Figure 5
From: Neurobiologically realistic neural network enables cross-scale modeling of neural dynamics
NBGNet captures and reconstructs the latent dynamics in the reaching-out task. (a) Schematic of protocol indicates the time window used for analysis. Probability of each target direction is uniform. (b) We predicted that the latent dynamics can be recovered. (c) Representative latent trajectories derived from the ground-truth screw ECoG (left) and reconstructed screw ECoG (right). Each color represents each target direction in (a). (d) Projection of average ground truth (blue trace) and reconstructed (red trace) latent trajectories for each target on the first mode. (e) Bar plot showing the strong magnitude of the correlations between the ground truth and reconstructed latent trajectories (error bars, s.e.m.; n = 68). (f) Temporal correlation trajectories for each neural mode (green trace when above the threshold as 0.4; grey trace as below the threshold; mean ± s.e.m.). (g) Same as (c) for the inverse model to reconstruct the latent trajectories derived from LFPs. (h) Same as (d) for the projection of average ground truth LFPs-derived (blue trace) and reconstructed LFPs-derived (red trace) latent trajectories. (i) Same as (e) for the correlation between the latent trajectories obtained from recorded LFPs and estimated LFPs. (j) Same as (f) for the inverse model (purple trace when above the threshold as 0.4; grey trace as below the threshold). *p < 0.05 using two-sided Wilcoxon’s rank-sum test. n.s. indicates no significant difference.
