The digital sphinx: Can a worm brain control a fly body?

  1. Bingni W. Brunton
  2. Elliott T.T. Abe
  3. Lawrence Jianqiao Hu
  4. John C. Tuthill

  • Curated by eLife

    eLife logo

    eLife Assessment

    What can a neural network trained to imitate animal behavior tell us about biology? This valuable work uses deep reinforcement learning to train an artificial neural network to transform the dynamics of a recurrent neural network based on the C. elegans connectome into a convincing adult Drosophila walking program in a physical model of the fly body, demonstrating that achieving plausible output dynamics does not in and of itself imply biologically meaningful simulation. Evidence for this basic claim is solid, but more extensive analyses, better methodological description, and a discussion of deeper fundamental challenges in the undertaking of biological brain modeling would strengthen the study. This result demands the attention of the practitioners of the growing field of connectome simulation for the purpose of gaining mechanistic understanding of nervous system function.

Reviewed by

Read the full article See related articles

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

Log in to save this article

Abstract

Animal intelligence is not purely a product of abstract computation in the brain, but emerges from dynamic interactions between the nervous system and the body. New connectome datasets and musculoskeletal models now enable integrated, closed-loop simulations of the neural and biomechanical systems of the fruit fly Drosophila , an ideal model organism to investigate embodied intelligence. However, many biological parameters of the nervous system and the body, as well as how they interface, remain unknown. To fill such gaps, researchers are turning to deep reinforcement learning (DRL), a data-driven optimization framework, to create virtual animals that imitate the behavior of real animals. Here, we provide a cautionary tale about the interpretation of such models. We constructed a virtual chimera of two phylogenetically distant species: a connectome of the C. elegans nematode worm and a biomechanical model of the fly body. The worm connectome receives sensory information from the fly body, and an artificial neural network is trained with DRL to map worm motor neuron activations to the fly’s leg actuators. The resulting digital sphinx produces highly realistic fly walking—yet it is biologically meaningless. This exercise teaches us nothing about either animal and exposes a core peril of connectome-body models: behavioral fidelity is achievable without biological fidelity, making such models easy to overinterpret. Done carefully, virtual animals can be powerful partners to biological experiments, but only if their components and interfaces are grounded in biology.

Article activity feed

  1. eLife

    eLife Assessment

    What can a neural network trained to imitate animal behavior tell us about biology? This valuable work uses deep reinforcement learning to train an artificial neural network to transform the dynamics of a recurrent neural network based on the C. elegans connectome into a convincing adult Drosophila walking program in a physical model of the fly body, demonstrating that achieving plausible output dynamics does not in and of itself imply biologically meaningful simulation. Evidence for this basic claim is solid, but more extensive analyses, better methodological description, and a discussion of deeper fundamental challenges in the undertaking of biological brain modeling would strengthen the study. This result demands the attention of the practitioners of the growing field of connectome simulation for the purpose of gaining mechanistic understanding of nervous system function.

  2. eLife

    Reviewer #1 (Public review):

    Summary

    The authors build a "digital sphinx" by stitching together two neural network models: (i) a recurrent network with fixed parameters derived from the C. elegans connectome and imputed physiological (e.g. neural input/output) functions, and (ii) a feedforward encoder-decoder model with learnable parameters intended to represent a central brain - to - motor interface, then harnessing the combined model to a Drosophila biomechanical model situated in a physics simulator, and finally using deep reinforcement learning (DRL) training to optimize the parameters of the encoder-decoder model to reproduce a set of spatiotemporal patterns of jointed limb activations that together produce the overall organismal behavior of walking, within the physics simulator.

    The primary intent of this paper is to dispel the recent grandiose claims made in the mainstream press by a private company, Eon Systems, to have achieved a major advance in biologically based brain simulation of the production of a set of ethologically relevant motor behaviors by the fly. Representatives of the company referred to this modeling and training process euphemistically and deceptively as "brain uploading". The authors proceed with a reduction-to-triviality exercise by constructing their own high-parameter dynamical brain-plus-body model situated in a physical simulation that produces, after training by reinforcement learning, satisfying ethological behavioral imitation in the same vein as the private company claim, but based on a clearly absurd and biologically unrealistic set of model assumptions.

    Secondarily, the paper provides two overall admonitions that they assert their computational demonstration illustrates: that training high parameter network models to imitate behavior, even if they possess some biological detail, will deliver little or no biological insight, and that models of behavioral generation must be built from detailed biological data and, crucially, developed in a hypothesis generation/falsification loop with experimental validation, in order to be scientifically useful.

