [Lemaire et al 2021].
Whatever the source, the interesting thing about the circuit is that it’s an indirect modulation of underlying taxis action paths. The action of the coronet is gating or modulatory. While this coronet circuit is not homologous to the basal ganglia, using it as an analogy may be useful. For example, dopamine is a sleep / wake signal for the basal ganglia [Vetrivelan et al 2010]. Because low dopamine reduces basal ganglia activity both at the striatum input layer and the Snr (substantia nigra pars reticulata) output layer, it’s an effective sleep controller.
Indirect chemotaxis
Consider indirect chemotaxis, where the animal is seeking toward the odor, but the underlying action path is phototaxis or geotaxis, like the ascidian circuit above. If the animal detects an odor, it increases the current direction. In other words, the current direction is toward or near a food odor. This strategy is like the e. coli tumble-and-run strategy, where the bacteria runs further when the odor gradient is increasing.
Consider the basal ganglia as an analogy. For example, Ob has some dopamine interneurons (Ob.sac – short axis cells) that project to S.ot (olfactory tubercle) [Burton 2017], a portion of the stratum focused on olfactory input. For the corollary of the phototaxis path, consider the Hb.m (medial habenula) phototaxis path [Zhang et al 2017].
Hypothetical indirect seek circuit where chemotaxis uses an underlying phototaxis to hunt for food. Hb (habenula), Ob (olfactory bulb), P (pallidum), R.ip (interpeduncular nucleus), R.rs (reticulospinal motor neurons), S (striatum), V.mr (median raphe).
When the odor is detected, Ob enables the basal ganglia, which enhances the phototaxis path. If the odor isn’t detected, the default semi-suppressed behavior means the direction is semi-random. This indirect control would allow for seeking odor when the underlying navigation is phototaxis and geotaxis.
Discussion
After writing this description. I think this model may be a bit sketch for something like chemotaxis, although it’s a reasonable model for sleep. Because I’m not sure the idea is likely to be productive, I’m holding off on doing any implementation, but writing down the description in case it makes sense later.
Let’s return to the task of essay 16 on give-up time in foraging, which covered food search with a timeout. At first the animal uses a general roaming search and if it smells a food odor, it switches to a targeted seek following the odor with chemotaxis. If the animal finds food in the odor plume, it eats the food, but if it doesn’t find food, it will eventually give up and avoid the local area before returning to the roaming search.
Search state machine. Roam is the starting state, switching to seek when it detects odor, and switching to avoid after a timeout.
For another attempt at the problem, let’s take the striatum (basal ganglia) as implementing the timeout portion of this task using the neurotransmitter adenosine as a timeout signal and incorporating the multiple action path discussion from essay 30 on RTPA. Adenosine is a byproduct of ATP breakdown and is a measure of cellular activity. With sufficiently high adenosine, the striatum switches from the active seek path to an avoidance path. These circuits are where caffeine works to suppress the adenosine timeout, allowing for longer concentration.
Mollusk navigation
As mentioned in essay 30, the mollusk sea slug has a food search circuit with a similar logic to what we need here. The animal seeks food odors when it’s hungry, but it avoids food odors when it’s not hungry [Gillette and Brown 2015].
Mollusk food search circuit, illustrating a hunger-modulated switchboard. When the animal is not hungry, the switchboard reverses the odor to motor links turning it away from food.
This essay uses the same idea but replaces the hunger modulation with a timeout. When the timeout occurs, the circuit switches from a food seek action path to a food avoid action path.
Odor action paths
Two odor-following actions paths exist in the lamprey, one using Hb.m (medial habenula) and one using V.pt (posterior tuberculum). The Hb.m path is a chemotaxis path following a temporal gradient. The V.pt path projects to MLR (midbrain locomotor region), but The lamprey Ob.m (medial olfactory bulb) projects to both Hb.m (medial habenula) and to V.pt (posterior tuberculum), which each project to different locomotor paths [Derjean et all 2010], Hb.m to R.ip (interpeduncular nucleus) and V.pt to MLR (midbrain locomotor region). The zebrafish also has Ob projections to Hb and V.pt [Imamura et al 2020], [Kermen et al 2013].
Dual odor-seeking action paths in the lamprey and zebrafish. Hb (habenula), Ob.m (medial olfactory bulb), V.pt (posterior tectum).
Further complicating the paths, the Hb.m itself contains both an odor seeking path and an odor avoiding path [Beretta et al 2012], [Chen et al 2019]. Similarly Hb.m has dual action paths for social winning and losing [Okamoto et al 2021]. So, this essay could use the dual paths in Ob.m instead of contrasting Ob.m with V.pt, but the larger contract should make the simulation easier to follow.
This essay’s simulation makes some important simplifications. The Hb to R.ip path is a temporal gradient path used for chemotaxis, phototaxis and thermotaxis. In a real-world marine environment, odor diffusion and water turbulence is much more complicated, producing more clumps and making a simple gradient ascent more difficult [Hengenius et al 2012]. Because this essay is only focused on the switchboard effect, this simplification should be fine.
Striatum action paths with adenosine timeout
The timeout circuit uses the striatum, which has two paths: one selecting the main action, and the second either stopping the action, or selecting an opposing action [Zhai et al 2023]. The two paths are distinguished by their responsiveness to dopamine with S.d1 (striatal projection with D1 G-s stimulating) or S.d2 (striatal projection with D2 G-i inhibiting) marking the active and alternate paths respectively. This model is a simplification of the mammalian striatum where the two paths interact in a more complicated fashion [Cui et al 2013].
Essay odor seek with timeout circuit. The seek path flows from Ob, through S.d1 to P.v to V.pt. The avoid path flows from Obj, though S.d2 to Pv. to Hb. Ad (adenosine), Hb (habenula), Ob (olfactory bulb), Pv (ventral pallidum), S.d1 (striatum D1 projection neuron), S.d2 (striatum D2 projection neuron), V.pt (posterior tuberculum)
As mentioned, the two actions paths are the seek path from Ob to V.pt and the avoid path from Ob to Hb. For the timeout and switchboard, the Ob has a secondary projection to the striatum. Although this circuit is meant as a proto-vertebrate simplification, Ob does project to S.ot (olfactory tubercle) and to the equivalent in zebrafish [Kermen et al 2013].
The timeout is managed by adenosine, which is a neurotransmitter derived from ATP and a measure of neural activity. The striatum has three sub-circuits for this kind of functionality, which I’ll cover in order of complexity.
S.d1 and adenosine inhibition
The first circuit only uses the direct S.d1 path and adenosine as a timeout mechanism. When the animal follows an odor, the Ob to S.d1 signal enables the seek action. As a timeout, ATP from neural activity degrades to adenosine and the buildup of adenosine is a decent measure of activity over time. The longer the animal seeks, the more adenosine builds up. Of the Ob projection axis contains an A1i (adenosine G-i inhibitory) receptor, the adenosine will inhibit the release of glutamate from Ob, which will eventually self-disable the seek action.
S.d1 action path inhibited by adenosine buildup as a timeout. A1i (adenosine G-i inhibitory receptor), Ad (adenosine), mGlu5q (metabotropic glutamate G-q receptor), Ob (olfactory bulb), S.d1 (D1-type striatal projection neuron)
In practice, the striatum uses astrocytes to manage the glutamate release. An astrocyte that envelops the synapse measures glutamate release with an mGlu5q (metabotropic glutamate with G-q/11 binding) receptor and accumulates internal calcium [Cavaccini et al 2020]. The astrocyte’s calcium triggers an adenosine release as a gliotransmitter, making the adenosine level a timeout measure of glutamate activity. The presynaptic A1i receptor then inhibits the Ob signal. The timeframe is on the order of 5 to 20 minutes with a recovery of about 60 minutes, although the precise timing is probably variable. Interestingly, the time-out is a log function instead of linear measure of activity [Ma et al 2022].
This circuit doesn’t depend on the postsynaptic S.d1 firing [Cavaccini et al 2020], which contrasts with the next LTD (long term depression) circuit which only inhibits the axon if the S.d1 projection neuron fires.
S.d1 presynaptic LTD using eCB
S.d1 self-activating LTD uses retrotransmission to inhibit its own input using eCB (endocannabiniods) as a neurotransmitter. Like the astrocyte in the previous circuit, S.d1 uses a mGlu5q receptor to trigger eCB release, but also require that S.d1 fire, as triggered by NMDA glutamate receptor. The axon receives the eCB retrotransmission with a CB1i (cannabinoid G-i inhibitory) receptor and trigger presynaptic LTD [Shen et al 2008], [Wu et al 2015]. Like the previous circuit, the timeframe seems to be on the order of 10 minutes, lasting for 30 to 60 minutes.
S.d1 LTD circuit. A coincidence of glutamate detection with mGlu5q and S.d1 activation with NMDA triggers eCB release, which activates CB1i leading to presynaptic LTD. CB1i (cannabinoid G-i inhibitory receptor), mGlu5q (glutamate G-q receptor), Ob (olfactory bulb), S.d1 (striatum D1-type projection neuron).
This circuit inhibits itself over time without using adenosine or astrocytes. In the full striatum circuit, high dopamine levels suppress this LTD suppression, meaning that dopamine inhibits the timeout [Shen et al 2008].
The next circuit adds the S.d2 path, which uses adenosine and self-activity to trigger postsynaptic LTD.
S.d2 postsynaptic LTP via A2a.s
Consider a third circuit that has the benefits of both previous circuits because it uses adenosine as a timer managed by astrocytes and is also specific to postsynaptic activity. In addition, it allows for a second action path, changing the circuit from a Go/NoGo system to a Go/Avoid action pair. This circuit uses LTP (long term potentiation) on the S.d2 striatum neurons.
Timeout circuit using postsynaptic LTD at the S.d2 neuron and adenosine as a timeout signal. As adenosine accumulates, it stimulates S.d2, which both disables S.d1 and drives the avoid path. A2a.s (adenosine G-s stimulatory receptor), Ad (adenosine), mGlu5q (glutamate G-q metabotropic receptor), Ob (olfactory bulb), S.d1 (striatum D1-type projection neuron), S.d2 (striatum D2-type projection neuron)
When the odor first arrives, Ob activates the S.d1 path, seeking toward the odor. S.d1 is activated instead of S.d2 because of dopamine. In this simple model, the Ob itself could provide the initial dopamine like c. elegans odor-detecting neurons or the tunicate’s coronal cells or the dual glutamate and dopamine neurons in Vta (ventral tegmental area).
