Granule cells reorient cortical trajectories to separate contexts

Key results

1. ShownWithin each task, GrC activity was about as low-rank as premotor L5PT output and reproduced its representational similarity structure. GrCs did not expand cortical dynamics into a high-dimensional code.

   Participation ratio, Reach: L5PT 4.1±0.2 vs GrCs 5.0±0.5 (p=0.001); virtual-reality (VR) task: 4.3±0.3 vs 4.1±0.2 (p=0.054). Single-trial PR was lower for GrCs in both tasks (p=0.0003). L5PT-GrC representational similarity analysis (RSA): Spearman ρ=0.83±0.03 within task vs 0.41±0.2 across tasks (p=0.0003). 18 session pairs from 9 mice. · Similar L5PT and GrC representations; GrCs preserve within-task L5PT geometry; Figure legends

2. ShownAcross tasks, single L5PT neurons and population modes generalized, while GrCs remapped in time through phase shifts and polarity inversions. L5PT-GrC communicating subspaces stayed similarly correlated in both tasks.

   Attenuation-corrected cross-task correlation, single cells: r=0.58±0.05 L5PT vs 0.23±0.1 GrCs (p=0.002). Top 2 joint-task PCs: 0.83±0.02 vs 0.24±0.1 (p=0.002). Same-task controls did not differ (p=0.3). Population modes raised L5PT consistency (0.58 to 0.83, p=0.0005) but not GrC consistency (0.23 vs 0.24, p=0.7). Canonical correlation strength was conserved across tasks (173 canonical variables, r=0.69). · L5PTs generalize but GrCs remap; Canonical Correlation Analysis (CCA); Figure legends

3. ShownGrC remapping was coherent: trajectories rotated within the plane or elongated out of it between tasks, and the size of these changes predicted decorrelation. L5PT-to-GrC mappings and GrC state decoders did not transfer across tasks.

   Procrustes rotation 54±12° GrCs vs 19±3° L5PT (p=0.007); elongation 46±9% vs 26±5% (p=0.02); no difference in same-task controls. Rotation and elongation predicted decorrelation (adjusted R^2=0.82). L5PT-to-GrC regression: r^2=0.72±0.2 same-task vs 0.2±0.1 cross-task (p=0.008). Cross-task linear discriminant decoding of delay vs reward: 90±2% L5PT vs 68±6% GrCs (p=0.0009). · GrC trajectories reorient between tasks; L5-GrC mapping reconfigures across tasks; GrC behavioral readout remaps; Figure legends

4. SuggestedExtra GrC cross-task separation, beyond L5PT, grew with training and was largest when mice licked predictively in both tasks, so GrC codes diverged as behaviour converged.

   ΔDecorrelation (GrC PC1 minus L5PT PC1) was modestly higher in trained than novice pairs (p=0.037; 18 trained, 9 novice). It correlated with the dual-task predictive licking score (ρ=0.6, p=0.001), and predictive pairs exceeded reactive pairs (p=0.0009, 26 pairs). Recomputed with all tracked cells, ρ=0.56. · GrC task separation grows with learning; Behavioral Analysis; Figure legends

5. SuggestedIn a simulated mossy fiber-GrC-readout network trained on two alternating prediction tasks, a task-specific rotation of the GrC input learned faster than sparse high-rank expansion and reduced interference about as well as expansion did. A plain geometry-preserving relay suffered catastrophic forgetting.

   100 mossy fiber inputs, 100-fold GrC expansion with 4 fixed random inputs per GrC, a trained linear readout, 20 seeds. Relay and Rotation learned faster than Expansion (1/τ from exponential loss fits). Relay forgot Task 1 while training on Task 2. Tangling the Relay inputs slowed learning toward the Expansion baseline. The text gives no numerical effect sizes. · Manifold rotation for optimal learning; Model architectures; Learning speed; Figure legends