What an atlas-fitted connectome can and cannot do: evolutionary search over interneuron stimulation in a whole-body C. elegans model
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
The C. elegans connectome is complete, but which behaviors its wiring can produce, and which it cannot, is unknown. We asked this question in a whole-body simulation whose only adaptive element is a learning agent outside the nervous system. The agent may inject current into a short whitelist of interneurons, never into motor neurons or muscles; the 302-neuron OpenWorm c302 network, with dynamics fitted once to a signal-propagation atlas, a literature-based motor layer and a viscoelastic body on agar do the rest. Every hypothesis was pre-registered with its judgment criterion. Stimulating five command and steering interneurons (AVB, AVA, SMDD, SMDV, RIV) was sufficient for chemotaxis to a source 2.5 mm away (91% of trials; four of five training seeds), whereas stimulating eight sensory neurons was no better than chance, reproducing at the behavioral level the silence of the sensory-to-command step in the fitted network. Evolution rediscovered the pirouette rule, reversing and executing a deep bend when concentration fell, and ablations showed that reversal, deep bend and steering were all necessary. Shuffling the wiring while preserving weights, counts and signs cut performance to a fifth; in the atlas-fitted dynamics this loss was carried by the gap-junction layer. In corridors four body widths wide, corners were passed by a deep bend pressed against the wall, so the dorsal or ventral direction of the bend fixed the direction of the turn. With ventral bends only, right turns almost never occurred and the worm entered the ventral arm of a T-maze first. Because temporal sensing cannot distinguish the arms at a junction, the learned strategy was to enter one arm and reverse when the odor faded; a left–right concentration difference was needed before the bend direction followed the source. The model yields three testable predictions for narrow-maze assays and separates what the wiring contributes to navigation from what sensing and the body must supply.
Author summary
Knowing every connection in a nervous system does not tell us what the wiring does. We took the complete wiring diagram of the worm C. elegans , kept every connection fixed, and let a learning algorithm discover what it could achieve by stimulating a handful of neurons, as an experimenter with optogenetic tools could. The algorithm was forbidden to touch the neurons that drive muscles. Through the wiring it learned to find an odor source, and it did so by rediscovering the strategy real worms use: reverse and turn sharply when the smell gets weaker. It failed when given only sensory neurons, which tells us that the sensory-to-command step is missing from the fitted wiring. Randomizing the wiring destroyed the behavior, and in our fitted model the electrical connections were the ones that mattered. In narrow mazes, the worm could only turn toward the side of its sharp bend, so a worm that bends to one side almost always turns that way and needs a costly reversal to reach food on the other side. These results separate what the wiring contributes from what the body and the senses contribute, and they make predictions that can be tested with worms in mazes.
