Visuomotor flexibility is embedded in the topography of frontal cortex

Reviewer #1 (Public review):

Summary:

Flexible natural behavior requires flexible sensory-motor mapping. In the visual domain, a visual stimulus at one location can guide a saccade toward another. How the receptive field (RF) and motor field (MF) properties of oculomotor structures support this flexibility is not known. Dotson and Reynolds address this question in the marmoset, using oblique Neuropixels penetrations across horizontal segments of the frontal eye field+, supplemented by electrical microstimulation. They report that visual RF and saccade MF vector angles each change smoothly with occasional abrupt jumps, that the two maps are organized as mosaics at distinct preferred spatial scales, and that a moiré interference pattern arising from a constrained spatial-scale mismatch between partially correlated mosaics reproduces the empirical distribution of RF-MF angular differences. They conclude that visuomotor flexibility is embedded in the geometry of mismatch and matches between visual and motor maps.

Strengths:

(1) The question is well-motivated. Sensory-motor mapping is known to be flexible, and asking whether the topographic relationship between the two maps itself supports that dissociation is a fresh reframing of a long-standing problem in oculomotor control.

(2) FEF+ lies on the smooth marmoset cortical surface, which permits high-density horizontal sampling that would be difficult in the macaque arcuate sulcus, and oblique penetrations are a sensible way to track tuning across the surface. The dataset is substantial by the standards of the field (39 sites of high-density recordings across two animals, several thousand isolated units).

(3) The data are thorough, and the convergence of three independent lines of evidence is the strongest feature of the paper. Unit recordings, electrical microstimulation, and two architecturally distinct generative models point to the same organization.

(4) The central idea is conceptually novel. The proposal that flexibility can reside in the geometry of the maps, rather than only in time-varying activity, is original, and it generates concrete, testable predictions for tasks that require flexible visuomotor routing.

Weaknesses:

Major concerns

(1) The analysis collapses each oblique penetration onto a single horizontal axis and pools angles across all cortical layers, treating cortical distance as purely tangential. Because the trajectory is angled, horizontal distance and depth are confounded, so some of the apparent RF-MF drift along a penetration could reflect a laminar transition, in addition to tangential mosaic structure.

(2) RFs and MFs are estimated from the same free-viewing sessions in temporally adjacent epochs, leaving each measurement open to contamination by the other. Activity near a saccade can reflect peri-saccadic remapping rather than the stable retinotopic RF, and saccade-aligned activity following a recent flash can carry a residual visual component, given the long-lasting visual responses in FEF+ (>500 ms). Residual cross-contamination of this kind would tend to make RF and MF angles look more similar than they are, inflating the apparent local coupling and biasing the RF-MF difference distribution that the moiré model is fit to.

(3) The paper claims that visual and saccade mosaics occupy distinct spatial scales, but the two preferred spatial frequencies are close, and the separation is summarized by overlapping "failed-test" bands rather than by a statistical test or confidence interval on the preferred frequency itself. The reliability of this separation is not established.

(4) It is not clear whether the moiré model is a better model than the non-mosaic alternative. The moiré models are shown to be consistent with the data through failure to reject a Kolmogorov-Smirnov null, which is a weak form of evidence, and they are not benchmarked against a non-mosaic alternative or null model. The AM/NM convergence demonstrates architecture independence, but not that a mosaic organization is required.

Minor concerns

(1) The link from topography to behavioral flexibility (such as anti-saccades and other context-dependent transformations) is presented as a prediction but is not tested with any task manipulation. The work establishes an organizational principle and a plausible generative mechanism; whether that organization is actually exploited during flexible behavior remains open, and the framing should make this clear so the functional claim is not over-read.

(2) It is unclear how relative depth (depth 0) is defined and how layer boundaries were assigned. The Methods mention common-average re-referencing for CSD and local field analyses, but no CSD or power-depth profile is shown to anchor the layer IV / depth 0 reference across penetrations.

(3) The Discussion is brief relative to the strength of the claims. It would be helpful to address the concerns and alternative explanations above, where these cannot be fully resolved by the data.

