To reverse engineer an entire nervous system

Gal Haspel, Edward S Boyden, Jeffrey Brown, George Church, Netta Cohen, Christopher
Fang-Yen, Steven Flavell, Miriam B Goodman, Anne C Hart, Oliver Hobert, Eduardo J
Izquierdo, Konstantinos Kagias, Shawn Lockery, Yangning Lu, Adam Marblestone, Jordan
Matelsky, Hanspeter Pfister, Horacio G Rotstein, Monika Scholz, Eli Shlizerman, Quilee Simeon,
Michael A Skuhersky, Vineet Tiruvadi, Vivek Venkatachalam, Guangyu Robert Yang, Eviatar
Yemini, Manuel Zimmer, Konrad P Kording

Abstract:
A primary goal of neuroscience is to understand how nervous systems, or assemblies of neural
circuits, generate and control behavior. Testing and refining our theories of neural control would
be greatly facilitated if we could reliably simulate an entire nervous system so we could replicate
the brain dynamics in response to any stimuli and different contexts. More fundamentally,
reconstructing or modeling a system is an important milestone in understanding it, and so,
simulating an entire nervous system is in itself one of the goals, indeed dreams, of systems
neuroscience. To do so requires us to identify how each neuron's output depends on its inputs,
within some nervous system. This deconstruction – understanding function from input-output
pairs – falls into the realm of reverse engineering. Current efforts at reverse engineering the
brain focus on the mammalian nervous system, but these brains are mind-bogglingly complex,
allowing only recordings of tiny subsystems. Here we argue that the time is ripe for systems
neuroscience to embark on a concerted effort to reverse engineer a smaller system and that the
nematode Caenorhabditis elegans is the ideal candidate system. In particular, the established
and growing toolkit of optophysiology techniques can non-invasively capture and control each
neuron’s activity and scale to hundreds of thousands of experiments, across a large population
of animals. Data across populations and behaviors can be combined because across individuals
neuronal identities are largely conserved in form and function. Modern machine-learning-based
model training should then enable a simulation of C. elegans’ impressive breadth of brain states
and behaviors. The ability to reverse engineer an entire nervous system will benefit systems
neuroscience as well as the design of artificial intelligence systems, enabling fundamental
insights as well as new approaches for investigations of progressively larger nervous systems.

Why should we reverse engineer a nervous system?

“What I cannot create, I do not understand,” one of Richard Feynman’s famous dictums, nicely
highlights the need to build systems to drive their understanding. Reverse engineering
behaviors by characterizing the input-output dependence of neurons and muscle cells may
enable us to both create and understand neural systems driving complex behaviors.

Let us first be clear about what we mean by successful reverse engineering. Independent of
variations in cellular biophysics, we consider each neuron and muscle cell as computing its
output as a function of the activity of its input cells and spontaneous activity. As such, it is
characterized by a spatiotemporal function mapping input traces into an output trace (i.e. an
input-output function), which represents both synaptic and non-synaptic effects (e.g.
neuropeptidergic signaling or spontaneous activity). In our framing, reverse engineering consists
of figuring out the input-output mapping for all neurons and muscle cells as well as the inputs
from the world (i.e. system identification), then reassembling the collection of input-output
functions into a robust, simulatable model that can exhibit key behaviors when connected to the
simulated body (Fig. 1). To be successful, this working model should recapitulate behavior
under a range of conditions, stimuli, and perturbations.

With a working model of a nervous system and its interactions with the body and world, we
could test hypothesized neuroscientific models and principles in silico inexpensively and rapidly.
The newfound knowledge could be used to inspire and develop candidates for therapeutic
approaches, tested through simulation. Such a model may also catalyze the design of new
information processing systems and intelligent signal processing systems that solve
well-defined goals. The results may inspire a new generation of artificial intelligence systems
that are orders of magnitude more efficient than current ones.

A central goal of systems and computational neuroscience is thus to model how brains convert
stimuli, spontaneous activity, and internal states into the coordinated muscle contractions which
are behaviors (also see (Harel 2003; Krakauer et al. 2017)). Indeed, the National Institute of
Health (NIH) and other funders started the BRAIN Initiative (Insel, Landis, and Collins 2013)
with a multi-billion dollar investment to develop new large-scale neurotechnologies. Major
initiatives by other funders include the Human Brain Project (Markram 2012), MICrONS (funded
by the Intelligence Advanced Research Projects Activity), and the Simons Global Brain
Collaborations. The investment of significant resources into reverse engineering the brain
reflects the value of this endeavor.

What will we learn by reverse engineering a nervous system?

If we could decompose a simulation of the nervous system from neuronal input-output functions,
we could predict downstream behavior in response to any sensory signals (past and present)
under any experimental manipulations and internal states. We could then predict the full
behavioral repertoire and what every neuron does as a function of brain state and stimulation.
This ability to build a simulation is one of the definitions used for understanding in the systems
neuroscience community (Kording et al. 2020) and is a natural product of reverse engineering
(Csete and Doyle 2002).

But prediction is only the first step, a means of falsifying or validating our in silico model.
Ultimately, reverse engineering a nervous system aims to build an explanatory model of the
dynamics of a complete nervous system that captures adaptation and plasticity at cellular,
circuit, whole-brain, and behavior levels. A validated model that correctly accounts for all our
data under some condition would provide a platform for running in silico experiments, to test –
precisely, specifically, and efficiently – the role of different model constituents (a neuron, a
compartment, a connection, even a neuromodulator or ion channel) in neural and circuit
dynamics. Such in silico experiments would then allow us to build understanding: to interpret the
dynamics in terms of computational concepts, from decision-making, memory, and sensory
integration to attention and coordination, and indeed to understand fundamental principles of
circuit structure and function. This link between neuronal dynamics (both in the form of
input-output relations and whole-brain states) and interpretable function is, arguably, the holy
grail of systems and computational neuroscience (Csete and Doyle 2002).

By reverse engineering any entire nervous system, we would gain important insights about the
scientific process of biological reverse engineering that could generalize to larger systems,
discovering which information matters, and which shortcuts are possible. We now know synaptic
patterns of connectivity (White et al. 1986; Cook et al. 2019; Witvliet et al. 2021; Brittin et al.
2021) poorly predict interactive patterns of neural activity (Bentley et al. 2016; Yemini et al.
2021; Susoy et al. 2021; Uzel, Kato, and Zimmer 2022; Beets et al. 2022; Ripoll-Sánchez et al.
2022; Dag et al. 2023; Atanas et al. 2023), so what are we missing to understand dynamic
operations in these circuits? We know that as in other animals, in C. elegans individual neurons
can compartmentalize signals and thus perform multiplex computations (Hendricks et al. 2012),
so what resolution do we need to distinguish these compartmentalized signals? Do millisecond
timescales matter or is it enough to have lower frequency signals? Can we solve these
problems with optical imaging only? Can the information describing the dynamics of nervous
systems be compressed into a small number of principles? Is it enough to reverse-engineer
using data on only some parts from each individual? We know that the biophysical properties of
different cell types matter (Dag et al. 2023), but to what extent must these be understood and
modeled? We now know that even in C. elegans with its invariant cell lineage, the wiring of
neural circuits varies across individuals (Brittin et al. 2021). How can individual variability be
taken into account and might data from many animals produce a meaningful general model? We
cannot currently answer these questions because we have not performed the required
experiments and nervous system modeling. Pioneering reverse engineering in C. elegans can
point out favorable approaches when we try to reverse engineer the more complex nervous
systems of rodents and other animals. Demonstrating reverse engineering of an entire nervous
system would clarify what kind of data we may need for future, increasingly ambitious
endeavors.

In reverse engineering a nervous system, we may also learn about failure modes. Can we easily
be misled and believe we understand how a nervous system works from partial recording? How
probable is it that the models we fit get the correlations right and the causality wrong (Tremblay
et al. 2022)? How much data of what kind is too little to reverse engineer systems? Answers to
these questions could guide research in all areas where the goal is to understand nervous
systems. Reverse engineering a system may also lead us to discover misleading “principles” in
past neuroscience research.

By reverse engineering a nervous system, we will galvanize and motivate the enrichment and
expansion of technologies that are critical for research across neuroscience. Neuroscience can
scale: we record from thousands of neurons where we used to record from a handful
(Stevenson and Kording 2011). We would optically image a million neurons where we previously
imaged hundreds (Abdelfattah et al. 2022). Similar progress happened in molecular techniques
where, for example, driving the human genome project ultimately made sequencing a cheap
tool used by virtually all labs for countless objectives, and analogously reverse engineering
nervous systems should produce broadly useful techniques. Reverse engineering requires the
development of hardware automation, data handling, and sharing, as well as algorithms and
software to extract and analyze neural activity from dense whole-brain recordings, yielding a
suite of interlocking technologies that can be scaled up and potentially generalized. Integral to
this vision would be an informatics infrastructure enabling an unprecedented and integrated
distribution of connectomics, neural activity, behavior, and computational models.
The existence and fact-checking of all these resources would jump-start a drive toward scaling
entire nervous systems reverse engineering and simulation.