As time goes on, adenosine from the astrocyte builds up, which activates the S.d2 A2s.a (adenosine G-s stimulatory receptor) until it overcomes dopamine suppression and increases the S.d2 activity with LTP [Shen et al 2008]. Once S.d2 activates, it suppresses S.d1 [Chen et al 2023] and drives the avoid path.
The combination of these circuits looks like it’s precisely what the essay needs.
Simulation
In the simulation, when the animal is hunting food and finds a food odor plume, it directly seeks toward the center and eats if it find food. In the screenshot below, the animal is eating.
Simulation showing the animal eating food after seeking the odor plume.
Satiation disables the food seek. This might sound obvious, but hunger gating of food seeking requires specific satiety circuits to any seek path that’s food specific, which means the involvement of H.l (lateral hypothalamus) and related areas like H.arc (arcuate hypothalamus) and H.pv (periventricular hypothalamus). And, of course, the simulation requires simulation code to only enable food odor seek when the animal is searching for food.
The next screenshot shows the central problem of the essay, when the animal seeks a food odor but there’s no food at the center.
Screenshot showing the animal stuck in the middle of the food odor plume before the timeout.
Without a timeout, the animal circles the center of the food odor plume endlessly. After a timeout, the animal actively leaves the plume and avoid that specific odor until the timeout decays.
Screenshot showing the animal escaping from the odor plume after the timeout.
This system is somewhat complex because of the need for hysteresis. A too-simple solution with a single threshold can oscillate, because as soon as the animal starts leaving the timeout decays, which then re-enables the food-seek, which then quickly times out, repeating. Instead, the system needs to make re-enabling of the food seek more difficult after a timeout.
But that adds a secondary issue because if food seek is a lower threshold, then the sustain of seek needs to raise the threshold while the seek occurs. So, the sustain of seek needs a lower threshold than starting seek. This hysteresis and seek sustain presumably needs to be handled by the actual striatum circuit.
Discussion
I think this essay shows that using the stratum for an action timeout for food seek is a plausible application. The circuit is relatively simple and is effective, improving search by avoiding failed areas.
However, the simulation does raise some issues, particularly hysteresis problem. If the striatum does provide a timeout along these lines, it must somehow solve the hysteresis problem. While the animal is seeking, the ongoing LTP/LTD inhibition should use a high threshold to stop seeking, but once avoidance starts, there needs to be a high threshold to return to seeking to avoid oscillations between the two action paths.
Because LTD/LTP is a relatively long chemical process (minutes) internal to the neurons, as opposed to an instant switch in the simulation, the delay itself might be sufficient to solve the oscillation problem. It’s also possible that some of the more complicated parts of the circuit, such as P.ge (globus pallidus) and its feedback to the striatum or H.stn (subthalamic nucleus) might affect the sustain of seek or breaking it and so control the hysteresis problem.
The simulation also reinforced the absolute requirement that action paths need to be modulated by internal state like hunger. For the seek paths, both Hb.m and V.pt are heavily modulated by H.l and other hypothalamic hunger and satiety signals.
As expected, the simulation also illustrated the need for context information separate from the target odor. While the food odor is timed out, the animal can’t search the other odor plume because this essay’s animal can’t distinguish between the odor plumes, and therefore avoids both odors. With a long timeout and many odor plumes, this delays the food search. A future enhancement is to add context to the timeout. If the animal can timeout a specific odor plume, it can search alternatives even if the food odor itself is identical.
I’m looking to improve the foraging algorithm with an idea from essay 17, which suggested that when the foraging fails, the animal should avoid the failed area. The foraging task uses an odor cue to seek food. Currently when the model gives up (times out), it disables seeking, but doesn’t actively avoid the current place, but returns to the wide-ranging roaming search.
For now, I’m still avoiding memory, but consider the alternating T-maze used in rodent behavior [Deacon and Rawlins 2006]. Mice are released at the base of the T and choose one of the directions to search for food. If the experiment repeats (by picking up the mice and restarting) mice will tend to explore the unexplored end first.
But for our foraging task, let’s use the same device for a different purpose. Instead of repeating the experiment by unnatural teleportation, consider the simpler problem of foraging with this device as an environment.
T-maze exploration. Food might be at either the red dot or the blue dot.
When rodents are foraging and reach one end, they will reverse and search the other end. Because rodents are far more advanced than the toy model, they can remember which arm of the maze they’ve already explored. But consider a simpler sub-strategy that uses RTPA (real-time place avoidance) where the animal temporarily avoids the current area or areas associated with food. By actively avoiding the already-explored area, the animal will save time by avoiding repeated searching.
A difficulty in finding the neural correlates of RTPA is the great diversity of reasons for RTPA, and circuits even in the brainstem. There are many reasons for place avoidance:
Startle: reflex escape
Escape from an imminent predator
Escape from an environment hazard (CO2, temperature)
Avoiding innate cues (predator odors)
Avoiding learned cues (CPA conditioned place avoidance)
Search optimization: avoiding already searched areas
Because this topic is large and the number of circuits is also large, I’ll start with a more abstract view to provide some context for a later dive into details. The two architectures will be a set of labeled path seek and avoidance circuits, and a secondary consensus circuit to coordinate the labeled paths.
Labeled path
A labeled path architecture uses individual circuit paths for each behavior and sense, as opposed to bringing all stimuli into a central node with a general decision algorithm [Helmbrecht 2018]. (Helmbrecht uses “labeled line,” which conflicts with the fish “lateral line” sense.) As least to some extend, the brainstem is designed around labeled paths, which is particularly evident if using the chimera model of the bilateral brain [Tosches and Arendt 2017].
The chimera model posits that brains of bilateral animals combine features from apical (unilateral) and bilateral (“blastoporal” in their terminology because they focus on zooplankton larvae). The apical mode is associated with the front of the brain, such as the hypothalamus, and its locomotion is temporally gradient based, like the tumble-and-run of bacteria. The bilateral mode is more reflexive, turning left if touched on the right, like Braitenberg vehicles [Braitenberg 1984]. Apical systems include olfactory search and phototaxis, while bilateral touch, lateral line, auditory and bilateral vision in a second system. For zebrafish one study describes multiple paths as a “high road” through Hb (habenula, apical) and a “low road” through OT (optic tectum, bilateral) [do Camo Silva et al 2018].
Some labeled paths for locomotion in vertebrates. H.l (lateral hypothalamus), Hb (habenula), MLR (midbrain locomotive region), N8 (acoustic-vestibular cranial nerve 8), OT (optic tectum), R.ip (interpeduncular nucleus), R.mcell (Mauthner-cell), R.rs (reticulospinal motor command)
The above diagram shows some vertebrate labeled paths, which is clearer in simpler vertebrates like the lamprey and zebrafish. In the zebrafish startle reflex, a sudden noise triggers a fast C-bend turn followed by rapid swimming. The trigger can be a noise, vestibular, or lateral line motion [Berg et al 2018]. The startle circuit is only three synapses from the original sensor to the muscle, from the N8 auditory/vestibular nerve to the giant M-cell (Mauthner cell in r4) to the motor neuron that drives locomotion. In young zebrafish larva, head touch neurons (N5 trigeminal) connect to M-cells and are later replaced by N8 [Kohashi et al 2012]. M-cells fire only once per escape to drive the initial turn. Interestingly the escape turn choice uses an axo-axonic repeater and amplifier [Guan et al 2021].
In a different path looming and dimming visual signals that represent predators or obstacles drive OT (optic tectum), which can drive escape that either uses or bypasses the M-cell depending on the threat level [Bhattacharya et al 2017]. OT also pre-programs the M-cell circuit by suppressing the left or right to avoid an obstacle [Zwaka et al 2022].
Phototaxis (seeking or avoiding light) uses a temporal gradient system composed of left Hb.m (medial habenula) and R.ip.d (dorsal interpeduncular nucleus), which projects to the R.rs (reticulospinal motor command) neurons via relays in V.mr (median raphe) and P.ldt (laterodorsal nuclei) [Chen and Engert 2014]. Food odor seeking uses the right Hb.m and R.ip.v (ventral interpeduncular nucleus) [Chen et al 2019].
In lamprey a distinct food-seeking path through V.pt (posterior tuberculum – possibly homologous to vertebrate Vta/Snc) to MLR (midbrain locomotor region) and finally to R.rs [Derjean et al 2010]. Zebrafish has a similar dual path through Hb.l (lateral habenula) through a midbrain TSN circuit [Koide et al 2018].
Slower escape uses a distinct prepontine (rhombomere r0-r1) circuit, which is suppressed by the M-cell escape circuits [Marquart et al 2019].
Some of these paths have shared elements, particularly at the motor control like MLR, but the general pattern is multiple labeled paths for each behavior. The paths already mentioned don’t include more complex food-seeking paths through the basal ganglia and hypothalamus.
Multiple labeled paths immediately raises the difficulty of coordination. How does the system juggle priorities? Even the simple startle reflex needs to be modulated because the animal shouldn’t startle if the loud sound is expected, such as near a waterfall. In contrast in a dangerous area with possible predators the animal should increase the reflex to a hair-trigger. Similarly if the threat is weak and the animal is hunting or eating and hungry, it might ignore the threat to continue eating. A second architecture, distinct from the label path, emphasizes the coordination of multiple paths, possibly using a consensus system to decide on an appropriate action.
Consensus loops
While the labeled paths have strong evidence, the consensus loop is only a thought experiment to tie the paths together. Multiple paths for food seeking and for avoiding is a distributed system, and distributed systems makes decision circuitry more complicated because they’re not central decision node. Every node needs to agree with the decision. Whether to avoid or approach needs to be agreed on by all the systems. It wouldn’t make sense for one system to believe the animal is approaching an object but another system believes the action is avoiding. Voting distributes the consensus.
Illustration of a consensus loop. Multiple drives or labeled paths vote for approval to drive motor output.
The above diagram shows the model. The different labeled paths of seeking or avoiding join a voting consensus system in a motivational look, which allows one path to drive motor output.
Consensus system showing one path. The driving sense or motivation tries to disinhibit itself by voting in the consensus loop.
A single labeled path has a sense or motivation drive that tries to act on motor output, but is inhibited by the consensus system. For example, if a predator odor arrives, the odor avoidance path votes to enable its own locomotion. If the consensus system agrees, it will disinhibit the odor avoidance path, letting the animal escape. Note that a high priority threat could bypass the consensus system.
Seek and avoid consensus system
This system can manage conflicts between seek and avoidance, such as animals continuing to eat if a predator threat exists but is low. Consider a simplified consensus system with only one seek node and one avoid node, using the consensus to select one when there’s a conflict.
Managing conflicts between seeking and avoiding.
