From systematic interneuron perturbations to connectome-constrained dynamics

Quantitative understanding of neural control over behavior requires characterizing computations performed along information-processing pathways. In principle, Caenorhabditis elegans is ideally suited for this goal since it generates non-trivial, context-dependent behavior with a relatively small nervous system. Existing anatomical and functional connectomes represent major advances but also miss important synaptic details and critical connectivities, respectively. Whole-brain functional imaging, another important approach, by itself only reveals correlations. A perturbation-based, causal atlas is needed to pin down neuronal computations in the context of behavior, including redundancies, synergies, and time constants. We sought to satisfy this need through systematic chemogenetic perturbation, combined with connectome-constrained generative modeling. Using reversible chemogenetic silencing combined with whole-brain calcium imaging, we silenced seven head interneurons individually and in pairs: two act as opposing controllers of forward-reverse states, interneuron activity predicts behavior at lags of 15-25 seconds, and pairwise silencing reveals structured non-additivity. In parallel, we developed a connectome-constrained recurrent dynamical system trained on freely-behaving whole-brain recordings, with three current layers (gap-junction, fast-synaptic, extrasynaptic) and multi-timescale kinetics. On a 40-animal cohort, our architecture beats unconstrained baselines, ties anatomy-masked models on neural prediction, and transfers across animals by relabeling parameters alone, while anatomy-masked baselines collapse to negative R^2. Three independent analyses recover the canonical locomotor circuit without being built in. The two approaches converge on the same interneuron, identified by silencing as a key modulator of reversal behavior. Our work demonstrates advantages of integrating cell-specific perturbation with connectome-constrained modeling, and motivates extending this approach to larger nervous systems.