Inferring Active Neural Circuits Using Diffusion Scores
New AI technique maps lag-specific brain interactions from calcium imaging data
A new preprint from researchers at Yale and Harvard introduces Score-Block Time Graphs (SBTG), a method for inferring active neural circuits from high-dimensional population recordings. The approach uses denoising score models to estimate joint-window scores over consecutive brain-state snapshots, then converts those scores into calibrated, directed edge tests via cross-block score products. The key insight is that these products recover the Jacobian of the transition map between brain states under nonlinear dynamics. By introducing minimal multi-block windows that condition on intermediate time points, SBTG avoids the omitted-lag bias that plagues pairwise analyses, cleanly separating lag-specific effects.
Applying SBTG to whole-brain calcium imaging data from C. elegans, the team recovered lag-specific circuit structure not resolved by current methods. They found improved alignment with independent connectomes, revealing cell-type-specific temporal organization and neuromodulatory profiles consistent with known receptor kinetics. This demonstrates SBTG's potential as a practical AI-for-science tool that turns noisy neural population data into statistically testable circuit hypotheses, offering a data-driven route to map dynamic neural interactions without assuming parametric dynamics.
- SBTG uses denoising score models and cross-block score products to recover the Jacobian of the transition map between brain states, enabling inference of directed interactions.
- Minimal multi-block windows condition on intermediate time points to avoid omitted-lag bias inherent in traditional pairwise analyses.
- Applied to whole-brain C. elegans calcium imaging, SBTG identified lag-specific circuit structure with improved connectome alignment, cell-type-specific timing, and known neuromodulatory kinetics.
Why It Matters
SBTG turns high-dimensional neural recordings into testable circuit hypotheses, enabling scalable AI-driven discovery of brain computation.