Inferring brain-wide interactions using data-constrained recurrent neural network models
TL;DR
A computational tool called CURBD uses recurrent neural networks to accurately infer the direction and magnitude of communication currents between multiple brain regions from neural recording data.
Problem / question
Understanding how distributed brain regions coordinate during behavior requires inferring directional communication between them, but current correlational methods cannot easily distinguish common inputs from directed interactions or scale to large, multi-region datasets.
Methods
The researchers developed current-based decomposition (CURBD), which trains a recurrent neural network (RNN) to reproduce experimental neural time-series data and uses the resulting directed interaction matrix to calculate inter-region currents. They validated CURBD on simulated ground-truth networks and applied it to in vivo calcium imaging and electrophysiology datasets from larval zebrafish (habenula, raphe nucleus, telencephalon), mice (V1, M2, PPC, RSC), macaques (amygdala, ACC, striatum), and humans (preSMA, dACC, hippocampus, amygdala).
Key findings
In simulated networks, CURBD accurately reconstructed ground-truth currents even when subsampling just 5% of the neural population. In larval zebrafish, CURBD revealed that stress-induced passivity is driven first by inputs from the raphe nucleus to the habenula, followed by ramp-like inputs from the telencephalon. In mice, currents from the secondary motor cortex (M2) to the primary visual cortex (V1) best predicted spontaneous running speed. In macaques and humans, CURBD identified context-specific directional currents, such as asymmetric striatum-to-ACC currents during water rewards and preSMA-to-hippocampus/amygdala currents during the retrieval of familiar memories.
Why it matters
CURBD provides a scalable, unbiased way to reverse-engineer directional, multi-region neural interactions from standard recording data, allowing researchers to uncover the specific communication dynamics underlying complex behaviors across different species.
Limitations
CURBD does not incorporate known anatomical constraints, struggles with sparse neural activity, fits only a single dynamical system that cannot natively adapt to temporal drift, and infers functional interactions that may reflect indirect polysynaptic pathways rather than direct anatomical connections.
Takeaway
By training recurrent neural networks to mimic recorded neural activity, CURBD can mathematically isolate the specific, directional currents flowing between distinct brain regions during behavior.