Fig. 2: Time- and trial-slicing components identify preparatory and kinematic information in motor cortical activity, respectively.

a, Behavioral and motor cortical trajectories (n = 182 neurons from M1 and PMd) during a classic center-out reaching task with straight reaches (top) and curved maze reaches (bottom; modified from ref. ⁵²). Different colors indicate different reach directions. Hand position: hand positions during the experiment. Trial-averaged raw data: condition-wise trial-averaged reaches (dashed lines) versus neural population activity (solid lines), projected onto the two-dimensional (2D) subspace that best matches hand trajectories. Raw data: raw population activity mapped onto hand trajectories at single-trial resolution. Neuron-slicing NMF: denoised population activity mapped onto hand trajectories (neuron-slicing NMF, 12 components; equivalent to NMF performed on the trial-concatenated data matrix). TCA: denoised population activity (TCA, 12 components) mapped onto hand trajectories. Trial-slicing NMF: denoised population activity (trial-slicing NMF, 12 components) mapped onto hand trajectories. b, Schematic of a sliceTCA model with multiple components of the same slice type versus a model with mixed slice types. c, Two example trial-slicing components, with neurons ordered by peak activation times of the first component. Sequential patterns distinguish specific reach conditions (here, upper left versus upper right straight reaches). d, The single time-slicing component, which displays a high temporal weight preceding movement onset. Condition-specific neural weights are shown in the slice. e, R² of fivefold cross-validated velocity decoding in each model (error bars represent the s.e.m. over n = 49 and n = 53 test trials for the maze and no-maze conditions, respectively, averaged over a fivefold cross-validation of 20 permutations of the trials). f, Correlations between neural weights on the time-slicing component in the PMd. Correlations were high for pairs of trials with similar reach direction and curvature and low for dissimilar reaches. g, Mapping of average activity in the time-slicing component before movement onset (from 0.75 to 0 s before onset) onto reach targets, revealing a strong association (R² = 0.95 and R² = 0.91, center-out versus curved reaches). h, Partially reconstructed activity from the time-slicing component, projected into a 3D subspace identified to maximally separate clockwise (CW) versus counterclockwise (CCW) movements and target x and y positions. Data points are clustered according to both reach direction and curvature, indicating that the time-slicing component encodes information about the dynamics of the upcoming movement (dots, clockwise reaches; triangles, counterclockwise reaches).