In contrast to artificial intelligence systems, the nervous system of C. elegans is compact and
incredibly low-energy consuming and yet able to allow the species to thrive all over the planet.
As such, we may hope to find design principles that we can generalize to AI systems. Of
particular interest here are the basic building blocks, figuring out which computational elements
are used in C. elegans, e.g. in terms of nonlinearities, promises to inform the design of new
low-energy AI systems as well as to make them more resilient in an adversarial world (Agarwal
et al. 2017). In analogy, being able to simulate the full nervous system of C. elegans may allow
us to design new biological systems to solve technical problems.

Why have we not yet reverse-engineered a nervous system?

Experimental limitations and theoretical realities mean that, despite early efforts (Sarma et al.
2018) we have yet to simulate the entire nervous system of C. elegans. We need many
parameters to describe C. elegans and astronomically many to describe mammals. Our
experiments so far have been too limited in breadth but also, focusing on correlations, have not
measured the causal parameters needed for simulations. Indeed, it remains an open and hotly
debated question whether, and how theories can supply the missing parameters.

One reason why these challenges appear insurmountable is that we focus on reverse
engineering rodents, with most of the circuit work in mice. However, we can only ever observe a
tiny part of their nervous system; neither can we record all inputs to a single neuron. There is no
doubt that the relevant models for mice are exceptionally complicated, both in terms of neuronal
and neural processing and in terms of learning during experimental procedures. Because
neuroscientific research takes place mostly on mammals, it is not known how good modern
techniques may be at reverse engineering these circuits (Jonas and Kording 2017). After all,
individual neurons are often weakly correlated with behavior and the enormous numbers of
neurons in mammals preclude a population-level view with the single-neuron resolution
necessary to understand their in-depth circuit dynamics. Human nervous systems have about
86 billion neurons; and even mice have 70 million (Herculano-Houzel, Mota, and Lent 2006),
compared to C. elegans’ 302 neurons (White et al. 1986). The Drosophila nervous system, with
about 130,000 neurons, has emerged as a premier model for systems neuroscience, revealing
important features of how circuits control complex behaviors, like flexible navigation or learning
and memory (Fisher 2022; Modi, Shuai, and Turner 2020). However, even with extensive
connectome data (Scheffer et al. 2020; Dorkenwald et al. 2023) and transgenic tools, it is not
yet feasible to record all neurons with knowledge of cellular identity or stimulate every neuronal
cell type in succession during brain-wide recordings. Such datasets may be essential to build
accurate simulations of a nervous system. Moreover, every cortical neuron in the rodent
receives input from thousands of neurons, on the same order of magnitude as all the neural
connections in the nematode nervous system. In our quest to understand complex nervous
systems we have started with some of the most complicated ones; maybe to get close to human
physiology as soon as possible. Not only do we not know how to simulate such nervous
systems, we do not know what we do not know on the path to simulate them.

We have had much of the C. elegans connectome since 1986 (White et al. 1986); further
annotated (Varshney et al. 2011; Chen, Hall, and Chklovskii 2006; Cook et al. 2019); and new
datasets were added (Brittin et al. 2021; Witvliet et al. 2021; Mulcahy et al. 2022). But
connectivity alone is insufficient because without knowing the strength and temporal properties
of all connections and neuronal properties (Kopell et al. 2014), “... it would not be possible to
simply go from the wiring diagram to the dynamics of even two neurons“ (Cornelia I. Bargmann
and Marder 2013). Despite impressive progress (Einevoll et al. 2019; Eliasmith and Trujillo
2014), we cannot yet simulate any entire nervous system.

No discussion of reverse engineering C. elegans can be complete without careful consideration

of the OpenWorm project (Szigeti et al. 2014)(https://openworm.org/). The OpenWorm project
impressively used relatively solid knowledge about the physical environment of the nematode,
the physics of swimming, and muscle properties (Boyle and Cohen 2008; Boyle, Berri, and
Cohen 2012). They used the very partial relevant published information (White et al. 1986;
Varshney et al. 2011) as well as a two-dimensional atlas (Hall and Altun 2008) of nematode
nuclei. Impressively, OpenWorm even began compiling a list of ion channel inventories for
neurons and models of their biophysical properties (Sarma et al. 2018). However, estimating
neuronal interactions from behavior, or even from ongoing neuronal activity, is essentially
impossible, as many different neuronal properties can produce the same behavior (Prinz,
Bucher, and Marder 2004). Moreover, in the absence of standardized data formats and norms
for data sharing, the transfer of neural data into published papers is equally an irreversible
process. By starting the simulation project, the OpenWorm project took a massive step in the
right direction. But missing are experiments explicitly aimed at producing the kind of data that
would allow a faithful simulation. Data was the limiting factor all along.

Rapidly developing advances in data-driven approaches, including artificial intelligence,

combined with growing computational power, provide much-needed tools for learning
multidimensional, potentially non-linear mappings for input-output relationships. Concurrently,
key experimental technologies, including microscopy, optophysiology, and data management
have been reaching the required capabilities. The tractability of C. elegans and the advent or
maturation of these tools for reverse engineering converge into a unique opportunity.

Why a small nervous system? Why C. elegans?

C. elegans has far fewer neurons and neuron-neuron connections making it a great starting
point to measure the causal dynamics, captured by its input-output functions, driving
stereotyped behavior. It has established optical techniques for recording and stimulating at a
scale (Nguyen et al. 2016; M. Liu et al. 2022; Bergs et al. 2023; Randi et al. 2022a). Importantly,
C. elegans allows us to record and combine information across animals. Until recently,
variations among nematodes were thought to be relatively unimportant, and physiological
parameters were considered relatively conserved (Randi et al. 2022a), supporting the pooling of
data across individuals. More recent evidence of variability in the neuronal wiring (Brittin et al.
2021) and of neuronal encoding (Atanas et al. 2023)across individuals suggests a role for
individuality. Put together, C. elegans offers a unique system to address these questions: by
analyzing many individuals, we can determine the extent of variation at the cellular, circuit, and
behavioral levels, and we can take these variations into account in building computational
models. However, there is still considerable disagreement about the remaining variation across
animals (Brittin et al. 2021). It may also be easier as electrical signals common in C. elegans
require lower temporal resolution of both stimulation and recording (M. B. Goodman et al. 1998;
Jiang et al. 2022; Q. Liu et al. 2018; Lockery and Goodman 2009). As a historical example of
this process, sequencing the C. elegans genome also significantly advanced biology and laid
the groundwork for sequencing the human genome (C. elegans Sequencing Consortium 1998).
Most of the benefits of reverse engineering a nervous system listed above can be realized by
reverse engineering a relatively small nervous system.

C. elegans is small enough to perturb and record calcium signals from each neuron, or even
groups of neurons. Continuous recording and stimulation on a single setup would reach about
50,000 specimens a year and can be parallelized across setups (Fig. 1AB). Undoubtedly this is
a large number, but it is only an order of magnitude larger than the number of animals whose
behavior was recorded for recent publications (Dag et al. 2023). Experimental advantages allow
a very large number of experiments, clearly enough for at least measuring the effect of all
pairwise stimulations (~90k pairs, or e.g., neuromodulator by neurotransmitter).
In C. elegans and only in C. elegans, this approach can sample a massive number of
combinations of presynaptic neurons. In other words, within a single year, we could densely
sample the space of inputs to all the neurons. Within a single year, a single targeted C. elegans
experimental paradigm that is run on multiple setups may obtain more data than all mouse
experiments taken together.

What will reverse engineering the C. elegans nervous system deliver?

The result of reverse engineering a brain is a dynamical model from which one could generate
testable predictions for all possible manipulations and experiments. In other words, we want a
causally correct ‘digital twin’ of the biological system. This twin should react to any kind of
stimulation in the same way as the real system. So what is the minimal aspect of what we mean
by wanting to reverse engineer it? Let us go through the outcomes we want to have.

The model should be able to describe the nervous system and its motor output (i.e. muscle
contractions that are behavior). This description will have to include the influence of each
neuron on other neurons and ultimately on behavior to predict stimulation effects. Importantly,
neurons interact through synapses with neurotransmitters, but also through gap junctions,
neuromodulators, and non-synaptic and non-neuronal paths such as glia (Raiders et al. 2021;
Mu et al. 2019) as well as muscle activations and body shape(Cohen and Denham 2019; Zhan
et al. 2023). We thus need a model that captures interactions of these kinds. It should be able to
simulate the entire trajectory of neural activities and thus predict activities and tuning curves at
all times and contexts (Hallinen et al. 2021).

While our conceptualization is built on there being little structured natural variability in wiring,
neural activity, and behavior among individuals, assays should be included in the model as the
measured variability in input-output functions and should replicate the variability in behavior.
Importantly, if there are enough experiments, we can not just measure average properties but
gain relevant insights into the overall distribution, or indeed identify multiple solutions. The
behavioral variability among hundreds of simulations drawn from the measured distribution
should be similar to that among hundreds of animals.