If there’s a food cue and no conflicting threats, the food vote passes easily and the animal seeks the food. Similarly a predator odor with no conflict will enable avoidance. If there’s a conflict, the system can weigh the costs and benefits of the threat and the food, possibly depending on hunger state or a more sophisticated threat assessment.
Keeping these general ideas of the labeled path and consensus systems in mind, let’s start working through several specific paths. The end goal is to organize the main brainstem locomotive areas into a simplified, unified model. The two major paths will be apical paths through Hb (habenula) using temporal gradients (klinotaxis) [Chen and Engert 2014] and bilateral paths through OT (optic tectum) using spatial gradients (tropotaxis).
Apical and bilateral avoidance
Because there are many labeled paths, dividing them up might help organize the model. An early division between labeled paths goes back to the bilaterian (worm-like, slug-like) ancestors, which added bilateral, dual-sensory navigation (tropotaxis, spatial gradient) to an older single-sensor navigation that used the animal’s movement to choose a direction (klinotaxis, temporal gradient), such as the simple tumble-and-run that even bacteria and simple radial zooplankton use for seeking odors (chemotaxis) and seeking or avoiding light (phototaxis). This chimera hypothesis [Tosches and Arendt 2013] considers bilateral animals as a fusion between the locomotive systems. The apical zooplankton larvae of bilaterian worms may have been a secondary development to escape predation [Mallatt 2021]. In vertebrates, apical klinotaxis is implemented by Hb (habenula) temporal gradient seeking and H.l (lateral hypothalamus) motivation. Bilateral tropotaxis navigation is implemented by several labeled path systems, typified by OT (optic tectum) and the M-cell start reflex.
A primitive apical example is the helical phototaxis of many annelid (marine worm) zooplankton larvae, and a primitive bilateral example is the mollusk sea slug navigation.
Zooplankton apical phototaxis
One type of zooplankton is essentially a globe with a fringe of cilia and an apical tuft for chemical processing, such as the Platynereis larva.
Phototaxis for this larva depends on its helical movement (helical klinotaxis). As it moves forward, the larva also rotates and wobbles, which means that parts of the equatorial band are nearer the light or further from the light depending on the rotation. If the upper cilia halt, the larva will steer toward the light. If the lower cilia halt, the larva will steer away from the light [Randel and Jékely 2016].
Phototaxis for an zooplankton larva.
The system depends on a directional eye, which uses a photoreceptor and a pigment cell that imposes directionality by shading the photoreceptor, because other cells of the larva are transparent. The photoreceptor compares the current brightness to its average brightness as the larva rotates. If it’s brighter than average, then it must be facing the light, and will signal the cilia to briefly halt, using ACh (acetylcholine) as a neurotransmitter. This trivial one-neuron circuit is sufficient for simple phototaxis [Randel and Jékely 2016].
Although this example larva uses two photoreceptors, it’s not truly bilateral and the two photoreceptors don’t communicate. Ablating one photoreceptor doesn’t abolish phototaxis, although it does reduce efficiency. Using three or four photoreceptor/pigment pairs would work, as well as removing all but one. This system is apical klinotaxis, not bilateral tropotaxis, which makes sense because the above zooplankton is not bilateral. While this zooplankton uses helical klinotaxis, another common form of klinotaxis is a side to side “casting” motion used by other simple animals like c. Elegans [Izquierdo and Lockery 2010].
If zooplankton phototaxis is an example of apical navigation, then the mollusk sea slug is an example of bilateral navigation.
Mollusk sea slug seek and avoid
The mollusk sea slug circuit is a pure bilateral circuit, almost directly a Braitenberg circuit [Braitenberg 1984], discussed in essay 14. The following shows a rough schematic of the sea slug seek and avoid. This circuit is interesting because with only a few neurons, the slug can switch from turning toward a food odor when hungry to turning away from the odor when not hungry [Gillette and Brown 2015].
Odor seek and avoid circuit for a sea slug. Hunger switches a food odor from seek to avoid.
In the diagram above, the central grey area is a switchboard circuit. Hunger reconfigures the switches connecting the odor to the turn motor neurons. When the slug is hungry, the right odor sensor connects with the left turn muscle, seeking the odor. But when the slug is sated, the right odor sensor connects with the right turn muscle, avoiding the odor. When the slug is hungry, it approaches food but when it’s not hungry, it avoids food odor cues.
A similar animal with a different circuit configuration uses serotonin to switch from avoidance to approach [Hirayama et al 2014].
For the goal of this essay, avoiding a failed food cue, this circuit is perfect because when the animal finds a false cue, it reversed movement from seek to avoid, which exactly fits the essay needs. Unfortunately, the vertebrate circuits aren’t nearly as straightforward. As a start for the vertebrate navigation paths, the startle reflex managed in vertebrates by the giant Mauthner cells is a simple starting point.
Amphioxus fast twitch reflex
The fast twitch startle reflex is a clear example of a bilateral labeled path avoidance circuit. A noxious sense on one side causes a fast turn away from the sense. The sense can be a touch on the head, such as running into an object, or a loud sound or a vestibular imbalance signal. This circuit predates vertebrates and a similar circuit exists in amphioxus, a filter-feeding chordate that looks something like a fish without a distinct head and without eyes, but with several photoreceptors including a frontal “eye.”
In amphioxus the startle reflex drives fast twitch muscle fibers, where normal swimming uses slow twitch fibers [Lacalli and Candiani 2017]. This circuit path is entirely distinct even to using the different muscles. The following diagram shows part of the amphioxus motor control circuit. (Because the neuron names are specific to amphioxus, they’re not hugely important for this essay.)
Amphioxus fast twitch escape uses LPN3, glutamate large paired neuron.
The diagram shows the LPN3 (large paired neuron) fast twitch escape path, and the PPN2 normal swimming match, including intermediary motor control neurons [Lacalli and Candiani 2017]. This amphioxus escape circuit resembles the zebrafish Mauthner cell escape.
Zebrafish Mauthner cell escape
Zebrafish have a pair of large M-cell (Mauthner cell) neurons that are specialized for auditory and vestibular startle escape. These are very fast reflexes on the order of 10ms, which can be modulated by higher context [Zwaka et al 2014] including OT. Although the M-cells perform a similar role to the amphioxus LPN3, it’s not clear that they’re homologous, which requires common descent, because the large escape neuron is a common pattern in non-chordate systems.
Zebrafish startle response at right in context with other labeled paths. M-cell (Mauthner R.rs cells in r4), N8 (acoustic/vestibular cranial nerve 8), OT (optic tectum), R.pp (prepontine avoidance in r0-r1), R.rs (reticulospinal motor control)
The primary input to M-cell escape is an auditory and vestibular signal from N8 (8th cranial nerve is auditory and vestibular). In water, sound and primitive vestibular sense have some similarities, because water motion produces not just sound but animal motion, depending on the frequency. The M-cell directly connects to motor neurons to muscles. The startle escape is only a three neurons and a clear, distinct labeled path.
A second zebrafish threat avoidance path uses neurons in R.pp (pre-pontine r0-r1) [Marquart et al 2019] for more distance threats. Unlike the M-cell circuit, this R.pp path is more than a reflex, but it’s still a hardwired path. A third threat circuit uses OT (optic tectum), for example the looming response. Most vertebrates will flee or freeze from a rapid and overhead expanding dark object, representing a potential predator or an obstacle. The mammalian startle circuit shares similarity with an acoustic projection to R.pn.c (caudal pontine reticular) neurons, in an analogous area to the M-cell [Kim et al 2017].
Some of these circuits do share sub circuits. For example, hindbrain locomotion and turning are distinct circuits that are used by both bilateral and apical avoidance circuits.
Hindbrain locomotion and turning
Senses are not the only source of distinct paths because actions can be split into parts like a car’s divided steering and acceleration. In vertebrates, accelerating and turning use distinct hindbrain circuits. Although both MLR (midbrain locomotive region) and OT.d (deep layer of the optic tectum) encode seeking and avoiding, they don’t encode left or right turns. Activating the left or the right MLR produces straight movement [Brocard et al 2010]. Turning is managed from OT.i (intermediate layer of the optic tectum) to distinct R.rs motor command neurons, marked by the chx10 transcription factory [Cregg et al 2020].
Hindbrain acceleration and turning circuits. R.rs (reticulospinal motor control)
The above diagram shows the basic idea. The upstream MLR can command forward movement without specifying details, because swimming is an oscillatory process with CPG (central pattern generators) in the spinal cord and the hindbrain. To turn, the chx10 neurons inhibit the swimming stroke in one direction [Cregg et al 2020], similar functionally to the apical zooplankton inhibition of cilia for phototaxis.
Splitting out turning can simplify the system by dividing labor, where OT.i is always responsible for obstacle avoidance, but a diverse set of labeled paths decode whether to seek or to avoid.
Optic tectum and dimming
The OT is named after its retinotopic visual map that is used for avoiding looming/dimming obstacles and predators, and also for seeking prey [Basso et al 2021]. For most vertebrates, OT is the primary visual area, and the visual cortex only provides abstract context, and amphibians and fish lack a proper visual cortex [Heap et al 2018]. For this essay, OT is less important for its sophisticated visual organization, but more because it also contains motor maps for seeking or prey and avoidance of looming objects, and dimming fields. Its motor map also contains drinking and licking [Liu et al 2022].
Looming/dimming path through optic tectum. OT.m (medial, deep optic tectum), R.rs (reticulospinal motor command)
OT processes looming and dimming objects and avoids them. Since the essay’s model lacks proper vision, the dimming is currently most important. Because OT also has obstacle avoidance, it’s a more sophisticated system than simply reflex. It’s likely that other avoidance systems will use OT for obstacle handling. Even in the case of the M-cell reflex, the OT.i pre-programs the M-cell, to avoid obstacles in case of a future startle [Zwaka et al 2014].
Optic tectum and turning
This division between turning and acceleration applies to OT itself. OT is a layered structure where the top layer is a visual map, the intermediate layer integrates other senses and produces turns, and the deepest layer includes actions such as avoiding and seeking [Liu et al 2022]. OT.d (deep OT) is a motor area for seek and avoid, connected with MLR and M.pag (periaqueductal grey) motor output, and integrates general dimming from the retina with distinct expansion calculation in OT itself [Heap et al 2018], to avoid looming objects. OT.i (intermediate OT) includes multi sensory integration and turning motor area, connected with LL (lateral line) electro sensation and water motion, somatosensory (whiskers in mice), auditory input from M.ic (inferior colliculus) and optic input from OT.s (superficial OT).
