Ethics statement

Data were provided by the Human Connectome Project [41], WU-Minn Consortium (1U54MH091657) funded by the 16 NIH Institutes and Centers that support the NIH Blueprint for Neuroscience Research, and by the McDonnell Center for Systems Neuroscience at Washington University. All participants provided written informed consent prior to data collection and all data were de-identified.

Data acquisition and participants

Data used in the present study were collected by the Washington University-Minnesota Consortium of the Human Connectome Project (HCP) [41] as part of the young adult (S1200) release. Participants all provided written informed consent in accordance with protocols approved by the ethics review board of Washington University. This dataset included resting- and task-state functional neuroimaging data (see following Methods sections for more details on each). We obtained the minimally preprocessed [44] neuroimaging data for N = 352 healthy young adults. Following the pipeline outlined Ito et al. [45], this N = 352 subset of the HCP participants were selected to ensure high quality scan data and no familial relations. All study procedures were approved by the Institutional Review Board of Rutgers University-Newark. Further details on participant recruitment can be found in Van Essen et al. [41].

We used a split-sample validation approach to minimize false discovery rate [46]. The cohort of HCP participants in the present study (N = 352) were quasi-randomly allocated to either a discovery (n = 176) or replication (n = 176) dataset. Participants were chosen randomly, but the sample size of both discovery and replication datasets were set at n = 176 to ensure equal split-halves. Each dataset was analyzed identically but independently. Participants in the discovery dataset (77 men, 99 women) had the following age ranges: 22–25 years (26.14%), 26–30 years (41.48%), 31–35 years (30.68%), and 36+ years (1.7%). Participants in the replication dataset (93 men, 83 women) had the following age ranges: 22–25 years (22.16%), 26–30 years (43.18%), 31–35 years (34.09%), and 36+ years (0.57%). Results presented in figures refer to the discovery dataset, with replication dataset results reported in-text.

Task paradigm

In a subset of analyses (see “Network analyses: Response profile network contributions”), we used data corresponding to seven HCP tasks and 24 task conditions that sampled diverse cognitive domains: emotion, gambling reward, language, motor, relational reasoning, social cognition, and working memory. In all other analyses we focused on the working memory/category-specific representation n-back task (see [47] for more details on all tasks). Prior studies suggest that the n-back task can be used to test hypotheses regarding the function(s) of specific brain areas [48]. Further, manipulating how far back a participant must remember (0-back versus 2-back) allows for the assessment of working memory maintenance (along with other contrasts, such as stimulus type). While working memory is not the main interest of the present study, specifically manipulating working memory has the added benefit of promoting task engagement.

The cognitive domain of interest in the present study was the processing of semantic categories [11,49,50]. Category-specific representations were embedded in the n-back task via blocks of trials that presented images of four reliably studied and distinct visual semantic categories [6,21]: (1) bodies [2,3]; (2) faces [4,51,52]; (3) places [5,53]; and (4) tools [49,54,55].

The specifics of the n-back task are detailed in Barch et al. [47]. Briefly, the n-back task included 2 runs with blocks of visual category trials (8 blocks, 10 trials each, 2.5 seconds each) and fixation blocks (4 total, 15 seconds each). Half of the blocks were 0-back and half were 2-back. During 0-back working memory tasks, participants determined via button-press whether presented stimuli (images of either bodies, faces, places, or tools) matched a constant target shown in the initial cue screen. During 2-back working memory tasks, participants were asked to indicate if the present stimulus matched a target from two screens back. Given that working memory load was not the main interest of the present study, results presented here reflect an average across 0-back and 2-back conditions (i.e., fMRI data was averaged after all preprocessing steps), except for in analyses examining “response profiles” (i.e., responsiveness of regions-of-interest to 24 HCP conditions, assessed independently). Additionally, preliminary analyses found no evidence that visual category selectivity was variable across the two working memory load conditions.

MRI parameters

All MRI data were collected at Washington University in St. Louis, Missouri, USA in a customized Siemens 3T “Connectome Skyra” scanner. Whole-brain, multiband, and echo-planar images (EPI) were acquired with a 32-channel head coil. The repetition time (TR) was 720 ms; the echo time (TE) was 33.1 ms; the flip angle was 52 degrees; the bandwidth (BW) was 2290 Hz/Px; in-plane field-of-view (FoV) was 208 x 180 mm; 72 slices; 2.0 mm isotropic voxels; and the multiband acceleration factor was 8. Whole-brain and high-resolution T1-weighted and T2-weighted anatomical scans were also acquired, with an isotropic voxel resolution of 0.7 mm. All imaging data were collected over two days for each participant. Each day included two resting-state fMRI scans, each lasting 14.4 minutes (for a total of approximately 29 minutes of rest data per day). During rest, participants were instructed to keep their eyes open and fixated. Task fMRI (30 minutes of task data per day) followed resting-state fMRI scans on each day. Each of the 7 tasks were performed over two consecutive runs. Further details on MRI protocols and parameters can be found in Van Essen et al. [41].

fMRI preprocessing

We acquired minimally preprocessed data [44] from the HCP database (www.humanconnectome.org). The minimal preprocessing pipeline is extensively detailed in Glasser et al. [44] and is openly available (https://github.com/Washington-University/HCPpipelines). In-brief, minimal preprocessing included: anatomic reconstruction and segmentation, motion correction, intensity normalization, and EPI reconstruction to a standard (surface-based) template. We also implemented HCP MSM-All registration (multimodal surface matching registration based on areal features from multiple sources, including: cortical folding/sulcal depth, myelin maps, resting-state networks, and visuotopic maps; [56]). Preprocessing was performed equivalently but independently on all functional runs, including resting-states and the 24 task-state conditions. The resulting data were in CIFTI grayordinate (i.e., vertex) space. We performed additional preprocessing steps on cortical vertices (59,412 vertices across the two hemispheres). On both resting- and task-state data, we removed the first five frames of each fMRI run, demeaned and detrended the timeseries, and performed nuisance regression. We followed a variant [45] on methods suggested by Ciric et al. [57] to model and regress out nuisance parameters, including motion and physiological noise. We performed nuisance regression to remove six motion parameters, their derivatives, and quadratics (24 total). Anatomical CompCor (aCompCor; [58]) was used on white matter and ventricle timeseries to model physiological noise. The first five principal components from white matter and ventricles were independently extracted. These parameters, their derivatives, and the quadratics of all regressors were removed (40 total). Altogether, 64 nuisance parameters were modeled, and associated variance removed from the data.

Global signal was not removed given evidence that it can artificially introduce negative relationships [59,60] and is non-standard for task activation analyses. However, the nuisance parameters modeled and removed with aCompCor contain components similar to the proposed global signal, with the added benefit of not removing gray matter signals [45,61].

Task-state activity estimation

Cortical activations related to task-state fMRI conditions (n-back visual semantic categories) were estimated by a standard general linear model (GLM), which convolved task timing (per block-design condition) with the SPM canonical hemodynamic response function [62]. This was performed per cortical vertex, each belonging to a parcel (or region) of the Glasser et al. [56] multimodal parcellation (MMP) atlas. Average coefficients of the GLMs were utilized for each region or functional complex (described further below in “Identification of functional complexes”) as estimates of visual category evoked brain activity. Note that the fully-distributed model (Fig 1A) was performed on the region-level data; the stimulus-driven (Fig 1B) and stimulus-driven plus visual network interactions (Fig 1C) models were performed on vertex-level data before averaging into functional complexes; and the stimulus-driven plus fully-distributed model (Fig 1D) was performed on vertex-initialized data in the first step and region-level data in the second step. The steps and rationale of each of these models are extensively detailed below.

Identification of functional complexes

To address the question of whether distributed network interactions can map [19,27,28] localized, category-selective functional responses, we focused on the functional brain regions identified in Osher et al. [21] given the similarity of the four visual semantic categories in that study to the HCP n-back task studied herein. Additionally, as suggested by Poldrack et al. [63], variables that will be used in predictive analyses should not be chosen in an empirically-driven manner (e.g., from post hoc cross validation and/or optimization procedures) because this has the potential to artificially inflate accuracy due to circularity. Thus, selecting functional regions of interest a priori has the benefit of controlling false positives in accuracy, particularly with our focus on the well-established visual category literature. Upon further review of the literature (at the time of originally conducting this study and compiling this manuscript: 2021–2022) as well as the regions belonging to the Glasser et al. [56] MMP atlas, we observed that these functional regions: (1) contained multiple MMP regions (e.g., the parahippocampal place area or PPA, which contained 6 MMP regions as in [53]); (2) did not have an exact terminological match in the MMP atlas; and/or (3) included two discrete locations (e.g., the retrosplenial complex, or RSC, often studied in conjunction with the PPA). To address this we first performed a literature search of each of the Osher et al. [21] functional regions with Google scholar (https://scholar.google.com/) and Neurosynth (https://neurosynth.org/), filtering for relevance, year, and citation count (when possible) (see Table 1 references in right-most column). We focused further on publications that had corresponding volumetric coordinates (or MMP surface region labels when possible) and generated a list of Montreal Neurological Institute (MNI) x/y/z volumetric coordinates for each functional region (with a minimum of 10 entries per region). We then used the average x/y/z coordinate set as a reference to identify corresponding MMP regions (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Specifications of functional brain areas.

https://doi.org/10.1371/journal.pcbi.1012507.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Average MNI coordinates for each functional complex from the literature.

https://doi.org/10.1371/journal.pcbi.1012507.t002

To convert from volumetric space to region space, we employed the following steps (using Connectome workbench, FreeSurfer, and AFNI/SUMA commands, per hemisphere): (1) convert the MMP atlas to FreeSurfer’s fsaverage space; (2) convert fsaverage to the surface-mesh MNI-152 template (which is also included in AFNI); (3) convert surface MNI-152 template to volumetric MNI-template and obtain coordinates. Additional commands were used to ensure that the templates aligned as best as possible, including zero-padding, thicker ribbons (that were later filled), and adding MMP sphere labels. To obtain x/y/z coordinates, we divided the template into equal portions perpendicular to the long axis of a given sphere label, identified the middle division, then calculated the centroid. We then found the MMP regions that were physically closest to the reference coordinate set for each functional region. We then referred to the Glasser et al. [56] supplemental results (as well as publications they based the MMP atlas upon; see references with asterisks in Table 1) to exclude regions that were physically close but functionally irrelevant (e.g., auditory regions near the pSTS, which was a potential ROI for face processing).

