Mesoscale hierarchical organization of primary somatosensory cortex captured by resting-state-fMRI in humans

Abstract

The primary somatosensory cortex (S1) plays a key role in the processing and integration of afferent somatosen-sory inputs along an anterior-to-posterior axis, contributing towards necessary human function. It is believed that anatomical connectivity can be used to probe hierarchical organization, however direct characterization of this principle in-vivo within humans remains elusive. Here, we use resting-state functional connectivity as a complement to anatomical connectivity to investigate topographical principles of human S1. We employ a novel approach to examine mesoscopic variations of functional connectivity, and demonstrate a topographic organisation spanning the region’s hierarchical axis that strongly correlates with underlying microstructure while tracing along architectonic Brodmann areas. Our findings characterize anatomical hierarchy of S1 as a ‘continuous spectrum’ with evidence supporting a functional boundary between areas 3b and 1. The identification of this topography bridges the gap between structure and connectivity, and may be used to help further current understanding of sensorimotor deficits.

1. Introduction

The primary somatosensory cortex (S1) is integral to the somatosensory system and important for many functions, such as tactile recognition (Drevets et al., 1995; Bensmaia et al., 2008; Pei et al., 2010), bodily perception (Kim et al., 2015), and motor control (Disamond et al., 2008; Lee et al., 2008). Hierarchical organization of S1 along the anterior-to-posterior axis supports these functions through sequential processing of afferent somatosensory inputs (Iwamura, 1998). A vast body of animal literature has supported the role of S1 integration for computations related to object localization (Kleinfeld and Deschênes, 2011) and texture decoding (Isett et al., 2018), and may have further implications in goal-directed somatosensory-related behaviours (Yamashita and Petersen, 2016). In humans, impairment to areas of S1 may lead to abnormal processing of somatosensory information and may contribute to sensorimotor related deficits commonly found in neurological disorders, such as stroke (Kim and Choi-Kwon, 1996), and Parkinson’s Disease (Conte et al., 2013). Despite the relevance of S1 in cognitive and clinical neuroscience, a method to characterize hierarchical organization of S1 in-vivo within human cortex remains elusive.

To date, magnetic resonance imaging (MRI) using myelin mapping techniques have made it feasible to delineate S1 into anterior-to-posterior architectonic subdivisions – Brodmann areas 3a, 3b, 1 and 2 - corroborating findings from cytoarchitecture (Brodmann 1909; Glasser and Van Essen, 2011; Fischl et al., 2008). While structural MRI can sufficiently delineate architectonic boundaries with one or more MRI contrasts, evidence from resting-state functional connectivity (RSFC) reveals further separation of S1 along its somatotopic boundaries following a ventral-to-dorsal axis (Yeo et al., 2011). In principle, measures of connectivity, in this case using macroscale RSFC patterns, should also align with architectonic S1 subdivisions along an anterior-to-posterior axis, as supported by non-human primate (NHP) anatomical studies (Krubitzer and Kaas, 1990; Pons and Kaas, 1986). However, a correspondence between mesoscale measures of structure and connectivity supporting notions of anatomical hierarchy on a millimetre scale have yet to be established in human S1. Here, mesoscale is defined as the spatial scale over which individual differences transition into species typicality.

Recent methodological developments have shown manifold learning as a viable tool to embed high dimensional RSFC data into a low dimensional space while preserving biologically meaningful structure, commonly referred to as RSFC gradients (Haak et al., 2018; Margulies et al., 2016). Macroscale gradients of the cerebral cortex have been shown to represent an embedding scheme that positions primary and transmodal cortices on opposite ends of a spectrum (Margulies et al., 2016). RSFC gradients estimated within specific cortical areas, referred to as ‘connectopic mapping’ (Haak et al., 2018), revealed biologically relevant interactions, such as linking hippocampus to microstructure (Vos de Wael et al., 2018), and striatum to goal-directed behaviours (Marquand et al., 2017). Additionally, this technique can be used to estimate multiple overlapping gradients, where each gradient may correspond to unique organizational principles of the cortical area. The aim of the current study was to extend the use of RSFC to investigate mesoscale hierarchical organization in human S1. We hypothesized that RSFC will enable characterization of S1 along an anterior-to-posterior axis in-vivo that spatially maps onto somatosensory Brodmann areas.

