Article Figures & Data
Figures
- Figure 1.
FMA implantation and task design. A–C, Array locations for animals B, S, and Z, respectively. Two arrays were placed in F5 on the bank of the arcuate sulcus (AS). Two additional arrays were placed in AIP toward the lateral end of the intraparietal sulcus (IPS). In animals B and Z, two more arrays were placed on the bank of the central sulcus (CS). The cross shows medial (M), lateral (L), anterior (A), and posterior (P) directions. Note that animal S was implanted in the left hemisphere and animals B and Z in the right hemisphere. D, Sketch of an animal in the experimental setup. The cues were presented on a monitor projected onto a mirror, making the light dots appear superimposed onto the grasping handle. E, Delayed grasping task with two grip types (Task 1). An example of each grip type can be seen during the movement epoch (top, power grip; bottom, precision grip). The handle was rotated to a supine orientation for demonstration purposes only. F, Delayed grasping task with two grip types and three decision conditions (Task 2). Free-choice trials were presented twice as often as each of the other conditions. Delayed-instructed trials contained a second grip cue turning a free-choice trial into a delayed-instructed trial. Trials were presented in a pseudorandom order.
- Figure 2.
Scatter plots of RT versus memory period length. A–C, The RT of animals B, S, and Z, respectively, as a function of memory period length for all task conditions and datasets. The solid line represents the mean, and error bars indicate SEM within nonoverlapping 50 ms bins.
- Figure 3.
Neural data and RT prediction methods visualized as low-dimensional trajectories. A, Neural data of both areas of an exemplar condition reduced to a low-dimensional representation of the trial course (determined by GPFA). Thick trace represents the mean of trials for one condition (instructed precision grip, dataset Z120829). Thin gray traces represent 10 random single trials. Shaded ellipses (90% confidence) represent the state of all selected single trials at the start of each epoch. B–D, High-dimensional RT prediction methods in a two-dimensional illustration. Thick red and green traces represents the mean of trials. Thin gray trace represents a single exemplar trial. α denotes the component used to predict RT for the projection method (B), Euclidean distance method (C), and velocity projection method (D).
- Figure 4.
Determination of the optimal reference time (Δt) relative to go cue on the mean trajectory. A, B, Results of the projection method in areas F5 and AIP, respectively. C, D, Results of the Euclidean distance method in areas F5 and AIP, respectively. E, F, Results of the velocity projection method in areas F5 and AIP, respectively. Thick traces are the mean of all conditions and datasets of each animal, thin traces are the SEM, and white circles are the optimal Δt used in all subsequent analysis. Insets in A–F show histograms of correlation coefficients between each neural predictor and RT over all conditions (2–6), datasets (6), and cross-validation folds (2). Black bars denote correlations with a p value <0.05. Arrows show the median together with the p value of significant difference from zero (Wilcoxon's signed-rank test). G, H, Correlation coefficient histograms of the AR at go method and the SCAR method, respectively, in F5. I, J, Same as G and H, but for neural data from AIP.
- Figure 5.
Average fraction of RT variance explained for all methods and datasets (averaged across conditions and cross-validation folds). A, Average fraction explained by F5 data. B, Average fraction for AIP. Note the clear advantage of area F5 over AIP. Chance-level calculation is based on shuffling neural data with respect to RTs repeatedly. The observed R2 values are then compared against the shuffled distributions to assess significance. Significant results are illustrated as solid bars, whereas the open bars show results that can be explained by chance (p ≥ 0.01).
- Figure 6.
Comparison of prediction performance and fraction of significant full/partial correlations between predictors and RT over all task conditions, datasets, and brain areas. A, Average fraction of RT variance explained by correlation. B, Average fraction of RT variance explained by partial correlation. Significant results are illustrated as solid bars, whereas the open bars show results that can be explained by chance (p ≥ 0.01). C, Fraction of conditions with significant correlations (p < 0.05). D, Fraction of conditions with significant partial correlations (p < 0.05).
- Figure 7.
Stability of the sign-correction vector determined at the go cue by the SCAR method. RT prediction is calculated using sign-corrected neural activity around each time point. A, B, SCAR method as a function of time for animals B and Z in areas F5 (A) and AIP (B). C, D, SCAR method as a function of time for animal S in areas F5 (C) and AIP (D). Dashed lines indicate the go cue and the median movement onset (Move). Note the difference in peak RT prediction in F5 between animals B and Z and animal S.
- Figure 8.
Selection of units by firing rate variance at the go cue for the two best performing methods in animals B and Z. A, B, Variance selection of units versus random selection using the SCAR method in F5 (A) and AIP (B). C, D, Variance selection of units versus random selection using the projection (before go cue) method in F5 (C) and AIP (D). Horizontal black bars on top represent unit percentages in which the variance selection performed significantly better than random selection (p = 0.05, Bonferroni's corrected, permutation test).
- Figure 9.
Low-dimensional representation (PCA) of neural state space illustrating day-to-day and animal-to-animal variability. Trajectories are plotted in the two principal components of area F5 explaining the most variance. A, Trajectory of a power grip from dataset B140509 in which RT is most variable along the projection axis. B, Trajectory of a precision grip from dataset B140515 in which RT is most variable orthogonal to the projection axis. C, Trajectory of a precision grip from dataset S120913 in which RT is most variable on the shared distance and projection axis. D, Trajectory of a power grip from dataset Z120921 in which RT is most variable along the projection axis.