Optic tectum layered structure, emphasizing turning and motion. LL (lateral line water motion), MLR (midbrain locomotor region), OT.s (superficial optic tectum), OT.i (intermediate OT), OT.d (deep OT), R.rs (reticulospinal motor command)
Because only the top layer is specifically optic, some neuroscientists use “tectum” (roof in latin) instead of OT to emphasize its multi sensory and motor function, not just the optic features. On argument suggests that the optic layer OT.s is a secondary layer, added to a more primitive OT.i and OT.d that are more connected with reticular areas like MLR and M.pag [Edwards 1980], [Basso et al 2021]. With that argument, OT is primarily a moving and turning structure, receiving turning and moving input from touch, lateral line, primitive dimming, and other directional senses and combining with seek and avoid decisions. When the visual system developed enough detail to support crude images like looming disks or moving prey-like dots, the OT integrated vision into its top layer.
On the other hand, since OT.d receives dimming information from H.lg (central lateral geniculate nucleus) for looming escape [Heap et al 2018], it’s also conceivable that the base OT function is visual escape from dimming, where the later expanding, looming visual processing is an optimization.
Optic tectum obstacle avoidance combined with MLR seek or avoid movement. MLR (midbrain locomotor region), OT.i (intermediate optic tectum), R.rs (reticulospinal motor neurons).
This separation of obstacle avoidance turning from seeking and avoiding greatly simplifies some other circuitry that doesn’t need to duplicate the obstacle avoidance. Since other circuitry from the apical path, like the Hb-R.ip (habenula – interpeduncular nucleus) has its own turning system, OT.i doesn’t have a monopoly on turning. But even in that case, OT.i obstacle avoidance can inform apical navigation.
Some of these avoidance circuits are from the bilateral part of the chimera, such as the M-cell and the looming OT circuits, and others are from the apical part, such as Hb.m phototaxis, chemotaxis, and thermotaxis. So, let’s now more from the bilateral avoidance circuits, explore the vertebrate apical navigation.
Tunicate helical swimming and phototaxis
Tunicates (including sea squirts) are the closest chordates to the vertebrates, but because they have evolved at a greater rate and in specialized directions, comparison with vertebrates is difficult [Stolfi and Brown 2016]. Ascidian tunicates (sea squirts) have a mobile tadpole stage that plants itself in under 24 hours and transforms into a sessile filter feeder, reforming the entire brain. In general, neuroscientists believe amphioxus more resembles the ancestral vertebrate, and that ascidians have lost too many ancestral structures for a reasonable comparison [Holland 2016]. But for the sake of exploration let’s run through a thought experiment as if the ascidian larva is a compressed and simplified version of the vertebrate ancestor, although possibly only the vertebrate larva.
Specifically, consider phototaxis in the apical helical klinotaxis mode that follows a temporal gradient, since both amphioxus and ascidian larva swim in a helical pattern. Even bacteria can follow odor gradients [Hengenius et al 2012] and as discussed above zooplankton phototaxis can move toward light with only a single photosensor [Randel and Jekely 2016]. Both amphioxus and ascidian larva have single unpaired eyes, amphioxus as a single frontal eye [Lacalli 2022] and ascidians with an asymmetrical eye paired with a second pigment cell used for geotaxis as a primitive vestibular sense [Hoyer et al 2024]. In both cases, the “eye” is directional with a pigment cell, but a non-image-forming collection of photoreceptors. The ascidian asymmetrical eye works because the ascidian tadpole swims in a helical pattern so the timing of the light on the eye matters more than its position [Ryan et al 2016].
Ascidian larvae swim in a helical pattern comprised of unilateral tail flicks and symmetrical swimming [Ryan et al 2017] and use the asymmetry of the photoreceptor and photopigment to swim toward light [Mast 1921], [Zega et al 2006]. Since helical swimming doesn’t need stabilizing fins or vestibular systems to manage roll, yaw, and pitch with 3d swimming, it’s available to evolutionarily simpler systems. Another advantage of helical phototaxis is that the photoreceptors are auto-calibrating by simply averaging the light in a rotation and requires less circuitry than a bilateral comparison of light [Randel and Jekely 2016].
However, unlike the trivial zooplankton circuit that directly connected the photoreceptor to arrest the cilia, ascidian larvae need to modulate the bilateral swimming in the primitive hindbrain, timing the muscle inhibition to achieve the same effect.
The ascidian ocellus (“eye”) has two types of photoreceptors with distinct responses. Type 1 has a pigment and lens and is directional (37 cells), while type 2 is non-directional (no pigment partner) [Salas et al 2018]. If the pigment is genetically deleted, the animal can’t use phototaxis but does respond to dimming with an escape response. In other words, the dimming response and phototaxis use distinct labeled paths with distinct input neurons [Kourakis et al 2019]. The following shows the circuit for the type 1 photoreceptors for phototaxis, where the boxes represent single neurons or small collections (5-8) of neurons, not large functions (from [Ryan et al 2016]).
Ascidian larva phototaxis (ocellus) and geotaxis (otolith) circuit. Ant2 (antenna geotaxis sensors), antRN (antenna relay neuron), mgIN (motor ganglion interneurons, left and right), MN (motor neurons, left and right), PR-1 (type-1 phototaxis photoreceptors), prRN (photoreceptor relay neuron)
The above diagram shows both geotaxis and phototaxis circuits, which are specific to right or left motor neurons respectively, because the ascidian larva neuron circuits are highly asymmetrical. Ascidian larva geotaxis swims upward and phototaxis swims away from light, generally downward. The combination encourages swimming to the underside of ledges, such as the underside of boats and harbor piers [Ryan et al 2016]. Because of the helical swimming, the left and right motor neurons aren’t left or right turns, but turns toward or away from the target. Although this circuit is more complicated than the purely apical zooplankton because of the interface to bilateral swimming, the helical swimming keeps the circuit relatively simple.
The above partial circuits are complicated by the coronet cells, another sensory cell that are paired with the photoreceptors, but with unknown function. The circuit connectivity is interesting, because coronet cells modulate both the phototaxis and geotaxis paths, but aren’t a path of their own. The phototaxis and geotaxis relay neurons above are partially bilateral. Only 70% of their connectivity is to the main side, but 30% of the connectivity is to the opposite side. In contract, the coronet-enabled neurons are 100% to the main connection [Ryan et al 2016].
Coronet cell modulation of phototaxis and geotaxis in the ascidian larva. ant2 (antenna geotaxis cell), ant-core (antenna-coronet relay neuron), antRN (antenna relay neuron), DA (dopamine), mgIN (motor ganglia interneuron, left and right), MN (motor neuron, left and right), PR-1 (photoreceptors type-1), pr-cor (photoreceptor-coronet relay neuron), prRN (photoreceptor relay neuron)
As a thought experiment (unsupported by scientific evidence) the main phototaxis path might be uncertain and stochastic, while the coronet-enabled path would be a certain, deterministic connection. If the coronet cells measured the certainty of the animal’s current direction, it could encourage sticking to the current path. For example, if the coronet cells were food-odor gradient sensors, they could fire when the animal was heading toward food, enabling a chemotaxis based on modulation of geotaxis and phototaxis.
Tunicate dimming response
The ascidian dimming response triggers locomotion with a strong turn as an escape response to predators [Kourakis et al 2019]. Unlike the phototaxis photoreceptor, the dimming photoreceptors are non-directional because’er not shaded by the pigment cell. There are 23 type-1 directional photoreceptors and 7 type-2 non-directional photoreceptors for dimming.
The above diagram shows the dimming path in context with the previous phototaxis path. Like the phototaxis path, the dimming path starts from the type-2 photoreceptors to a relay neuron and to the control neurons in the motor ganglion. Unlike the phototaxis path, the dimming path is modulated by ascending motor signals from AMG (ascending motor ganglion) and from the phototaxis path [Ryan et al 2016], presumably so the normal helical phototaxis doesn’t trigger a dimming response.
Cement gland and attachment
The ascidian larva hatches before dawn, swims upward for a few hours because geotaxis is enabled before phototaxis neurons attach, and then swims away from the light, settling on a lively rock, preferring a ledge to settle under if possible. Larva do not feed [Ryan et al 2016]. The larva will attach with a cement gland on the front of its head, a trio of palms, and then transforms into the adult sessile filter feeder. The palp sensors trigger the attachment circuit, which stops all swimming and begins the metamorphosis [Anselmi et al 2024].
Although the full details of the above circuit [Ryan et al 2016] aren’t critical, the PN (palp neuron) senses the animal bumping into a rock modulated by chemical senses that avoid toxic area, and triggers a swimming shutdown by inhibiting the motor neurons and interneurons [Hoyer et al 2024]. Like the previous diagrams, the boxes represent individual neurons or small group, not large functional regions.
While the ascidian cement gland is permanent, several fish [Pottin et al 2010] and amphibians [Rétaux and Pottin 2011] have a homologous cement gland used for larvae, not adults. For example, frog tadpoles can attach to the bottom of leaves or to the water surface to avoid predators until they are large enough to hunt [Jamieson et al 2000], [Yoshizawa et al 2008]. Because of the widespread cement gland among many fish species and amphibians as well as the tunicates, it’s likely the original vertebrates had a similar cement gland [Rétaux and Pottin 2011]. Whether the gland was larva-only like in vertebrates or also used for adults as in tunicates is unknown. In either case, the cement gland circuit that inhibits locomotion must have been part of the original vertebrate.
Vertebrate analogies to the ascidian circuits
Because the ascidians are so specialized and reduced from the common ancestor with vertebrates, including major losses in genes, cells and structures, comparing the two is essentially impossible to be homologous (shared descent) [Holland 2016]. However, for the sake of exploration, I’m ignoring that advice, and looking for analogous vertebrate circuits to the ascidian larva.
The ascidian behavior each have distinct circuit paths that mostly only come together at the motor control neurons. The exception is the feedback from the AMG (ascending motor ganglion) neurons, which do feedback to the midbrain neurons, but the main paths are separate forward paths. Each of the geotaxis, phototaxis, dimming, cement gland attachment, and bilateral escape are circuit paths that are distinct until the motor command neurons.
A vertebrate analogy to the ascidian phototaxis gradient path might be the path from the retina to Hb.m (medial habenula) to R.ip (interpeduncular nucleus) and V.mr (median raphe), which then project to R.rs (reticulospinal motor command). Like the ascidian path, the Hb-R.ip phototaxis path is relatively isolated from the other paths, although Hb.m does receive large modulation from the hypothalamus. Although R.ip is mostly descending, like ascidian mgIN, V.mr is both ascending and descending like mgIN and AMG.
