This study introduces Current-Based Decomposition (CURBD), a computational framework that fits data-constrained recurrent neural network (RNN) models to multi-region neural recordings to infer the magnitude and directionality of inter-region currents. It was validated on simulated ground-truth networks, then applied to calcium imaging in mice (4 cortical regions, spontaneous running), electrophysiology pseudopopulations in macaques (amygdala, ACC, striatum; Pavlovian conditioning), and single-unit recordings in 5 human epilepsy patients (hippocampus/amygdala, preSMA, dACC; memory retrieval).
CURBD accurately recovered ground-truth inter-region currents in simulations down to 5% neuronal subsampling (VAF up to 0.98 for key current sources). In mice, M2-to-V1 currents best decoded running speed. In macaques, striatum-to-ACC currents were strong after water reward but ACC-to-striatum currents only emerged bidirectionally after juice reward—an asymmetry consistent across both monkeys and 5 RNN initializations. In humans, preSMA-to-H/A currents specifically increased for familiar images, while novel images recruited large sustained currents across the whole network.
- Inferred directed interactions reflect functional, not anatomical, connectivity and cannot establish causality. - The model assumes a single static directed interaction matrix across the entire recording period, which may miss learning- or state-dependent changes. - Human and monkey datasets relied on pseudopopulations (non-simultaneously recorded neurons averaged across trials), which may not capture single-trial dynamics.
CURBD is not a clinical intervention but is a freely available computational tool (github.com/rajanlab/CURBD) that neuroscientists and clinicians studying multi-region brain dynamics—including in human intracranial recordings—can apply to infer directional inter-area communication without requiring simultaneous recordings or known anatomical connectivity.
Explore related topics