We should be able to produce all the basic behaviors the animal performs, including
spontaneous, responsive, and coupled behaviors. The hermaphrodite C. elegans has an
impressively rich behavioral repertoire (Hart 2006; Atanas et al. 2023), constructed of first-order
building blocks including: 1) locomotion behaviors, including forward, backward, change of
speed, steer, turn, and halt; 2) feeding behaviors, more specifically occurrence, rate, and
coordination of pharyngeal pumping (used to collect and crush organic material such as
bacteria); 3) egg laying; 4) defecation; 5) fixed action patterns for mating. These behaviors can
all be studied effectively with minutes-long recordings, and hence the above list excludes
behaviors with a longer timescale to currently doable (and scalable) experiments (e.g. learning).
We envision a simulation that replicates these building blocks of all behaviors in a single model.
To do so requires a simulation of the whole body, the inputs from the world around it, and how
its effectors in turn affect the world that is around them, at least at a level that allows
interpretation of motor output to behavior. These are the minimal objectives, in our opinion, for
producing acceptable models of the nervous system of C. elegans. We expect that such a
simulation should replicate the more complex behaviors that are made out of the first-order ones
(Ghosh et al. 2016; Dekkers et al. 2021). Examples of these more complex behaviors include
finding and attracting mates, fleeing from predators, avoiding problematic chemicals or
temperatures, collective behavior, or avoiding parasites. In other words, a good simulation
should cover all the ethologically relevant high-order behaviors that are described for the
nematode.

Feasibility: What do we need to do to reverse engineer a nervous system?

To close the loop we need a model of how the nematode’s sensors translate the state of the
nematode’s body and the environment into neural activity, how neurons interact, and how neural
activities through muscle contraction (effectors) influence the nematode’s body and the
environment. We also need a model for the world because the world is an important part of the
solution to many behaviors. So we need a model of the world, the sensors, and the effectors
(Fig. 1C). With a model of the body and environment in hand, we should be able to calculate the
information going into the animal at all times and how the action of the effectors produces its
future dynamics (Fig. 1D). To get there, we primarily need a sequence of experiments and
modeling that require the components below. Because each component can be achieved by
multiple technologies and approaches, we will describe the need and give a concrete example.
 Figure 1: An overview of the proposed approach, demonstrating how imaging,
recording, perturbing, modeling, and simulation interplay.

(1) Staging
Needed: Any approach we can think of will require the use of many animals, as we will need
many thousands of hours of experiments to obtain the statistical power needed to meaningfully
simulate all neurons. We thus need a way of getting nematodes in and out of set-ups with high
throughput. If we were to do that manually, we would require large amounts of human labor.
Another complication would be to standardize a manual workflow, so automatic staging is a
better option.
Potential approaches: Depending on the experimental design, animals could be immobilized,
restricted, or freely moving. Each option introduces challenges in collecting and interpreting the
data but all past experiments have used extensive human labor. Automation could include
robotic manipulation (Li et al. 2022), microfluidics approaches that either move nematodes in an
aqueous solution or small bubbles of oil (San-Miguel and Lu 2013), and machine-vision-based
focusing and positioning (Li et al. 2022). Neuronal calcium imaging data recorded in freely
moving animals (Nguyen et al. 2016; Venkatachalam et al. 2016; Susoy et al. 2021; Atanas et
al. 2023), by their nature, do not reveal causality. The publicly available data (Atanas et al.
2023) can be used to test and refine simulated models constructed from input-output functions.
A concrete potential solution: Microfluidics can be used to position nematodes onto a
microscope stage. Such a system will include a reservoir of age-synchronized animals, in which
animals are introduced to chemogenetics reagents, tubing and channels will then place ten
chemogenetically paralyzed adult hermaphrodites next to each other to fill a rectangular field of
view (Mondal et al. 2016). Microfluidic technology (San-Miguel and Lu 2013) and automation (Li
et al. 2022) are well-established in C. elegans and appear to pose no relevant risk to the
reverse engineering project.

(2) Microscopy
Needed: The nervous system of C. elegans is distributed along its body. About two-thirds of the
302 neuronal cell bodies are in a head ganglion and the rest split between a tail ganglion and a
ventral nerve cord. Animals are about 1 mm in length and the maximal thickness of the nervous
system is about 50 microns. This volume should be imaged at a sampling rate of at least 10 and
preferably 100 Hz, requiring some progress relative to today’s typically slower rates (Atanas et
al. 2023). We will need one microscope for each set-up (although we could place multiple
animals in the same field of view) with high enough resolution to be able to resolve individual
neurons as we need to identify and match them across experiments. Importantly, we need to
overcome field-of-view limitations, where typical microscopes have a field-of-view that is too
small for fitting the entire worm. Fortunately, neuronal cell bodies are large enough that
diffraction-limited imaging should be sufficient to resolve the relevant signals, in particular, if the
relevant indicators can be localized to the soma. Importantly, we will have to design the setup
specifically for the goal of automated, large-scale systems identification.
Potential approaches: Because diffraction-limited microscopy of fluorescent indicators is
expected to be sufficient, there are many possible configurations (Lemon and McDole 2020;
Balasubramanian et al. 2023). This includes various light-sheet and spinning disk confocal
designs that can scan this volume at the required sampling rate. It also includes the more
sophisticated two-photon setups that allow exceptional precision in space, e.g., for stimulation.
A concrete potential solution: An inverted SCAPE2.0 microscope (Voleti et al. 2019) with a
microfluidic stage would be suitable for the optogenetic and imaging approach. The inverted
design provides access to liquid handling and stimuli and the SCAPE2 can scan the required
volume (1 x 1 x 0.05 mm for ten immobilized animals) at about 10 Hz. This volume is sufficient
to image ten animals.

(3) Neuron Identification

Needed: For each neuron that we record from we need to be able to assign it reliably to one of
the 302 neuron identities. To do so, we will need to automatically identify each neuron from
optical stacks after recording.
Potential approaches: A large number of solutions are possible. High-resolution, potential
super-resolution, stacks, e.g. from expansion, would allow solving the problem by machine
learning from large datasets. NeuroPAL (Yemini et al. 2021), a multicolor transgene allows for
nervous-system-wide neuronal identification using a combination of reporters and colors to
generate an invariant color map across individuals radically simplifying the identification
process. Alternatively, machine learning may enable identification from large sets of multimodal
data (Kirillov et al. 2023). This field of identification is quickly maturing.
A concrete potential solution: Neuronal identification could build on the tried and tested
NeuroPAL (Yemini et al. 2021) strain of transgenic animals. Techniques to automate using
NeuroPAL are quickly becoming standard but will still require additional software development
(Skuhersky et al. 2022) and the development of novel methods for aligning neurons across
thousands of animals. Importantly, it is sufficient to solve the identification problem in most
animals.

(4) Stimulation and Recording

Needed: The core of reverse engineering a nervous system is figuring out how interactions
among components (here neurons) shape neural dynamics and behavior; we need to know how
a neuron’s output is caused by all its input cells (including those acting through
neuromodulators) and how a neuron’s output affects other neurons and muscle cells. In other
words, we need the interactome which is a generalization of the connectome. This interactome
is complicated because there are both synaptic interactions and neuropeptide interactions
(Ripoll-Sánchez et al. 2022; Beets et al. 2022), but there have been recent attempts at reverse
engineering a larger number of interactions (Randi et al. 2022b; Uzel, Kato, and Zimmer 2022).
These approaches have, however, only revealed the average linear neuron-to-neuron
interactions instead of the full nonlinear interactions. We need to know how stimulating (or
inactivating) combinations of neurons affect the activity of each neuron. If we had this
information, we would be able to calculate each neuron’s output given its inputs.
Potential approaches: There are many ways of stimulating nematode neurons. We can stimulate
them through direct physical effects (Suzuki et al. 2003; O’Hagan, Chalfie, and Goodman 2005;
Ramot, MacInnis, and Goodman 2008; Miriam B. Goodman and Sengupta 2019), stimulate
them electrically (M. B. Goodman et al. 1998), or stimulate them through optogenetic means
(Husson, Gottschalk, and Leifer 2013; Nagel et al. 2005). We can do so with single-photon or
multi-photon approaches, trading of price and precision. All of these approaches are well
established, but optogenetic methods are particularly scalable. There are also many ways of
recording from nematode neurons. We can record from them electrically (M. B. Goodman et al.
1998), with calcium (Kerr et al. 2000), or by voltage imaging with genetically encoded sensors
(Azimi Hashemi et al. 2019). Neurons are the most obvious target for stimulation but a complete
interactome might require stimulating other cells such as glia (Perea et al. 2014). There are a
large number of interactions to measure. The main question is which kinds of approaches can
be applied to very large datasets derived from large numbers of individuals.
A concrete potential solution: To simplify the optogenetic access and experimental design,
transgenic nematodes could be generated that stochastically express an optogenetic activator
in a random subset of neurons (Aoki et al. 2018; Pospisil, Aragon, and Pillow 2023). Given the
experience with Brainbow, this is doable and presents minimal risk. For example, if each neuron
has a 1/150 chance of expressing the activator, we expect most (~80%) of the animals to
express the activator in one, two, or three neurons. A few animals will express it in more than
three neurons, while 10-15% of animals will not express the activator in any neuron and will
serve as an internal control group. Providing a single-photon optogenetic stimulation on only
one channel makes experiments far simpler and inexpensive than 2-photon targeted
stimulation. To avoid polysynaptic recruitment and overinterpretation (e.g. Penfield and Boldrey,
1937), light activation should be as brief as possible and then as low in intensity as possible.
Another approach to reducing intrinsic activity and polysynaptic activation is to subtly inactivate
all neurons and muscle cells by making use of transgenic animals expressing a histamine-gated
chloride channel (Pokala et al. 2014). Genetically encoded calcium indicators such as GCaMP
are a reasonable proxy for neuronal activity. The specific calcium indicator should be selected
for spectral separation from NeuroPAL and the optogenetic illumination.