The final set of regions, including their proposed category selectivity, network affiliations, and references are listed in Table 1. The term “functional complex” indicates firstly that more than one MMP region comprised each complex, and secondly serves to distinguish an MMP atlas region from a category-selective (based on prior literature) complex in the present study. Further, in some cases there were two functionally relevant complexes in the literature (e.g., PPA and RSC, both relevant to place processing, as in [21]), therefore both were included. When appropriate, we applied the Cole-Anticevic brain-wide network partition (CAB-NP, [25]), which was based on HCP resting-state fMRI data and was generated from a Louvain community detection algorithm that assigned MMP regions to 12 functional networks. Table 1 lists the network assignments of each region, which is further reflected in Results figures by color labeling matching throughout all relevant figures. The reference sets of volumetric coordinates (x/y/z) are listed in Table 2.

Resting-state functional connectivity estimation

Functional connectivity (FC) was modeled as a series of target dependent variables (i.e., each brain region), each with multiple source independent variables (i.e., other brain regions) (see following section “Activity flow mapping: a generative model”). As in Cole et al. [27], source vertices excluded all vertices within a 10 mm dilated mask of the target region (as well as excluding vertices of the target region itself). This reduced potential circularity related to spatial autocorrelation by avoiding nearby vertices in predicting the activity of the target. Modified source sets of vertex-level timeseries were then averaged together based on their Glasser et al. [56] MMP region assignment. FC was then estimated with a method proposed by Sanchez-Romero and Cole [125] termed combinedFC. CombinedFC integrates the standard methods of bivariate correlation, partial correlation, and multiple regression together to account for causal confounds better than each method alone. In Sanchez-Romero and Cole [127], combinedFC improved precision in a series of simulated networks, and retained robust network structure with empirical fMRI data (see [127] for a detailed report of this method). Briefly, each participant’s resting-state data (pre-processed) was firstly assessed with partial correlation (amongst each pair of regions, conditioning on all other regions) to find conditional associations, which were examined for statistical significance. This step addressed potential confounders and causal chains [128] and built an initial connectivity architecture by adding in significantly correlated edges only. Then, bivariate correlation estimates (i.e., Pearson correlation) for each pair of connected regions (from the initial partial correlation step) were used to remove spurious edges resulting from conditioning on colliders (specifically, edges where r was statistically equal to zero). Lastly, a multiple regression procedure was performed to scale the estimated edge weights. For each target region functioning as a response variable, all other connected source regions functioned as regressors to estimate the final combinedFC weight.

Activity flow mapping: a generative model

Activity flow mapping is a generative modeling approach that maps held-out brain activity based upon the inherent link between activation patterns and functional connectivity [20,34,129,130]] (i.e., an empirically-parameterized extension of generative network models outlined in Betzel and Bassett [131]). Activity flow was developed by Cole et al. [27] (see Cocuzza et al. [28] for further details), and is described by Formula 1 (Fig 2): (1)

Where j is a held-out brain region, and i indexes all other regions in a given parcellation. Note that the standard application of activity flow mapping given by Formula 1 is consistent with the “fully distributed” mapping of Fig 1A (i.e., mapping that is used to generate visual category selectivity, see “Category selectivity” below) because source regions (i) refer to all other cortical regions. Variations on this are outlined in later Methods sections (see the following sections below titled “Stimulus-driven visual category selectivity” and “Category selectivity from stimulus-driven network interactions further shaped by fully distributed network interactions”). This method induces the bias that brain activations are propagated and shaped via distributed network interactions. Throughout the present study, the term “distributed” means globally connected brain systems that are specified by functional brain network organization (i.e., functional connectomes). Applied to research questions on functional processing, activity flow maps task-evoked (functionally relevant) activity in held-out target regions as the sum of the task-evoked activity in all other source regions scaled by FC estimates between the target region and corresponding source regions (Formula 1). The activity flow mapping procedure explicitly tests the hypothesis that the sum of empirically-based source region activity, parameterized by empirical patterns of brain connectivity, is equal to target region activity. Using a simulated model of neural processing, Cole et al. [27] found that activity flow mapping was most accurate when a global coupling parameter was high and the strength of self-coupling (local) connections was low (Fig 2B). Critically, this suggests that the successful application of fully distributed activity flow mapping (Fig 1A) to empirical brain data reflects similar high global coupling properties being preserved in real (rather than simulated) data. Thus, the success of activity flow mapping can be taken as a proxy for how much signals over distributed network interactions contribute to a given activation (see “Category selectivity” below for more details).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Activity flow procedure to map task activity in held out brain regions.

(A) Activity flow mapping toy diagram and formula. Task activity for held-out region j (purple) is mapped as the sum of task activity of all other cortical regions, i (coral) (n = total number of regions), weighted by their connectivity estimates with j (gray). (B) Activity flow simulation results (reproduced with permission from [27]) demonstrating that activity flow mapping is most successful when distributed processing mechanisms are high and localized processing mechanisms are low. (C) Example of activity flow mapping with empirical data (steps 1–6) (reproduced with permission from [28]). For a given target region j, estimates of intrinsic (e.g., resting-state) connectivity between j and all other source regions (step 1) are multiplied by all other regions’ actual task activations (step 2). This yields an activity flows map quantifying the contribution of all other regions’ activity flow upon the held-out region, j, for the task of interest (step 3). These are summed to equal the mapped task activity of j (step 4). This procedure is iterated over all regions to generate activity-flow-mapped task activations across the whole cortex (step 5). This is compared with the actual whole-cortex map of task activations via Pearson’s r, MAE, and R² to estimate accuracy (step 6). Importantly, this approach is flexible to different estimates of connectivity. Source vertices not included in the analyses (10 mm from the target region j; see Methods) are depicted in green.

https://doi.org/10.1371/journal.pcbi.1012507.g002

To map category-selective responses to bodies, faces, places, and tools we applied the activity flow approach to each of these categories independently. Actual task activations used were beta estimates generated by task GLMs (described above in “Task-state activity estimation”), with a focus on the regions comprising functional complexes as the held-out target regions (described above in “Identification of functional complexes”). Functional connectivity estimates between targets (functional complex regions) and sources (all other connected regions) were based upon resting-state timeseries and estimated via combinedFC (described above in “Resting-state functional connectivity”). To assess accuracy, activity-flow-mapped category responses were compared to actual (i.e., empirical) category responses via Pearson r correlation coefficient (NumPy corrcoef function in Python), mean absolute error (MAE), and the coefficient of determination (R²) (sklearn r2_score function in Python).

Each of these indices quantified a different aspect of accuracy, altogether providing a thorough evaluation of activity flow mapping. In brief, Pearson’s r estimated the linear correlation between actual and mapped brain activations, with a potential range from -1 to 1 and the property of scale invariance. MAE measured the arithmetic mean of absolute errors between actual and mapped brain activations, with a sensitivity to scale. The coefficient of determination (R²) estimated the percent of unscaled variance in the actual brain activations that was accounted for in the mapped activations. Note that R² maintains a range of negative infinity to 1; from 0 to 1 R² can be interpreted as 0 to 100% variance, but negative values indicate incorrectly scaled predictions. Thus, each index had clear advantages and disadvantages, but altogether provided a complementary, full characterization of mapping accuracy. In all cases, these comparison statistics were taken per participant, then averaged across participants to provide a random effects estimate.

Category selectivity

To assess how selective each functional complex of interest was for a given visual category, we measured the ratio of category activation amplitude divided by non-category activation amplitude, for both activity-flow-mapped data and actual data, per participant. For example, for the functional complex of the extrastriate body area (EBA) and fusiform body area (FBA) (see above: “Identification of functional complexes”), we quantified body responses divided by non-body responses. Importantly, non-category responses included only other n-back conditions, and not any other HCP task condition, such that the estimation of visual category selectivity only considers comparable task manipulations. Per participant and functional complex, we used Formula 2: (2)

We will henceforth refer to the ratio in Formula 2 as “category selectivity”, which can be interpreted as how many times larger category responses were than non-category responses. For example, if a given participant’s body-category selectivity score (for the EBA and FBA complex) was 1.5, then their activations to body images were 1.5 times larger than their activations to non-body images (i.e., face, place, and tool images in the n-back task). A category selectivity score of 1.0 reflected the hypothesized null mean given that it quantifies category activations as equal to non-category activations. The main advantage of utilizing a ratio versus a difference score is that it is less sensitive to beta value scale differences across brain regions.