A central part of this investigation is to use accurate regions of interest (ROI) to characterize the anatomical hierarchy of S1 using RSFC. Here, we take advantage of the Human Connectome Project’s multimodal parcellation, which subdivides the cerebral cortex into 360 cortical areas, four of which represent S1’s architectonic subareas (i.e., Brodmann areas 1, 2, 3a, and 3b) (Glasser et al., 2016). Although not part of the parcellation, Glasser and colleagues also propose further subdivision of S1 into their somatotopic subareas that relate directly to distinct body parts (Penfield and Rasmussen, 1950). With the goal to discover a mesoscale hierarchical gradient in S1 we explore other overlapping gradients, and the dominant gradients of S1 somatotopic subareas to see whether evidence of RSFC heterogeneity along the anterior-to-posterior axis exists.

Given these considerations, we chose to use measures of RSFC to investigate structural organization of S1. RSFC provides a measure of function that closely follows principles guided by anatomical connectivity (Wang et al., 2013; Jbabdi et al., 2015). Additionally, RSFC can predict interindividual differences of task-based fMRI activations across a wide variety of cognitive paradigms, supporting the notion that cognitively relevant functional interactions are preserved at rest (Tavor et al., 2016). Here, we used resting-state data from the WU-Minn Human Connectome Project (Van Essen et al., 2013; Smith et al., 2013) and S1 architectonically-defined regions as a proxy for anatomical hierarchy taken from Glasser and colleagues multi-modal parcellation of the cerebral cortex (Glasser et al., 2016). To characterize mesoscale structural organization of S1, we explore the principal gradient derived using RSFC of S1 somatotopic subareas and (1) demonstrate its use as a proxy for anatomical hierarchy, (2) its link to underlying tissue microstructure, and (3) its correspondence to Brodmann area boundaries – here, we specifically found evidence for a distinct functional division that exists between Brodmann areas 3b and 1, rather than four architectonic subareas. Finally, we demonstrate the application of this gradient scheme as a means to achieve a more comprehensive characterization of human thalamocortical connectivity profiles based on what is known in NHP literature. A characterization of S1 topography that follows governing principles of anatomical hierarchy and microstructure may provide insight into studying the interplay between mesoscale structure and function in humans. In doing so, this may further offer an interpretative framework for studying sensorimotor-related deficits across a wide range of neurological disorders. Together, this work provides new insight into the use of RSFC to characterize mesoscale structural organization of human S1.

2. Methods

2.1. Resting-state fMRI dataset and preprocessing

This 3T resting-state dataset was taken from N = 100 unrelated participants from the Human Connectome Project (HCP) of the 1200-participant release (Van Essen et al., 2013). Each participant underwent four 14.4-min rs-fMRI scans (TR = 0.72 s) acquired across two days (two scans per day, acquired in opposite phase-encode direction [left-right/right-left]). Each scan was preprocessed through the HCP with details described in Smith et al., 2013, and includes spatial distortion, and head-motion correction, registration to a T1 weighted structural, resampling to a 2 mm Montreal Neurological Institute (MNI) space, global intensity normalisation, high-pass filtering (cut-off at 2000 s), and ICA-based artefact removal (FSL-FIX [Griffanti et al., 2014; Salimi-Khorshidi et al., 2014]). In addition to HCP minimal preprocessing, mean white matter and ventricular signal was regressed from the data, followed by smoothing with a 5 mm FWHM Gaussian kernel respecting the natural geometry of the brain. Specifically, surface-base smoothing was applied to the cortical ribbon, whereas volumetric-base smoothing was applied to the subcortex. All scans from a single subject were Z-score normalised to zero mean and unit standard deviation and concatenated into a single one-hour rs-fMRI scan. The concatenated rs-fMRI scan was used in subsequent connectopic mapping of contralateral S1 subregions.

2.2. Motion exclusion criteria

To mitigate biases from the effects of motion we only include participants who fell within our necessarily stringent motion exclusion criterion (Power et al., 2012). Specifically, participants who had a mean framewise displacement (FD) greater than 0.2 mm in any of their four scans were excluded from analyses. This resulted in N = 65 (FD; mean = 0.138 mm, SD = 0.023 mm) subjects used across all analyses.

