Journal of Neural EngineeringSource publication:
Loading content…
Models of cross-region neural dynamics have largely stayed linear, while each region's recorded activity mixes local dynamics with dynamics shared across regions. CroP-DYNO sets learning directly on predicting target-population activity from source activity, so shared dynamics are extracted first, and it lets each model component be specified as linear or nonlinear. On macaque S1–M1 and rat CA1–CA3 data it predicts target activity better than a non-prioritized nonlinear model and a prioritized linear model, and it localizes nonlinear structure and yields interaction strengths consistent with known circuit anatomy.
Cross-region modeling can now isolate shared dynamics first and discuss nonlinearity second. Earlier cross-region models either did not prioritize shared dynamics or restricted all parameters to linear functions, so they could not satisfy prioritization and nonlinearity together. Across four region pairs spanning macaque S1 and M1 and rat CA1 and CA3, CroP-DYNO predicted target-population activity better than a non-prioritized nonlinear RNN and the prioritized linear model CroP-LDM across latent state dimensions.
The model can place nonlinearity in a specific component rather than calling the whole system nonlinear. Fully linear or fully nonlinear models cannot separate nonlinearity in the evolution of shared latent dynamics from nonlinearity in the mapping from latent states to target activity. In the sensorimotor data, making only the mapping from shared latents to target activity nonlinear matched the prediction accuracy of flexible nonlinearity selection; in the hippocampal data, cross-region nonlinearity could sit in the mapping or the latent dynamics, and localization differed between CA1 and CA3 within-region models.
Interaction strength can be quantified and lines up with known anatomical direction. Raw prediction accuracy is confounded by target noise and intrinsic predictability, so it cannot compare interaction strength across targets. Using partial R², which measures the source population's contribution beyond what the target explains about itself, M1→S1 exceeded S1→M1, CA3→CA1 was slightly stronger than the reverse, and within-CA3 interactions were strongest, matching sensorimotor projections, the Schaffer collateral pathway, and CA3 recurrent circuitry.
The method is validated on simultaneous recordings from macaque S1–M1 during a movement task and rat CA1–CA3 during a working-memory task, covering both cross-region and within-region interactions, and at inference time it estimates shared latent states from source-region activity alone. Nonlinearity localization stays at the model level: it indicates where nonlinear structure best helps predict cross-population activity rather than identifying the biological source in the circuit.
The work states that causal filtering only ensures source activity precedes target activity in time, and that the quantified interaction strengths reflect correlation rather than functional causality, since shared inputs or common task variables could produce the same predictive relationship. The authors propose testing these hypotheses with perturbation experiments such as electrical or optogenetic stimulation, and list extension to multiple populations, nonstationary interactions, and other neural modalities as future directions.