Prior to calculating the category selectivity score (Formula 2) it was necessary to normalize the range of activation values. This was necessary for two reasons: (1) avoiding negative numbers (which would limit interpretability of the selectivity values) and (2) rescaling activation values such that every participant’s data was in the same range (which allows for group level comparisons). We used a standard normalization procedure termed min-max normalization (also termed feature scaling). We used Formula 3: (3) Where a and b are the lower and upper bounds of the rescaled data, respectively. We used a = 0 and b = 1.0, which were selected to increase interpretability of the category selectivity ratio since all ratios were positive and included a possible minimum of 0. The minimum and maximum values in Formula 3 were calculated based on the actual activations of a given region set (e.g., EBA and FBA across-complex average for body selectivity) across all 24 task conditions (x) for each participant separately. Note that results were highly similar if minimum and maximum values were calculated separately for mapped and actual activations. Additionally, we report the cross-complex average results herein given that no functional complex was substantially driven by a particular region (i.e., regions in a given functional complex had similar category selectivity scores). We additionally identified and removed outlier participants (in category selectivity) following the methods described by Leys et al. [132] (median absolute deviation), using a highly conservative threshold of -/+ 5 deviations relative to the median. Given this high threshold, outlier removal excluded a small number of participants (an average of n=5 of 176 were removed across analyses, or approximately 2.8% of participants) with category selectivity scores that were not representative of the group average, which is how category selectivity was reported (which had a large sample size of n = 176 per discovery and replication datasets). Note that statistical significance (see “Experimental design and statistical analysis”) was not impacted by removal of outlier participants.

In assessments of the fully distributed network interaction scheme (Fig 1A), we estimated the contribution of distributed network interactions to a given category-selective response by computing the percent of actual category selectivity (distributed plus localized processes) captured by activity-flow-mapped category selectivity (i.e., a model based upon distributed processes only, Fig 2B). We propose that the percent leftover (i.e., not captured via activity flow mapping) is due to either local processes (not isolated in the model) or other sources of model error (e.g., noise in the data). We used Formula 4: (4)

If the percentage given by Formula 4 was statistically greater than 50%, we inferred that category selectivity in that functional complex was generated primarily via distributed activity flow processes. Note that these estimates reflect a lower bound on the contribution of distributed processes, since they can be driven lower not only by local processes but by other sources of model error.

Network analyses

In assessments of the fully distributed network interaction scheme of Fig 1A, we tested the contributions of each large-scale CAB-NP functional network ([25] Fig 3B) to activity-flow-mapped responses. We did this at both the levels of: (1) mapped category-specific responses, and (2) mapped response profile.

Control analyses

Stimulus-driven visual category selectivity

We sought to test the possibility that stimulus-driven activity flow processes conferred by V1 are capable of generating visual category selectivity (Fig 1B). Note that analyses that examined stimulus-driven (described first below; Fig 1B) and stimulus-driven plus visual network interactions (described second below; Fig 1C) were performed with vertex-level data, then averaged into functional complexes after estimating category selectivity. The final model, stimulus-driven plus fully-distributed network interactions (Fig 1D) were initialized with vertex-level data in a first step, then assessed across region-level averaged data, as detailed in the next section titled “Category selectivity generated by stimulus-driven network interactions further shaped by fully distributed network interactions”. Firstly, we estimated rsFC between all visual system (i.e., VIS1 and VIS2 network vertices from Ji et al. [25], see “Network analyses” above) vertices, and refer to this visual sub-system as “VIS” throughout corresponding Results and Figures. Note that we included all functional complexes in this VIS sub-system. FC was estimated with combinedFC [127] as in prior analyses (see “Resting-state functional connectivity estimation”), except regularized regression was used instead of multiple regression to scale edge weights, given that there were more vertices than time points in vertex-level resting-state data. For each participant and target vertex (i.e., dependent variable or vertex timeseries that connections were to be estimated with), we used L1 regularized lasso regression with cross-validation across resting-state runs. The penalty term was chosen from a wide sweep of 100 log-spaced terms. We then conducted a multi-step activity flow mapping procedure. We report step 1 as the V1, stimulus-driven model, and steps 2+ as the extended VIS model.

In step 1, only activation patterns of source vertices from V1 (both hemispheres), shaped by their connectivity fingerprints with each VIS vertex (see “Activity flow generative model” and Formula 1), were used to map the task-evoked activation patterns of each VIS vertex (Fig 1B). Resulting maps for functional complex vertices were the focus of the step 1 analysis, but the entire VIS sub-system was mapped because it was later used in steps 2+. Note that cortical region V1 was identified based on the V1 region identified by Glasser et al. [56] using multimodal methods. Cortical V1 vertices were all those inside of the V1 border, with 10 mm dilated masking applied to activity flow mapping (with VIS) as described previously (see “Resting state functional connectivity estimation” and “Activity flow generative model”, Fig 2) to control for the effects of spatial autocorrelation. When mapping held-out target vertices of each functional complex, the cross-vertex (i.e., across V1) mean activation value for the corresponding category-specific condition was subtracted from each V1 source vertex to account for the (unlikely but possible) potential confound that there may still be category-specific information fed back to V1 from the complex of interest (at the temporal resolution of fMRI). These mappings were then used to generate activity in each functional complex, which was then tested for visual category selectivity (see “Category selectivity” and Formula 2). Note that at this point in the analysis, category selectivity (actual and activity-flow-mapped) was estimated for each vertex in each functional complex, then averaged (i.e., average selectivity of vertices in each complex). Thus, actual category selectivity in corresponding Results slightly varied from the region-level estimates based on fully distributed mappings (Fig 1A) (which was estimated for each region, then averaged). Additionally, we removed outliers (using a conservative threshold of 5 median deviations) as before (see “Category selectivity”). The degrees of freedom are indicated in corresponding Results that report hypothesis testing statistics.

We also performed a control analysis testing whether each functional complex’s unique connectivity fingerprint with V1 determined its visual category selectivity better than a null rsFC network model. Here, we randomly permuted the connectivity architectures (per participant) 100 times but maintained edge strength and degree [136,153,154]. Note that this was used instead of rsFC substitution (see “Control analyses” above) because there was a variable number of vertices in each functional complex, and thus they could not be used to substitute each other’s connectivity fingerprints. Nonparametric estimates of statistical significance given by Max-T (see “Experimental design and statistical analyses”) were used to compare real mapped category selectivity to the null network mapped category selectivity.

Then, in steps 2+ (extended VIS) we used the mapped activation patterns from step 1 as the source activation patterns in activity flow mapping (Formula 1) and extended this to include all other VIS sub-system source vertices (Fig 1C). We repeated this until a “settling threshold” was reached. This was to simulate potential feedback and/or recurrent processes (initialized with stimulus-driven processes) that may occur over VIS network interactions. We defined a settling threshold as the step when the mapped activation patterns no longer changed within four decimal point precision, which can be thought of as the beginning of a horizontal asymptote in mapped activation values. We then performed all analyses as we did in step 1, to assess whether stimulus-driven activity flow processes further shaped by VIS network interactions improve the generation of visual category selectivity.

Category selectivity generated by stimulus-driven network interactions further shaped by fully distributed network interactions

Expanding upon the analyses in the previous section (“Stimulus driven category selectivity”), we sought to test whether V1-initialized activity flow mappings (step 1) that are then processed over all cortical network interactions (step 2) (Fig 1D) – as opposed to a step 2 that incorporates only visual subsystems – would accurately generate visual category selectivity. We henceforth refer to this as the “stimulus-driven + fully distributed” mapping. Here, modulations conferred by all other functional networks (e.g., possible top-down effects from source regions in the dorsal attention network; DAN) were allowed to contribute to activity flow mapping in a second step. As before, we firstly initialized activity flow mapping with stimulus-driven interactions with V1 (as in “Stimulus-driven visual category selectivity” above). This first step of activity flow mapping was initialized with vertex-level data as in “Stimulus-driven visual category selectivity” above. Then we conducted an additional step that included all cortical source regions. Given the computational cost of vertex-level regularized regression (with cross-validation) across the whole cortex, this second step was performed at the region level. Following prior analyses, to control for multiple comparisons we used max-T nonparametric permutation testing (see “Experimental design and statistical analysis”). We used 100,000 permutations (i.e., p of 0.00001) and report statistical significance in terms of p < 0.00001.

Experimental design and statistical analysis

Each functional complex that was proposed to be highly responsive to the visual categories of bodies, faces, places, and tools (see above: “Identification of functional complexes”) had estimates for actual activity and activity-flow-mapped activity. For each of these estimates and each visual category, we contrasted category responses to non-category responses (e.g., body vs non-body), hypothesizing that a functional complex would exhibit higher activity (in actual and mapped data) to the category vs the non-category. To test these hypotheses, we used one-tailed, paired samples t-tests on cross-participant data. For each category of interest, non-category activity was the average of the three remaining categories. For example, if body was the category of interest, responses to faces, places, and tools were averaged for non-body activations. This provided a consistent contrast across all four sets of analyses (i.e., always category vs non-category).

In every set of analyses involving a comparison, we used the Max-T nonparametric permutation approach (shuffling conditions over 100,000 permutations – i.e., p-values of 0.00001 – unless otherwise stated) to correct for multiple comparisons [155]. For example, the Max-T approach was used when testing whether category selectivity scores (Formula 2) revealed for each functional complex were greater than 1.0 (the null mean), as well as testing whether estimates of distributed processing contributions (Formula 4) were greater than 50% (see “Category selectivity” for more on these measures). We likewise used Max-T (here with 10,000 permutations for computational tractability) to address multiple comparisons in the network-based analyses (i.e., comparing each network-level result to every other network-level result). When comparing estimates based on each network to every other network, the max-T threshold statistic that was surpassed for significance is reported, instead of a lengthy list of observed t-statistics. The max-T procedure was used in any analysis where we could not assume a normal distribution a priori. Note that for category selectivity and percent distributed processing analyses, degrees of freedom were based on sample size with outlier participants removed (see “Category selectivity”). To assess accuracy of the activity-flow-mapped (Formula 1) results, we compared the actual and activity-flow-mapped activations in each of the four visual category analyses with Pearson’s r, MAE, and R² (Fig 2). All findings are reported for cortical regions in the right hemisphere for the discovery dataset. Left hemisphere and replication dataset results are reported in S1 Table through S15 Table.