2.3. Region of interest definitions of S1 anatomical hierarchy

All regions of interest were obtained from a surface-based multimodal parcellation of the cerebral cortex (Glasser et al., 2016). We used Brodmann areas 3a, 3b, 1, and 2, to define an anatomical hierarchy of S1. The Glasser atlas also proposes five somatotopic subregions within S1: lower limb (LL), trunk (T), upper limb (UL), ocular and face. For this current study, we only consider somatotopic regions that traverse all S1-related Brodmann areas to best illustrate anatomical hierarchy. As such, only LL, T and UL are included going forward (i.e., the ocular region includes only Brodmann area 3, and the face region includes areas 3 and 1). The UL and LL region were guided by left/right hand, and feet task contrasts, respectively, whereas, the definition of the trunk region was interpreted and localized between the UL and LL representations due to the absence of trunk-related contrasts.

In summary, each somatotopic subregion under consideration (i.e., LL, T, and UL) can be further delineated by its hierarchical structure (i.e., Brodmann areas 3a, 3b, 1, and 2) permitting investigations of anatomical hierarchy in S1. Human S1 hierarchy was annotated using a hierarchical organization scheme as proposed in macaques (Felleman and Van Essen, 1991) with supporting evidence found in humans (Bodegård et al., 2001).

The thalamus was also considered for seed-based connectivity analyses to explore and validate anatomical hierarchy of S1. Here, the thalamus was defined using the Harvard-Oxford atlas, and further parcellated into thalamic nuclei using the Morel atlas. The lateral geniculate nucleus (LGN) and inferior pulvinar (PuI) were excluded from all thalamocortical connectivity analyses due to the lack of overlap between the thalamic ROI from the Harvard-Oxford atlas and the Morel atlas. We considered thalamocortical connectivity ipsilaterally, therefore seed-based analyses of left hemisphere S1 regions only aimed to probe connectivity to the left thalamic nuclei, and vice versa.

2.4. Connectopic mapping

Connectopic mapping is a data-driven technique which can be used to spatially quantify RSFC patterns within a region of interest. Full details of the procedure are described in Haak et al. 2018 (only the Laplacian eigenmap technique used to embed voxel-wise RSFC patterns were used in the present study). In brief, RSFC was calculated between each ROI voxel (A_t×ν: t=time,ν=number of voxels in the ROI) and the rest of the brain (B_t×ν′: ν′=number of voxels in the rest of the brain (i.e., cortical and subcortical voxels)). To maintain computational tractability, the cortex and subcortex timeseries matrix (B_t×ν′) was projected onto a subspace spanned by its principal components $\left({\tilde{B}}_{t\times t-1}\right)$. A RSFC matrix (C_ν×t−1) was computed which describes an ROI’s RSFC pattern. Next, similarity between inter-voxel RSFC patterns were computed resulting in a similarity matrix (S_ν×ν).

The Laplacian eigenmap algorithm is applied on the graph Laplacian of the similarity matrix to obtain a low dimensional manifold representation of the data, or gradients. The graph Laplacian is denoted as follows:

$$L=D-W$$

where W is a graph representation of S (sparsity is enforced such that the graph is connected), D is the degree matrix defined as $D_{i,i}={\displaystyle \sum }_i{W}_i$, and L is the graph Laplacian. Solving the generalized eigenvalue problem Ly = λDy yields m eigenvectors {y₁, …, y_m} corresponding to the smallest m non-zero eigenvalues {λ₁, …, λ_m}. Here, we focus on y₁, defined as the region’s dominant gradient and reflects the greatest changes in RSFC connectivity over a ROI, whereas higher order gradients, denoted by y_n, 1 < n < m, reflects more subtle changes in RSFC.

This procedure is performed to generate dominant gradients of each region (i.e., left and right LL, T, and UL) for each subject. The direction, or sign of all generated gradients are often-times ambiguous and may be dependent on the selected dataset, and ROI-choice. To ensure consistency, the direction of all generated gradients was matched to have the same orientation. Specifically, the anterior-to-posterior and ventral-to-dorsal gradients corresponded to increases in gradient value. Additionally, each subject’s dominant gradients were normalized between 0 and 1 to ensure consistency in scale across the cohort. Dominant gradients for each ROI were used to investigate relationships to anatomical hierarchy and geodesic distance in subsequent analyses.