(5) Quality control

Needed: Countless variables affect the results of the kinds of experiments we sketch here. This
ranges from the genetic background of the nematodes (C. I. Bargmann 1998; Hobert 2013) to
the properties of the microscopes (Marblestone et al. 2013), to the expression patterns of the
indicators (Schmitt et al. 2012). As such, careful monitoring of the overall pipeline is essential.
To be able to pool results, the results must be all comparable.
Potential Approaches: There are two kinds of well-established approaches. One, as in the case
of the International Brain Lab (Wool and International Brain Laboratory 2020) is to establish
standards that are held up in many labs working in parallel. The other, as in the Allen Institute, is
to establish one central location where large-scale experiments are made with a team that is
very much centered around quality control.

(6) Annotated Connectomes

Needed: To further understand the interactome data and possibly predict neuronal function, we
want to have a specific molecularly annotated connectome (Bhattacharya et al. 2019; Taylor et
al. 2021). In other words, we want to know the whole structure of the nervous system of animals
from which we collected input-output data, but also where key molecules are expressed. This
kind of information can significantly cut down on the set of models we need to consider.
Potential approaches: Serial electron microscopy and several super-resolution approaches can
be adapted to image the morphology of the nervous system of collected and fixed animals.
These are mostly low-throughput methods. Expansion microscopy is the only approach we are
aware of that could provide molecular annotation in addition to morphological data and at a
relatively high throughput (Alon et al. 2021; Shen et al. 2020).
A concrete solution: After stimulation and recording, a fraction of the animals could be extruded
from the microfluidic device, and expanded (Yu et al. 2020), to identify a host of molecules in 3
dimensions, along with the full connectome, which, in turn, will require advances in computer
vision such as automatic proofreading.

(7) Model fitting

Needed: We need to be able to fit models to the data that are sufficient to faithfully simulate
outputs and behaviors under perturbations, and that generalize to other contexts.
Potential approaches: There are two major branches for such fitting. There is a school focusing
on detailed models (in the sense of cell properties and cable equations, (Hines and Carnevale
1997; Kim, Leahy, and Shlizerman 2019; Kunert, Shlizerman, and Kutz 2014) and embodiment
(Kim et al. 2019)). There is another school focusing on setting up machine learning models, e.g.
by fitting deep learning models(Tanaka et al. 2019; Benjamin et al. 2018; Yamins and DiCarlo
2016). In all cases, we would have to fit models where we estimate neuron outputs 𝑦𝑗(𝑡) as a
function of neuron inputs 𝑥𝑖(𝑡). Subsequently, we could build in biological prior knowledge, e.g.
the stronger interactions between nearby synapses, e.g. by adapting the standard attention
function used in transformer models.
A concrete solution: Starting with a deep learning system based on transformers (Vaswani et al.
2017) for every neuron and tokenizing the input state as a set of discrete descriptors for each
input neuron’s state 𝑥𝑖(𝑡) learned using a Vector-Quantised-Variational AutoEncoder (VQ-VAE)
(Van Den Oord, Vinyals, and Others 2017). The overall system can then be simulated by
interactions between these fitted local models, overall sensory inputs, and movement outputs.
(8) Open Science
Needed: Such a project will require a concentration of effort and resources that can be best
supported by an open community. A community may also provide quality assurance - the data
may have problems that we do not anticipate.
Potential solutions: To produce a governance structure that minimizes that risk or to be
absolutely open.
A concrete solution: Doing both. It is important to engage the relevant community to guide the
approaches and open science has been the standard approach in the C. elegans field. This
effort can be an opportunity to practice and demonstrate radically open science in which code
and data are made available within a month of the time at which they are developed or collected
and analyzed.

(9) Diverse Science

Needed: This project will require the integration of scientists with expertise in instrumentation,
genetic analysis, molecular tools, behavioral neuroscience, cellular biophysics, computation,
data science, and theory.
Potential solutions: Recruit scientists broadly and empower team members to cross-train in
multiple disciplines. Intentionally build an inclusive community of scientists and embed a
tradition of design and data review into teamwork.

Power calculations
It is notoriously difficult to estimate statistical power in the context of machine learning
applications. However, there are two things that we can do. (1) We can produce a simulation
that models the relevant problem and see how well we can reverse engineer that simulation to
obtain useful intuition. (2) We can use special cases, such as linear models with nonlinear basis
functions, to obtain closed-form estimates.

(1) One simulation strategy is to produce virtual connectomes of varying resolution (different
numbers of neurons and different densities of connections between them) and varying
complexity. There are of course many ways to vary simulated neuron complexity, but one
valuable approach is to change the number of input parameters that a neuron uses in its
activation transfer function. A second valuable dimension is how much of its own and its
synaptic partners’ history a neuron takes as input — where a perfectly Markovian neuron takes
only the latest state of the network as input and an unrealistically complex neuron takes as input
all of the complete history of the entire network. Of course, real neurons fall in the middle of this
range, though we do not know where. It is not hard to empirically evaluate how much recording
time is required to analyze the complexity of a neuron (Fig. 2).

(2) We find with considerable amounts of uncertainty (because we do not yet know how
complex C. elegans neurons are) that we will need to continuously record from the entire C.
elegans nervous systems for the equivalent of about 900,000 seconds or 250 days (see
Appendix 1: analytical power calculations), e.g., 10-second long experiments in 90,000 animals.
As such, the proposed experiments are doable, even if we use conservative calculations.

Fig. 2. A simple linear model of connectome dynamics can be

reverse-engineered in the manner described here. Predictions on a
100-neuron model benefit somewhat from longer-duration recording time but are
dramatically improved by inducing artificial perturbations during recording. The
prediction error is most effectively reduced through a combination of the two.
The dynamics of this system are described in Appendix 2.

Beyond C. elegans
Ultimately, the value of reverse engineering a nervous system is that it will make reverse
engineering nervous systems accessible. Being able to compare nervous systems is, arguably,
much more interesting. How do neural dynamics change during disease in the nervous system?
How can we scale this approach to larger and more nervous systems? How should the methods
be adapted and how do the resulting models differ across species? Just like the sequencing of
the first genome enabled the Human Genome Project and the sequencing of the genomes of
many humans, reverse engineering a nervous system promises to open the floodgates to
reverse engineering many other nervous systems (Vogelstein et al. 2016).

Generalizing to larger nervous systems

Reverse engineering the C. elegans nervous system should just be a step in the path toward
understanding nervous systems that are more like our own. There are of course many
differences, in terms of size (human brains with about 86 billion neurons vs 302 in C. elegans),
in terms of computational primitives (most cells in C. elegans do not spike at the millisecond
timescale; (M. B. Goodman et al. 1998; Jiang et al. 2022; Q. Liu et al. 2018; Lockery and
Goodman 2009), and in terms of complexity of brain organization. However, we can sketch what
a path toward larger brain understanding can look like. Power calculations make the
identification of the input-output function of cortical neurons in humans that have of the order of
10,000 inputs from stimulation, seem infeasible. However, through the work that we argue
needs to happen in this paper, we will not merely establish causal interactions between C.
elegans cells, we will also establish how to predict them from annotated connectomes.
Annotated connectomes can readily be scaled up to much bigger nervous systems. We also
expect that there are some principles (some known, most probably still unknown) that may carry
over from small nervous systems to larger ones. The important thing is that the work on reverse
engineering C. elegans will provide a Rosetta-stone-like translation between the language of
annotated connectomes and functional properties. This insight into the causal interactions and
this ground truth-based approach can then be generalized. For example, we could analyze the
kinds of interactions in human brain organoids producing ground truth causality and then scale
up from there. But we will need the whole nervous system scaling to know how to put together
such local information with other factors to get at the causal flow of information in larger nervous
systems.