Localized response profiles are captured by activity flow processes

Our preliminary hypothesis was that activity flow processes could accurately map responses in extensively investigated category-selective visual cortex regions. Firstly, we extended prior work [20,27,28] by implementing a resting-state functional connectivity (rsFC) method [127] that estimates a more causally plausible network (Fig 3A and 3B), and by analyzing a large cohort (N = 352) split into discovery and replication datasets. These updates serve to improve the inferential validity of activity flow mapping results [18]. Secondly, we quantified response profiles (also termed population receptive fields) of functional complexes identified as integral to four visual categories (Tables 1 and 2). A response profile (rows of Fig 3D) was defined as the task-evoked activity levels across all 24 HCP conditions. Establishing response profiles allowed us to verify that each complex was responsive to their respective visual categories (outlined in black: Fig 3D, right) and verify that activity flow mapping (Fig 3D, left) was reliable across a variety of cognitive domains.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Mapped activations across all HCP conditions yielded response profiles for four functional complexes of interest.

(A) Causally confounding graph patterns in standard FC estimation methods (adapted with permission from [127]). CombinedFC incorporates both bivariate and multiple regression measures such that confounders, chains, and colliders are accounted for. (B) The cross-participant (n = 176) average resting-state connectivity matrix (estimated via combinedFC) of 360 MMP regions [56], ordered along each axis per the Cole Anticevic brain-wide network partition (CAB-NP [25]; color-coded on each axis to match panel C). This was the functional network organization utilized in the present study for activity flow mapping. Note that our implementation of combinedFC used multiple regression as the final step, and therefore FC estimates were given by beta coefficients (see Methods). (C) Cortical schematic of the CAB-NP and its 12 functional networks from Ji et al. [25], reproduced with permission. (D) Response profiles (across all 24 HCP conditions) of four complexes of interest to the present study (indicated along the y-axis; note that “r” stands for right hemisphere); mapped versus actual (left and right respectively; mean across participants depicted in each panel). Black boxes highlight the n-back conditions that maintained visual semantic category embeddings germane to a given functional complex (e.g., 0-back bodies and 2-back bodies for the right EBA and right FBA). Activity-flow-mapped response profiles were highly accurate, suggesting that mapped activation patterns of the functional complexes of interest were reliable across multiple cognitive domains.

https://doi.org/10.1371/journal.pcbi.1012507.g003

Activity-flow-mapped response profiles maintained high accuracy in both hemispheres (right: r = 0.92, MAE = 3.93, R² = 0.80; left: r = 0.92, MAE = 3.96, R² = 0.79; replication dataset: S1 Table), which corroborates prior findings that activity flow processes (specified by connectivity patterns) can accurately map task-evoked activity in held out brain regions across diverse cognitive domains [27]. Moreover, this suggests that intrinsic (i.e., resting-state) functional network organization (Fig 3C) plays a prominent role in generating localized response profiles. Subsequent analyses examined the processes involved in category-selective responses (i.e., focusing on black-outlined conditions in Fig 3D), as generated by activity flow mapping.

Localized visual category selective responses are generated via empirically-derived models of fully distributed network interactions

We next tested the hypothesis that the classically observed selectivity for visual categories in select visual cortex regions is shaped by their unique connectivity fingerprints distributed across cortex (Fig 1A). An emerging literature suggests that network interactions and/or distributed activation patterns are integral to visual category processing [11,14,93,129]. However, the nature of the network interactions underlying the generation of visual category selectivity remains unclear, as well as the extent to which those distributed interactions account for localized category selectivity (relative to local, within-region processes). Here, we sought to test the generative capacity of distributed network interactions by modeling activity flow processes over intrinsic functional connections [27]. The four visual category models are reported in separate sections below with similar formatting to aid cross-comparison. The functional complexes identified based on prior literature (Fig 4 and Tables 1 and 2) were anatomically positioned along occipital (anterior and lateral to V1) and temporal (tending posterior and ventral to A1) cortices, which reflects the associative processing demands of visual category computations [49,50]. Functional complex regions belonged primarily to the secondary visual (VIS2) network, with additional locations in dorsal attention (DAN), default mode (DMN), and language networks (LAN) (Fig 4) [25].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Visual categories and identified functional complexes.

(A) Exemplar stimuli from each of the four visual semantic categories shown to participants across the n-back task (bodies, faces, places, and tools). Please note that the face example is a photograph of the author of this manuscript, meant to be representative (but not an exact match) of the original face stimuli while protecting privacy concerns of the original models. In each category, stimuli varied to sample a wide array of representations. Body images included whole bodies (no faces) and isolated body parts (excluding nudity); face images included diverse ages and facial expressions; place images included indoor and outdoor scenes (and combinations, e.g., a patio); and tool images included distinct items (e.g., plier versus drill) as well as similar variants (e.g., two drills: one white, one blue). See Barch et al. [47] and www.humanconnectome.org for further details. (B) A schematic of the cortical surface in three gross orientations, with MMP regions [56]. outlined in silver and functional complexes relevant to processing body features (see Tables 1 and 2, Methods) colored in based on CAB-NP network affiliations ([25], Fig 1C) and identified with text labels and arrows. EBA = extrastriate body area; FBA = fusiform body area. VIS2 = secondary visual network; DAN = dorsal attention network. “r” denotes right hemisphere; “l” denotes left hemisphere. (C) Functional complexes related to face processing. FFA = fusiform body area; pSTS = posterior superior temporal sulcus; LAN = language network; DMN = default mode network. (D) Functional complexes related to places and scenes. RSC = retrosplenial complex; PPA = parahippocampal place area. (E) Functional complexes related to processing images of tools and objects. LOC = lateral occipital complex.

https://doi.org/10.1371/journal.pcbi.1012507.g004

We tested the hypothesis that fully distributed activity flow processes (Fig 1A) – shaped by the intrinsic functional connectivity fingerprint of each functional complex (Fig 4) – are sufficient to generate category-selective responses without the need for additional within-region processing. Before this, we began with two confirmations essential for subsequent tests of this hypothesis. We first constrained prior activity flow mappings [27,28] to body, face, place, and tool conditions and compared actual versus activity-flow-mapped responses across the whole cortex (S1 Fig). We then focused on activations in each functional complex (Fig 4) and in each hemisphere and quantified “benchmark” contrasts to corroborate the literature by comparing the category of interest versus non-category responses (e.g., body versus non-body) in the actual and activity-flow-mapped data (note that category selectivity is reported subsequently). Whole-cortex, activity-flow-mapped responses to visual categories were highly accurate (bodies: r = 0.89, MAE = 5.27, R² = 0.78; faces: r = 0.86, MAE = 5.83, R² = 0.72; places: r = 0.88, MAE = 5.85, R² = 0.77; tools: r = 0.89, MAE = 5.62, R² = 0.78; S1 Fig). Significant benchmark contrasts in each functional complex were also observed in both the actual (bodies: t(175) = 26.65, Cohen’s d = 2.01, faces: t(175) = 20.37, Cohen’s d = 1.54; places: t(175) = 39.53, Cohen’s d = 2.89; tools: t(175) = 8.49, Cohen’s d = 0.64; all p < 0.0001) and activity-flow-mapped data (bodies: t(175) = 17.91, Cohen’s d = 1.35; faces: t(175) = 10.06, Cohen’s d = 0.76; places: t(175) = 26.06, Cohen’s d = 1.88; tools: t(175) = 9.94, Cohen’s d = 0.75; all p < 0.0001) (S2 Fig) (left hemisphere and replication: S1 and S2 Tables). In each set of benchmark contrast analyses, multiple comparisons correction was performed with nonparametric permutation testing (i.e., the “max-T” procedure described in the Methods; here, with 10,000 permutations for each set of analyses).

Distributed activity flowing over intrinsic connectivity generates localized body category selectivity.

Processing of human body images (Fig 4A) is associated with the fusiform body area (FBA) and the extrastriate body area (EBA) [2,3,73], which are located in the lateral occipito-temporal cortex (Tables 1 and 2 and Fig 4B). EBA computations are thought to discriminate between bodily attributes and parts [156] and provide postural information to frontoparietal regions [74], indicating a pivotal role in action planning. The FBA is thought to process images of whole bodies [3] and pairs of body parts [157]. Extending the fully distributed hypothesis to the processing of body images, we hypothesized that activity flowing via the intrinsic connectivity fingerprint of the EBA and FBA (henceforth: EBA/FBA) (Fig 5A and 5B) — its unique pattern of distributed cortical connections—is likely the primary determinant of its body-selective responses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Distributed activity flows account for the majority of localized selectivity to bodies.

