Avalilação PREreview de Multiplexing behavioral signals in sensory representations

Review of Mayer and Mlynarski 2025 (https://doi.org/10.1101/2025.09.04.674251)

Review by Kyu Hyun Lee (kyuhyun.lee@ucsf.edu)

Please note that this review will also be posted on PREreview.

It has been long appreciated that even the primary sensory areas of the neocortex are influenced by nonsensory factors, such as locomotion, arousal, and attention. Many of these take the form of gain modulation, which is thought to shift the tuning curve of the sensory neurons up or down. Although these modulations are widely observed, little is known about their computational roles. One could even argue that they degrade the fidelity of sensory codes; for example, the responses of a neuron in the primary visual cortex (V1) to an optimal stimulus while the animal is stationary and a suboptimal stimulus during locomotion may be similar, introducing ambiguities. Perhaps it is a bug rather than a feature?

In this work, Mayer and Mlynarski investigate how the gain modulation may be a feature after all by enabling multiplexed encoding of both sensory and behavioral information. They focus on orientation-selective neurons in the mouse visual cortex. They construct models of these neurons, in which the orientation selectivity emerges by combining direction-selective "subunits" whose preferred directions are 180 degrees apart. They then show that if the gain modulation (acting multiplicatively on the sensory tuning) is applied differentially to these subunits before they are summed, then a population of such neurons can encode both the stimulus direction and the behavioral modulating variable, such as running speed. By analyzing publicly available datasets from Allen Brain Observatory, they demonstrate that many orientation-selective neurons in the mouse visual cortex indeed show such differential gain modulation for the two direction-selective subunits. They go on to characterize the surprising diversity of these modulations, which range from ones that are not monotonic to ones that are not even correlated much with running speed. These observations are maintained from primary to higher order visual cortical areas in the mouse brain.

Overall, the paper makes an important contribution to our understanding of how behavioral factors influence sensory coding. The results strongly suggest that the naive characterization of running speed as a gain modulator acting monotonically and uniformly on all neurons needs to be revised. The main idea for how these modulators enable encoding of both sensory and behavioral information is well presented and verified convincingly with a large experimental dataset that is publicly available. The descriptions of model implementation and statistical analyses are sufficient. The figures are beautiful and very clear. The following are all minor points of discussion, provided in case the authors find them helpful.

·       The stimulus property that the authors focus on in the manuscript is the direction of moving gratings. However, given the large number of orientation-selective neurons in the mouse visual cortex, it may be that a more important stimulus feature to encode is the stimulus orientation. This might change the narrative associated with Fig 1d, in which the neuron model whose subunits are equally modulated always give the largest decoding error. I would suggest that the authors consider this possibility in the Discussion section.

·       On a related note, it seems the direction-selective neurons in visual cortex are quite sufficient for encoding both stimulus direction and running speed by virtue of having diverse gain modulation curves. Indeed the authors find many such neurons in the Allen Brain Observatory data. The main advantage of the differentially modulated orientation-selective neurons over direction-selective neurons seems to be that you need fewer of them to achieve a given level of decoding error. While this may be the case, the difference is rather small. To increase the importance of this claim, the authors are encouraged to find other advantages of the differentially modulated orientation-selective neurons over direction-selective ones. Perhaps they are good at encoding both the orientation and direction at various running speeds? This may require simulating the response to flashed gratings (rather than gratings that move in a particular direction) with a spatiotemporal receptive field.

·       The authors' model assumes that orientation selectivity in visual cortex arises from combining direction-tuned subunits. They further speculate that these subunits could correspond to LGN neurons. But are the LGN neurons typically direction-selective? What about the orientation-selective neuron in the higher-order visual areas that do not have direct access to LGN? Are the large number of direction-selective neurons that are already present in the visual cortex not involved in generating orientation selectivity? In general, how does the proposed model compare to the current best models of orientation selectivity informed by experimental data (e.g. synaptic connectivity)? Making such a comparison to justify the validity of this assumption would make the model more convincing.

·       An important question relates to the mechanism underlying the diverse modulation of single subunits by running speed. Given that the modulators are applied multiplicatively, do the authors have any ideas about the circuit level implementations? Where does the diversity of modulation originate? The paragraph in Discussion section on this could be expanded.

·       The range of velocity considered by the authors is 0 to 20 cm/s. Is 20 cm/s the maximum velocity for the head-fixed mouse in the experimental condition? If not, perhaps an additional justification for restricting the velocity to this range would be informative.

·       It seems the data the authors chose to analyze used moving gratings with a single spatial frequency (0.04 cpd). Are the results dependent on this parameter? If the gain modulation from running speed is different for stimuli with different directions, perhaps the same applies to other stimulus features. The authors are encouraged to explore this possibility. Similar questions could be asked about the temporal frequency, which seems to be represented in the data.

·       The authors report using in total 1693 neurons from 100 recording sessions after removing neurons that didn't meet their criteria. What fraction of the total number of neurons was this? Are the neurons that were removed not gain modulated by running speed?

·       Some parts of the Methods section would benefit from more descriptions. For example, in "Data preprocessing", it says data used must have "sufficient data in each bin", but what is meant by sufficient here?

·       The authors are advised to add a supplementary table detailing which of the Allen Brain Observatory datasets (including the experiment IDs) were used in the analysis, so that future work can easily reproduce the results.

·       The code that implements the model or the fits to data is not publicly available. I think it is imperative that the code be available to the public along with the manuscript (which is on bioRxiv) so that the results can be verified.

·       Stylistic suggestions:

o   In figure panels that show the oriented grating stimuli (e.g. Fig 1bc, Fig 7bc etc) the two pictured grating stimuli are actually not offset by 180 degrees, which means the neuron that pools them is technically not orientation-selective. Given that the authors mostly focus on orientation-selective neurons, it may be better to use a different icon to denote these stimuli (e.g. grating with an arrow indicating the movement direction).

o   Fig 1: Instead of "joint", "uniform" may be better, since "joint" may be taken to just means that both subunits are modulated together without reference to how this is done.

o   Page 7, first paragraph: Does "averaging modulators of each subunit" mean averaging over speed bins? It would be better to say this explicitly.

o   Page 12, "Simulated populations of gain-modulated neurons": the variable r is used for both tuning curve and response. I would use different variables.

o   References: Including the DOI for all references may make it easier for readers to identify them.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.