2.5. Structural MRI measures

We considered cortical thickness, and T1w/T2w (Glasser and Van Essen, 2011) as structural MRI surrogates of S1 anatomical hierarchy and compare them against our RSFC-derived gradients. Due to the effects of surface folding and curvature on cortical thickness, a curvature-corrected cortical thickness measure is used in subsequent analyses. Similarly, to mitigate the effects of B₁ inhomogeneity, a bias field corrected T1w/T2w measure is used.

2.6. Geodesic distance

We considered geodesic distance to probe how the dominant gradient varies while traversing the anterior-to-posterior axis of S1 (i.e., a geodesic distance of zero indicates the most anterior portion of BA3a, and higher values indicate more posterior areas, moving towards BA2). In addition, geodesic distance may be able to capture inter-individual differences and may be relevant for interpreting subject-level behaviour. Here, geodesic distance was measured as the shortest vertex-to-vertex path across a cortical surface connecting the middle vertex of the Brodmann area boundaries (i.e., vertex-to-vertex path adjoining areas 3a & 3b, 3b & 1, and 1 & 2). The middle vertex of the anterior border of BA3a, and posterior border of BA2 was hand selected due to ambiguity of where these regions begin and end, respectively. Vertex-to-vertex paths were generated for left and right, LL, T and UL ROIs separately and used to extract dominant gradient values (based on RSFC) along this trajectory for all subjects. We also repeated this procedure considering other vertex-to-vertex paths, specifically considering 1^st quantile, and 3^rd quantile vertices.

2.7. Evidence of a functional boundary

We define a functional boundary to exist if it separates two discrete parcels. In theory, this means a sharp transition in RSFC should be expected at the interface of a functional boundary. Each boundary between Brodmann areas 3a & 3b, 3b & 1, and 1 & 2 is considered separately: a ‘parcel line’ is drawn to represent parcellation of each pair of regions using the vertex-to-vertex path described in the previous section – i.e., vertices of the anterior Brodmann area is given a value of 1, and vertices of the posterior Brodmann area is given a value of 0 (the line is smoothed with a 5 mm FWHM Gaussian kernel to match rs-fMRI preprocessing procedures). The RSFC values were extracted from the vertex-to-vertex paths for each pair of Brodmann areas (3a & 3b, 3b & 1, and 1 & 2) and normalized between 0 and 1 to match the scaling of the ‘parcel line’. Evidence for a functional boundary is defined by computing the L2 norm between the parcel line and RSFC values extracted from the dominant gradient. A value of zero suggests a perfect overlap between the dominant gradient and parcel line to be used as evidence for a functional boundary.

2.8. Seed selection

K-means clustering was performed on the dominant gradient map to assess the degree of overlap between clusters and Brodmann areas. The number of clusters considered were K = 2, 3, 4. If the dominant gradient reliably clusters well into the architectonic Brodmann areas, then this may further substantiate structure and function relationship of cortical S1. We justify the use of a clustering scheme that optimizes overlap with Brodmann area regions. Furthermore, we used this clustering scheme to define seed ROIs in subsequent thalamocortical connectivity analyses.

Choice of cluster number (K) was optimized by calculating an average dice similarity coefficient (DSC) between each of the K-clusters and a configuration of the Brodmann areas such that the average DSC is maximized. The DSC measures the similarity between a set X and Y (i.e., a K-means cluster and its corresponding ground truth label), if the sets are identical (i.e., they contain the same vertices), the coefficient is equal to 1, while if X and Y have no vertices in common, then it is equal to 0, otherwise the DSC falls somewhere between 0 and 1 (Dice, 1945). A silhouette analysis was conducted to assess the choice of K clusters. Silhouette coefficients are calculated for each vertex in a clustered region, providing a similarity measure ranging between -1 and 1. A high coefficient value indicates that the vertex is well matched to its cluster, a value of 0 indicates the vertex is between two clusters, and a negative value indicates a possible incorrect cluster assignment.

2.9. Statistical methods

2.9.1. Multiple comparisons corrections

All statistical comparisons were conducted using an α-level of 5% fully Bonferroni corrected for the number of comparisons.

2.9.2. Dependent correlation test

As a statistical test for the difference between similarities of the measures, we employed the dependent correlation test. This test provides a nonparametric test to compare the Spearman correlations of two variables against a common dependent variable using a bootstrapping approach (Wilcox, 2016). The dependent correlation test was used to compare Spearman rank correlation of different metrics to anatomical hierarchy.

