A decentralised neural model explaining optimal integration of navigational strategies in insects

  1. Xuelong Sun
  2. Shigang Yue
  3. Michael Mangan

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Insect navigation arises from the coordinated action of concurrent guidance systems but the neural mechanisms through which each functions, and are then coordinated, remains unknown. We propose that insects require distinct strategies to retrace familiar routes (route-following) and directly return from novel to familiar terrain (homing) using different aspects of frequency encoded views that are processed in different neural pathways. We also demonstrate how the Central Complex and Mushroom Bodies regions of the insect brain may work in tandem to coordinate the directional output of different guidance cues through a contextually switched ring-attractor inspired by neural recordings. The resultant unified model of insect navigation reproduces behavioural data from a series of cue conflict experiments in realistic animal environments and offers testable hypotheses of where and how insects process visual cues, utilise the different information that they provide and coordinate their outputs to achieve the adaptive behaviours observed in the wild.

Article activity feed

  1. Version published to 10.7554/elife.54026 on eLife
  2. Version published to 10.1101/856153v3 on bioRxiv
  3. Version published to 10.1101/856153v2 on bioRxiv
  4. eLife

    ###Reviewer #3

    This manuscript describes a complete model of robust insect navigation. The originality of this remarkable work relies on a clear endeavour to describe the neural basis of each function involved in the homing behaviour of the ant. This paper focuses on the neural processing related to various theoretical hypotheses in terms of signal processing. Several previous studies replicated the route following behaviour but did not account for visual homing, i.e., the ability of the ant to return to familiar regions from novel locations. The proposed model extends the one proposed by Webb in 2019 to account for two very challenging points: the ability of the ants to home from new locations and the ability of the ant to switch between strategies according to the context.

    Major points:

    • I was very surprised by the slow velocity of the simulated ant (Vo = 1cm/s) compared to the real one (about 50cm/s). Why is the speed so slow? This point must be discussed. Is there any fundamental reason?
    • Concerning the path integration strategy, the distance does not seem to be measured (odometer) or included in the model.
    • What would happen to the simulated ant if an obstacle was placed on the familiar route? What is the robustness of the Zernike-based moment algorithm to the unpredicted presence of an obstacle that could appear during the homing? I suggest doing additional simulations in this sense that could show the robustness of the proposed navigation model. These new simulations could be in line with the well-known experiments proposed by Wehner and Wehner (Insect navigation: use of maps or ariadne's thread?).

    Page 16, lines 417: would it be possible to plot Crf with respect to angular orientation of the simulated ant in various places (every 10° steps for example)?

  5. eLife

    ###Reviewer #2

    The beautifully illustrated manuscript by Sun et al is a challenging but highly rewarding, interesting and intellectually stimulating modeling study that proposes a unified model of insect navigation, which, at least in large parts, is constrained by neuroanatomical and physiological data. It elegantly combines previous models of path integration of the central complex and visual learning in the mushroom body (underlying visual homing) and proposes a third model for habitual route following. In the end, all three models are integrated and mapped onto known neural structures of the insect brain, most notably the central complex and the mushroom body. The information extracted from the environment is decomposed using a novel method that separates rotationally invariant feature information from rotational variant directional information. While the first is used to carry out visual homing based on image familiarity, the second is used to follow habitual routes. The important novelty in the paper is that this new information processing strategy allows to integrate all mentioned navigational modules. Moreover, it does so using previous biologically constrained models and expands this basis towards a full system that can replicate numerous behavioral data from ants, including difficult experiments, in which ants have to trade off different strategies against each other. I highly welcome this paper as an important addition to both the literature on the insect central complex, as well as to more theoretical navigational work, in particular as many predictions can be made based on the presented models. Nevertheless I have several points that need to be addressed.

    Major comments:

    1. Accessibility to a broad readership. While the general text is written very well and the content is highly interesting for a life science (in particular insect neuroscience) audience, the methods section and some aspects of the reasoning behind the model are very technical. Being an insect neurobiologist myself, I struggle to follow large parts of the methods and had admittedly never heard of Zernike moments. Given that the mathematical model and the concepts of frequency analysis are the foundations of the paper, I suggest to add some more intuitive and broadly accessible language that would allow a biologist to grasp at least the key principles of what is done by those initial analyses of the visual information in the model (of course, the math is needed for a computational audience and essential for replication of the model, but a few additions might go a long way for biologists). A schematic illustration as to what Zernike moments are, maybe combined with some simple examples might help a lot. This is important as the paper is not only directed towards computational biologists, but is highly relevant also for physiologists, anatomists and behaviorists, most of whom (extrapolating from my own mathematical ignorance) probably fail to grasp the essence of the new principles presented.
    1. Neuroanatomical correspondence of model details: The paper claims that the model is in most parts biologically constrained and that most elements can be mapped onto known neurons. Where this was not possible (route following) the authors speculated about the possible implementations. While on the levels of neuropil groups this is all quite true, the details, especially in the central complex, are less clear and many of the proposed circuits have no known counterpart in any insect brain to date. This is not saying that those parts of the model are not realistic or interesting, but that the claim that they correspond to existing neurons in the central complex, is slightly misleading. I will list a series of obvious mixups of cell types below, which need to be corrected (2.1), but additionally, it should be clearly stated where the model does not (yet) have a solid grounding in biology (see point 2.2). Finally, the speculative route following implementation seems at odds with neurophysiological data from various species and alternative pathways and implementations seem more likely (point 2.3).

    2.1)

    • Line 126: CPU3 neurons are supposed to be a mirrored TB1 ring attractor network? I'm not sure if this is what the authors want to say, as CPU3 neurons are known in locusts (Heinze and Homberg, 2008), but connect the PB with the FB as columnar cells. If the authors mean CPU4 cells, these neurons are also not forming a ring-network (even though they could receive shifted compass information from TB1 cells by some means). Most simply, would not a parallel set of TB1 cells be optimally suited for this task? There are four TB1 cells for each column in the PB, potentially enough for four parallel ring attractors. These cells are neurochemically distinct and could function independently (see Beetz et al, 2015).
    • There is no known direct connection between the EB and the FB (proposed in figure 4)
    • There is no direct connection from the OL to the CX (indicated in caption of figure 1 as underlying PI).
    • line 348: CL2 neurons should be CL1 (CL2 correspond to fly P-EN neurons, not E-PG)
    • In the PI section of the methods, sometimes TN cells are referred to as TN2 cells or just as TN cells. TN2 is one of two types of TN cells (tangential noduli neurons) and was the one primarily used for the standard model of Stone et al 2017. Please be consistent. Also, the tuning cells of the visual homing circuit are called TN cells. This is very confusing and should be changed.

    2.2) There are no known ring attractors in the FB. The only ring attractor shown experimentally is the one in the EB/PB, which employs recurrent feedback loops with the PB (E-PG/P-EN/P-EG cells; equal to CL1a, CL2, and CL1b) and inhibitory neurons in the PB (TB1 or delta7 cells). While a similar recurrent connection pattern is thinkable in the FB as well, using unknown types of columnar cells, there is no experimental support for that. Pontine cells might also form local connections that could result in a RA, but that is even more speculative. Please clearly state that the numerous RAs required by the model are hypothetical and have not yet any biological correspondence in the form of identified cell types. Also, I suppose not all the neuron rings drawn in the figures are ring attractors. I suggest making that distinction clearer (the many abbreviations for the different neuron rings do not make this easier to follow either).

    2.3) The authors assume a second compass system in the PB that is fed directly from the OL via the posterior optical tract. There is no evidence for this beyond a single cell type from locusts that connects the accessory medulla (circadian clock) to the POTU, which is also innervated by TB1 neurons. However, there is no connection to the visual part of the OL, and no physiological data exists on the AME->POTU connection. In contrast, the anterior optic tract via the AOTU has been shown in Drosophila to contain many neurons that respond to visual features and they converge on the head direction cells in the EB via a recently resolved mechanism. It seems odd to ignore this known compass pathway and propose another one for which no evidence exists. That said, the authors use the anterior pathway to construct a desired heading via an ANN residing in the AOTU/BU pathway, information that is then used to feed into an EB ring attractor that then connects to additional attractors in the FB. Whereas the EB attractor (in conjunction with the PB) exists, there is no evidence for FB based ring attractors and there is no known direct connection between the EB and the FB. While this all results in a really nice figure, it unfortunately is misleading and based on not enough evidence to show it so prominently (readers might easily take it for factual).

    If I may, I would like to point out that there is an alternative solution for at least the compass problem: There are four individual CL1 cells in each column of the EB in locusts as well as in flies (EPG/PEG cells). While they are identical in their projection patterns, some connect the PB to the EB and others connect the EB to the PB, so that there are in theory enough cells to form two parallel recurrent loops (needed to maintain a head direction signal). One of them could be driven by landmarks, while the other could be driven by global compass cues. Whereas the current idea is that both inputs converge on a single head direction signal (celestial and local cue based), this might not be true, given that local cues have been tested in Drosophila and global cues in locusts and some other species. These neurons are neurochemically distinct and most likely play different functional roles.

