Item

Abstract

Decoding is a popular approach for assessing the information the activity of a neural population contains about externally accessible variables. Decoding analyses are typically limited to examining a small fraction of the brain at high temporal resolution (e.g. using electrophysiology), or a large fraction of the brain at low temporal resolution (e.g. fMRI). Here we use data from hundreds of neuropixel penetrations covering hundreds of brain regions in the Allen atlas to decode task and behavioral variables at unprecedented spatial and temporal resolutions in mice performing a perceptual decision-making task.
In the International Brain Lab (IBL) task, mice are presented with a visual grating stimulus on one side of a screen, and report whether this was left or right by turning a steering wheel; this results in a reward if the chosen side matches the stimulus side. Mice maximize rewards by exploiting a blockwise prior probability governing stimulus side. We decode task variables (blockwise prior probability; stimulus identity; reward) and behavioral variables (choice; wheel speed; whisker movements) from neural activity using maximum likelihood linear and logistic regression. We report decoding measures (R2 or accuracy) on held-out test trials using multi-fold cross validation, assessing statistical significance by comparison with bespoke null distributions.
Preliminary results indicate substantial variability in decoding performance of all variables across sessions and brain regions. Despite this, we find several broad trends in the data. The reward signal, motion energy of the whisker pad, and wheel movements are represented across many brain regions. Notably, wheel speed is better decoded than wheel velocity across all brain regions considered. This result indicates that much of the movement-related information in brainwide neural activity is not specific to the exact kinematics of the movement.
We can also decode blockwise prior probability, stimulus, and choice across a number of cortical and subcortical regions. Decoding of pre-stimulus blockwise prior probability, however, is more modest than for stimulus and choice decoding. We find, for example, strong stimulus decoding in VISp, VISpm, and ZI and strong choice decoding in MOp, MOs, CP, and MRN.
In future work, we aim to compare our results with encoding and dimensionality reduction analyses. We also hope to build on our decoding of the blockwise prior probability to better understand how prior probabilities are represented across brain regions.