2.9.3. Polynomial regression

Trend lines in Fig. 2a-c were calculated using polynomial regression, a form of regression analysis in which the dependent variable is modelled as an nth degree polynomial (in this case we choose n = 2 to account for the U-shape observed in the data).

Fig. 2. Associations between anatomical hierarchies of S1 and metrics.

(A) Correlation between anatomical hierarchy of somatotopic S1 subregions and their corresponding dominant gradients derived from RSFC data. (B) Correlation between anatomical hierarchy of somatotopic S1 subregions and T1w/T2w. (C) Correlation between anatomical hierarchy of somatotopic S1 subregions and cortical thickness. (D) Absolute spearman rank correlation between different metrics and anatomical hierarchy of S1 for each somatotopic subregion (***P < 10⁻⁴ for all metrics and ROIs). (E) Comparison between metrics correlation to S1 anatomical hierarchy, across all somatotopic ROIs demonstrating RSFC gradient performs on par with T1w/T2w. Error bars in (A-C) indicate the SEM. Hierarchical levels (1, 2, 3, 4) correspond to (Brodmann areas 3a, 3b, 1, 2).

2.9.4. Thalamocortical connectivity analyses

We considered each ipsilateral LL, T, and UL as its own functional unit, as such, thalamocortical connectivity analyses were performed for each of these ROIs independently. The average timeseries were extracted from architectonic subdivisions of S1 determined from the K-means clustering analysis. Partial correlations maps of the unilateral thalamus were calculated using the extracted timeseries. Thalamocortical connectivity between Brodmann areas 3a & 3b, and 1 & 2 of somatotopic S1 and the thalamus was quantified by calculating the average partial correlation in the thalamus. Next, dominant connectivity of architectonic subregions of S1 to the thalamus was assessed using a nonparametric one-sample t-test (Nichols and Holmes, 2002). Significant voxels are considered as areas demonstrating dominant connectivity (P < 0.05, correcting for multiple comparisons after threshold-free cluster enhancement). Using the t-statistic image obtained from the one-sample t-test permutation test, and the Morel atlas of the thalamus, we quantified which thalamic nuclei demonstrated peak thalamocortical connectivity in each ROI (Krauth et al., 2010). Lastly, differences in the magnitude of partial correlation scores (or thalamocortical connectivity) of Brodmann parcels to each thalamic nucleus were assessed (P < 0.05, nonparametric Wilcoxon signed-rank test).

2.9.5. Data and code availability statement

All relevant MRI data are publicly available at https://db.humanconnectome.org. The processed data including the S1, and somatotopic subarea gradients, in addition to K-means clustering analyses can be found in the following repository: https://github.com/gngo4/S1_RSFC_Gradients.

The code for connectopic mapping is available in the following repository: https://github.com/koenhaak/congrads. Note, connectopic mapping was performed to resting-state fMRI data in CIFTI format. To generate RSFC gradients, the CIFTI formatted resting-state fMRI data must first be converted to NIFTI format. For further requests please contact the corresponding author.

2.9.6. Ethics statement

This study used publicly released HCP dataset and was approved by the Washington University Institutional Review Board. Furthermore, the present study follows the WU-Minn HCP Consortium Open Access Data Use Terms.

3. Results

3.1. The dominant gradient of resting-state functional connectivity captures S1 anatomical hierarchy and microstructure

To enable the study of anatomical hierarchy in S1 using RSFC, our first aim was to embed high-dimensional RSFC data of S1 into a lower dimensional space. Specifically, the Laplacian eigenmap algorithm was computed on voxel-wise RSFC patterns to obtain gradients of S1 where each gradient represents a one-dimensional embedding of a region of interest. Previous work looking at the dominant gradient, or the gradient corresponding to the lowest non-zero eigenvalue of precentral gyri (i.e. primary motor cortex as defined using FreeSurfer (M1)) revealed somatotopic organization (Haak et al., 2018). In line with this previous work, we replicated the observation of somatotopy in left and right S1 cortices derived from group data. The dominant gradient reflected somatotopy, as observed by a gradual change in RSFC moving from lower limb to upper limb areas, whereas a clear boundary distinguishing the face from the upper limb region was observed. Furthermore, higher order gradients 2 and 3 of S1 revealed subtler differences in connectivity between the S1 subfields (i.e., anterior-to-posterior), notably observed in LL, T, and UL (Fig. 1b). Together, these observations suggest variation in RSFC is sensitive to biologically plausible functional organization, in this case, of S1. Motivated by this, we applied this technique to investigate functional organization within each somatotopic region of S1 (i.e., left and right hemisphere, LL, T and UL regions) to tease apart subregion-specific RSFC differences, and we hypothesized that the dominant gradient will provide a non-invasive predictor for anatomical hierarchy in S1.

