Skip to main content
Back to timeline
bioRxivSource publication:

A digital twin of freely moving mouse V1 reproduces classic orientation and cell-type tuning when frozen and replayed with gratings

Synopsis

Activity in mouse primary visual cortex (V1) during free movement can now be predicted by a multimodal digital twin from retinocentric video plus eye and head movements, locomotion speed and pupil diameter. Across 223 neurons from three mice the model reached a mean prediction correlation of 0.31; frozen and replayed with drifting gratings, it recovered canonical orientation, spatial-frequency and contrast tuning and the systematic differences between broad- and narrow-waveform cells, and without retraining it matched the population representational geometry of an independent head-fixed natural-movie dataset as closely as two recorded mice match each other.

Source-provided article image: A digital twin of freely moving mouse V1 learns a general visual code bridging classical physiology and active vision
Figure 3 · bioRxiv · Page 4

Interpretation

The digital twin predicts V1 spiking in freely moving mice from the animal's own world-view video together with six behavioral variables, not from visual input alone. V1 tuning and behavioral modulation had been measured almost entirely under head fixation with passive viewing of experimenter-controlled stimuli, which removes the animal's own shaping of its retinal input. Across 223 neurons from three mice, predicted and recorded spike counts correlated at a mean of 0.31; on 2-second-smoothed responses the mean correlation was 0.59, matching the earlier digital twin trained on the same recordings.

A visual-behavioral core is shared across animals and paired with a separate linear readout for each mouse's neurons; the shared core is pretrained on some mice and adapted to a held-out mouse. Training the same architecture from scratch on the held-out mouse needed substantially more data to reach comparable accuracy. With 3.3 minutes of recording from a new mouse, the pretrained model reached 0.21, 60% higher than training from scratch on the same data (0.13); the two converged at 33 minutes (0.32 versus 0.31).

Freezing the model and replaying drifting gratings reads out canonical V1 tuning in silico, including systematic differences between putative excitatory and inhibitory cells. That tuning had been established mainly with parametric stimuli in head-fixed experiments, yet the model was never trained on such stimuli and was given no cell-type labels. Broad-waveform units had a mean orientation selectivity index of 0.31 versus 0.20 for narrow-waveform units (permutation test p = 5×10⁻⁵); direction selectivity was 0.13 versus 0.09 (p = 9×10⁻⁴); the contrast-response exponent was 4.45 versus 3.40 (p = 0.007).

Clamping behavioral variables one at a time shows locomotion and pupil dilation share a gain increase, while only locomotion redistributes sensitivity toward higher spatial frequencies. In free behavior these variables co-vary naturally, making their separate contributions to visual encoding hard to isolate. Across ten networks retrained with different random seeds, only locomotion and pupil reliably raised evoked firing (mean +1.09 and +0.53 Hz); only locomotion's high-spatial-frequency bias differed from zero (mean +0.22, same direction in nine of ten).

Perspective

The demonstrated scope is V1 recordings from three freely moving mice, in silico clamp experiments on the frozen model, and a representational-geometry comparison with head-fixed natural-movie data from the Allen Brain Observatory; the behavioral effects are model predictions rather than direct measurements.

The behavioral manipulations are in silico clamps, predictions of how V1 would respond that still need testing against closed-loop experiments; the locomotion-specific spatial-frequency redistribution reproduces in nine of ten retrained networks but remains untested in vivo; the freely moving data come from a single dataset of three mice; and the freely moving and head-fixed regimes differ in several respects at once, so the generalization cannot yet be attributed to active self-generated sampling itself.