The dimming path from the type-2 photoreceptors resembles the dimming input to the vertebrate OT (optic tectum). Although existing vertebrates have more sophisticated eyes that can distinguish expanding objects, the dimming input to OT is still important and used for escape directionality [Fotowat and Engert 2023], [Heap et al 2018]. Retina dimming cells reach AF6 and AF8 [Temizer et al 2015], which are thalamic arborization fields before reaching OT.d. Although more complicated expanding looming response in vertebrates is better studied, expansion detection requires an image-supporting eye, and OT.d receives the simpler dimming input. Like the ascidian dimming pr-AMG (photoreceptor – ascending motor ganglion) neuron, OT.d receives multiple ascending and descending inputs that modulate the dimming response. In particular Ppt (pedunculopontine nucleus) and P.ldt (laterodorsal nucleus) both receive OT.d output and forward to R.rs, functionally similar to mgIN (motor ganglion interneuron), and send ascending feedback from R.rs to OT.d, resembling the AMG (ascending motor ganglion) functionality.
Because the cement gland exists in vertebrates, the circuit should be available, and studies do show that N5 (head touch trigeminal nerve) innervates it automatically [Pottin et al 2010], but I haven’t read any study that says this this specific group of trigeminal neurons connects to. As a through experiment, consider H.stn (subthalamic nucleus) as a choice for the cement gland, because H.stn halts ongoing action, and because H.stn receives direct input from C.i (insular cortex) and C.ss (somatosensory cortex), which are more sophisticated versions of the chemo / mechanosensory palp neurons.
The coronet path enhances taxis confidence, reducing stochastic choice, and is a set of dopamine neurons. The striatum circuit and dopamine’s role has a similar function. Without dopamine, the basal ganglia suppress weak input, and allow stochastic action. With dopamine, the basal ganglia suppresses the randomness and keep action on track. This path resembles rheotaxis food seeking, where a fish approaches a food odor by swimming upstream [Coombs et al 2020]. The “what” signal (odor) differs from the “how” signal (water current cues). Like rheotaxis, the coronet cells enhance the existing phototaxis and geotaxis, reducing the default stochastic noise.
Hb.m Medial habenula
Of the ascidian labeled paths above, the Hb (habenula) phototaxis path will be a useful anchor for the upcoming consensus circuit. Like the ascidian asymmetrical phototaxis neurons, the vertebrates Hb.m (medial habenula) is also governed by Nodal asymmetry [Roussigne et al 2009], where Nodal is a developmental genetic transcription factor. In zebrafish the left Hb.m support phototaxis, and the right Hb.m supports chemotaxis [Chen et al 2019]. Hb.m phototaxis receives both “on” and “off” neurons from the retina with a relay either in H.em (pre thalamic eminence) [Zhang et al 2017] or T.a (an area in the anterior thalamus) [Cheng et al 2017], where the connections are debated. Although Hb.m does receive dimming input from the adjacent photoreceptive pineal gland, the retina photoreceptors are more important for phototaxis [Dreosti et al 2014].
As an anatomical note, the zebrafish Hb.m is actually dorsal and therefore named Hb.d. Similarly the zebrafish Hb.m is ventral and named Hb.v, but as a simplification I’ve used the mammalian name.
The output path from Hb.m is through R.ip (interpeduncular nucleus), which projects to several areas including R.gc (pontine central era), V.mr (median raphe – serotonin), V.dr (dorsal raphe – serotonin), and P.ldt (laterodorsal nucleus – ACh) [Quina et al 2017]. The V.mr glutamate and GABA neurons may be more important for this circuit than the serotonin neurons, which they outnumber. Also, note that V.mr is located in the same hindbrain rhombomeres (r2-r5) as some of R.rs, but are more ventral, and are reciprocally connected. In other words, V.mr is highly action and motor associated.
As described above, the Hb.m-R.ip path is a klinotaxis path for phototaxis [Chen and Engert 2014], chemotaxis and thermotaxis, where the klinotaxis is temporal from the animals movement, but not the helical movement of the ascidian larvae. The Hb.m-R.ip klinotaxis has multiple inputs for lamprey, including light, odor, and lateral line (water movement) [Stephenson-Jones et al 2011].
Habenula klinotaxis for lamprey for light, odor, and lateral line. Hb.m (medial habenula), LL (lateral line), R.ip (interpeduncular nucleus)
Although I’ve focused on Hb.m as an avoidance gradient circuit, it’s also a food odor seeking circuit [Chen et al 2019]. The Hb.m klinotaxis for light and odor also applies to temperature, using input from Po.m (medial preoptic nucleus) [Palieri et al 2024] and social seek and avoidance [Okamoto et al 2021], [Chou et al 2016].
Because Hb.m has several sub-nuclei and genetic clusters, it likely represents different labeled paths, supporting multiple distinct seek and avoidance paths. A binary seek vs avoid circuit is likely an oversimplification, because studies have found at least 5-6 olfactory Hb.m clusters in the larval zebrafish [Jetti et al 2014], [Beretta et al 2014]. Hb.m is asymmetrical, like the ascidian larva. Odors from either olfactory bulb activate the right Hb.m [Chen et al 2019]. Hb.m neurons have at least 262 neuropeptide receptors [Ables et al 2023] as well as morphine receptors [Gardon et al 2014], [Boulos et al 2020] including neuropeptides modulating hunger or social motivation from hypothalamic areas like H.l and H.pv.
R.ip interpeduncular nucleus
Since I’ve already covered some of the R.ip klinotaxis function in essay 24 and essay 25, I’m going to focus on the R.ip connectivity, particularly the ascending connectivity. R.ip descending efferents don’t target R.rs directly, but instead use intermediaries like R.gc (pontine central gray), V.mr (median raphe) and P.ldt (laterodorsal tegmental area) [Lima et al 2017], [Quina et al 2017].
The ascending afferents of R.ip also work through intermediaries, particularly P.ldt and V.mr [Quina et al 2017], although other connectivity studies report R.ip as directly producing ascending connectivity [Lima et al 2017]. Because V.mr is directly caudal to R.ip, the disagreement is essentially about the boundaries between R.ip and V.mr.
The ascending R.ip connectivity will become important in the next section on the consensus circuit because it completes the consensus loop, where other labeled path connectivity is descending. The ascending role is analogous to the ascidian AMG (ascending motor ganglion) neurons. For a consensus circuit to work, all nodes need to be informed of the consensus decision.
Consensus circuit narrative
Let’s now consider the consensus circuit and how it might develop from a strict labeled path system. This is just a thought experiment as a narrative explanation for the Hb.l (lateral habenula) system.
For simplicity, let’s restrict the narrative to apical systems only, ignoring bilateral systems like OT, and let’s start from a labeled path system. In the lamprey, odor information from Ob (olfactory bulb) splits into multiple paths. One path reaches Hb.m directly and another contacts V.pt (posterior tuberculum), considered a homologue of Vta / Snc (substantia nigra pars compacts), which then contacts MLR in lamprey [Derjean et al 2010] and zebrafish [Kermen et al 2013].
In this example, these two paths are distinct with threat odors going through the Hb.m – R.ip circuit and using P.ldt as an apical version of the MLR (which in lamprey may not be distinct from Ppt MLR, since the lamprey doesn’t have distinct Ppt, P.ldt and M.cnf (cuneiform nucleus)). The animal seeks food using the V.pt to MLR path.
These two system can come into conflict. For the above simple system, suppose conflicts are resolved in R.rs itself, as a hard-coded priority where threats always win. But now consider a system where the conflict is resolved earlier in the stream by adding Hb.l as a lateral inhibition relay.
As a first step consider lateral inhibition of threat odor suppressing food seeking. Here the lateral inhibition path uses a relay from Hb.m to Hb.l [Gouveia and Ibrahim 2022] in a primitive Hb.l that then suppresses the V.pt path. This lateral inhibition duplicates the earlier lateral inhibition in R.rs, but is more specific because it inhibits earlier in the two paths.
In the above diagram, the blue lines represent new connections. Notice the gating pattern for V.pt resembles the gating for the consensus circuit where the action nodes are V.pt and V.mr. Hb.l then becomes the vote accumulator for the consensus circuit. Also notice the similarity with the sleep model from essay 29, where Hb inhibits food seeking for sleep. An alternative narrative might repurpose the sleep inhibition into a path inhibition [Hikosaka 2010].
Both Hb.l and Hb.m are tonically acting, meaning that without any input their resting output is a middle value, not a binary output. This means Hb.l can gate V.pt seek and also gate its opposing avoidance circuit in V.mr and P.ldt.
For the next step, let’s add both a bidirectional selection and also add some internal state management, because the animal shouldn’t seek food if it’s sated.
H.l hunger modulation
H.l (lateral hypothalamus) has access to hunger and satiety information by sensing blood levels directly and from connections from R.pb (parabrachial), which has signals from the digestive system via N10 vagus nerve through R.nts (solitary tract nucleus). When Ob (olfactory bulb) senses a food odor , H.l can modulate it with the current hunger sense. This means H.l as gating input to Hb.l is more effective than the simple lateral inhibition from Hb.m If the animal is sufficiently hungry, it might ignore weak threats. Note the similarity to the mollusk sea hare circuit, where hunger changed food odor from seeking to avoidance depending on the internal state.
Adding H.l hunger modulation to the decision between the threat odor avoidance path and the food odor seek path. H.l (lateral hypothalamus), Hb.l (lateral habenula), Hb.m (medial habenula), MLR (midbrain locomotor region), P.ldt (laterodorsal tegmentum), R.ip (interpeduncular nucleus), R.rs (reticulospinal), V.mr (median raphe), V.pt (posterior tuberculum).
In addition to hunger, other internal states can modulate Hb such as hypothalamic threat signaling ([Wagle et al 2022]. This step also adds control of the threat path, taking advantage of the Hb.l tonic activity to either inhibit food seeking or threat avoidance.
Place avoidance without a threat
Suppose we take the above circuit, but ignore or disable the threat avoidance path via Hb.m. Even without the threat path, there is an avoidance path from Hb.l to V.mr and P.ldt, where Hb.l not only disinhibits threat avoidance, but can produce place avoidance without a threat.
The above diagram shows deletion of the threat path, while retaining the abstract place avoidance path. If place avoidance is triggered, the animal will avoid the current location without needing a specific threat to avoid. This means that H.l stimulation by itself can trigger real-time avoidance [Stamatakis et al 2016]. In mammals the H.l to Hb.l connection has at least 6 clusters [Calvigioni et al 2023], which suggests multiple paths even in the abstract place avoidance.
S.v ventral striatum digression
This model of the seek vs avoid circuit can be extended to S.v (ventral striatum aka nucleus accumbens) and P.v (ventral pallidum). Consider S.v / P.v as a generalization of H.l, providing more general context beyond hunger. This basal ganglia extension allows for a positive feedback loop. which enables multiple rounds of voting, integrating values, such as with drift diffusion.