Conclusions
Here we have outlined how reverse engineering the entire sensory-neuro-motor system of C.
elegans may be possible, producing, for the first time, an understanding of how an entire brain
computes and drives behavior. Reverse engineering the causal interactions in the nervous
system of C. elegans promises to establish in silico simulations as a way of accelerating
neuroscience. It promises to teach us how to run experiments that reveal causal mechanisms of
neural circuits. We may even learn how to build powerful, error-robust,
energetically-economical, synaptically-compact AI systems. A concerted, open effort would
galvanize the development of approaches to help understand the brains of more complicated
animals, like our own, promising a new era of causal neuroscience. The time has come to
reverse engineer the C. elegans nervous system.
References

Abdelfattah, Ahmed S., Sapna Ahuja, Taner Akkin, Srinivasa Rao Allu, Joshua Brake, David A.
Boas, Erin M. Buckley, et al. 2022. “Neurophotonic Tools for Microscopic Measurements
and Manipulation: Status Report.” Neurophotonics 9 (Suppl 1): 013001.
Agarwal, Nikita, Neil Mehta, Alice C. Parker, and Karam Ashouri. 2017. “C. Elegans
Neuromorphic Neural Network Exhibiting Undulating Locomotion.” In 2017 International
Joint Conference on Neural Networks (IJCNN). IEEE.
https://doi.org/10.1109/ijcnn.2017.7966349.
Alon, Shahar, Daniel R. Goodwin, Anubhav Sinha, Asmamaw T. Wassie, Fei Chen, Evan R.
Daugharthy, Yosuke Bando, et al. 2021. “Expansion Sequencing: Spatially Precise in Situ
Transcriptomics in Intact Biological Systems.” Science 371 (6528).
https://doi.org/10.1126/science.aax2656.
Aoki, Wataru, Hidenori Matsukura, Yuji Yamauchi, Haruki Yokoyama, Koichi Hasegawa, Ryoji
Shinya, and Mitsuyoshi Ueda. 2018. “Cellomics Approach for High-Throughput Functional
Annotation of Caenorhabditis Elegans Neural Network.” Scientific Reports 8 (1): 10380.
Atanas, Adam A., Jungsoo Kim, Ziyu Wang, Eric Bueno, Mccoy Becker, Di Kang, Jungyeon
Park, et al. 2023. “Brain-Wide Representations of Behavior Spanning Multiple Timescales
and States in C. Elegans.” Cell 186 (19): 4134–51.e31.
Azimi Hashemi, Negin, Amelie C. F. Bergs, Christina Schüler, Anna Rebecca Scheiwe, Wagner
Steuer Costa, Maximilian Bach, Jana F. Liewald, and Alexander Gottschalk. 2019.
“Rhodopsin-Based Voltage Imaging Tools for Use in Muscles and Neurons of
Caenorhabditis Elegans.” Proceedings of the National Academy of Sciences of the United
States of America 116 (34): 17051–60.
Balasubramanian, Harikrushnan, Chad M. Hobson, Teng-Leong Chew, and Jesse S. Aaron.
2023. “Imagining the Future of Optical Microscopy: Everything, Everywhere, All at Once.”
Communications Biology 6 (1): 1096.
Bargmann, C. I. 1998. “Neurobiology of the Caenorhabditis Elegans Genome.” Science 282
(5396): 2028–33.
Bargmann, Cornelia I., and Eve Marder. 2013. “From the Connectome to Brain Function.”
Nature Methods 10 (6): 483–90.
Beets, Isabel, Sven Zels, Elke Vandewyer, Jonas Demeulemeester, Jelle Caers, Esra Baytemur,
William R. Schafer, Petra E. Vértes, Olivier Mirabeau, and Liliane Schoofs. 2022.
“System-Wide Mapping of Neuropeptide-GPCR Interactions in C. Elegans.” bioRxiv.
https://doi.org/10.1101/2022.10.30.514428.
Benjamin, Ari S., Hugo L. Fernandes, Tucker Tomlinson, Pavan Ramkumar, Chris VerSteeg,
Raeed H. Chowdhury, Lee E. Miller, and Konrad P. Kording. 2018. “Modern Machine
Learning as a Benchmark for Fitting Neural Responses.” Frontiers in Computational
Neuroscience 12 (July): 56.
Bentley, Barry, Robyn Branicky, Christopher L. Barnes, Yee Lian Chew, Eviatar Yemini, Edward
T. Bullmore, Petra E. Vértes, and William R. Schafer. 2016. “The Multilayer Connectome of
Caenorhabditis Elegans.” PLoS Computational Biology 12 (12): e1005283.
Bergs, Amelie C. F., Jana F. Liewald, Silvia Rodriguez-Rozada, Qiang Liu, Christin Wirt, Artur
Bessel, Nadja Zeitzschel, et al. 2023. “All-Optical Closed-Loop Voltage Clamp for Precise
Control of Muscles and Neurons in Live Animals.” Nature Communications 14 (1): 1939.
Bhattacharya, Abhishek, Ulkar Aghayeva, Emily G. Berghoff, and Oliver Hobert. 2019. “Plasticity
of the Electrical Connectome of C. Elegans.” Cell 176 (5): 1174–89.e16.
Boyle, Jordan H., Stefano Berri, and Netta Cohen. 2012. “Gait Modulation in C. Elegans: An
Integrated Neuromechanical Model.” Frontiers in Computational Neuroscience 6 (March):
10.
Boyle, Jordan H., and Netta Cohen. 2008. “Caenorhabditis Elegans Body Wall Muscles Are
 Simple Actuators.” Bio Systems 94 (1-2): 170–81.
Brittin, Christopher A., Steven J. Cook, David H. Hall, Scott W. Emmons, and Netta Cohen.
2021. “A Multi-Scale Brain Map Derived from Whole-Brain Volumetric Reconstructions.”
Nature 591 (7848): 105–10.
C. elegans Sequencing Consortium. 1998. “Genome Sequence of the Nematode C. Elegans: A
Platform for Investigating Biology.” Science 282 (5396): 2012–18.
Chen, Beth L., David H. Hall, and Dmitri B. Chklovskii. 2006. “Wiring Optimization Can Relate
Neuronal Structure and Function.” Proceedings of the National Academy of Sciences of the
United States of America 103 (12): 4723–28.
Cohen, Netta, and Jack E. Denham. 2019. “Whole Animal Modeling: Piecing Together
Nematode Locomotion.” Current Opinion in Systems Biology 13 (February): 150–60.
Cook, Steven J., Travis A. Jarrell, Christopher A. Brittin, Yi Wang, Adam E. Bloniarz, Maksim A.
Yakovlev, Ken C. Q. Nguyen, et al. 2019. “Whole-Animal Connectomes of Both
Caenorhabditis Elegans Sexes.” Nature 571 (7763): 63–71.
Csete, Marie E., and John C. Doyle. 2002. “Reverse Engineering of Biological Complexity.”
Science 295 (5560): 1664–69.
Dag, Ugur, Ijeoma Nwabudike, Di Kang, Matthew A. Gomes, Jungsoo Kim, Adam A. Atanas,
Eric Bueno, et al. 2023. “Dissecting the Functional Organization of the C. Elegans
Serotonergic System at Whole-Brain Scale.” Cell, May.
https://doi.org/10.1016/j.cell.2023.04.023.
Dekkers, Martijn P. J., Felix Salfelder, Tom Sanders, Oluwatoroti Umuerri, Netta Cohen, and
Gert Jansen. 2021. “Plasticity in Gustatory and Nociceptive Neurons Controls Decision
Making in C. Elegans Salt Navigation.” Communications Biology 4 (1): 1053.
Dorkenwald, Sven, Arie Matsliah, Amy R. Sterling, Philipp Schlegel, Szi-Chieh Yu, Claire E.
McKellar, Albert Lin, et al. 2023. “Neuronal Wiring Diagram of an Adult Brain.” bioRxiv : The
Preprint Server for Biology, July. https://doi.org/10.1101/2023.06.27.546656.
Einevoll, Gaute T., Alain Destexhe, Markus Diesmann, Sonja Grün, Viktor Jirsa, Marc de
Kamps, Michele Migliore, Torbjørn V. Ness, Hans E. Plesser, and Felix Schürmann. 2019.
“The Scientific Case for Brain Simulations.” Neuron 102 (4): 735–44.
Eliasmith, Chris, and Oliver Trujillo. 2014. “The Use and Abuse of Large-Scale Brain Models.”
Current Opinion in Neurobiology 25 (April): 1–6.
Fisher, Yvette E. 2022. “Flexible Navigational Computations in the Drosophila Central Complex.”
Current Opinion in Neurobiology 73 (April): 102514.
Ghosh, D. Dipon, Tom Sanders, Soonwook Hong, Li Yan McCurdy, Daniel L. Chase, Netta
Cohen, Michael R. Koelle, and Michael N. Nitabach. 2016. “Neural Architecture of
Hunger-Dependent Multisensory Decision Making in C. Elegans.” Neuron 92 (5): 1049–62.
Goodman, M. B., D. H. Hall, L. Avery, and S. R. Lockery. 1998. “Active Currents Regulate
Sensitivity and Dynamic Range in C. Elegans Neurons.” Neuron 20 (4): 763–72.
Goodman, Miriam B., and Piali Sengupta. 2019. “How Senses Mechanical Stress, Temperature,
and Other Physical Stimuli.” Genetics 212 (1): 25–51.
Hall, David H., and Zeynep F. Altun. 2008. C. Elegans Atlas. CSHL Press.
Hallinen, Kelsey M., Ross Dempsey, Monika Scholz, Xinwei Yu, Ashley Linder, Francesco
Randi, Anuj K. Sharma, Joshua W. Shaevitz, and Andrew M. Leifer. 2021. “Decoding
