Proactive distractor suppression in early visual cortex

Reviewer #1 (Public review):

Summary:

The authors investigated if/how distractor suppression derived from statistical learning may be implemented in early visual cortex. While in a scanner, participants conducted a standard additional singleton task in which one location more frequently contained a salient distractor. The results showed that activity in EVC was suppressed for the location of the salient distractor as well as for neighbouring neutral locations. This suppression was not stimulus specific - meaning it occurred equally for distractors, targets and neutral items - and it was even present in trials in which the search display was omitted. Generally, the paper was clear, the experiment was well-designed, and the data are interesting. Nevertheless, I do have several concerns mostly regarding the interpretation of the results.

(1) My biggest concern with the study is regarding the interpretation of some of the results. Specifically, regarding the dynamics of the suppression. I appreciate that there are some limitations with what you might be able to say here given the method but I do feel as if you have committed to a single interpretation where others might still be at play. Below I've listed a few alternatives to consider.

(a) Sustained Suppression. I was wondering if there is anything in your results that would speak for or against the suppression being task specific. That is, is it possible that people are just suppressing the HPDL throughout the entire experiment (i.e., also through ITI, breaks, etc., rather than just before and during the search). Since the suppression does not seem volitional, I wonder if participants might apply a blanket suppression to HPDL until they learn otherwise. Since your localiser comes after the task you might be able to see hints of sustained suppression in the HPDL during these trials.

(b) Enhancement followed by suppression. Another alternative that wasn't discussed would be an initial transient enhancement of the HPDL which might be brought on by the placeholders followed by more sustained suppression through the search task. Of course, on the whole this would look like suppression, but this still seems like it would hold different implications compared to simply "proactive suppression". This would be something like search and destroy however could be on the location level before the actual onset of the search display.

(2) I was also considering whether your effects might be at least partially attributable to priming type effects. This would be on the spatial (not feature) level as it is clear that the distractors are switching colours. Basically, is it possible that on trial n participants see the HPDL with the distractor in it and then on trial n+1 they suppress that location. This would be something distinct from the statistical learning framework and from the repetition suppression discussion you have already included. To test for this, you could look at the trials that follow omission or trials. If there is no suppression or less suppression on these trials it would seem fair to conclude that the suppression is at least in part due to the previous trial.

Reviewer #2 (Public review):

The authors of this work set out to test ideas about how observers learn to ignore irrelevant visual information. Specifically, they used fMRI to scan participants who performed a visual search task. The task was designed in such a way that highly salient but irrelevant search items were more likely to appear at a given spatial location. With a region-of-interest approach, the authors found that activity in visual cortex that selectively responds to that location was generally suppressed, in response to all stimuli (search targets, salient distractors, or neutral items), as well as in the absence of an anticipated stimulus.

Strengths of the study include: A well-written and well-argued manuscript; clever application of a region of interest approach to fMRI design, which allows articulating clear tests of different hypotheses; careful application of follow-up analyses to rule out alternative, strategy-based accounts of the findings; tests of the robustness of the findings to detailed analysis parameters such as ROI size; and exclusion of the role of regional baseline differences in BOLD responses.

The report might be enhanced by analyses (perhaps in a surface space) that distinguish amongst the multiple "early" retinotopic visual areas that are analysed in the aggregate here. Furthermore, the study could benefit from an analysis that tests the correlation over observers between the magnitude of their behavioural effects and their neural responses.

The study provides an advance over previous studies, which identified enhancement or suppression in visual cortex as a function of search target/distractor predictability, but in less spatially-specific way. It also speaks to open questions about whether such suppression/enhancement is observed only in response to the arrival of visual information, or instead is preparatory, favouring the latter view. The theoretical advance is moderate, in that it is largely congruent with previous frameworks, rather than strongly excluding an opposing view or providing a major step change in our understanding of how distractor suppression unfolds.

Author response:

We thank the editor and the reviewers for the positive evaluation of our manuscript and the thoughtful comments. Below we provide a provisional reply to the reviewers’ comments, which we will address in more detail in the revised manuscript.