In the above, I’ve split the V.pt of the lamper into an ascending Vta (ventral tegmental area) dopamine area from mammals, but left the V.pt to represent the descending glutamate / GABA portion of Vta, despite mammals lacking a distinct V.pt. If there’s a food cue when hungry, H.l to Vta stimulation will generate high DA in S.v, enabling it, which will disinhibit V.pt to enable food seeing. Here, S.v / P.v is acting as the consensus circuit and the V.pt path is the action for food seeking.
As with the smaller Hb.l circuit, S.v / P.v is also part of a sleep / wake circuit using dopamine as a wake signal, as used in essay 29. If the animal is currently seeking food, it shouldn’t fall asleep, and the high dopamine signals to stay away. Again, from a narrative sense, this circuit could have been repurposed from a wake circuit, as opposed to a path conflict system.
In zebrafish Hb.l only projects to V.mr and does not project to any DA [Amo et al 2014], while in the more primitive lamprey Hb.l projects to both V.mr and DA [Stephensen-Jones et al 2011], which suggests that the V.mr projection is more functionally critical to this circuit than the Vta projection, or that the Vta circuit is a later development. The zebrafish V.pt has descending dopamine but the existence of significant projections to the striatum is questioned [Yamamoto and Vernier 2011].
Note that H.l retains its central role, where the S.v circuit generalizes the base H.l function without replacing it. Stimulating H.l.g (H.l GABA neurons) can trigger seeking through its projection to Vta [Nieh et al 2016], and stimulating H.l.glu (glutamate H.l neurons) can trigger place avoidance through the H.l.glu projection to Hb.l [Stamatakis et al 2016].
Hippocampus digression
For place preference and place avoidance E.hc (hippocampus) plays a natural because E.hc represents context and place such as place cells, and H.hc projection strongly to both the hypothalamus and S.v. If we add the H.hc projections to H.l via S.ls, the seek / avoidance circuit looks something like the following.
Hippocampus modulation of H.l place seek and avoid. DA (dopamine), E.hc (hippocampus), H.l (lateral hypothalamus), Hb.l (lateral habenula), P.v (ventral pallidum), S.ls (lateral septum), S.v (ventral striatum), Vta (ventral tegmental area).
H.l has neurons that represent food zones and non-food zones [Jennings et al 2015], presumably using E.hc place information, although possibly using P.bst (bed nucleus of the stria terminals) as an intermediary.
H.sum completing consensus loop
The consensus circuits needs to return the final action and motor choice back into the early layers, otherwise the motivation circuit wouldn’t know if a lower-level startle or OT looming escape took priority of the seek path. With analogy to the ascidian larva, this role resembles the AMG (ascending motor ganglia) neurons, which I associated with P.ldt and V.mr. For this consensus narrative, I’m taking H.sum (supramammillary) as the primary feedback node with an assist from Poa (preoptic area) to complete the loop to Hb.m and M.pag (periaqueductal gray).
H.sum as completing the consensus loop, linking the habenula output back to habenula input. H.l (lateral habenula), H.sum (supramammillary nucleus), Hb.m (medial habenula), MLR (midbrain locomotor region), M.pag (periaqueductal gray), P.ldt (laterodorsal tegmentum), Poa (preoptic area), R.ip (interpeduncular nucleus), R.rs (reticulospinal), V.mr (median raphe), V.pt (posterior tuberculum).
H.sum has several sub circuits with different functions, which studies are only starting to untangle. H.sum tac1 (neurotransmitter aka substance P) is strongly associated with upcoming locomotion [Farrell et al 2021]. H.sum’s Poa projection is specifically associated with threat avoidant locomotion [Escobedo et al 2023].
V.mr and P.ldt are connected with R.rs and the bilateral OT circuit, and therefore have information about the selected action at the level of the hindbrain and motor afferent copies. Both are strongly connected to H.sum. H.sum also connects with M.pag (periaqueductal gray) and H.sum activates when M.pag.d is stimulated [Pan et al 2004]. H.sum also activates when the H.vm (ventromedial hypothalamus) threat nuclei are stimulated.
H.sum is immediately rostral to Vta and highly connected with it (not shown in the diagram.) H.sum contains some DA neurons itself, which are sometimes considered as an extension of A10, the Vta dopamine neuron area, although the neuron types differ [Yetnikoff et al 2014], [Menegas et al 2015].
H.sum is strongly connected with E.hc (hippocampus) and is one of the few external input to both E.dg (dentate gyrus) and E.ca2 (cornu ammonia), and is a major theta source to P.ms (medial septum), which drives E.hc theta. Its link to E.hc are important for both novel object exploration [Chen et al 2020], [Takahashi et al 2023] and social memory [Qin et al 2022]. Although I’m not yet adding E.hc to the essays, the novel object detection will be important soon to avoid repeated exploration of the same object.
Note that Poa has already participated in the Hb.m to R.ip circuit because Poa drives thermotaxis [Palieri et al 2024] as part of the original Hb aversive apical path.
M.pag tetrapod complications
In a sense, the vertebrate brain is designed around fish navigation, exemplified by the simple M-cell startle circuit that requires only three neurons between the acoustic sense and the swimming muscles. Although the direct Braitenberg-like connections to R.rs work for fish locomotion, tetrapod locomotion is more complex. M.pag (periaqueductal grey) is a central grey area surrounding the midbrain ventricle (“periaqueductal”), and it an inner ring to OT, which is immediately dorsal to it. Naming it “OT.dd” (deep, deep layer of OT) would not be unreasonable. Among other tasks like vocalization [Jürgens 1994] and hunting [Marín-Blasco et al 2020], M.pag provides a similar to R.rs but at a higher level, like syllables to phonemes. So in the following examples, M.pag can be viewed as similar functionality to R.rs.
Unlike R.rs, M.pag can access more sophisticated navigation. Where the M-cell can only turn left or right, M.pag can use OT for obstacle avoidance and even higher navigation of the hippocampus using H.pm.d (dorsal premammillary nucleus) [Wang et al 2021].
M.pag flight
M.pag implements innate behaviors, including flight, freezing, hunting, grooming, and vocalizations. The following diagram shows some of the looming flight circuitry [Zhou et al 2019]. As before, OT.m primarily processes the looming signal and OT.m sends input to M.pag.d as an integrated threat signal, where M.pag.d computes a threshold for responding to the threat [Evans et al 2018].
M.pag flight for the looming circuit. M.pag.d (dorsal periaqueductal gray), OT.m (medial, deep optic tectum), R.rs (reticulospinal), S.a (central amygdala), Vta.g (gaba neurons of the ventral tegmental area).
In the diagram, the second interesting path is through Vta.g (Vta GABA neurons) and S.a (central amygdala). Because OT.m and M.pag.d directly output to R.rs neurons, the projects to Vta.g and S.a aren’t required for motor control, but because of the distributed consensus system, other systems need to be informed of the looming response. S.a modulates defense, hunting, and eating systems, and Vta.g also inhibits the current action by suppressing dopamine, back to the consensus loop, suppressing any current seek action.
M.pag.vl avoidance
While M.pag.d is strongly associated with fast escape, M.pag.vl is more complicated with diverse functions including hunting [Franklin 2019], [Marín-Blasco et al 2020], vocalization [González-García et al 2024], and laughter [Klingbeil et al 2021]. Since this essay focuses on avoidance, where avoidance here isn’t the high speed predator escape of M.pag.d.
As discussed above, H.l is a central motivational node, filling a similar role to the central hunger node in the mollusk sea hare navigation. However, H.l is much more complicated than a simple hunger node. One developmental paper divided H.l into nine distinct regions [Diaz et al 2013], but that anatomical division understates the complexity. A genetic transcription analysis finds 15 glutamate and 15 GABA clusters [Mickelson et al 2019]. Interestingly, the Diaz study identifies their H.l.1 area with H.sum.l, treating H.sum.l as part of H.l.
In general, H.l.glu produces place avoidance and H.l.g enables seeking, but as mentioned above with at least 15 genetic types and 9 regions, this division is almost certainly an oversimplification.
H.l seek and avoid efferents. E.ca1.v (ventral hippocampus), H.l (lateral hypothalamus glutamate and GABA), Hb.l (lateral habenula), M.pag (periaqueductal gray), S.ls (lateral septum), Vta.g (ventral tegmental area GABA).
The H.l.glu to M.pag connection is certainly capable of driving motor avoidance. Interestingly, a different H.l population is part of the M.pag hunting circuit. Both Vta.g and Hb.l enter the motivation loop. I’ve added the E.ca1.v (ventral hippocampus CA1) input to H.l because E.hc.v (ventral hippocampus) is strongly associated with place, and E.hc.v specifically with aversive context.
R.pb peribrachial nucleus
R.pb (peribrachial nucleus) is a pain, alarm, feeding, and respiration hub in the prepontine isthmus area (r0-r1). As an alarm center [Campos et al 2018], R.pb is connected with escaping and avoiding circuits. As covered in essay 29 speed, it includes a high Co2 trigger that drives place avoidance. It also includes pain triggers for escape. R.pb has multiple functions defined more by chemical markers than topology. One study explored R.pb’s role in escape and avoidance behavior [Chiang et al 2020].
R.pb avoidance circuits. dyn (dynorphin neurotransmitter), H.vm (ventromedial hypothalamus), M.pag.l (periaqueductal gray), P.bst (bed nucleus of the stria terminalis), R.pb (peribrachial nucleus), RTPA (real-time place avoidance), S.a (central amygdala), tac1 (tachykinin 1 / substance P neurotransmitter)
R.pb.dl (dorsolateral R.pb) and R.pb.el are adjacent R.pb areas that are associated with alarm and pain responses. R.pb.dl receives direct N5 (trigeminal – head, jaw) and N.sp (spinal) pain input, including pain input marked by tac1 (tachykinin 1 peptide aka substance P). Relevant to this essay, the outputs divide between direct escape behavior with not learning and indirect avoidance behavior with learning. The M.pag.l projection produces flight and jumping. The S.a (central amygdala) and P.bst (bed nucleus of the stria terminalis – extended amygdala) projections produce real-time place avoidance and are capable of CPA (conditioned place avoidance) [Chiang et al 2020]. The R.pb example is useful because it combines a direct locomotion to M.pag with output to the slower consensus circuit.
Preoptic area
Poa (preoptic area) is a multifunctional area directly anterior to the hypothalamus and often considered part of the hypothalamus, although genetic markers suggest it’s more closely related to the forebrain. Like other brainstem areas, its functionality is more organized by genetic markers than topology.