    Finally with respect to the desired heading, a short term plasticity based, associative mechanism linking the phase of the head direction signal and the local environment was recently demonstrated in Drosophila (Fisher at al. 2019 and Kim et al, 2019). The authors state that several of these phases can be stored and retrieved in each respective environment. To me this sounds very close to what the authors of the current study suggest for routes in ants. Please consider these points and revise the proposed circuit identity accordingly.

    1. The overall layout of the model is not fully clear to me from the paper. The authors present many (nicely illustrated) parts of the model, but I fail to reconcile some of the partial models with one another and have no immediate way of seeing how many neurons there are overall, or what their complete connectivity patterns are. I assume this is all obvious from the code itself, but being a neuroanatomist and physiologist, I struggle to get an intuition for the circuits based on Python code. This hinders independent interpretation and finding alternative solutions for mapping the model onto anatomical neural circuits once newly discovered neurons become available in the future. I suggest including (at least in the supplements) a full graphical depiction of the model with all existing neurons and their connections. Maybe using a force directed graph diagram like used by the authors of Stone et al. 2017 for their path integration model results in a model illustration that is intuitively understandable for researchers who think more in terms of anatomy. But even if it turns out to be somewhat messy, it would still be helpful.
  6. eLife

    ###Reviewer #1

    This is an interesting and timely study on a topic of considerable interest: computational strategies used by insects to perform their remarkable navigational feats. The authors identify shortcomings in existing models – specifically, that they do not account for the entire range of capabilities and the flexibility that the most accomplished of insect navigators display – and integrate and build upon prior models to successfully fill these gaps. The integrated model pins specific computational functions on specific anatomical structures, making it, in principle, testable in the near-medium term. The figures are well-made and the writing is compact but readable. Here are a few specific concerns:

    1. It is entirely reasonable that the authors combine experimental and modeling work from a range of different insect species to build different pieces of their own model. By and large they are careful to state which is which. However, they could make it clearer which assumptions are based on experimental data and which are based on prior models (i.e., not actual data). As an example, although the mushroom body has been suggested by numerous modeling studies and conceptually driven reviews to be involved in visual navigation, the experimental evidence for this is lacking, and their precise role is far from well-established.

    2. I commend the authors for integrating useful components from prior models to construct their integrated model, but, although the figures go some way towards clarifying how the different pieces might fit together, it would be useful to make even clearer what is entirely novel here and what is derived/integrated from previous work. In addition, although the authors make a testable case for the involvement of the fan-shaped body in a series of different navigational computations, controlled by the mushroom body, the figures are still somewhat complex and confusing. Please try and further clarify them.

    3. The authors could derive more constraints from the fly physiology literature than they do. As examples, Fisher et al., Nature, 2019 and Kim et al., Nature, 2019 have relevant findings relating to plasticity in mapping visual stimuli onto a compass representation. Turner-Evans et al., eLife, 2017 has a data-driven ring attractor model that is relevant, and Turner-Evans, bioRxiv, 2019 features data demonstrating that the fly compass for current heading relies on visual input from the anterior optic tubercle, contrary to the authors' assumption deriving from an anatomical pathway from the posterior optic tubercle to the protocerebral bridge (175-176). On a somewhat related note, the fly heading system does not necessarily show 'bar following' in open loop (line 164): the experiments cited (Seelig & Jayaraman, 2015) were performed in closed loop, with the animal controlling bar position.

  7. eLife

    ##Preprint Review

    This preprint was reviewed using eLife’s Preprint Review service, which provides public peer reviews of manuscripts posted on bioRxiv for the benefit of the authors, readers, potential readers, and others interested in our assessment of the work. This review applies only to Version 1 of the preprint: https://www.biorxiv.org/content/10.1101/856153v1

    ###Summary

    This is an original, focussed study that offers a new model to explain the neuronal “computation” that underlies insect navigation. The authors identify shortcomings in existing models – specifically, that they do not explain the entire range and flexibility of insect navigational capabilities – and integrate and build upon prior models to successfully fill these gaps. The integrated model is particularly valuable because it relates specific computational functions to specific anatomical structures, most notably the central complex and the mushroom body. It is an important addition to both the literature on the insect central complex, as well as to theoretical work on insect navigation. Many testable predictions can be made based on the presented models. The figures are well made and the writing is compact. Nevertheless, several points need to be addressed.

  8. Version published to 10.1101/856153v1 on bioRxiv