    Appraisal

    The authors are well justified in challenging the non-rigorous claims of "uploading" or even the delivery of a neurobehavioral simulation with potential scientific utility, in unison with the vocal criticisms of many other researchers in the fields of AI and neuroscience, and it is an important message to deliver to the world. However, the authors' own modeling counter-exercise, while clever and vivid in imagery, suffers from its own lack of rigor, both in disclosure of implementation and in scientific case-making. Some sacrifice of clarity and thoroughness in the interest of brevity is inevitable under the brief format of this manuscript; however, we suggest that crucial additions and modifications should be made to avoid falling into a similar trap of non-rigorous sensationalism.

    Because the private company claims were not accompanied by a scientific paper, preprint, code repository, or much methodological disclosure of any kind, the authors have the particular challenge of building a refutation case against an undefined target. As a consequence, the authors chose their own task, model structure, and training paradigm.

    The authors argue that brain models need to be built from biological data to be useful for yielding biological insight. We agree with the overall principle; however, in practice, this procedure is fraught with epistemological difficulty. Biological modeling suffers from a unique challenge within the larger endeavor of scientific/physical modeling, which is that it is generally unclear as to precisely what biological quantities should be measured and at what level of detail they should be measured. Additionally, biological data will by necessity be incomplete and noisy, and thus decisions of coarse-graining must be made at the outset of large-scale data collection projects, and some, possibly a substantial, level of data imputation will have to be performed in order to build testable models in our lifetimes. Despite the astonishing success of scaling (in both parameter count and corpus size) in engineered neural networks for certain human-like tasks, it is not at all clear that simply adding more detail to biological models will produce deeper scientific insight, or whether cataloging parameters from snapshot data will yield functional simulations. The failed Blue Brain mega-project should provide a lesson, as well as Marder's longstanding work on parameter variation in neural systems. The coupled, pernicious questions of choosing measurement detail and modeling detail represent a deep, unsolved challenge area for the field, and this context should be raised in the text.

    The message about overinterpreting models trained with deep reinforcement learning, while valid and important, should be broadened to be a message about overinterpreting trained high-parameter models in general, in their ability to fit data or reproduce simple behavior. Other parameter optimization/learning procedures for building underdetermined and/or high-parameter models risk the same misinterpretation. The prescription of building models in conjunction with experimental prediction and validation is an important point.

    The authors leave out an additional important and underappreciated challenge of brain-model-building, which is that imitating a time segment of behavior is a computational task of unspecified, and possibly low complexity. Successful recapitulation of behavioral time series may simply not be considered cognitively interesting, even if the model is built entirely on biological data. While quantifying task complexity is another open area of computational and neuroscientific research, the authors should, at a minimum, describe their particular task data in explicit mathematical terms and preferentially provide some complexity analysis. In the absence of task complexity analysis, at a minimum, computational controls should be applied to demonstrate the necessity of whatever structure or data is being asserted in the model. This epistemological practice is glaringly absent in much, if not most, of the neurobehavioral modeling literature. This paper would be a good opportunity to set an example of rigor.

    Finally, the authors' description of prior work in the field of whole-organism neurobiological simulation feels incomplete and skewed toward work in Drosophila versus other model organisms. An internet search reveals many published efforts to build neurobehavioral models at varying levels of detail in C. elegans, of which only two are referenced.

    We do feel this work constitutes an illustrative scientific exercise and important counterpoint to the sensationalism building around efforts in neurobiological simulation. It should inspire further work in defining a rigorous and scientifically productive epistemological framework for these kinds of brain modeling efforts.

    Further Comments

    (1) The authors oversell the completeness and quality of connectome datasets and what they lack.

    Language such as "complete wiring diagrams," "nearly comprehensive connectomes" neglects the well-appreciated gaps in biological data that most practitioners believe necessary for useful, detailed models to be built. There is a brief mention that biological parameters "remain unknown" and that interfaces are "incompletely characterized", but beyond that, the authors do not explain which parameters are missing, why these parameters might matter, and what still needs to be addressed in order to make any plausible whole-brain emulation claims. This may also inadvertently bolster the sensationalist claims that the manuscript is trying to deflate by giving the impression that neurobiological and physiological data collection is a near-complete exercise.

    (2) Prior work in C. elegans neurobehavioral modeling should be more acknowledged, if nothing else, for why it has been largely unsatisfying.

    C. elegans is rarely discussed, while Drosophila is primarily focused on. The status of C. elegans connectomics, physiological mapping, biomechanics, and neurobehavioral modeling is worth more treatment.