The above diagram shows some of the Pom (medial preoptic area)functions. Temperature management has been discussed with a connection through Hb.m gradient following. Threat avoidance from signals from H.sum, H.pv (periventricular hypothalamus), or S.ls (lateral septum) can lead to RTPA through a M.pag projection [Escobedo et al 2023]. Local exploration, a RTPP function, also uses a M.pag projection [Shin et al 2023], and Pom can also enable hunting [Park et al 2018], although through a M.pag projection. The recent genetic research tools will likely unravel more of its functionality.
Poa has a strong projection to both Hb.m and Hb.l, suggesting that it’s an important node in the locomotion consensus circuit. In the thought experiment I’ve outlined above, Poa is part of the feedback system through H.sum, but Poa also receives E.hc.v (ventral hippocampus) input through S.ls (lateral septum), so it may be an important node in its own right.
Ppt / P.ldt
The ACh (acetylcholine) neurons near the midbrain-hindbrain boundary Ppt (pedunculopontine tegmentum) and P.ldt (laterodorsal tegmentum) are the core of the MLR. In simpler vertebrates like the lamprey, the MLR is only a single area, generally named P.ldt. In mammals, not only are P.ldt and Ppt split, but a chunk of locomotive action is in a different nucleus M.cnf (cuneiform). Although M.cnf is more of a direct locomotive area, the locomotive neurons don’t respect the anatomical boundary, but are a group of glutamate neurons spanning from Ppt to M.ncf, where Ppt and M.cnf are neighbors [Caggiano et al 2018]. Tetrapod locomotion is more complex than fish swimming, which may be a partial reason for the expansion and division.
Ppt is strongly reciprocally connected with the deeper layers of OT: OT.i for turning and sensory integration, and OT.d for seek and avoid. Its connections resemble the R.pgb (parabigeminal aka nucleus isthmi) which sustains attention for the OT.s (superficial OT) [Knudsen 2011] and covered in essay 19. R.pgb, Ppt, and P.ldt are sibling nuclei that develop from the same area in r1 that also produces R.pb and cerebellum granule cells [Pose-Méndez et al 2023].
Some P.ldt connections, emphasizing that Vta connections are collaterals of R.rs. H.sum (supramammillary), P.ldt (laterodorsal tegmentum), R.rs (reticulospinal), Vta (ventral tegmental area)
P.ldt is complicated by the relative lack of recent studies of its descending projections since [Cornwall 1990] and an over-focus on its Vta connection. Because neuron tracing in [Zhao et al 2023] suggests that P.ldt has more descending connections to R.rs than Ppt and that all Vta connections are collaterals of R.rs connections, studies like [Coimbra et al 2021] and [Liu et al 2022] that find locomotion through Vta projections could be produced by its R.rs projection. P.ldt has reciprocal connections with H.sum.
Vta
Although Vta (ventral tegmental area) is most studied for its ascending dopamine projections to S.v (ventral stratum) and F.pfc (prefrontal cortex), it also contains glutamate and GABA projections, including descending connections. Non-tetrapods like fish and lamprey have a homologous V.pt (posterior tuberculum) with prominent descending locomotor connections to MLR [Ryczko et al 2017], [Derjean et al 2010]. The earlier thought experiment for the development of a locomotor consensus split out an ancient V.pt from the mammalian Vta as a way of describing the old descending functionality.
Vta glutamate and GABA connections. H.l.glu (lateral hypothalamus glutamate), Hb.l (lateral habenula), M.pag (periaqueductal gray), OT (optic tectum), P.bst (bed nucleus of the stria terminalis), S.a (central amygdala), S.am (medial central amygdala), S.msh.pv (medial shell of the ventral striatum, parvalbumin neurons), Vta.da (ventral tegmental area, dopamine), Vta.g (Vta GABA), Vta.glu (Vta glutamate)
The above diagram shows some of the connections of the glutamate and GABA Vta [Taylor et al 2014], including projections to M.pag and to Hb.l that are direct locomotor for seek and avoid. The Vta is a main dopamine source for S.v and F.pfc with multiple distinct areas. Vta.m, which projects to S.msh (medial shell of S.v) is aversive, while Vta.l, which projects to S.lsh (lateral shell of S.v) and S.core (core of S.v) promotes seek [Szőnyi et al 2019]. Vta.m is non-reinforcing, as opposed to Vta.l, which is well-studied for reinforcement.
P.v ventral pallidum
P.v is a main output of S.v and the only output of S.ot (olfactory tubercle). As essay 29 covered, it’s an important sleep/wake node. For this essay, the important bit is a split between RTPP and RTPA depending on its output.
Calvigioni D, Fuzik J, Le Merre P, Slashcheva M, Jung F, Ortiz C, Lentini A, Csillag V, Graziano M, Nikolakopoulou I, Weglage M, Lazaridis I, Kim H, Lenzi I, Park H, Reinius B, Carlén M, Meletis K. Esr1+ hypothalamic-habenula neurons shape aversive states. Nat Neurosci. 2023 Jul;26(7):1245-1255.
Chen S, He L, Huang AJY, Boehringer R, Robert V, Wintzer ME, Polygalov D, Weitemier AZ, Tao Y, Gu M, Middleton SJ, Namiki K, Hama H, Therreau L, Chevaleyre V, Hioki H, Miyawaki A, Piskorowski RA, McHugh TJ. A hypothalamic novelty signal modulates hippocampal memory. Nature. 2020 Oct;586(7828):270-274.
Farrell JS, Lovett-Barron M, Klein PM, Sparks FT, Gschwind T, Ortiz AL, Ahanonu B, Bradbury S, Terada S, Oijala M, Hwaun E, Dudok B, Szabo G, Schnitzer MJ, Deisseroth K, Losonczy A, Soltesz I. Supramammillary regulation of locomotion and hippocampal activity. Science. 2021 Dec 17;374(6574):1492-1496.
Szőnyi A, Zichó K, Barth AM, Gönczi RT, Schlingloff D, Török B, Sipos E, Major A, Bardóczi Z, Sos KE, Gulyás AI, Varga V, Zelena D, Freund TF, Nyiri G. Median raphe controls acquisition of negative experience in the mouse. Science. 2019 Nov 29;366(6469):eaay8746.
The original impetus for this sleep essay was the idea that the basal ganglia could best be understood as a sleep and wake circuit [Kazmierczak and Nicola 2022]. After reviewing the rest of the brainstem sleep circuitry, it’s time to tackle the original problem.
Snr as a sleep/wake gate
Snr (substantia nigra pars reticulata) is the output node of the basal ganglia. It’s a set of GABA neurons that tonically suppress the majority of all brainstem motor areas including MLR (midbrain locomotor region), OT (optic tectum), and R.rs (hindbrain reticulospinal motor command) with corollary discharge to the thalamus. Snr can inhibit initiation of eating and motion [Rossi et al 2016], but don’t disrupt ongoing actions [Liu et al 2018]. Disruption of Snr can cause hyperactivity and insomnia [Geraschenko et al 2006]. The caudal Snr derives from hindbrain r1 (rhombomere r1 near the midbrain-hindbrain boundary) [Achim et al 2012], [Lahti et al 2015], [Partanen and Achim 2022], suggesting it may be evolutionarily old, possibly older than other basal ganglia regions.
Sleep as gating motive from action or sleep from action. Wake as disinhibiting sleep. Snr (substantia nigra pars reticulata).
As described in part 1 this essay, sleep suppresses senses, motivation and action. To implement this suppression, sleep could disconnect senses and motivation neurons from action neurons. In the above diagram, the gate is conceptual. The circuit could also inhibit the sense or action nodes directly instead of requiring specific gating neurons. This gating architecture has the advantage of simplicity because the sleep circuit can be localized in the gate, while the senses and actions can be mostly free of sleep circuitry.
As a preview, sleep neurotransmitters and peptides in BG (basal ganglia) include AD (adenosine), enk (enkephalin), MOR (μ-opioid receptor), and wake neurotransmitters include DA (dopamine), tac1 (tachykinin 1 aka neurokinin 1 aka substance P), dyn (dynorphin), and DOR (δ-opioid receptor).
If the vertebrate brain follows this architecture, Snr is well-placed to control that gate. Snr.m (medial Snr) projections have many collaterals to distinct motor areas and suppressing the wake-promoting areas covered earlier in this essay, which suggests widespread suppression as opposed to fine-grained control.
Snr.m gad2 connectivity. 60% of Snr.m inputs are from motor, motivation and wake areas. H.l (lateral habenula), H.stn (subthalamic nucleus), H.zi (zona incerta), M.pag (periaqueductal gray), OT.m (medial optic tectum), P.ge (external global pallidus), Ppt (pedunculopontine nucleus), R.rs (reticulospinal motor command), S.d (dorsal striatum), Snr.m (medial substantia nigra pars reticulata).
As the above diagram illustrates, despite its description as basal ganglia output, 60% of the gad2 (genetic marker), Snr.m inputs are outside of the basal ganglia, particularly from the midbrain (30%) and hypothalamus (10%) [Liu et al 2020]. Snr.m has two independent neuron types marked by gad2 and pv (parvalbumin), which are topographically organized with gad2 in Snr.m and pv in Snr.l (lateral Snr). While Snr.l.pv seems to be strictly motor related, Snr.m.gad2 are sleep related [Liu et al 2020]. However, [Lai et al 2021] reports Snr.l as sleep related.
Functional sleep and action requirements. Any ongoing action should suppress sleep, and sleep should suppress all actions.
Snr’s widespread motor and motivation connectivity suggests a possible primitive role in sleep. Sleep needs to suppress all actions, but any ongoing action needs to suppress sleep, because an animal shouldn’t fall asleep while eating or moving. It seems plausible that a primitive proto-vertebrate could have used Snr for sleep regulation without needing the rest of the basal ganglia.
Because astrocytes can integrate inputs spatially and temporally and are associated with sleep, it’s plausible that Snr astrocyte would be involved in this circuit. Interestingly Snr astrocytes are sensitive to dopamine and become hyperactive in the absence of dopamine [Bosson et al 2015] and are sensitive to glutamate from H.stn [Barat et al 2015].
Dopamine D2.i sleep / wake circuit
Although the independent Snr circuit is a functional sleep / wake gating circuit, it tonically inhibits the sense to action circuit, adding noise. An improvement to the circuit enables the gate when a signal is available, using the striatum to selectively open the gate. This circuit uses dopamine to open and close the gate. High dopamine is a wake signal and low dopamine is a sleep signal.
