Research map / Probabilistic and scientific modeling / Molecules and proteins

Discovering Mechanistic Models of Neural Activity: System Identification in an in Silico Zebrafish

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative latent models for single-cell data · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds an in silico zebrafish (simZFish) testbed with known ground truth and shows LLM-guided tree search finds accurate but OOD-failing autoregressive models, while wiring-diagram structural priors yield models that generalize and recover effective connectivity and impulse responses.

Problem

Neural system identification lacks ground truth, so models are ranked only by predictive error and may exploit statistical shortcuts instead of recovering true circuit mechanisms.

Approach

Use a neuromechanical simulator of the larval zebrafish optomotor circuit as a verifiable testbed, and run LLM-guided PUCT tree search over code-based transition models, comparing unconstrained search with search given only wiring-diagram existence priors (no weights).

Key results

• Tree-search model ts[422] outperforms NLinear, TSMixer and DeepAR by almost an order of magnitude on average over prediction horizons on test MAE, but shows a large generalization gap on the OOD holdout.
• Structure-constrained model sts[445] is more than an order of magnitude better on OOD holdout than baselines at longer horizons, and achieves over 30-fold lower effective connectivity error and over 5-fold lower impulse response error than ts[422].
• Ground-truth circuit with history only (gt[h]) has higher error than the naive mean baseline, showing the autoregressive task is non-identifiable without sensory drive.

Surprising findings

• The true mechanistic circuit model conditioned on history only performs worse than a naive mean-persistence baseline.
• The best unconstrained tree-search model achieves top test accuracy yet fails on OOD stimuli, indicating statistical shortcuts.

Limitations

• simZFish covers only a simplified visuomotor circuit and behavior repertoire.
• Tree search is computationally intensive.
• Evaluated only in a fully observable, noise-free regime; partial observability and low SNR not studied.

Experimental setup

Datasets: simZFish simulated dataset (400,000 samples)

Benchmarks: simZFish in silico testbed (Test MAE, Holdout MAE)

Metrics: MAE Holdout MAE (OOD) effective connectivity Frobenius error (L_Jac) impulse response MAE (L_IR)

Baselines: NLinear TiDE TSMixer DeepAR mean baseline ground-truth circuit

Builds on: PUCT simZFish ZAPBench AlphaEvolve

Closest papers by meaning

• Partial observation can induce mechanistic mismatches in data-constrained models of neural dynamics (NeurIPS 2024 · similarity 0.86)
• Discovering plasticity rules that organize and maintain neural circuits (NeurIPS 2024 · similarity 0.86)
• BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.85)
• Model Based Inference of Synaptic Plasticity Rules (NeurIPS 2024 · similarity 0.85)
• Intrinsic Goals for Autonomous Agents: Model-Based Exploration in Virtual Zebrafish Predicts Ethological Behavior and Whole-Brain Dynamics (NeurIPS 2025 · similarity 0.85)
• Characterizing control between interacting subsystems with deep Jacobian estimation (NeurIPS 2025 · similarity 0.85)