    (3) Critiques of Eon Systems announcements also, by and large, apply to more detailed and disclosed efforts in neurobehavioral modeling using RL for parameter imputation, and this should be recognized.

    By way of reference to a tweet in the first paragraph, the authors are responding to a recent claim made by a startup that they have fully "uploaded" a fly brain, a significant advance vis-à-vis prior work in neurobehavioral modeling in Drosophila, such as references [3 and 9], which are mentioned as background in the paper but left out of the methodological critique. But one of the central warnings of the paper is around the challenge of interpretability when using reinforcement learning to optimize model parameters. The authors also should acknowledge that the use of RL has been justified by building neurobehavioral model builders as a proxy for the learning and tuning processes thought to occur during animal development.

    (4) Substantiate the reservoir computing explanatory claim with appropriate computational controls.

    The reservoir computing idea is the only piece of hypothesizing a necessary function for the central brain component model in the paper. This claim could be substantiated with some basic computational controls rather than just hypothesized. We suggest the following possibilities as additions to the model: (a) replace the connectome with an RRNN, (b) shuffle the connectome, or (c) use other simple dynamical systems in place of the worm brain model.

    Specific Manuscript Comments

    (1) Abstract

    "New connectome datasets and musculoskeletal models now enable integrated, closed-loop simulations of the neural and biomechanical systems of the fruit fly Drosophila, an ideal model organism to investigate embodied intelligence."
    This sentence could mislead non-specialists into thinking all current simulations are novel because the connectome datasets are new. In fact, FlyWire (2024), NeuroMechFly (2022), and other connectomes have already been available for some years now. We believe that this sentence is a chance to make the opposite point that these resources have existed for a while, and that many simulations have been built before.

    "However, many biological parameters of the nervous system and the body, as well as how they interface, remain unknown."
    Some examples of specific parameter/physiological data types that are missing and thought to be critical, such as neuronal input/output functions, are warranted. See below for a comment on the confusing construct of "interface" as a distinct entity from the neural network.

    (2) Introduction

    "Among animals that walk, the integration of brain wiring and body models is perhaps closest to fruition in Drosophila, due to the recent completion of multiple complete wiring diagrams (known as connectomes) of the fly nervous system." ...and... "The fly is the only animal with legs for which nearly comprehensive connectomes of its brain and nerve cord exist."
    The walking qualifier allows the authors to skirt around the substantial and decades-long work on connectomes in C. elegans, which crawls and does not walk. Yet sinusoidal crawling is a multidimensional, adaptive behavior, so it seems this exclusion was for narrative convenience rather than contextual accuracy.

    "Despite this progress, closed-loop integration of biomechanical and neural models remains far from straightforward."
    Work (and shortcomings) in C. elegans neurobehavioral modeling should also be stated here alongside the fly.

    "Where interfaces between brains and body models are missing or only partially characterized, one approach is to train an artificial neural network (ANN) to approximate these interfaces with deep reinforcement learning (DRL)."
    The choice of "interface" as a distinct, well-defined neurobiological entity is somewhat confusing and may mislead non-practitioner readers. If neuronal and muscular (and their interactions) physiology are incorporated into a neurobehavioral model, then in principle there is nothing left to call an "interface". It would be clearer to explain that prior neurobehavioral models have often inserted a trainable multilayer feedforward network between sensory and central brain and between the central brain and motor effectors in order to have a substrate for learning, and that this insertion may render the entire biological modeling exercise scientifically pointless, or at a minimum require a set of computational controls.

    "In building virtual animal models, a motor policy is commonly learned by DRL so that the integrated, closed-loop virtual body successfully mimics the detailed kinematics of real animal behavior."
    The authors could define "motor policy" in simple terms and give a brief example.

    "Additional realism is added when the motor policy network is constrained by a connectome dataset. However, many biophysical parameters for individual neurons and synapses remain un-measured."
    "motor policy network" is confusing; this is referring to the entire network model here, presumably.

    (3) Methods

    "We used the adult hermaphrodite C. elegans nematode connectome dataset [15, 16, 5], including the identities of its 302 neurons and their synapses (Fig. 1A)."
    We believe the authors should specify the dataset type, which is a structural, unsigned connectome lacking grounding in physiological function.

    "The policy network was trained in closed loop using PPO as implemented by MIMIC-MJX"
    The authors should define "PPO" and "MIMIC-MJX" in simple terms and explain why they were used.

