Journal: Current biology : CB

Article Title: Rat orbitofrontal ensemble activity contains multiplexed but dissociable representations of value and task structure in an odor sequence task

doi: 10.1016/j.cub.2019.01.048

Figure Lengend Snippet: (A) The firing rates of all single neurons on each trial constituted a high-dimensional vector (360 vectors or data points in a 1078- dimensional space). The firing rates of all neurons at the odor time were linearly projected to a principal component subspace with 80% variance explained, then to an LDA space with labels about the current reward. Each LDA component combined a weighted sum of inputs from all the neurons. The LDA transformation was supervised by trial-type labels that only separated current value (reward vs. non-reward) so that the LDA could find components that best separated the two classes. Comp., Component. (B) The first but not the second LDA component perfectly separated the two trial types (p = 1.0 × 10−3 and 1.0, respectively; two-sided permutation test, 1000 bootstrap samples). (C) An ROC-based value-selectivity index (2 × |AUC – 0.5|) ranging from 0 (low selectivity) to 1 (high selectivity) was used to test current value selectivity for each individual LDA component. The first LDA component showed perfect value selectivity (1.0; p = 1.0 × 10−3; permutation test; 1000 bootstrap samples). But, none of the remaining 150 LDA components were selective for the current value (< 0.05; p = 1.0 for all components; two-sided permutation test; 1000 bootstrap samples). (D) Value discriminability was used to test whether value was distributed across components (0 – 1 indicates the level of value discriminability by population components). The true discriminability was compared with that from the label-shuffled data. The first LDA component showed significantly higher value discriminability than the shuffled data (1.0 vs. 0.05; p = 1.0 × 10−3; one-sided permutation test; 1000 bootstrap samples), but the remaining LDA components did not show significantly higher value discriminability than the shuffled data (0.13 vs. 0.1; p = 1.0; one-sided permutation test; 1000 bootstrap samples). Curr. Val. Discrim., Current Value Discriminability. Error bars are standard deviations (SDs). (E) A dendrogram using all different LDA components contained both value and state information (left). A dendrogram that only used the first LDA component only contained value information without detailed state information (center), while a dendrogram that only used the remaining LDA components contained state information without current value (right). (F) Decoding of 24 states with the first LDA component (reconstructed to 151 PCs before the decoding analysis). (G) Comparison of decoding accuracy for each state (represented by each dot) between all LDA components and the first LDA component being used. Dec. Decoding; LCs, LDA Components. (H) Confusion matrix at odor time was binarized at thresholds 0%, 5%, and 20%. (I) Correlation coefficients compare the similarity between hypothesized “current value” and “current location” matrices and the actual confusion matrices (obtained by using the first LDA component) at different filtering thresholds. (J) Decoding of 24 states with the remaining 150 LDA components (reconstructed to 151 PCs). (K) Comparison of decoding accuracy for each state between all LDA components and the remaining LDA components (the first one was left out) being used. (L) Confusion matrix at odor time was binarized at thresholds 0%, 5%, and 20%. (M) Correlation coefficients compare the similarity between hypothesized “current value” and “current location” matrices and the actual confusion matrices (obtained by using the remaining LDA components) at different filtering thresholds.

Article Snippet: We trained a linear discriminant analysis (LDA) algorithm (MATLAB function: fitcdiscr ) to classify 24 trial types or locations for each one of six task events.

Techniques: Plasmid Preparation, Transformation Assay, Comparison