Reviewer 1 highlights three important alternative interpretations of our results: (1) sustained suppression, (2) enhancement followed by suppression, and (3) priming. We believe that these alternatives need to be addressed to improve the conclusions we can draw from the available data.

(1) Sustained suppression: As outlined by R1, it is possible that participants suppressed the HPDL throughout the entire experiment, instead of proactively instantiating suppression on each trial. While possible, we believe that this account is unlikely to explain the present results, given the utilized analysis approach, a voxel-wise GLM fit to the BOLD data per run (see Materials and Methods for details). Specifically, we derived parameter estimates from this GLM per location to estimate the relative suppression. Sustained suppression would modulate BOLD responses throughout the run, i.e. also during the implicit baseline period used to estimate the contrast parameter estimates. Hence, a sustained suppression should not result in a differential modulation between locations, as the BOLD response at the HPDL during the baseline period would be equally suppressed as during the trial. We will discuss this important aspect in the revised manuscript.

(2) Enhancement followed by suppression: R1 correctly points out that BOLD data, given the poor temporal resolution, do not allow for the detection of potential transient enhancements at the HPDL followed by a later and more pronounced suppression (akin to “search and destroy”). We agree with this assessment. However, we would also argue that a transient enhancement followed by sustained suppression before search onset constitutes proactive suppression in line with our interpretation, because suppression would still arise proactively (i.e., before search and hence distractor onset). Whether brief enhancement precedes suppression cannot be elucidated by our data, but we believe that it constitutes an interesting avenue for future studies using time-resolved and spatially specific recording methods. We will address this important addition in the updated manuscript.

(3) Priming: It is possible that participants particularly suppress locations which on previous trials contained a distractor. This account constitutes a different perspective than statistical learning integrating across many trials. We believe that it is likely that both accounts contribute to the observed effect to some degree, as both the distant (but often repeated) and the most recent past should inform our priors. Indeed, arguably recent trials should be particularly informative for our predictions as natural environments vary across time, and hence the statistical learning system should remain sensitive to potential changes in the environment. In short, we agree with R1 that the n-1 trial may impact suppression, and therefore charting the potential contributions of this type of priming compared to statistical learning is a relevant addition to the manuscript. We will perform the suggested analysis; however, we also note that dividing trials based on the n-1 trial will significantly reduce the reliability of the parameter estimates (e.g. only ~1/3 of trials follow omissions).

Reviewer 2 had two valuable suggestions to advance the inferences we can draw from the available data. In particular, R2 proposed two additional analyses, which we will consider during revision.

First, R2 suggests separating the utilized early visual cortex (EVC) ROI mask into the three retinotopic areas comprising EVC (V1, V2, V3) and to perform the key analyses in surface space for each ROI separately. We agree that exploring distractor suppression across V1, V2 and V3 separately is an interesting extension to our results. Our reasoning to combine early visual areas into one mask was two-fold: First, we did not have an a priori reason to expected distinct neural suppression between these early ROIs. Therefore, we did not acquire retinotopy data to reliably separate V1, V2 and V3, instead opting to increase the number of search task trials. The lack of retinotopy data naturally limits the reliability of the resulting cortical segmentation. However, we believe that separating EVC into its constituent areas using anatomical data is nonetheless a promising addition to our primary analyses. Therefore, during revision we will explore the main suppression analyses split into V1, V2, and V3.

Second, R2 highlights that behavioral facilitation and neural suppression could be correlated across participants. The rationale is that should neural suppression in EVC relate to the facilitation of behavioral responses, we may expect a positive relationship between neural suppression at the HPDL and RTs across participants. We agree with R2’s suggestion and will perform the analysis accordingly. However, we note that any results should be interpreted with caution, as the present sample size of n=28 is small for an across participant correlation analysis involving neural and behavioral difference scores.

In summary, we believe that addressing the reviewers' suggestions will substantially improve our manuscript, particularly regarding the interpretation and scope of our findings.