(A) The activity flow mapping procedure (steps match Fig 2C) generating body responses in EBA/FBA (black), projected onto cortical schematics (right hemisphere). Green: source vertices excluded from analyses. Step 2 was not blacked out for visual comparison with step 4, however it was held out in-analysis. Step 4 color scale shows maximum of all regions’ mapped activations to body images for visual comparison with step 2. (B) The connectivity fingerprint [13,15] of the right EBA/FBA via rsFC (black lines). Radial lines: source regions connected to EBA/FBA, clustered by functional network assignments [25] (colored per legend; Fig 3C). 95% confidence interval: across participants. (C) Right EBA/FBA body selectivity: activity-flow-mapped (purple) and actual (coral). Gray dots: individual participants’ scores. Dashed line: no selectivity (1.0); used for comparison in statistical tests. Significant t-statistics are indicated with an asterisk (p<0.00001; see Methods). (D) Estimated contribution of distributed network interactions to body selective responses in EBA/FBA. (E) The activity flows (as in A3) of each source region contributing to mapped EBA/FBA responses to body images (statistical significance asterisks at network-mean level). (F) Variance explained by each network-restricted activity flow model (unmixed partial R² in gray) of EBA/FBA’s response profile. Black lines: 95% confidence interval across participants. Asterisks: statistical significance versus each other network. E-F suggests that activity flowing over interactions with VIS2 and DAN represents a general network coding mechanism for EBA/FBA. See Methods for full details of each analysis.

https://doi.org/10.1371/journal.pcbi.1012507.g005

We assessed body selectivity in the EBA/FBA—quantified as the ratio of body to non-body category activations—in both activity-flow-mapped (Fig 5C, purple) and actual (Fig 5C, coral) data. As expected, mapped and actual body selectivity were statistically significant relative to a selectivity ratio of 1.0 (i.e., greater than the null hypothesis of equivalent responses to bodies and non-bodies; p < 0.00001; see Methods): activity-flow-mapped mean body selectivity = 1.35, t(166) = 16.59, Cohen’s d = 1.29; actual mean body selectivity = 1.67, t(166) = 20.01,Cohen’s d = 1.55; Fig 5C). We next estimated the contribution of distributed activity flow processes to body selectivity via the percentage of actual body selectivity captured by mapped body selectivity (Fig 5D). Given that the fully distributed activity flow approach maps task-evoked activity based on distributed features ([27]; Fig 2B), this ratio approximates the extent that distributed network interactions contribute to the generation of body selectivity in the EBA/FBA (relative to local, within-complex processes and/or error). As expected, the estimated contribution of distributed activity flow processes was 81%, which was statistically greater than 50% (p < 0.00001; see Methods) (Cohen’s d = 2.52, t(166) = 32.49, Fig 5D) (left hemisphere and replication: S3 Table).

Next, we examined how each large-scale functional network (Fig 3C [25]) contributed to the observation that activity flowing over a fully distributed intrinsic network architecture (connectivity fingerprint: Fig 5B) shaped body selectivity in the EBA/FBA. The activity flows map (Fig 5A, step 3) — where source activity was weighted by rsFC with the (held out) target region—was an entry point to assess such network contributions. Firstly, we averaged the estimated activity flows based on network assignment (Fig 5E). Secondly, we used dominance analysis to identify each functional network’s unmixed contribution (partial R², see Methods) to the activity-flow-mapped response profile (i.e., activations across all 24 conditions, Fig 3D) of the EBA/FBA (Fig 5F). Secondary visual network (VIS2) activity flows exhibited the largest contribution (versus each other network) to EBA/FBA’s body-evoked activations (Fig 5E, all t-statistics above the max-T threshold (175) = 3.39, p < 0.0001; left hemisphere and replication: S4 Table). DAN activity flows were also significantly higher (same max-T thresholds, see Methods) than the other networks, except VIS2 and posterior multimodal network (PMM). VIS2 also accounted for most of the explained variance in EBA/FBA’s response profile (partial R² = 0.51, 61% of the full model R², Fig 5F). DAN accounted for the next highest amount of explained variance (partial R² = 0.09, 10% of the full model R², Fig 5F; left hemisphere and replication: S5–S7 Tables). Thus, for both body-specific conditions and across all conditions (i.e., response profile), activity flowing over VIS2 and DAN exhibited substantial contributions to EBA/FBA activation patterns.

As hypothesized by the fully distributed network interaction model (Fig 1A), we found evidence that activity flowing over resting-state functional connections distributed throughout cortex shaped body selectivity in the EBA/FBA. Results also suggested that distributed processes chiefly specified localized body-selective responses in the EBA/FBA. The total explained variance in the EBA/FBA’s 24-condition response profile was significantly greater than 50% (total R² = 0.83, versus 0.5: t(175) = 66.07, p = 1.14 x 10⁻¹²⁵; left hemisphere and replication: S5–S7 Tables), suggesting that distributed activity flow processes predominantly influence EBA/FBA responses across a variety of cognitive domains.

Distributed activity flowing over resting-state connectivity generates localized face category selectivity.

The fusiform face area (FFA) is a hallmark region exhibiting localized category selectivity given its well-replicated relationship with face processing [4, 51]. The posterior superior temporal sulcus (pSTS) is thought to be a multisensory processing region responsive to the changeable aspects of faces [21,158]. The FFA and pSTS are loci of specialization in a two-pathway model [89,93], representing facial identity (relatively invariant) and facial expression (dynamic), respectively [11,94]. While both the FFA and pSTS are part of the ventral temporal cortex, the pSTS is more dorsal and lies adjacent to the temporoparietal junction; the FFA is more posterior and typically identified on the ventral aspect of the cortex (i.e., the fusiform gyrus) (Tables 1 and 2 and Fig 4C). Extending the fully distributed hypothesis (Fig 1A) to the processing of face images, we hypothesized that activity flow processes via the whole-cortex, resting-state connectivity fingerprint of the FFA and pSTS (Fig 6A and 6B) (henceforth: FFA/pSTS) is sufficient to determine its face selective responses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Distributed activity flows account for the majority of localized selectivity to faces.

All figure specifications follow Fig 5. (A) The activity flow procedure mapping activations to face categories, projected onto cortical schematics (right hemisphere). Right FFA/pSTS was the held-out target complex. (B) The connectivity fingerprint [13,15] of the right FFA/pSTS via whole-cortex rsFC (black lines). Radial lines: source regions connected to the FFA/pSTS, clustered by functional network assignments [25] (colored per legend and Fig 3C). (C) Face category selectivity exhibited by the right FFA/pSTS. Significant t-statistics are indicated with an asterisk (p<0.00001; see Methods). (D) Estimated contribution of distributed activity flow processes to face selectivity exhibited by the right FFA/pSTS. (E) Activity flows (as in A step 3) of each source region contributing to the mapping of FFA/pSTS responses to face images. VIS2 regions contributed most to the FFA/pSTS mapped activation magnitude to faces. (F) Variance explained by each network-restricted activity flow model (unmixed partial R² via dominance analysis; Methods) of the right FFA/pSTS’ response profile. VIS2 accounted for the most variance, altogether suggesting that activity flowing over VIS2 regions represents a general network coding mechanism for FFA/pSTS processing. DAN and DMN regions also accounted for a nontrivial amount of variance at the response-profile level suggesting that, across diverse cognitive domains, FFA/pSTS processing is impacted by activity flowing over DAN and DMN regions, in addition to VIS2 (in the face-specific case).

https://doi.org/10.1371/journal.pcbi.1012507.g006

We followed the foregoing pipeline for body selectivity exactly, except analyzing face images and the FFA/pSTS (Fig 6A, black regions). Overall, as expected, the pattern of results observed for body processing (EBA/FBA) extended to face processing (FFA/pSTS). Significant face selectivity was observed in both the actual and activity-flow-mapped data (mapped mean face selectivity = 1.33, t(170) = 10.78, Cohen’s d = 0.83; actual mean face selectivity = 1.37, t(170) = 11.67, Cohen’s d = 0.89; p < 0.00001; Fig 6C), and the estimated contribution of distributed activity flow processes was 96% (t(170) = 29.61, Cohen’s d = 2.27, p < 0.00001; Fig 6D) (left hemisphere and replication: S3 Table). We found that face activations in the FFA/pSTS were most influenced by activity flows over VIS2 connections (max-T threshold (175) = 3.39, p < 0.0001; Fig 6E; left hemisphere and replication: S4 Table). VIS2 also accounted for the most variance in FFA/pSTS processing at the response profile level (40%; left hemisphere and replication: S5–S7 Tables).

These results suggest that face selectivity in the FFA/pSTS was significantly shaped by activity flow processes over a fully distributed intrinsic network architecture. These processes were most prominently influenced by activity flowing over VIS2, similar to the EBA/FBA. Across all conditions, activity flowing over dorsal attention (DAN) and default mode (DMN) network connections were predictive as well (DAN: 11%, DMN: 9%). The total explained variance in this response profile model (Fig 6F) was R² = 0.89 (versus 0.5: t(175) = 92.61, p = 1.32 x 10⁻¹⁵⁰; left hemisphere and replication: S5–S7 Tables), suggesting that distributed processes chiefly influenced FFA/pSTS responses across many cognitive domains.

Distributed activity flowing over intrinsic connectivity generates localized place category selectivity.

We next focused on place images (Fig 4A), sometimes termed scenes, environment, or topography. Place-specific regions include the parahippocampal place area (PPA) and the retrosplenial cortex (RSC) (Tables 1 and 2 and Fig 4D), which are thought to act cooperatively toward a cohesive percept [5,99,103]. The PPA is thought to compute viewpoint-specific discrimination [159], as well as mediating (or binding) contextual associations pertinent for episodic memory [160]. The RSC is thought to provide the medial temporal lobe with visuospatial information [161], and to integrate viewpoint-invariant information for navigation and learning [5]. The PPA consisted of previously identified regions [53], and the dorsal RSC corresponded to Brodmann areas 29 and 30 [162] (Fig 4D). Extending the fully distributed hypothesis (Fig 1A) to the processing of place images, we hypothesized that activity flow via the whole-cortex, resting-state connectivity fingerprint of the PPA and RSC (Fig 7A and 7B) (henceforth: PPA/RSC) significantly determines its place-selective responses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Distributed activity flows account for the majority of localized selectivity to places.