Fig. 1. Data-driven characterization of RSFC patterns in somatosensory cortex and somatotopic subregions.

(A) ROI definitions mapped onto the cortical surface. (B) Group level gradients derived from S1 to whole brain and subcortex RSFC patterns (N = 65). The top three gradients are shown where each gradient represents a similarity embedding, as such, areas of similar colours express similar connectivity patterns. The first gradient matches somatotopic organization, and the second and third gradient suggests separation of S1 along the anterior-to-posterior axis. (C) Group level gradients derived from somatotopic S1 subregions (i.e., LL, T, and UL) to whole brain and subcortex RSFC patterns (N = 65). All dominant gradients demonstrate a change in RSFC along the anterior-to-posterior axis. Gradient directionality is arbitrary; orientations have been matched to ensure consistency of interpretation.

The dominant gradient obtained of left and right LL, T and UL regions, to some degree, revealed an anterior-to-posterior axis (Fig. 1c) which may be an indicator of S1 hierarchical organization. The dominant gradients across all somatotopic regions were stable across individual subjects: the mean pairwise Pearson correlation between subjects’ dominant gradients were [0.89, 0.96, 0.99, 0.98, 0.79, 0.62] for left and right LL, T, and UL, respectively. In the case of right UL, we observed a clear change in RSFC traversing from anterior-to-posterior (i.e., Brodmann areas 3a to 2). However, it was also noted that the maximum and minimum values of the gradient occurred in a dorsal-to-ventral manner within the region. This may be due to an imprecise definition and over estimation of the right UL region in the most ventral portion of the ROI. Such a definition may include voxels of S1 that encodes somatosensory face information, reflected by voxels with differing RSFC patterns, which in turn, may skew the right UL embedding. Subsequent cropping of the right UL region by removing the ventral-most area (which appears to overlap with the somatosensory face information) revealed a gradient spanning the region’s anterior-to-posterior axis (Supplementary Fig. 1). Specifically, the cropped right UL region was generated by removing 10% of the voxels with the lowest gradient values. As a result, this improvement was matched by a mean pairwise Pearson correlation between subjects’ dominant gradient of 0.76 (compared to 0.62).

Next, we considered the ability of the dominant gradients to estimate anatomical hierarchical levels by comparing them to alternative surrogate measures derived from structural MRI. Based on previous work, measures of cortical thickness (Wagstyl et al., 2015) and T1w/T2w (i.e., considered as a proxy for microstructure/myelination) were considered: cortical thickness generally increases in areas higher up along the anatomical hierarchy, whereas the inverse is true for measures of myelination. Although all three measures were strongly associated with hierarchical levels of S1 (***P < 10⁻⁴; Spearman correlation between all measures, and all somatotopic regions, see Fig. 2a–c), we found that the dominant gradients with the exception of right UL were more strongly correlated to hierarchical levels in S1 compared to cortical thickness (***P < 10⁻⁴; dependent correlation test, see Fig 2d–e). Dominant gradients of left and right, LL and T performed better than T1w/T2w, whereas left UL performed the same, and right UL performed worse (presumably due to improper ROI definition) (***P < 10⁻⁴; dependent correlation test, see Fig. 2d–e). Overall, we found that the dominant gradient was more strongly correlated to our annotated hierarchy scheme of S1 compared to structurally-derived surrogates. This finding provides evidence for the utility of RSFC-based gradients for characterizing anatomical hierarchy of S1.

Interestingly, the dominant gradient across all somatotopic regions strongly correlated to T1w/T2w intensity (***P < 10⁻⁴; Pearson correlation, see Fig. 3) providing evidence of structure-function relationships in S1. With the exception of right UL, correlations were high across all regions (left LL: r=.84; left T: r=.90; left UL: r=.93; right LL: r=.91; right T: r=.87; right UL: r=.44). Although significance was observed between right UL’s dominant gradient and T1w/T2w intensity, we attribute its low correlation values due to improper definition of this region as was previously suggested. In fact, an improved correlation value of r=.82 (compared to r=.44) was observed when using the truncated-right UL (Supplementary Fig. 2).