Locomotion from Population Neural Activity in Moving C. Elegans.” eLife 10 (July).
https://doi.org/10.7554/eLife.66135.
Harel, David. 2003. “A Grand Challenge: Full Reactive Modeling of a Multi-Cellular Animal.” In
Hybrid Systems: Computation and Control, 2–2. Lecture Notes in Computer Science.
Berlin, Heidelberg: Springer Berlin Heidelberg.
Hart, Anne. 2006. “Behavior.” WormBook: The Online Review of C. Elegans Biology.
https://doi.org/10.1895/wormbook.1.87.1.
Hendricks, Michael, Heonick Ha, Nicolas Maffey, and Yun Zhang. 2012. “Compartmentalized
 Calcium Dynamics in a C. Elegans Interneuron Encode Head Movement.” Nature 487
(7405): 99–103.
Herculano-Houzel, Suzana, Bruno Mota, and Roberto Lent. 2006. “Cellular Scaling Rules for
Rodent Brains.” Proceedings of the National Academy of Sciences of the United States of
America 103 (32): 12138–43.
Hines, M. L., and N. T. Carnevale. 1997. “The NEURON Simulation Environment.” Neural
Computation 9 (6): 1179–1209.
Hobert, Oliver. 2013. “The Neuronal Genome of Caenorhabditis Elegans.” WormBook: The
Online Review of C. Elegans Biology, August, 1–106.
Husson, Steven J., Alexander Gottschalk, and Andrew M. Leifer. 2013. “Optogenetic
Manipulation of Neural Activity in C. Elegans: From Synapse to Circuits and Behaviour.”
Biology of the Cell / under the Auspices of the European Cell Biology Organization 105 (6):
235–50.
Insel, Thomas R., Story C. Landis, and Francis S. Collins. 2013. “Research Priorities. The NIH
BRAIN Initiative.” Science 340 (6133): 687–88.
Jiang, Jingyuan, Yifan Su, Ruilin Zhang, Haiwen Li, Louis Tao, and Qiang Liu. 2022. “C. Elegans
Enteric Motor Neurons Fire Synchronized Action Potentials Underlying the Defecation
Motor Program.” Nature Communications 13 (1): 2783.
Jonas, Eric, and Konrad Paul Kording. 2017. “Could a Neuroscientist Understand a
Microprocessor?” PLoS Computational Biology 13 (1): e1005268.
Kato, Saul, Yifan Xu, Christine E. Cho, L. F. Abbott, and Cornelia I. Bargmann. 2014. “Temporal
Responses of C. Elegans Chemosensory Neurons Are Preserved in Behavioral Dynamics.”
Neuron 81 (3): 616–28.
Kerr, R., V. Lev-Ram, G. Baird, P. Vincent, R. Y. Tsien, and W. R. Schafer. 2000. “Optical
Imaging of Calcium Transients in Neurons and Pharyngeal Muscle of C. Elegans.” Neuron
26 (3): 583–94.
Kim, Jimin, William Leahy, and Eli Shlizerman. 2019. “Neural Interactome: Interactive Simulation
of a Neuronal System.” Frontiers in Computational Neuroscience 13 (March): 8.
Kim, Jimin, Julia A. Santos, Mark J. Alkema, and Eli Shlizerman. 2019. “Whole Integration of
Neural Connectomics, Dynamics and Bio-Mechanics for Identification of Behavioral
Sensorimotor Pathways in Caenorhabditis Elegans.” bioRxiv. bioRxiv.
https://doi.org/10.1101/724328.
Kirillov, Alexander, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete
Xiao, et al. 2023. “Segment Anything.” arXiv [cs.CV]. arXiv. http://arxiv.org/abs/2304.02643.
Kopell, Nancy J., Howard J. Gritton, Miles A. Whittington, and Mark A. Kramer. 2014. “Beyond
the Connectome: The Dynome.” Neuron 83 (6): 1319–28.
Kording, Konrad P., Gunnar Blohm, Paul Schrater, and Kendrick Kay. 2020. “Appreciating the
Variety of Goals in Computational Neuroscience.” arXiv [q-bio.NC]. arXiv.
http://arxiv.org/abs/2002.03211.
Krakauer, John W., Asif A. Ghazanfar, Alex Gomez-Marin, Malcolm A. MacIver, and David
Poeppel. 2017. “Neuroscience Needs Behavior: Correcting a Reductionist Bias.” Neuron 93
(3): 480–90.
Kunert, James, Eli Shlizerman, and J. Nathan Kutz. 2014. “Low-Dimensional Functionality of
Complex Network Dynamics: Neurosensory Integration in the Caenorhabditis Elegans
Connectome.” Physical Review. E, Statistical, Nonlinear, and Soft Matter Physics 89 (5):
052805.
Lemon, William C., and Katie McDole. 2020. “Live-Cell Imaging in the Era of Too Many
Microscopes.” Current Opinion in Cell Biology 66 (October): 34–42.
Liu, Mochi, Sandeep Kumar, Anuj K. Sharma, and Andrew M. Leifer. 2022. “A High-Throughput
Method to Deliver Targeted Optogenetic Stimulation to Moving C. Elegans Populations.”
PLoS Biology 20 (1): e3001524.
Liu, Qiang, Philip B. Kidd, May Dobosiewicz, and Cornelia I. Bargmann. 2018. “C. Elegans AWA
Olfactory Neurons Fire Calcium-Mediated All-or-None Action Potentials.” Cell 175 (1):
57–70.e17.
Li, Zihao, Anthony D. Fouad, Peter Bowlin, Yuying Fan, Siming He, Meng-Chuan Chang,
Angelica Du, et al. 2022. “A Robotic System for Automated Genetic Manipulation and
Analysis of Caenorhabditis Elegans on Agar Media.” bioRxiv.
https://doi.org/10.1101/2022.11.18.517134.
Lockery, Shawn R., and Miriam B. Goodman. 2009. “The Quest for Action Potentials in C.
Elegans Neurons Hits a Plateau.” Nature Neuroscience 12 (4): 377–78.
Marblestone, Adam H., Bradley M. Zamft, Yael G. Maguire, Mikhail G. Shapiro, Thaddeus R.
Cybulski, Joshua I. Glaser, Dario Amodei, et al. 2013. “Physical Principles for Scalable
Neural Recording.” Frontiers in Computational Neuroscience 7 (October): 137.
Markram, Henry. 2012. “The Human Brain Project.” Scientific American 306 (6): 50–55.
Matelsky, Jordan K., Elizabeth P. Reilly, Erik C. Johnson, Jennifer Stiso, Danielle S. Bassett,
Brock A. Wester, and William Gray-Roncal. 2021. “DotMotif: An Open-Source Tool for
Connectome Subgraph Isomorphism Search and Graph Queries.” Scientific Reports 11 (1):
13045.
Modi, Mehrab N., Yichun Shuai, and Glenn C. Turner. 2020. “The Mushroom Body: From
Architecture to Algorithm in a Learning Circuit.” Annual Review of Neuroscience 43 (July):
465–84.
Mondal, Sudip, Evan Hegarty, Chris Martin, Sertan Kutal Gökçe, Navid Ghorashian, and Adela
Ben-Yakar. 2016. “Large-Scale Microfluidics Providing High-Resolution and
High-Throughput Screening of Caenorhabditis Elegans Poly-Glutamine Aggregation
Model.” Nature Communications 7 (October): 13023.
Mulcahy, Ben, Daniel K. Witvliet, James Mitchell, Richard Schalek, Daniel R. Berger, Yuelong
Wu, Doug Holmyard, et al. 2022. “Post-Embryonic Remodeling of the C. Elegans Motor
Circuit.” Current Biology: CB 32 (21): 4645–59.e3.
Mu, Yu, Davis V. Bennett, Mikail Rubinov, Sujatha Narayan, Chao-Tsung Yang, Masashi
Tanimoto, Brett D. Mensh, Loren L. Looger, and Misha B. Ahrens. 2019. “Glia Accumulate
Evidence That Actions Are Futile and Suppress Unsuccessful Behavior.” Cell 178 (1):
27–43.e19.
Nagel, Georg, Martin Brauner, Jana F. Liewald, Nona Adeishvili, Ernst Bamberg, and Alexander
Gottschalk. 2005. “Light Activation of Channelrhodopsin-2 in Excitable Cells of
Caenorhabditis Elegans Triggers Rapid Behavioral Responses.” Current Biology: CB 15
(24): 2279–84.
Nguyen, Jeffrey P., Frederick B. Shipley, Ashley N. Linder, George S. Plummer, Mochi Liu,
Sagar U. Setru, Joshua W. Shaevitz, and Andrew M. Leifer. 2016. “Whole-Brain Calcium
Imaging with Cellular Resolution in Freely Behaving Caenorhabditis Elegans.” Proceedings
of the National Academy of Sciences of the United States of America 113 (8): E1074–81.
O’Hagan, Robert, Martin Chalfie, and Miriam B. Goodman. 2005. “The MEC-4 DEG/ENaC
Channel of Caenorhabditis Elegans Touch Receptor Neurons Transduces Mechanical
Signals.” Nature Neuroscience 8 (1): 43–50.
Perea, Gertrudis, Aimei Yang, Edward S. Boyden, and Mriganka Sur. 2014. “Optogenetic
Astrocyte Activation Modulates Response Selectivity of Visual Cortex Neurons in Vivo.”
Nature Communications 5: 3262.
Pokala, Navin, Qiang Liu, Andrew Gordus, and Cornelia I. Bargmann. 2014. “Inducible and
Titratable Silencing of Caenorhabditis Elegans Neurons in Vivo with Histamine-Gated
Chloride Channels.” Proceedings of the National Academy of Sciences of the United States
of America 111 (7): 2770–75.
Pospisil, Dean, Max Aragon, and Jonathan Pillow. 2023. “From Connectome to Effectome:
Learning the Causal Interaction Map of the Fly Brain.” bioRxiv : The Preprint Server for
 Biology, November. https://doi.org/10.1101/2023.10.31.564922.
Prinz, Astrid A., Dirk Bucher, and Eve Marder. 2004. “Similar Network Activity from Disparate
Circuit Parameters.” Nature Neuroscience 7 (12): 1345–52.
Raiders, Stephan, Erik Calvin Black, Andrea Bae, Stephen MacFarlane, Mason Klein, Shai
Shaham, and Aakanksha Singhvi. 2021. “Glia Actively Sculpt Sensory Neurons by
Controlled Phagocytosis to Tune Animal Behavior.” eLife 10 (March).
https://doi.org/10.7554/eLife.63532.
Ramot, Daniel, Bronwyn L. MacInnis, and Miriam B. Goodman. 2008. “Bidirectional
Temperature-Sensing by a Single Thermosensory Neuron in C. Elegans.” Nature
Neuroscience 11 (8): 908–15.
Randi, Francesco, Anuj K. Sharma, Sophie Dvali, and Andrew M. Leifer. 2022a. “Neural Signal
Propagation Atlas of C. Elegans.” https://doi.org/10.48550/ARXIV.2208.04790.
———. 2022b. “Neural Signal Propagation Atlas of $\textitC. elegans$.” arXiv [q-bio.NC]. arXiv.
http://arxiv.org/abs/2208.04790.
Ripoll-Sánchez, Lidia, Jan Watteyne, Haosheng Sun, Robert Fernandez, Seth R. Taylor, Alexis
Weinreb, Mark Hammarlund, et al. 2022. “The Neuropeptidergic Connectome of C.
Elegans.” bioRxiv. https://doi.org/10.1101/2022.10.30.514396.
San-Miguel, Adriana, and Hang Lu. 2013. “Microfluidics as a Tool for C. Elegans Research.”
WormBook: The Online Review of C. Elegans Biology, September, 1–19.
Sarma, Gopal P., Chee Wai Lee, Tom Portegys, Vahid Ghayoomie, Travis Jacobs, Bradly Alicea,
Matteo Cantarelli, et al. 2018. “OpenWorm: Overview and Recent Advances in Integrative
Biological Simulation of.” Philosophical Transactions of the Royal Society of London. Series
B, Biological Sciences 373 (1758). https://doi.org/10.1098/rstb.2017.0382.
Scheffer, Louis K., C. Shan Xu, Michal Januszewski, Zhiyuan Lu, Shin-Ya Takemura, Kenneth J.
Hayworth, Gary B. Huang, et al. 2020. “A Connectome and Analysis of the Adult Central
Brain.” eLife 9 (September). https://doi.org/10.7554/eLife.57443.
Schmitt, Cornelia, Christian Schultheis, Navin Pokala, Steven J. Husson, Jana F. Liewald,
Cornelia I. Bargmann, and Alexander Gottschalk. 2012. “Specific Expression of
Channelrhodopsin-2 in Single Neurons of Caenorhabditis Elegans.” PloS One 7 (8):
e43164.
Shen, Fred Y., Margaret M. Harrington, Logan A. Walker, Hon Pong Jimmy Cheng, Edward S.
Boyden, and Dawen Cai. 2020. “Light Microscopy Based Approach for Mapping
Connectivity with Molecular Specificity.” Nature Communications 11 (1): 4632.
Skuhersky, Michael, Tailin Wu, Eviatar Yemini, Amin Nejatbakhsh, Edward Boyden, and Max
Tegmark. 2022. “Toward a More Accurate 3D Atlas of C. Elegans Neurons.” BMC
Bioinformatics 23 (1): 195.
Stevenson, Ian H., and Konrad P. Kording. 2011. “How Advances in Neural Recording Affect
Data Analysis.” Nature Neuroscience 14 (2): 139–42.
Susoy, Vladislav, Wesley Hung, Daniel Witvliet, Joshua E. Whitener, Min Wu, Core Francisco
Park, Brett J. Graham, Mei Zhen, Vivek Venkatachalam, and Aravinthan D. T. Samuel.
2021. “Natural Sensory Context Drives Diverse Brain-Wide Activity during C. Elegans
Mating.” Cell 184 (20): 5122–37.e17.
Suzuki, Hiroshi, Rex Kerr, Laura Bianchi, Christian Frøkjaer-Jensen, Dan Slone, Jian Xue,
Beate Gerstbrein, Monica Driscoll, and William R. Schafer. 2003. “In Vivo Imaging of C.
Elegans Mechanosensory Neurons Demonstrates a Specific Role for the MEC-4 Channel in
the Process of Gentle Touch Sensation.” Neuron 39 (6): 1005–17.
Szigeti, Balázs, Padraig Gleeson, Michael Vella, Sergey Khayrulin, Andrey Palyanov, Jim
Hokanson, Michael Currie, Matteo Cantarelli, Giovanni Idili, and Stephen Larson. 2014.
“OpenWorm: An Open-Science Approach to Modeling Caenorhabditis Elegans.” Frontiers
in Computational Neuroscience 8 (November): 137.
Tanaka, Hidenori, Aran Nayebi, Niru Maheswaranathan, Lane McIntosh, Stephen Baccus, and
 Surya Ganguli. 2019. “From Deep Learning to Mechanistic Understanding in Neuroscience:
The Structure of Retinal Prediction.” Advances in Neural Information Processing Systems
32.
https://proceedings.neurips.cc/paper/2019/hash/eeaebbffb5d29ff62799637fc51adb7b-Abstr
act.html.
Taylor, Seth R., Gabriel Santpere, Alexis Weinreb, Alec Barrett, Molly B. Reilly, Chuan Xu,
Erdem Varol, et al. 2021. “Molecular Topography of an Entire Nervous System.” Cell 184
(16): 4329–47.e23.
Tremblay, Sébastien, Camille Testard, Jeanne Inchauspé, and Michael Petrides. 2022.
“Non-Necessary Neural Activity in the Primate Cortex.” bioRxiv.
https://doi.org/10.1101/2022.09.12.506984.
Uzel, Kerem, Saul Kato, and Manuel Zimmer. 2022. “A Set of Hub Neurons and Non-Local
Connectivity Features Support Global Brain Dynamics in C. Elegans.” Current Biology: CB
32 (16): 3443–59.e8.
Van Den Oord, Aaron, Oriol Vinyals, and Others. 2017. “Neural Discrete Representation
Learning.” Advances in Neural Information Processing Systems 30.
https://proceedings.neurips.cc/paper/2017/hash/7a98af17e63a0ac09ce2e96d03992fbc-Abs
tract.html.
Varshney, Lav R., Beth L. Chen, Eric Paniagua, David H. Hall, and Dmitri B. Chklovskii. 2011.
“Structural Properties of the Caenorhabditis Elegans Neuronal Network.” PLoS
Computational Biology 7 (2): e1001066.
Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez,
Łukasz Kaiser, and Illia Polosukhin. 2017. “Attention Is All You Need.” Advances in Neural
Information Processing Systems 30.
https://proceedings.neurips.cc/paper/7181-attention-is-all.
Venkatachalam, Vivek, Ni Ji, Xian Wang, Christopher Clark, James Kameron Mitchell, Mason
Klein, Christopher J. Tabone, et al. 2016. “Pan-Neuronal Imaging in Roaming
Caenorhabditis Elegans.” Proceedings of the National Academy of Sciences of the United
States of America 113 (8): E1082–88.
Vogelstein, Joshua T., Katrin Amunts, Andreas Andreou, Dora Angelaki, Giorgio Ascoli, Cori
Bargmann, Randal Burns, et al. 2016. “Grand Challenges for Global Brain Sciences.” arXiv
[q-bio.NC]. arXiv. http://arxiv.org/abs/1608.06548.
Voleti, Venkatakaushik, Kripa B. Patel, Wenze Li, Citlali Perez Campos, Srinidhi Bharadwaj,
Hang Yu, Caitlin Ford, et al. 2019. “Real-Time Volumetric Microscopy of in Vivo Dynamics
and Large-Scale Samples with SCAPE 2.0.” Nature Methods 16 (10): 1054–62.
White, J. G., E. Southgate, J. N. Thomson, and S. Brenner. 1986. “The Structure of the Nervous
System of the Nematode Caenorhabditis Elegans.” Philosophical Transactions of the Royal
Society of London. Series B, Biological Sciences 314 (1165): 1–340.
Witvliet, Daniel, Ben Mulcahy, James K. Mitchell, Yaron Meirovitch, Daniel R. Berger, Yuelong
Wu, Yufang Liu, et al. 2021. “Connectomes across Development Reveal Principles of Brain
Maturation.” Nature 596 (7871): 257–61.
Wool, Lauren E., and International Brain Laboratory. 2020. “Knowledge across Networks: How
to Build a Global Neuroscience Collaboration.” Current Opinion in Neurobiology 65
(December): 100–107.
Yamins, Daniel L. K., and James J. DiCarlo. 2016. “Using Goal-Driven Deep Learning Models to
Understand Sensory Cortex.” Nature Neuroscience 19 (3): 356–65.
Yemini, Eviatar, Albert Lin, Amin Nejatbakhsh, Erdem Varol, Ruoxi Sun, Gonzalo E. Mena,
Aravinthan D. T. Samuel, Liam Paninski, Vivek Venkatachalam, and Oliver Hobert. 2021.
“NeuroPAL: A Multicolor Atlas for Whole-Brain Neuronal Identification in C. Elegans.” Cell
184 (1): 272–88.e11.
Yu, Chih-Chieh Jay, Nicholas C. Barry, Asmamaw T. Wassie, Anubhav Sinha, Abhishek
 Bhattacharya, Shoh Asano, Chi Zhang, et al. 2020. “Expansion Microscopy of.” eLife 9
(May). https://doi.org/10.7554/eLife.46249.
Zhan, Xu, Chao Chen, Longgang Niu, Xinran Du, Ying Lei, Rui Dan, Zhao-Wen Wang, and Ping
Liu. 2023. “Locomotion Modulates Olfactory Learning through Proprioception in C.
Elegans.” Nature Communications 14 (1): 4534.