All formatting and figure specifications follow Fig 5 (A) Activity flow mapping of activations to place categories, projected onto cortical schematics (right hemisphere only). Right PPA/RSC was the held-out target complex. (B) The connectivity fingerprint of the right PPA/RSC, as in Passingham et al. [13] except with whole-cortex rsFC (black lines). Radial lines: source regions connected to the PPA/RSC, clustered by CAB-NP functional network assignments [25] (colored per legend and Fig 3C). (C) Place category selectivity exhibited by the PPA and RSC in the right hemisphere. Significant t-statistics are indicated with an asterisk (p<0.00001; see Methods). (D) Estimated contribution of distributed activity flow processes to the emergence of place selective responses in the right PPA/RSC. (E) Activity flows (as in A step 3) of each source region contributing to the mapping of PPA/RSC responses to place images. VIS1, VIS2 and DAN contributed most to the right PPA/RSC mapped activation magnitude to place categories. Note that VIS1, VIS2, and DAN were all statistically greater than each other network, except for each other. (F) Variance explained by each network-restricted activity flow model (partial R² via dominance analysis; Methods) of the right PPA/RSC’s response profile. VIS2, DAN, and DMN accounted for the most variance, suggesting that activity flowing over regions in these networks represents a general network coding mechanism for PPA and RSC processing, while VIS1 contributes to place-specific responses.

https://doi.org/10.1371/journal.pcbi.1012507.g007

Significant place selectivity was observed in the PPA/RSC (mapped mean place selectivity = 1.8, t(170) = 19.15, Cohen’s d = 1.49; actual mean place selectivity = 2.59, t(170) = 22.42, Cohen’s d = 1.74; p < 0.00001; Fig 7C), and the estimated contribution of distributed processes was 69% (versus 50%: t(170) = 14.87, Cohen’s d = 1.15, p < 0.00001; Fig 7D) (left hemisphere and replication: S3 Table). Activity flowing over VIS1, VIS2, and DAN provided the largest contributions to PPA/RSC’s place activations (all significant except when compared to each other: max-T threshold (175) = 3.4, p < 0.0001; Fig 7E) (left hemisphere and replication: S4 Table). A similar set of networks accounted for the most variance at the response profile level (Fig 7F), including VIS2 (44%), DAN (13%) and DMN contributions (21%) (left hemisphere and replication: S5–S7 Tables).

As hypothesized by the fully distributed model of Fig 1A, the whole-cortex connectivity fingerprint of the PPA/RSC significantly shaped its place selectivity. PPA/RSC responses were influenced by activity flowing over VIS2 and DAN. Additionally, VIS1 was particularly important for place-specific activity; and DMN for cross-domain activity (Fig 7E and 7F), suggesting that PPA/RSC’s activity flow processes were more heterogeneous than prior models. The total variance in the PPA/RSC response profile explained by activity flow processes (Fig 7F) was greater than 50% (total R² = 0.68; t(175) = 23.93, p = 4.51 x 10⁻⁵⁷; left hemisphere and replication: S5–S7 Tables), which provides evidence that distributed processes predominantly influenced PPA/RSC activations to a variety of cognitive domains.

Distributed activity flowing over intrinsic connectivity generates localized tool category selectivity.

The final visual category included tool images (Fig 4A), sometimes termed inanimate objects. Following an extensive literature on object recognition [54,163], we hypothesized that tool selectivity is exhibited by the lateral occipital complex (LOC) [21,164], which is posteriorly located and wraps around the cortex in the ventromedial direction (Fig 4E). The LOC is thought to represent higher-level information of objects, as opposed to low level visual features [165], suggesting a role at the category level of visual processing. However, reports vary on the degree of semantic content processed by the LOC [166]. Thus, the link between the LOC and tool selectivity was the least clear (of all four models) a priori. Extending the fully distributed hypothesis (Fig 1A) to the processing of tool images, we hypothesized that activity flow via the whole-cortex, resting-state connectivity fingerprint of the LOC (Fig 8A and 8B) significantly determines its tool selective responses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Distributed activity flows account for the majority of localized selectivity to tools.

All formatting and figure specifications as in Fig 5. (A) Activity flow mapping of activations to tool categories in the held-out target—the right LOC—projected onto cortical schematics (right hemisphere). (B) The connectivity fingerprint of the right LOC, as in Passingham et al. [13] except with whole-cortex rsFC (black lines). Radial lines: source regions connected to the LOC, clustered by CAB-NP functional network assignments [25] (colored per legend and Fig 3C). (C) Tool category selectivity exhibited by the LOC in the right hemisphere. Significant t-statistics are indicated with an asterisk (p<0.00001; see Methods). (D) Estimated contribution of distributed activity flow processes to the emergence of tool selective responses in the right LOC. (E) Activity flows (as in A step 3) of each source region contributing to the mapping of LOC responses to tool images. VIS2 contributed most to the right LOC mapped activation magnitude to tool categories (F) Variance explained by each network-restricted activity flow model (partial R² via dominance analysis; Methods) of the right LOC’s response profile. VIS2 accounted for the most variance, suggesting that activity flowing over regions in these networks represents a network coding mechanism for LOC processing.

https://doi.org/10.1371/journal.pcbi.1012507.g008

Tool selectivity in the LOC was statistically significant (mapped tool selectivity = 1.12, t(174) = 11.65, Cohen’s d = 0.88; actual tool selectivity = 1.15, t(174) = 12.52, Cohen’s d = 0.95; p < 0.00001; Fig 8C), and the estimated contribution of distributed activity flow processes to tool selectivity in the LOC was particularly high at 97% (t(174) = 111.44, Cohen’s d = 8.45; Fig 8D) (left hemisphere and replication: S3 Table; given evidence in the literature that tool selectivity is left lateralized, future work building on activity-flow-mapped tool selectivity should pay special attention to the tool processing results reported in S2 Table through S15 Table). VIS2 activity flows demonstrated a strikingly high contribution to LOC’s tool-specific responses (max-T threshold (175) = 3.4, p < 0.0001; Fig 8E) (left hemisphere and replication: S4 Table) and accounted for the majority of the variance in LOC’s response profile (79%; Fig 8F) (left hemisphere and replication: S5–S7 Tables).

As hypothesized, tool selectivity was strongly influenced by distributed activity flow processes (97–98%). As in all other models (bodies, faces, and places) the total explained variance that activity flow mapping captured for the LOC response profile was greater than 50% (total R² = 0.90 vs 0.5: t(175) = 115.55, p = 3.71 x 10⁻¹⁶⁷). Given that the standard application of activity flow mapping [27,28] (i.e., following the fully distributed network interaction scheme of Fig 1A) is a distributed processing model (Fig 2B), this suggests that distributed processes may be the primary mechanism generating category-selective LOC activations.

Control analyses

Null connectivity fingerprints reduce visual category selectivity.

To corroborate that each functional complex’s unique whole-cortex resting-state connectivity fingerprint (Figs 5B–8B) — its placement in the brain’s intrinsic network architecture—determined its visual category selectivity (Figs 5C–8C), we built null models based on FC substitution. We hypothesized that a null connectivity fingerprint – here defined as a fingerprint based on the wrong functional complex – would confer significantly lower activity-flow-mapped category selectivity.

As hypothesized, visual category selectivities were significantly greater when based upon the true (i.e., empirically-based) whole-cortex connectivity fingerprints (Table 3, reporting right hemisphere functional complexes and the discovery dataset). This assessment is more stringent than a null model that randomly scrambles FC because there is likely some similarity in visual processing (and therefore activity flow patterns) across the four functional complex’s fully distributed network interactions. However, no substituted connectivity fingerprint was sufficiently similar to the true fingerprint to generate the original mapped category selectivity. Each of these results were corroborated in the left hemisphere and in the replication dataset (S8 and S9 Tables). These results corroborate that the fully distributed, resting-state connectivity fingerprint unique to each functional complex significantly shaped its visual category selective response. Given that connectivity fingerprints are the bases of activity flow processes (Fig 1A, Formula 1) that were used to generate mapped category selectivity, these results further support the proposition that fully distributed network interactions are sufficient to generate localized visual category selectivity in all four functional complexes tested.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Empirical (true) resting-state connectivity fingerprints generate visual category selectivity significantly better than null model connectivity fingerprints.

https://doi.org/10.1371/journal.pcbi.1012507.t003

Activity-flow-mapped category selectivity was not driven by category responsiveness in other parts of the brain.

A potential confound in our analysis pipeline was that the criteria for identifying functional complexes (see Methods) excluded some brain regions that also exhibit a degree of visual category selectivity and/or responsiveness (Fig 9A). Therefore, it is possible that the findings in the preceding sections – that fully distributed activity flow processes are capable of generating localized visual category selectivity – were driven by these other, excluded regions (which remained as sources of task-evoked activity). To address this, we held out select regions (Fig 9A) from the source sets of each visual category selectivity analysis and compared such source-set-controlled findings to the fully-distributed model findings (Figs 1A and 5–8). Individual differences in selectivity were preserved across all visual categories (Fig 9B), suggesting that activity flow processes were capable of generating visual category selectivity in each functional complex without the influence of other visual category responsive regions. Note that the following sections on stimulus-driven network interactions more fully address this circularity concern, while also testing refined versions of our core hypothesis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Other regions associated with visual category responses in the literature do not drive the fully-distributed model results.

