Neural recordings obtained from different individuals show substantial variability even when the recorded behavior is broadly shared. This variability arises because the specific neurons sampled differ across subjects and because the same behavioral events often occur at different times across recordings. Standard cross-subject analyses typically require either matched time points across recordings or anatomical correspondence between measured signals. These requirements exclude many datasets in which behaviors are comparable but not temporally or anatomically aligned.
The authors frame this issue as a barrier to discovering conserved neural–behavioral relationships across subjects and motivate a method that does not rely on cross-subject temporal correspondence or strict anatomical matching.
The authors introduce Shared Representation Discovery (ShaReD), a method designed to identify neural–behavioral relationships conserved across subjects. ShaReD operates by jointly learning a single shared behavioral projection together with subject-specific neural projections. By optimizing these projections together, the method aims to reveal dimensions of neural activity that map consistently to behavior across subjects while allowing each subject to retain its own projection from recorded neural variables into that shared space.
The formulation is intended to handle datasets where comparable behaviors occur without precise temporal alignment across subjects and where the recorded neural populations differ. The preprint describes development of the method and its intended role in extending multi-subject analyses to such challenging datasets.
The authors evaluated ShaReD on synthetic datasets to test its ability to recover common structure under controlled conditions. On these synthetic data, ShaReD recovered underlying shared components across a range of conditions, including different noise levels, sample sizes, and numbers of subjects. The method was also able to separate components that were confined to particular subsets of subjects, demonstrating specificity for shared versus subject-specific structure. The preprint does not report detailed quantitative metrics or parameter values in the source text provided here; those details are not reported in the source.
ShaReD was applied to neural recordings from non-human primate motor cortex to test whether it could identify kinematic representations that generalize across individuals. The authors report that ShaReD identified kinematic representations that generalized across animals and across reaching tasks that had different movement statistics. The source text does not provide specific task names, numbers of subjects, or quantitative measures of generalization; those details were not reported in the provided source content.
This application illustrates ShaReD's potential for revealing conserved sensorimotor mappings when animals perform similar behaviors that differ in timing or statistics across sessions or subjects.
The method was also tested on rat recordings collected during a spatial alternation task. Using these data, ShaReD isolated behavior-aligned directions within a communication subspace spanning hippocampal CA1 to prefrontal cortex (PFC). The authors present this as evidence that ShaReD can identify conserved, behaviorally relevant dimensions within anatomically and functionally connected circuits across subjects. The source does not include further methodological specifics, numbers of animals, or quantitative thresholds in the provided excerpt.
The authors conclude that ShaReD extends multi-subject neural data analysis to datasets where comparable behaviors occur without cross-subject temporal correspondence. By jointly learning a shared behavioral projection and subject-specific neural projections, ShaReD can reveal conserved neural–behavioral relationships across subjects and separate components that are present in only subsets of subjects.
Implications of this approach include enabling cross-subject discovery in datasets previously excluded by standard alignment requirements, facilitating comparisons across individuals with differing recorded populations, and isolating conserved communication subspaces across brain regions. The preprint text provided here does not enumerate limitations in detail, nor does it report specific performance metrics or implementation parameters; those details were not reported in the source excerpt.
The work received funding from the National Institute of Neurological Disorders and Stroke (R01NS125298). The authors declared no competing interests.