Do spontaneous brain-wide rotational dynamics constrain cue-evoked behavior and learning?

Rotation phase before a cue predicts the response

The paper asks whether ongoing, spontaneous brain activity contains directed rotational flow, and whether the state of that flow at the moment an unpredictable cue arrives shapes what the animal does. The answer it reports is yes on both counts. In mice, a rotational mode identified from cortex-wide calcium activity recorded at rest predicted cue-evoked behavior: the phase of rotation when a randomly timed cue arrived predicted whether the animal subsequently pulled the lever.

That result is the paper's strongest behavioral claim, and its scope matters. The cue was randomly timed, so its arrival sampled different phases of an ongoing cycle rather than a fixed, trained moment. The prediction concerns a discrete outcome, a lever pull, not the timing or vigor of a continuous movement. The authors describe the correlation as surprisingly strong and then try to account for it mechanistically rather than treating it as a curiosity.

Why variance-based methods can miss rotation

Rotational Component Analysis, or RCA, is the dimensionality-reduction technique the authors built for this problem. It takes a high-dimensional time series such as spike trains or calcium traces and searches for a two-dimensional plane onto which the data project with the most pronounced directed circulation. The quantity it maximizes is an asymmetry production rate, estimated from the mean and variance of accumulated angular current across short trajectory segments: persistent directed rotation raises the mean, while rotation-irrelevant fluctuations raise the variance.

This objective differs from principal component analysis, which selects directions of maximal variance. A rotational mode can sit in a low-variance plane roughly orthogonal to the leading principal component, and PCA would then miss it. The authors also compare against jPCA, a PCA-based method that first projects onto a leading principal-component subspace and then fits a skew-symmetric linear dynamical system. Because RCA optimizes circulation directly and uses no task variables or behavioral labels, it can be applied to continuous recordings or pooled individual trials without event alignment or trial averaging.

Single-trial detection separates RCA from jPCA

The authors benchmarked RCA on motor-cortical spiking during reaching, a setting where rotational population dynamics are already well established. When both methods were fitted to condition-averaged activity, RCA and jPCA both recovered rotational trajectories in macaque and human data. The informative comparison came next: fitting directly to individual trials pooled across reach conditions, without averaging first.

In the macaque data, the RCA plane fitted to individual-trial activity retained coherent loops across reach conditions, whereas the corresponding jPCA plane did not show coherent rotational organization. In the human data, both planes contained visible rotational trajectories, so the authors compared consistency across trials using a normalized cumulative angular current and a trial-to-trial signal-to-noise ratio. The reported values were 0.89 for RCA versus 0.05 for jPCA in the macaque condition, and 1.90 versus 0.98 in the human condition. These are trial-level consistency measures within specific conditions, not measures of decoding accuracy or behavioral prediction.

How an ongoing rotation could gate a response

The mechanism the authors propose is a dynamical constraint, not a simple correlation. A dynamical state specifies both the current activity pattern and its direction of evolution, so a cue arriving at one phase of the rotation combines with the intrinsic flow to make a response-generating trajectory more or less accessible. On this view the cue triggers a readout of a continuously evolving latent process: activity may cycle continuously, while behavior occurs only when cue and ongoing flow jointly carry the system toward a response-producing regime.

The authors support this interpretation with two models. A recurrent neural network developed a spontaneous rotation even though no rotational structure was explicitly imposed. A minimal two-dimensional model of response learning treats the intrinsic rotational mode as a constraint on both response and learning. The learning argument is that reinforcement could preferentially strengthen cue-evoked drives that cooperate with the existing flow rather than requiring cortex to construct a trajectory opposing it. This is the authors' interpretation, consistent with prior evidence they cite that animals learn activity patterns within existing neural manifolds more readily than trajectories violating intrinsic dynamics; the paper's own data establish the predictive relationship, while the learning account rests on the models.

Two rotation centers separate preparation from task

Beyond a single rotation, the authors identify double-rotation modes containing two centers of rotation, extracted with a variant called DRCA. In mouse cortex-wide calcium activity, pre-cue and post-cue trajectories occupied different parts of the double rotation. A classifier using the pre-cue DRCA state reached an area under the curve of 0.612 on held-out sessions, the cue-onset RCA state reached 0.638, and combining them reached 0.677, against chance performance on the diagonal. The double-rotation plane explained 0.8% of total variance in the mouse data.

The same structure appeared in human fMRI from the HCP-YA social task, where the paradigm alternated between task blocks and fixation blocks. Block-aligned trajectories occupied different parts of the double rotation: fixation dynamics circulated around one center and task dynamics around the other, with the mode explaining 2.1% of total variance. The authors emphasize that these are latent-space rotations, defined in a learned state space whose axes are weighted combinations of recorded regions or sensors, not physical waves traveling across the cortical sheet. A loop in this latent space does not by itself imply spatial propagation.

Rotational irreversibility tracks early psychosis

The clinical result concerns a different question from behavioral prediction. In human data, projections were learned from 41 randomly selected participants with psychosis and evaluated in the remaining 41 participants and all 53 controls. PCA standard deviation showed only a modest group difference (P = 0.0275), whereas the asymmetry production rate showed a more pronounced difference (P = 8.29 × 10⁻⁵), with group comparisons using two-sided Welch's t-tests.

The authors note that the asymmetry production rate is related to thermodynamic irreversibility but should not be equated with entropy production rate: entropy production quantifies all detectable time-reversal asymmetry, whereas this current-based measure isolates the component expressed through a current odd under both time reversal and the chosen spatial transformation. The practical point is that a variance-based summary nearly missed a group difference that a rotation-sensitive measure captured. This is a group-level statistical comparison in one psychosis dataset, not a diagnostic claim about individuals.

Rotation as a coordinate of dynamical readiness

The paper's unifying claim is that ongoing brain-wide activity is not merely background variability around an evoked response. Its directed flow defines coordinates that shape how an input is converted into action, how rewarded dynamics are reinforced during learning, and how preparatory and task-related states are organized. The mouse result is the clearest instance: a rotation present before the cue, sampled at random phases, predicted a discrete response, and the two-dimensional description makes that link simpler than earlier analyses of spontaneous-to-evoked coupling.

The most decision-relevant comparison in the paper is therefore not rotation versus no rotation but rotation-sensitive versus variance-based measurement. In the macaque single-trial benchmark, RCA retained coherent loops where jPCA did not, with trial-level signal-to-noise ratios of 0.89 versus 0.05; in the psychosis comparison, the asymmetry production rate separated groups far more sharply than PCA standard deviation. Readers weighing these methods should note the conditions: the behavioral prediction is established in mice with randomly timed cues and a lever-pull outcome, the double-rotation separation appears in mouse calcium imaging and human fMRI, and the clinical difference is a group-level finding in one dataset. Within those bounds, the paper's priority is that low-variance rotational structure carries behavioral and clinical information that variance-maximizing reductions can discard.

Sources used in this answer