Reviewer #2 (Public review):

Summary:

The authors asked how the visuomotor system can keep visual selection and saccade targeting related but not rigidly coupled-the flexibility required for tasks like anti-saccades. They recorded visual receptive fields (RFs) and saccade motor fields (MFs) from individual neurons in marmoset frontal eye fields and adjacent premotor eye fields (FEF+) using obliquely inserted Neuropixels probes that traverse horizontal segments of the smooth marmoset cortex. They reported that visual and motor vector angles each changed smoothly with occasional abrupt jumps (a mosaic, rather than retinotopic, organization), that the two maps drifted with respect to one another, and that they were best described by mosaic-map models tuned to different preferred spatial frequencies. They then proposed that the offset in spatial scale, combined with partial shared structure between the maps, produced a moiré interference pattern, in which the distribution of local visual-motor angle differences matched the data.

Strengths:

(1) Unlike in the macaque brain, where FEF is buried in the arcuate sulcus, the marmoset cerebral cortex is lissencephalic (smooth). The authors used this feature to their great advantage and sampled horizontal mesoscale structure with oblique penetrations of ultra-high-density Neuropixels probes.

(2) The mosaic framing was grounded in previous studies on direction maps in ferret V1 and MT, and the rate-of-change analysis (Supplementary Figure 2) plausibly reproduced the fracture-line phenomenon of those maps.

(3) The authors confirmed the robustness of the finding by reaching the same conclusions using two architecturally distinct generators: the Fourier-based annulus model and the Gaussian-noise model.

(4) They additionally provided an independent confirmation of the mosaic saccade-vector organization using electrical microstimulation. They reconciled the lower microstimulation spatial frequency by matching the spatial-averaging footprint (Supplementary Figure 6).

(5) The Noise Mosaic (NM) model was used thoughtfully to decouple two properties that the Annulus Mosaic (AM) model confounds - spatial-scale offset and inter-map correlation - and to show that both an intermediate correlation (ρ ≈ 0.6) and a scale offset are required. Conceptually, a structural (topographic) substrate for visuomotor flexibility is a fresh alternative to the standard account in which flexibility lives entirely in time-varying activity on a single map.

Weaknesses:

The two claims here are not of the same strength. The first claim that RF and MF angles are organized as mosaics at distinct spatial scales was well supported. The second claim that a moiré interference pattern is the substrate for visuomotor flexibility was an inference rather than a direct observation, and several features of the design contribute to how strongly it can be held.

First, the recordings were one-dimensional. Each oblique penetration yielded a line through the cortex, so the two-dimensional moiré pattern (Figure 4A) existed only in simulations; it was not reconstructed from the data. What the data provided was the marginal distribution of local angular differences, and the model was accepted when its simulated distribution was statistically indistinguishable from the empirical one. Matching a low-dimensional summary statistic is necessary but not strongly sufficient - multiple underlying architectures could produce similar 1-D difference distributions - so the moiré interpretation is best read as a plausible and parsimonious interpretation of the data rather than a confirmed mechanism.

Second, the inferential logic needs to be strengthened. The "preferred" spatial frequencies were those at which a two-sample Kolmogorov-Smirnov test fails to reject equality between model and data. Failure to reject is not confirmation, and the width of the accepted band depends on statistical power, which depends on sample size. The authors did show that most parameter combinations were rejected, so the test did discriminate. That said, a continuous goodness-of-fit landscape with confidence intervals on the preferred SF, and a direct test that the visual and saccade preferred SFs differ, would better support the "distinct spatial scales" claim than visual inspection of two overlapping troughs.

Third, the dataset was from two male marmosets, with 18 of 39 sites contributing to the core analyses, and the angular-difference distributions were pooled across penetrations and animals. This is standard for primate electrophysiology, but it means the spatial statistics were assumed stationary across the region, and individual variability in map layout was averaged over. The oblique-penetration geometry also added some uncertainty: the spatial-frequency estimates (in cycles/mm) are only as accurate as the reconstructed penetration angles (18.2{degree sign} {plus minus} 8.3{degree sign}), and angle error would propagate directly into the inferred scales.

Fourth, while the authors suggested the moiré interference pattern can serve flexible routing for behaviours such as anti-saccades, this was not directly tested. Instead, they used free viewing and natural saccades, so the paper demonstrated a candidate substrate without testing whether behaviour employs it. This does not undercut the main findings, but readers should treat the functional narrative in the introduction and discussion as a set of predictions rather than results.