DynISys splits multi-area recordings into per-area dynamics, with PMd-to-M1 traffic confined to three dimensions
Synopsis
Separating a brain area's own dynamics from the inputs driving it is hard, because the same recorded activity can arise from many combinations of the two. DynISys models each recorded area as its own low-dimensional dynamical system, lets areas exchange signals only through a few explicit channels, and maps latent states to firing rates through an invertible readout. On simultaneous PMd and M1 recordings from two monkeys, eight latent dimensions per area matched the PSTH reconstruction of MR-LFADS at 128 per area, and the inferred PMd-to-M1 communication timing was consistent with target information traveling from PMd to M1.
Interpretation
Simultaneous multi-area recordings can now be decomposed into each area's own dynamics, explicit inter-area communication, and external inputs from unrecorded sources, rather than fitted as one mixed system. Earlier multi-area dynamics models typically carried excess latent dimensions and could mistake modeling artifacts for circuit features; at the roughly 256 dimensions needed for comparable reconstruction, about 30% of MR-LFADS latent variance fell in the null space of its linear readout, activity absent from predicted firing rates. On simultaneously recorded PMd and M1 single units from two monkeys performing a continuous reaching task, DynISys with 8 latent dimensions per area reached PSTH R² of 0.60 in PMd and 0.67 in M1, versus 0.59 and 0.68 for MR-LFADS at 128 dimensions per area; CURBD fits on the same data were not substantially better than predicting each neuron's mean firing rate.
The model writes each area as a low-dimensional vector field generated by a neural ordinary differential equation, lets areas exchange signals only through a few communication channels, and infers drive from unrecorded sources separately as external input. An injective normalizing-flow readout maps latent states to high-dimensional firing rates, so the latent state stays recoverable from predicted rates and the model is limited in inventing structure along directions the rates cannot express. On the PMd and M1 recordings, the time course of next-target decoding from DynISys latents closely matched decoding from the recorded spikes; the IntLFADS control made target information decodable from its latents well before it was detectable in recorded activity.
Fixed-point searches on the fitted M1 vector field yield locations set by inputs rather than initial states, and those locations carry information about the upcoming reach. Of 821 fixed-point searches, 797 converged to locations with speed below 10⁻⁶, and all 797 were locally stable; swapping in another trial's initial condition left the fixed point nearly unchanged, while swapping in another trial's inputs moved it far away. Linear decoders from fixed-point location gave mean R² of 0.45 for starting position, 0.64 for reach angle and 0.34 for final target position; perturbations injected along the target-encoding directions of the PMd-to-M1 channel shifted fixed points in the same direction as the perturbation.
Perspective
The demonstrated scope is intracortical PMd and M1 recordings from two monkeys: one performing a continuous reaching task, and one performing a random-target reaching task across five sessions spanning about six weeks; training uses no task variables or behavior. The authors propose pathway-specific perturbation experiments as the next step to test the model's central predictions.
The authors state that the causal relationships identified here hold within DynISys and should be read as experimentally testable hypotheses rather than established properties of the biological circuit, and that their generality must be evaluated across animals, tasks and brain areas. The multi-session result comes from one public dataset, and its session-alignment read-in and read-out matrices are precomputed under behavioral supervision and frozen during training.