Does the connectome help predict C. elegans brain activity?

Score your model on a frozen benchmark: whole-brain calcium imaging of 38 freely moving worms (Atanas et al. 2023), 89 identified neurons, predicting activity 2.4 s (4 frames) ahead on animals the model never saw in training.

The reference models are one graph neural network trained with different edge sets. With real connectome edges it scores the same as with degree-preserving shuffles of the connectome, and slightly below nearest physical neighbours. Can your model use the wiring diagram, or anything else, to do better?

Reference results

How to submit

1. Download data/benchmark.npz (standardized traces X/<animal> and validity masks M/<animal>, each (T, 89)) and data/benchmark_meta.json (neuron order, horizon, and the frozen fold of every animal).
2. For each fold f in 0–4, train only on animals whose fold is not f, then predict the animals in fold f.
3. Save one .npz with a (T, 89) float array per animal, keyed by animal id. Row t is your prediction of row t + 4. Rows you cannot predict may be NaN.
4. Drop it above. make_persistence_submission.py writes a valid example.

Scoring: per animal, pooled R² over entries where the target is valid and that neuron's own input at time t was valid (the rule every reference model follows), so every submission is scored on the same entries. The summary is the mean over animals with a 2,000-draw bootstrap; comparisons are paired by animal. The Python reference scorer is scoring.py; its bootstrap intervals can differ from this page in the fourth decimal. The targets are public data, so the split is on the honour system: please report results only from fold-respecting training.

Data and citation

Data: Atanas, Kim et al. 2023, Cell 186:4134, DANDI:000776 v0.241009.1509 (CC-BY-4.0). Connectome edges for the reference models: White et al. 1986 and Witvliet et al. 2021. Benchmark: Rohit Ravi, 2026. Split sha256 ; preprocessing .