(A) While our procedure for identifying functional complexes from the literature was systematic (see Methods), it was possible that excluded, category-responsive regions were driving the results in Figs 5–8. Such regions were excluded from functional complexes because they were either: (1) not supported by 10 or more peer-reviewed studies at the time of preparing this manuscript, (2) studies used experimental stimuli too distinct from the n-back visual categories, or (3) studies did not provide spatial information systematically consistent with standard volumetric and/or surface-based topography. These regions, held-out from source sets in this control analysis, were as follows (from left to right in panel A): the dorsal visuomotor stream for body selectivity; the occipital face area (OFA), inferior frontal gyrus (IFG), and orbitofrontal cortex (OFC) for face selectivity; the transverse occipital sulcus (TOS) for place selectivity; the intraparietal sulcus (IPS) and middle frontal gyrus (MF) for tool selectivity. Colors of regions are consistent with the functional network assignment used throughout all other figures. (B) For each body, face, place, and tool (left to right) selectivity analysis, individual participant’s (dots) visual selectivity scores were maintained between the results in Figs 5–8 (x-axes) and the same analyses with the regions in panel A held-out from the source set. This suggests that other category responsive regions in the literature did not drive the fully-distributed model findings.

https://doi.org/10.1371/journal.pcbi.1012507.g009

Stimulus-driven activity flow processes directly from V1 are sufficient for generating category selectivity

We next sought to test the refined hypothesis that stimulus-driven activity flowing over a functional complex’s connectivity fingerprint with V1 only is sufficient to generate localized visual category selectivity (Figs 1B and 10A). This does not negate the “fully distributed” hypotheses tested above (Fig 1A), but instead specifies that processes instantiated in V1 are key to our prior observations that fully distributed activity flows are sufficient for generating localized visual category selectivity. Alternatively, it is possible that network interactions with V1 account for only a small proportion of variance in visual category selectivity generated by the fully distributed network interaction models (Figs 5–8). For example, it remains possible that the category selectivities observed in the fully distributed model were generated via (1) local (e.g., within-region) computations in regions outside the functional complexes, which (2) together flow into the functional complexes, concentrating selectivity within the complexes. Another possibility is that visual category selectivity is only minimally influenced by direct stimulus-driven inputs from V1, given evidence that visual information processing (which may not be strictly hierarchical; see [167, 168]) is modulated by attention [169–171], prior expectations [172], task goals/context [173–175], and synthesis-related processes (i.e., continuously maintaining a percept from a noisy or ambiguous visual scene) [176–178]. However, it is unclear if these top-down modulations (and other non-visual contributions, such as emotional modulation; as in [179]) are indeed just modulations, or requirements for visual category selectivity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Activity flows directly from V1 are sufficient to generate visual category selectivity.

(A) Theoretical schematic of stimulus-driven activity flow processes generating visual category selectivity (as in Fig 1B). Given prior literature, mapped activity flow processes (gray arrow) have a refined inference: from V1 to later visual regions, we inferred that activity flow processes were primarily stimulus driven. This contrasts with the prior whole-cortex models, which also included top-down and likely recurrent influences. (B) Activity flow mapping procedure for the stimulus-driven model, conducted at the vertex level. V1 sources were used to map targets across VIS. Note that the usage of “step 1” serves as a prelude to later steps tested in an extended visual system model (Fig 11). (C) The null connectivity fingerprint (rsFC) model used for control analyses. The top depicts the true VIS network, and the bottom depicts pseudo-random (edge degree and strength preserved) network architectures over 100 permutations. For visualization purposes, networks are shown at the region level, but analyses were conducted at the vertex level. (D) Actual (coral) and mapped (purple) visual category selectivity exhibited by the right EBA/FBA, FFA/pSTS, PPA/RSC, and LOC (left to right). Category selectivity exhibited by V1 (for each respective image category) is shown to demonstrate that activation patterns in V1 alone do not account for mapped visual category selectivities (V1 selectivity scores were all nonsignificant; see main text for statistics). Most importantly, these results demonstrate that the activity flow mapping process increased the category selectivity of every functional complex relative to V1 activity patterns, despite source activity originating solely from V1. Category selectivity generated by V1-initialized activity flow processes were significant in each functional complex (see main text for statistics). Significant t-statistics are indicated with an asterisk (p<0.00001; see Methods); see Figs 5–8 for significance of actual category selectivity of functional complexes.

https://doi.org/10.1371/journal.pcbi.1012507.g010

We additionally sought to address a theoretical limitation borne out of whole-cortex, undirected network estimation from fMRI data. Namely, a limited ability to isolate whether activity flow processes contributing to a given complex’s category-selective response also contain some information from that complex itself, or from other brain regions responsive to visual categories. For example, do fully distributed activity flows generating FFA/pSTS face selectivity (Fig 6) contain some information (potentially via feedback from areas downstream of FFA/pSTS) from the FFA/pSTS itself? This type of causal circularity is a theoretical concern for the ongoing goal of delineating the extent that network interactions contribute to the generation of localized processes with refined detail. However, given that regions within each functional complex (e.g., the FFA and pSTS for the FFA/pSTS complex) were excluded as activity flow sources, the most likely culprits of this sort of feedback circularity were already accounted for. Thus, it is likely that the fully distributed network models (Figs 5–8) accurately reflect the generation of category selectivity, just at a broad scale and with potential inferential limits.

To these ends, we developed an intrinsic-connectivity-based, generative model of category-selective responses that reduces the impact of (potential) causal circularity by constraining activity flows to V1 (sources) and functional complexes (targets) (Fig 10A and 10B). Given robust evidence that activity in V1 represents retinotopic mapping and simple visual feature detection [42,43], this model better ensures that mapped activation patterns are primarily stimulus-driven. We first estimated rsFC between all vertices in the visual system (including both the VIS1 and VIS2 networks and all functional complexes) (Fig 10C, top). We will henceforth refer to this sub-system as “VIS”. In the following steps of this analysis, we focused on the interactions between V1 sources and functional complex targets, but it was important to include the entire visual system in the initial rsFC estimation step. This allowed us to infer – based on our use of multiple regression for rsFC estimation – the likely direct connectivity between each V1 vertex and each functional complex vertex [27, 127].

We found that for each functional complex, stimulus-driven, activity-flow-mapped category selectivity was significantly above 1.0 (p < 0.00001 for all complexes; see Methods): right hemisphere EBA/FBA mean body selectivity = 1.12, t(174) = 5.64, Cohen’s d = 0.45); right FFA/pSTS mean face selectivity = 1.14, t(172) = 5.07, Cohen’s d = 0.41); right PPA/RSC mean place selectivity = 1.13, t(173) = 5.27, Cohen’s d = 0.42); right LOC mean tool selectivity = 1.07, t(172) = 5.41, Cohen’s d = 0.3) (Fig 10D; left hemisphere and replication dataset: S10 Table). In each of these four models, mapped category selectivity was statistically significant, suggesting direct activity flows from V1 (the primary visual input to cortex) are sufficient to produce selectivity for these four visual categories. Note, however, that selectivity in these four models was lower than when mapped via network interactions of the whole cortex. This suggests that additional selectivity is generated via distributed activity flow processes over the whole cortex (Figs 5C–8C).

As in prior control analyses (Table 3), we tested whether each functional complex’s unique resting-state connectivity fingerprint (here, with V1) determined its visual category selectivity by using a null rsFC model. Note that V1 itself did not exhibit significant selectivity for any of the four visual categories (V1 body selectivity = 0.63, t(175) = -17.3, p = 1, Cohen’s d = -1.32; face selectivity = 0.87, t(175) = -12.9, p = 1, Cohen’s d = -0.99; place selectivity = 0.64, t(175) = -13.6, p = 1, Cohen’s d = -1.04; tool selectivity = 0.66, t(175) = -12.8, p = 1, Cohen’s d = -0.97) (Fig 10D; replication dataset: S11 Table), thus downstream selectivity could not be driven by activation patterns in V1 without additional connectivity-based transformation (also supported by [115,180–182]). Given the varied number of vertices in each complex, we used randomly permuted connectivity architectures (Fig 10C) – maintaining edge strength and degree [136,153,154] – instead of rsFC substitution. Body, face, and place selectivity were significantly greater than the aggregate of 100 null rsFC models (body: t(175) = 6.4, p = 6.6 x 10⁻¹⁰, Cohen’s d = 0.49; face: t(175) = 11.6, p = 8.2 x 10⁻²⁴, Cohen’s d = 0.89; place: t(175) = 3.5, p = 2.5 x 10⁻⁴, Cohen’s d = 0.27), but not tool selectivity (t(175) = -1.15, p = 0.8, Cohen’s d = -0.09) (left hemisphere and replication dataset: S12 Table). This suggests that stimulus-driven category selectivity exhibited by EBA/FBA, FFA/pSTS, and PPA/RSC were significantly shaped by their unique connectivity patterns with V1.

These results suggest that direct stimulus-driven activity flows from V1 are a key step in the generation of visual category selectivity in visual cortex (Fig 1B). For some functional complexes, such as the LOC, early information patterns across V1 can select category-selectivity-generating activity flow processes in a manner less dependent on its connectivity fingerprint (but note that LOC’s connectivity fingerprint with V1 appears more critical in the whole-cortex model; Table 3). For other complexes such as the FFA/pSTS, the stimulus-driven account of category selectivity appears to be influenced to a large extent by its connectivity fingerprint with V1, given its strong statistical significance over the null rsFC model.