    (4) Discussion

    "Its role in the movement policy could be fulfilled equally well by a randomly connected RNN, akin to reservoir computing [20], since all the learning happens in the black-box ANN motor decoder."
    See above - this computational exercise should actually be performed.

    "Looking further ahead, swapping brain and body models of related species may one day yield real insights into how their brains and bodies diverged through evolution. However, far more model development and experimental validation is needed before we can learn anything from such a digital sphinx."
    These two sentences about future possible cross-species chimeras feel superfluous and unsubstantiated, and weaken the main argument of the paper about whole-brain emulation.

  3. eLife

    Reviewer #2 (Public review):

    Summary:

    The authors use DRL to train a C. elegans connectome-based ANN to control stepping in a D. melanogaster body model. The resulting system can walk. This shows that one needs further constraints to derive biologically meaningful results from this approach.

    Strengths:

    The authors perform a very simple experiment with a clear outcome. The interpretation (or lack of interpretation) is a striking cautionary tale.

    Weaknesses:

    There is little analysis of precisely how robust this result is to parameter variation and network wiring. The worm also undulates in an oscillatory fashion. Thus, it is possible that the network is tapping into biologically meaningful motifs to generate oscillations for walking. As well, it would be useful to examine which heuristics one can use to determine whether modeling efforts are sufficiently constrained (i.e., how much biological data will be necessary to start obtaining fruitful, interpretable outcomes from DRL task optimization). For example, their "solution" using the worm connectome is not sparse (i.e., it uses many neurons). Perhaps a signature of a biologically-meaningful, interpretable result is one that is sparse?

  4. eLife

    Reviewer #3 (Public review):

    Summary:

    The authors construct a computational chimera by attaching a C. elegans connectome to a Drosophila body biomechanical model and use deep reinforcement learning to link neural activity to motor output. The model is able to produce walking, but is considered a priori to be scientifically meaningless, and the work is treated as a cautionary tale in complex interpretation layers unconstrained by experiment or data.

    Strengths:

    In a period of increasing excitement about linking AI and neuroscience, I respect very much that the authors work through a nontrivial example of nonsense results, rather than just making a theoretical case. It offers a clear and memorable existence proof that matching outputs of complex trained networks does not mean the internal dynamics are themselves emulated.

    Weaknesses:

    While I understand that the work was a rapidly produced comment on science-by-press-release, the message seems too important to be treated in quite as pithy a manner as it is. In particular, because the computational experiment is so memorable, it is worth getting the message right to avoid a set of readers who take from it that they should dismiss this category of neuroAI wholesale (which the authors absolutely do not imply!).

    One part of me reads this work and thinks that by intentionally wiring up the sensory feedback in a particularly nonsense way, the authors have just made a bad model, and sometimes bad models can still generate sensible outputs, especially when expressive models are optimized to fit those sensible outputs. But I think this work is trying to say something more specific than this, and I would like it to be a bit clearer about that. The authors do a fairly good job of sharing a view about what should have been done instead, but this message would benefit from having some more concrete suggestions to avoid a simplistic interpretation. A few thoughts:

    (1) It's not entirely obvious to me that the model is "scientifically meaningless." As the authors know extremely well, Drosophila walking is thought to be driven by simple central pattern generators coupled to leg-specific implementations. The C. elegans neural circuit is clearly capable of producing rhythmic activity as well. A version of the model they ran could have identified biologically valid rhythmic activity in the C elegans circuit and mapped it via the DRL to the right locomotor behavior in the fly. While this would not be a good emulation of the fly, it's not a concept devoid of scientific meaning. Similarly, if the ANN is converting a rhythmic signal to coordinated walking, it's not obvious to me that there aren't useful principles to identify in how it achieves this - it's basically the equivalent of that post-CPG circuitry, no?

    (2) Similarly, is this outcome going to be relatively specific to rhythmic behaviors? I suspect that it would be harder to push the C. elegans connectome to produce some behaviors than others - for example, adding in visual navigation and other motor patterns, or a ring attractor. Rhythmic circuits arise in many places, and both biology and dynamical systems tell us they can come from numerous configurations of elements and interactions.

    (3) Aside from the nonsense formulation of the problem, I would have liked to know more about what the authors should have done to know their model was useless. Put another way, if the authors hadn't known that their model was bad from the beginning (e.g., if they had stuck a fly brain in the middle of it, gotten the sensory feedback right), would there have been some way to figure out if it was meaningful or meaningless based on the results of the trained model itself?

  5. Version published to 10.64898/2026.03.20.713233 on bioRxiv