Appendix 1: analytical power calculations

General setting. We need to describe the dependency of the neuron’s output on its inputs which
we can formalize as: 𝑦 = 𝑓(𝑥). Let there be 𝐿 noisy observations, abstracting away time, of the
2
form 𝑦𝑖 = 𝑓(𝑥𝑖) + η𝑖 with isotropic Gaussian noise η𝑖~𝑁(0, σ 𝐼) as neural observations generally
2
have channel noise. In this context, 𝐼 represents the identity matrix and σ denotes the noise
variance. In this setting, we want to ask how well we can approximate 𝑓(𝑥) by function fitting on
limited amounts of data. We want to describe this function so that we can predict it for all
possible inputs 𝑥. This thus covers all possible behaviors of the nervous system for any stimuli
and behaviors as well as the response to any internal perturbations. This setting protects us
from having a fragile model that only performs well inside the specific contexts studied in an
experiment.

Analytical power calculations

Linear setting
Instead of the general setting, let us operate in one in which we can describe the function as a
sum of 𝐾 terms of relevant basis functions:
𝐾
𝑓(𝑥) = ∑ 𝑊𝑘𝑔𝑘(𝑥)
𝑘=1
with a set of appropriate basis functions 𝑔𝑘(𝑥), which may, e.g. describe synapse-synapse
interactions or local dendrite interactions, and weights 𝑊𝑘 which describe how important each of
the basis functions is. Now, to be clear, real neuron transfer functions can not meaningfully be
written in this form, for example, because they have an output nonlinearity. However, while
output nonlinearities can e.g., set half of the outputs to be zero, they are unlikely to massively
affect the power calculations because real neural output nonlinearities tend to be relatively
simple smooth functions(Kato et al. 2014). Importantly, in this scenario, due to the linear nature
of the identification problem, we can use the well-established theory for linear systems
identification to obtain solid intuitions.

How many such basis functions should we need for a neuron? No one knows. In a super
simplistic world, neurons could mostly be linear, in which case we would just have one basis
function for each synapse and we could get away with K=30. In a super complicated world,
every synapse could be multiplied together into a basis function and we would need K>>1
million. However, in reality, complexity will be way higher than linear as we know that the
neuromodulators have major modulating effects. But clearly, we do not expect all combinations
to meaningfully interact, e.g. because many synapses are far away from one another. As such,
we may believe that the right K will be somewhere between K=1,000, allowing interactions
between any pair of synapses, and K=100,000 allowing many three-way interactions. We will
thus use K=10,000 as our estimate, knowing full well that uncertainty about this number of
parameters needed to describe neurons is high.

Because noise is isotropic in our system, we can use whatever basis function system we like to
do our analysis. Importantly, in the basis function of the singular vectors (components) of the
𝐸[𝑥𝑥'] system (~the system discovered by PCA), the dimensions stop being coupled to one
another, and we can view the identification problem as 𝐾 uncoupled estimates. Instead of
estimating the weights in the inconvenient original coordinate system of 𝑋 where all dimensions
have the same noise but complex covariance structure, we will thus estimate all our weights in
the convenient coordinate system of the relevant components, where all dimensions are
uncorrelated but have different noise levels. We will now call the transformed input activities with
means subtracted 𝑥𝑖 and whenever we use indices i we imply that we are in this transformed
coordinate system. Importantly they are now whitened, and have unit variance and zero
covariance with one another.
In that coordinate system we have our weight estimates:
^
β𝑖 = < 𝑥𝑖𝑦 > / < 𝑥𝑖𝑥𝑖 >

Now, in this estimate, errors come from misestimations of the first and the second term.
However, when whitening the signals, the 𝑥𝑖 are being divided by 𝑃𝐶𝑖. And for the difficult to
estimate dimensions with 𝑘 >> 1, our error estimate is entirely dominated by the second term
in the standard error expansion equation. This term has a noise level of
σ
σ𝑖 ≈
𝑃𝐶𝑖 𝐿

If all PCs are of the same size (say 1), then we obtain the well-known result:
^ 𝐾
||β − β||2 = σ
𝐿
But if the PCs are distinct as they always are in practice, we instead obtain a variance correction

In practice, PC spectra are extremely heavy-tailed. If we record even just a few hundred
neurons, we generally find that the smallest PCs are smaller than the largest PCs by a factor of
thousands. As such, estimation becomes very difficult, in particular given that σ ≈ 𝑃𝐶1in most
systems. The intuition for all this is simple, small singular values have only a tiny bit of
associated variance. For example, if we have two convergent input neurons that are strongly
correlated with one another then the weight associated with the first PC (the average of the two
neurons) is easy to estimate while the weight associated with the second PC (the difference of
their contributions to their shared target) is hard to estimate as it hard to assign credit to either
of the neurons.

This is where stimulation comes in. If we randomly stimulate in the space of the 𝑔𝑖 then we are
effectively adding the identity matrix to the covariance matrix. Stimulation thus makes all
singular values become similar to one another and reduces the difference between the large
and the small singular values. In relevant simulations, stimulation can typically get this
correction factor (condition number c) to be roughly 1 (or at least not larger than 3), see below.
Let us go back to the example of the two correlated input neurons. The stimulation will make
them less correlated, making it easier to assign credit to either of the inputs.

So the calculations of power are relatively simple from a mathematical perspective. Suppose an
analysis of one neuron with just one input will take, say, 10 seconds (enough for order 100
observations, which we will call ∆𝑡0) to identify the transfer function. If a neuron has K
2
parameters we will instead need 𝑐 𝐾∆𝑡𝑜. We assume a worst-case c=3 and hence that we would
need 900,000 seconds = 15,000 minutes ≈ 250 days. In other words, we would need massive
but tractable amounts of data.

Appendix 2: empirical power calculations

Here we compute empirical results of a timeseries linear dynamical model on a large sample (N≈1000) of
randomly generated connectomes. Dynamics at time-step t+1 are defined as:

𝑌𝑡 = 𝑙𝑜𝑔𝑖𝑠𝑡𝑖𝑐((𝐴 × 𝑋𝑡 + σ𝑡) − µ)

Where A is the (static) connectome adjacency matrix, 𝑋 is the input-state at time t, σ is a variability
𝑡
(standard deviation) which is generated anew at each timestep, and μ, a threshold/offset value
which remains constant during the simulation.

Taking from biology the understanding that such dynamics operate at the millisecond timescale,
we can calculate that one timestep in our simulation corresponds roughly to 20 ms of real-world
wallclock time, and therefore a simulation of 1,000 timesteps corresponds to 50 seconds (the
purple line in Figure 2).

We randomly sample the number of perturbations and compute the recording time to evenly
space these perturbations across the duration. Connectivity (density of A) is sampled as a
uniform value between .01 and 0.5, broadly accommodating the synaptic density of C. elegans
(0.06) (Matelsky et al. 2021). The σ parameter is randomly sampled along the log distribution
from [0.01-10.0). μ was set to 0.5 for these simulations, and the scale of the perturbations, as a
ratio of the number of neurons randomly influenced in each simulation (one value per
simulation) was randomly uniformly sampled on the interval [0.01-0.5).