In the above diagram, Snr and S.d2 (D2.i associated striatum projection neurons) are sleep-promoting regions and S.d1 (D1.s associated striatum projection neurons) is a wake-promoting region. D2.i (inhibitory Gi-protein dopamine receptor) disconnects inputs, as opposed to inhibiting a neuron directly. When DA is available, S.d2 is disconnected, and S.d1 inhibits Snr, opening the gate. When DA is low, S.d2 is active, which inhibits S.d1, disinhibiting Snr, closing the gate and producing sleep. The D2i between S.d2 and S.d1 is from [Dobbs et al 2016].
The idea of the circuit is that the sense signal disinhibits itself during wake, but sleep prevents sense from disinhibiting itself. The minimal system only requires D2i circuits [Oishi et al 2017]. Wake enables the gate, and sleep disables the gate. Although I’ll cover D1s later, D2i is more fundamental because disabling D1s can be reversed by sufficient arousal, but disabling D2i can’t [Kazmierczak and Nicola 2022].
Note the diagram is somewhat incorrect, because direct S.d2 to S.d1 connection is weak [Tepper 2008]. Instead, S.d2 GABA inhibits S.d1 input at distal dendrites as opposed to inhibiting the neuron soma itself.
P.v ventral pallidum and S.core
While S.d2 neurons in model above suppresses motor for sleep, S.d2 in S.core (ventral striatum core aka nucleus accumbens) can produce sleep pressure by inhibiting the wake supporting P.v (ventral pallidum) [Oishi et al 2017]. P.v is a tonically active, wake-promoting nucleus, primarily inhibiting sleep areas or disinhibiting wake areas.
Sleep/wake control adding P.v as a tonic wake producing node. DA (dopamine), D2i (inhibitory Gi-coupled dopamine receptor), H.l (lateral hypothalamus), Hb.l (lateral habenula), M.pag (periaqueductal gray), Ppt (pedunculopontine nucleus – ACh), P.v (ventral pallidum), S.d1 (D1-associated striatum projection neuron), S.d2 (D2-associated striatum projection neuron), Snr (substantia nigra pars reticulata), V.mr (median raphe – serotonin), Vta (ventral tegmental area – dopamine).
P.v fill a similar wake-promoting role as S.d1, but unlike S.d1 it’s tonically active and affects the motivation loop of H.l, Hb.l, and Vta instead of gating sense from action. Where P.v supports general wake, S.d1 supports specific wake for an action. Like the previous basal ganglia sub-circuit, this sub-circuit only requires D2i receptors.
P.v promotes wake by inhibiting Hb.l sleep-producing system [Li et al 2023]. It also promotes wake through Vta by disinhibiting GABA interneurons [Li et al 2021]. (It could also disinhibit H.l orexin but I don’t have a reference).
In the model above, stimulating S.d2 inhibits wake-producing P.v, which disinhibits sleep-producing areas like Hb.l and inhibits wake-producing areas like H.l and Vta through GABA interneurons. Conversely, stimulating the D2i receptor by high DA inhibits S.d2, which disinhibits Pv, allowing it so promote wake. Disabling the D2i receptor activates S.d2, promoting sleep even with high dopamine [Qu et al 2010].
Note that S.d1 also connects to P.v and can produce wake [Zhang et al 2023]. P.v has multiple sub-populations with opposing functions. For example, it has both a hedonic hot spot for liked food and a cold spot for disliked food [Castro et al 2015]. For the sake of simplicity the diagram only shows a sleep-promoting path through S.d2, but there may be a wake-promoting path through S.d2 to an opposing P.v subpopulation.
D1s – stimulator dopamine receptors
Although using only D2i as a mode switch to the sleep path is functional, it can be improved by also enhancing the wake path with D1s (stimulatory Gs-coupled dopamine receptor).
D1s as enhancing the basal ganglia wake path. DA (dopamine), D1s (stimulatory Gs-coupled dopamine receptor), D2i (inhibitory Gi-coupled dopamine receptor), S.d1 (D1-associated striatum projection neuron), S.d2 (D2-associated striatum projection neuron), Snc (substantia nigra pars compacta – dopamine), Snr (substantia nigra pars reticulata).
The improved circuit works exactly like the D2i-only circuit but enhances the wake path when DA is available. Dopamine boosts both the signals from the sense to S.d1 and the signal from S.d1 to Snr [Salvatore 2024], [Kliem 2007], [Rice and Patel 2015]. When dopamine is available, it boots the sense to S.d1 signal with D1s, which more strongly disinhibits the gate by inhibiting Snr, which is also boosted by D1s.
The D1s in Snr and dopamine may be more important for motor suppression than dopamine in the striatum [Salvatore 2024]. In Parkinson’s disease and also normal aging, bradykinesia (slow movement) correlates with dopamine in Snr more closely than dopamine in the striatum. Motor symptoms in Parkinson’s disease don’t generally occur until striatal dopamine is reduced by 80%, but the effect on Snr is more immediate with only a small drop of dopamine.
Note that the Snc (substantia nigra pars compacta) to Snr dopamine comes from somatodendritic broadcast, not from an axon synapse. Snc dendrites in Snr produce dopamine to enhance the S.d1 to Snr connection.
Although the previous diagrams show the basic logic of the circuit, the basal ganglia use adenosine as a sleep-producing neurotransmitter, competing with dopamine.
Adenosine in striatum sleep
Adenosine is a product of the energy molecule ATP and is produced by neural activity, and also as a astrocyte transmission molecule. Although adenosine can accumulate in a circadian manner, particularly in P.bf (basal forebrain), it’s typically a shorter term sleep pressure. Caffeine is wake promoting by suppressing adenosine receptors.
Dopamine and adenosine are paired, opposing neurotransmitters in the basal ganglia: dopamine produces wake and adenosine promotes sleep. As an opposing signal to dopamine, the adenosine circuit is a flip version of the dopamine circuit.
Parallel adenosine sleep circuit in the basal ganglia. AD (adenosine), A1i (inhibitory Gi-coupled adenosine receptor), A2a.s (stimulatory Gs-coupled adenosine receptor), S.d1 (D1-associated striatum projection neuron), S.d2 (D2-associated striatum projection neuron), Snr (substantia nigra pars reticulata).
When adenosine is active in the above circuit, it cuts off S.d1 input and output and enhances S.d2’s suppression of S.d1. With S.d2 fully suppressed, Snr is free to suppress the gate and therefore suppress sleeping action.
Since adenosine is low in the morning, sleep is suppressed, which is enhanced by high ultradian morning dopamine. If A2a.s (stimulating Gs-coupled adenosine receptor) are stimulated in the striatum, the animal is more likely to sleep even in the morning [Yuan et al 2017], specifically in S.core not S.sh (ventral striatum shell aka nucleus accumbens) [Oishi et al 2017].
The dual signal system allows for interesting combinations at the boundary between sleep and wake. If adenosine is high with sleep pressing, then a large amount of dopamine motivation is required to continue wake. In fact, sleep deprivation down regulates D2i receptors, moving from the neuron membrane to the interior [Volkow et al 2012], which tips the balance toward sleep by diminishing the D2i-mediated wake signal. Caffeine inhibits both the A1i (inhibitory Gi-coupled adenosine receptor) and A2a.s receptors, tipping the balance to dopamine wake.
Dorsal striatum indirect path
The full S.d (dorsal striatum) path includes an indirect path, but this path may be more related to pure motor control, not sleep. As mentioned above, Snr divides into two populations Snr.l with pv neurons and Snr.m with gad2 neurons, and the Snr.l neurons are motor related, not sleep related [Liu et al 2020]. Similarly, the indirect path including P.ge (external globus pallidus) and H.stn (sub thalamic nucleus) may not be sleep related. Nevertheless, I’ll include it here, in case it is sleep related.
S.d model with indirect path included. DA (dopamine), D1s (stimulatory Gs-coupled dopamine receptor), D2i (inhibitory Gi-coupled dopamine receptor), H.stn (subthalamic nucleus), P.ge (external globus pallidus), S.d1 (D1-associated striatum projection neuron), S.d2 (D2-associated striatum projection neuron), Snc (substantia nigra pars compacta), Snr (substantia nigra pars reticulata).
Note that both P.ge and H.stn are tonically active, and they oscillate together at beta frequencies (roughly 10hz), which suppresses action. An excessive beta oscillation in this P.ge and H.stn circuit is a Parkinson’s disease symptom that suppresses motion and can also interrupt sleep. D2i receptors in H.stn mean that dopamine suppresses H.stn output [Shen et al 2012].
One significant experiment showed that lesioning P.ge increased wake by 40%, particularly eliminating normal circadian night-time sleep, replacing it with day-time like napping [Qiu et al 2016], which would suggest that P.ge is a major sleep center like Po.vl (ventrolateral preoptic area) [Vetrivelan et al 2010]. Note that this analysis would suggest that my basal ganglia sleep diagram is entirely wrong, because P.ge as a sleep center is basically incompatible with its position in the circuit.
P.ge – external globus pallidus
Lesioning P.ge increases wake by 40%, almost entirely eliminating circadian sleep [Qiu et al 2016]. However, this produces hyperactive chewing, weight loss, abnormal motor behavior and death in 3-4 weeks [Vetrivelan et al 2010]. Other manipulations of P.ge produce hyperactivity, abnormal movement, and odd stereotypical behavior [Gittis et al 2014]. So, it’s unclear to me that P.ge is a sleep center, but removing P.ge produces excessive action which then suppresses sleep.
In addition, P.ge is a heterogenous area with at least three major cell types with distinct projections and roles. Arkypallidal neurons project strongly and exclusively to the striatum. Lhx6 neurons project strongly to Snc and to some areas of H.stn, excluding the center. Pv neurons project to all of H.stn and also to T.pf (parafascical thalamus) [Gittis et al 2014].
Distinct projection neuron types of P.ge. H.stn (subthalamic nucleus), P.ge (external globus pallidus), Snc (substantia nigra pars compacta), Snr (substantia nigra pars reticulata), Spn (striatal projection neuron), Spv (pv marked striatum interneuron), T.pf (parafascicular thalamus).
With three projection types, it’s possible that they have entirely separate functions. For example, the lhx6 projections are functionally compatible with a sleep promoting role, and lhx6 neurons in H.zi (zona incerta) are sleep promoting [Liu et al 2017].
Liu K, Kim J, Kim DW, Zhang YS, Bao H, Denaxa M, Lim SA, Kim E, Liu C, Wickersham IR, Pachnis V, Hattar S, Song J, Brown SP, Blackshaw S. Lhx6-positive GABA-releasing neurons of the zona incerta promote sleep. Nature. 2017 Aug 31;548(7669):582-587. doi: 10.1038/nature23663. Epub 2017 Aug 23.
