Introduction
Recent studies have investigated the geometric similarity of task structure by developing novel approaches to decode of brain states across animals [1,2,3]. The extent to which neuronal representations are similar within an animal at disparate times or between different animals performing the same task is not well understood. Quantifying the representational similarity of brain states will be critical for understanding disorders that involve impairment of neuronal dynamics. Here, we employed canonical correlation analysis (CCA) to quantify similar network states across time and between animals in recordings of rodent anterior cingulate cortex (ACC) during a decision-making task in which an animal must select one of two choices.
Methods
CCA identifies the strongest overlapping patterns between two different datasets by providing correlation coefficients (CCs) ordered by magnitude. These optimal CCs are identical to the singular values of the cross-covariance matrix calculated after orthogonalizing the data via QR decomposition (described in [2]).
We processed multi-unit neuronal recording into a set of fixed-length firing rate time series; each was synchronized to the time that the animal indicated its choice within a trial. CCA was computed for each pair of trials across all recording sessions to obtain CCs from each comparison.
Results
Within- and between- session trial-to-trial recurrence matrices were constructed using the 1st (i.e. the maximal) CC from each CCA comparison.
Using the within-session recurrence matrix for each session, we clustered trials (KMeans; k=2 clusters) and used cluster labels to decode the animal’s choice. Decoding performance for each session was quantified with Dice distance and evaluated via permutation tests against surrogate data from shuffled cluster labels. In 36/52 sessions, this metric indicated choice decoding was better than chance.
Using the within and between-session recurrence matrices, we clustered all trials from all sessions (KMeans; k=2), and choice decoding was better than chance in 30/52 sessions.
Discussion
We used CCA for pairwise trial comparison to align neural data within and between sessions, using these alignments to build trial-to-trial recurrence matrices that reveal representational similarities in neuronal activity. Notably, the successful decoding of choice from these neural metrics demonstrates that rodent ACC network states exhibit a common, low-dimensional structure across different animals.
References
1. Melbaum, S., Russo, E., Eriksson, D., Schneider, A., Durstewitz, D., Brox, T., & Diester, I. (2022). Conserved structures of neural activity in sensorimotor cortex of freely moving rats allow cross-subject decoding. Nature communications, 13(1), 7420.
2. Gallego, J., Perich, M., Chowdhury, R., Solla, S., & Miller, L. (2020). Long-term stability of cortical population dynamics underlying consistent behavior. Nature neuroscience, 23(2), 260–270.
3. Safaie, M., Chang, J., Park, J., Miller, L., Dudman, J., Perich, M. & Gallego, J. (2023). Preserved neural dynamics across animals performing similar behaviour. Nature, 623(7988), 765–771.
Acknowledgement
This work was supported by grants to CCL from NIH (AA029970, AA029409).