Multi-step activity flows across the visual system improves response profile accuracies, but not category selectivity

Given evidence that higher-level visual representations (such as category selectivity) involve information processing along the ventral visual stream (Fig 11A) [88,183,184], we sought to extend our stimulus-driven model and test the hypothesis that category selectivity mappings (which were significant but below empirically-observed levels in the V1 model) improve when accounting for all VIS network interactions (Fig 1C). We conducted this analysis using the mapped activation patterns from the stimulus-driven V1 model (Figs 10B and 11B) – weighted by rsFC across VIS – to model later visual cortex activity flow processes (Fig 11B, Step 2). This step was repeated to model potential bidirectional (i.e., reciprocal) and/or recurrent processes within the visual system [168,184] (Fig 11B, Steps 3+), until a settling threshold was reached (i.e., where mapped activation patterns no longer changed, Fig 11C; also see Methods).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 11. Visual category selectivity generated by activity flow processes over the extended visual system.

(A) Stimulus driven activity flow processes, further shaped by later visual interactions, generate localized visual category selectivity (Fig 1C). From V1 to later visual regions, we inferred that activity flow processes are stimulus driven (Figs 1B and 10). Within the visual system, we inferred that activity flow processes are bidirectional and/or recurrent. (B) Activity flow mapping procedure for the extended visual system model. All steps were conducted at the vertex level. Step 1: V1 sources were used to map targets across the visual cortex (VIS) (Fig 10). Step 2: mapped VIS activation patterns from Step 1 (weighted by connectivity estimates as in all activity flow models; see Methods) were used as sources to map held-out targets across visual cortex. Steps 3+: step 2 was repeated until a settling threshold was reached – or the point at which mapped values stopped changing (see Methods). (C) The settling threshold was reached at step 3. All further analyses only included steps 1–3 from panel B. (D) Accuracy of mapped activation patterns across all conditions (left: an average of all visual regions; right: functional complexes studied herein). Across visual cortex, explained variance tended to increase with each step, indicating that the extended visual system model was improving mapped response profiles across the cortex. This pattern was also observed across functional complexes, with some exceptions, such as the PPA/RSC (which appears most accurate at step 2). (E-H) Actual (coral) and visual-system-mapped (purple) category selectivity (see Methods) exhibited by the right EBA/FBA (E), FFA/pSTS (F), PPA/RSC (G), and LOC (H). Mapped category selectivity is shown for each step in panel B. Activity-flow-mapped category selectivity estimates were not consistently significant (see main text for statistics), suggesting that extending V1-initialized activity flow processes across the visual system alone does not stably generate localized visual category selectivity. In panels E-F: significant t-statistics are indicated with an asterisk (p<0.00001; see Methods); see Figs 5–8 for significance of actual category selectivity of functional complexes.

https://doi.org/10.1371/journal.pcbi.1012507.g011

Cross-condition (i.e., response profile) mapping accuracy (of all VIS vertices) increased with each step (Fig 11D), suggesting that this extended model improved the specification of visual system responses. However, category selectivity was not improved consistently across functional complexes of interest. Mapped body selectivity in the right EBA/FBA was improved in step two (body selectivity = 1.18, t(170) = 9.49, p < 0.00001, Cohen’s d = 0.75) but became nonsignificant in step three (body selectivity = 1.01, t(174) = 0.58, Cohen’s d = 0.03) (Fig 11E). Face selectivity diminished in step 2 (face selectivity = 1.03, t(174) = 1.82, not significant, Cohen’s d = 0.14) and step 3 (face selectivity = 0.9, t(167) = -9.23, not significant, Cohen’s d = -0.71) (Fig 11F), suggesting that within the confines of the visual system, stimulus-driven activity flows from V1 (step 1) were the best predictors of right FFA/pSTS face selectivity. Place selectivity in the right PPA/RSC improved in step 2 (place selectivity = 1.28, t(172) = 16.9, p < 0.00001, Cohen’s d = 1.43), but decreased to the original level in step 3 (place selectivity = 1.15, t(171) = 11.5, p < 0.00001, Cohen’s d = 0.88) (Fig 11G). Tool selectivity in the right LOC improved in step 2 (tool selectivity = 1.28, t(170) = 15.85, p < 0.00001, Cohen’s d = 1.5), but decreased in step 3, albeit to a level higher than the initial V1 step (tool selectivity = 1.16, t(171) = 12.5, p < 0.00001, Cohen’s d = 1.23) (Fig 11H; left hemisphere and replication dataset: S13 and S14 Tables). Interestingly, tool selectivity generated by step 2 was greater than the actual tool selectivity exhibited by the LOC (Fig 11H, coral), suggesting that extended visual system modulations to activity flow processes initially over-specified tool selectivity, and that step 3 modeled the best-performing network interactions for mapping LOC tool selectivity. Thus, for two complexes (EBA/FBA and PPA/RSC), one additional step of extended visual system network interactions best specified category selectivity. For one complex (FFA/pSTS), the V1 stimulus-driven (step 1) model was best, and for another complex (LOC) three steps were best.

Therefore, even though mapped response profiles tended to improve when modeling bidirectional and/or recurrent processes across the visual system, visual-system-mapped category selectivities did not consistently improve beyond what was specified by stimulus-driven activity flows from V1. Additionally, given that all observations were most robust in the fully distributed model (Fig 1A) (including response profile accuracies, which explained a maximum of 45% variance in the extended visual model, Fig 11D, and up to 92% in the whole-cortex model, Fig 3D), the visual system is likely further modulated by other systems (e.g., DAN interactions; also supported by Keane et al. [185]) to provide the full set of distributed activity flow processes that stably generate localized category selective responses (as in Figs 5–8).

Adding fully distributed network interactions to stimulus-driven activity flows further enhances category selectivity

Given evidence that stimulus-driven network interactions with V1 generated visual category selectivity (Fig 10), but to a lesser extent than the fully distributed mappings (Figs 5–8), and that this inconsistently improved when extended to later visual system interactions (Fig 11), we next sought to extend the stimulus-driven model to fully distributed network interactions (i.e., a “stimulus-driven + fully distributed” mapping) (Figs 1D and 12A). As in prior analyses, fully distributed network interactions refer to activity flow processes over all cortical source regions (Formula 1). Note that the stimulus-driven + fully distributed model differs from the initial cortex-wide model (Fig 1A) by constraining inputs to V1 – reducing the chance for causal circularity (e.g., from recurrent feedback to each functional complex) and excluding possible temporally extended activity (in non-V1 regions) originating prior to stimulus onset (e.g., attentional or task set top-down biases). First, using the extended mapping approach as in Fig 11B, we initialized activity flow mapping with V1 network interactions, then incorporated all other network interactions in a second step (i.e., with regions beyond just VIS, see Methods for full details). Across all functional complexes, we found that activity-flow-mapped category selectivity was not only statistically significant but was also remarkably close to selectivity generated by fully distributed network interactions (Figs 5–8), and in some cases exhibited an improvement. Further, selectivity was significantly improved for all functional complexes relative to the stimulus-driven V1-initiated model, demonstrating the importance of additional brain-wide activity flows in increasing category selectivity beyond direct-from-V1 activity flows.

Body selectivity in the EBA/FBA was statistically significant (mapped body selectivity = 1.41, t(166) = 10.65, Cohen’s d = 0.83, p < 0.00001; Fig 12B). Here, activity-flow-mapped body selectivity generated by stimulus-driven + fully distributed network interactions was an improvement upon selectivity generated from V1 only (Fig 10) (related samples t-tests across participants’ activity-flow-mapped body category selectivity scores: t(175) = 4.75, p = 4.18 x 10⁻⁶). Category selectivity in the FFA/pSTS (mean mapped mean face selectivity = 1.26, t(165) = 8.44, Cohen’s d = 0.57, p < 0.0000; Fig 12C), the PPA/RSC (mapped place selectivity = 1.94, t(166) = 14.54, Cohen’s d = 1.0, p < 0.00001; Fig 12D), and the LOC (mapped tool selectivity = 1.09, t(168) = 6.5, Cohen’s d = 0.58, p < 0.00001; Fig 12E) were all also significantly improved from the stimulus-driven (V1 only) mappings (faces: t(175) = 3.41, p = 8.04 x 10⁻⁴; places: t(175) = 8.41, p = 7.05 x 10⁻¹⁵; tools: t(175) = 2.16, p = 0.03).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 12. Adding fully distributed network interactions to stimulus-driven activity flows further enhances category selectivity.

(A) A schematic of V1-initialized activity flow processes that are further propagated across all cortical network interactions (as in Fig 1D). Here, fully distributed network interactions are initially established by activity flowing over the connectivity fingerprints of each functional complex with V1. We used the multistep mapping procedure of Fig 11B (but with only two steps), except for at the region level (Methods). (B) Mapped (purple) and actual (coral) body selectivity exhibited by the right hemisphere EBA/FBA (gray dots: individual participants; boxplot line: median). Statistical significance is reported in the main text. (C-E) Same as B, but: (C) face selectivity in FFA/pSTS, (D) place selectivity in PPA/RSC, and (E) tool selectivity in LOC. Across all functional complexes, stimulus-driven + fully distributed mappings generated visual category selectivity that was greater than the stimulus-driven alone model of Fig 10 and closely matched (and in some cases improved upon) the fully distributed alone model of Figs 5–8. This suggests that activity flow processes initialized in V1, that are further processed over all cortical network interactions, are capable of generating highly accurate visual category selectivity, with reduced causal confounds (see main text for full rationale). Panels B-D: significant t-statistics are indicated with an asterisk (p<0.00001; see Methods); see Figs 5–8 for significance of actual category selectivity of functional complexes.

https://doi.org/10.1371/journal.pcbi.1012507.g012