Data constrained models for mechanism discovery in Neuroscience

Kanaka Rajan, Harvard

Video Recording

Abstract:
Understanding how brain-wide circuit mechanisms govern slow behavioral state transitions—such as the shift from active to passive coping under persistent, inescapable stress, a behavioral hallmark related to learned helplessness and major depressive disorder—is fundamentally challenged by the partial observability of large mammalian brains and the limitations of standard correlational analyses. While experimental recordings capture neural activity outputs, they cannot directly isolate the underlying causal inputs or distinguish locally generated recurrent dynamics from signals inherited across distributed brain regions. To bridge this gap, we introduce a computational framework centered on data-constrained multi-region Recurrent Neural Networks (RNNs) trained via recursive least squares to directly match experimentally recorded single-neuron time series across species, ranging from larval zebrafish with near-complete brain observability to rodents, non-human primates, and humans. From the inferred directed interaction matrices of these trained networks, we develop CURBD (Current-Based Decomposition), a method that demultiplexes recorded neural activity into constituent within-area and inter-area source currents without requiring prior anatomical constraints.

Applying CURBD to whole-brain calcium imaging data from larval zebrafish subjected to inescapable behavioral stress, we uncover the multi-timescale causal mechanisms driving the transition from active to passive coping across the habenula, raphe nucleus, and telencephalon. Whereas conventional output-based analyses reveal only a monotonic ramp in ventral habenular activity and concurrent raphe suppression, source-current decomposition demonstrates that early active coping is driven externally by fast directed currents from the raphe to the habenula, while late-stage passivity is maintained by slower, plastic strengthening of recurrent connectivity within the habenula and projections from the telencephalon. Furthermore, by extending this approach to whole-brain recordings via chained epoch training and tensor decomposition, we extract low-dimensional compositional modes that capture brain-wide spatiotemporal motifs—including conserved ramping, oscillatory, and decaying modes alongside individual-specific baseline signatures. Under behavioral stress and pharmacological perturbation with ketamine, these unsupervised modes automatically isolate habenular hyperactivation and reveal that ketamine delays passive coping not by restoring baseline control dynamics, but by shifting brain-wide source-current trajectories into a distinct intermediate state, demonstrating how data-constrained generative models serve as a transparent virtual laboratory for causal mechanism discovery and therapeutic intervention design.

Bio:
Kanaka Rajan, PhD, is a computational neuroscientist, Associate Professor at Harvard Medical School, and a founding faculty member of the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. Her research seeks to understand how important cognitive functions — such as learning, remembering, and deciding — emerge from the cooperative activity of multi-scale neural processes. Using data from neuroscience experiments, she applies computational frameworks derived from machine learning and statistical physics to uncover integrative theories about the brain that bridge neurobiology and artificial intelligence.