Fig. 3. Association between dominant gradient and T1w/T2w across left and right somatotopic S1 subregions.

(A) Dominant gradient (left) and T1w/T2w intensities (right) for left and right, lower limb, trunk, and upper limb. (B) Correlation between the dominant gradient and T1w/T2w across all somatotopic S1 subregions (***P < 10⁻⁴ for all ROIs).

3.2. Characterising the anterior-to-posterior axis of S1 and correspondence to Brodmann areas

It is unclear how RSFC changes across the anterior-to-posterior axis in S1. Specifically, does RSFC within somatotopic S1 regions have discrete boundaries or does it gradually change over the space of the ROI? We considered the shortest vertex-to-vertex trajectory connecting the midpoint of adjacent Brodmann area boundaries to one another (i.e., a path joining Brodmann areas 3a, 3b, 1, and 2; see Fig. 4a) and used this path to investigate this question. We used geodesic distance away from the beginning of the trajectory to see how the RSFC embedding changes along this path, that is, a geodesic distance of zero corresponds to the most anterior portion of Brodmann area 3a. Fig. 4b shows that RSFC gradually changes moving along this trajectory with gradual changes observed within each Brodmann area as represented by the black points. Clear inter-individual variability in RSFC across the trajectory was also observed.

Fig. 4. Evidence of functional boundaries between Brodmann areas.

(A) Left and right S1 separated by somatotopic regions and colour coded by architectonic Brodmann areas. Black lines in each somatotopic region indicates the vertex-to-vertex path (or trajectory) used to extract values from the RSFC-derived gradients. A geodesic distance of 0 indicates the most anterior vertex of the trajectory. (B) Left and right somatotopic regions’ RSFC connectivity pattern plotted against geodesic distance as defined by a rostro-to-caudal vertex-to-vertex trajectory. Black points represent the average gradient value across all participants (N = 65). Overlaid are three ‘parcel’ definition lines drawn for each boundary between 3a & 3b, 3b & 1, and 1 & 2 (blue, green, and purple) smoothed with a 5 mm FWHM. Corresponding bar plots shows the L2 norm between parcel definition and RSFC connectivity pattern with a value closer to 0 suggesting more evidence for a functional boundary. Error bars in (B) indicate the standard deviation for all plots. *P < .05, ***P < 10⁻⁴

Next, we aimed to investigate whether discrete functional boundaries existed between Brodmann areas. To this end, boundaries between each adjacent Brodmann area were considered, such that a sharp transition exists at the interface between the two areas. These boundaries were additionally smoothed with a 5 mm FWHM Gaussian kernel to ensure consistency with the smoothing criterion that was used in the rs-fMRI preprocessing steps. Three parcels boundaries were drawn as coloured lines adjoining areas 3a-to-3b, 3b-to-1, and 1-to-2 (i.e., blue, green, and purple, respectively; Fig. 4b), and the black points represent data from the dominant gradient. Any observed overlap between the coloured lines and points would suggest evidence for a discrete functional boundary. Comparisons of the average L2 norm between the definition of a parcel boundary, and dominant gradient suggests the most evidence (mean L2 norm closer to zero) for a discrete boundary between Brodmann areas 3b and 1. Specifically, a significantly lower mean L2 norm was observed for the Brodmann area 3b-to-1 boundary, compared to boundaries between areas 3a-to-3b, and 1-to-2 in all somatotopic regions, with the exception of right UL and LL (*P < 0.05; nonparametric Wilcoxon signed-rank test, see bar plots in Fig. 4b). In the case of right UL, the evidence for a boundary was observed equally between areas 3b-to-1 and 1-to-2, and for right LL, evidence for a boundary was observed between areas 1-to-2. Similar observations were also observed when considering two other vertex-to-vertex paths (Supplementary Fig. 3 & 4). However, in both vertex-to-vertex paths, evidence for a boundary in the right UL (lowest L2 norm) were observed between Brodmann areas 3b-to-1. Overall, these results provide evidence for a discrete functional boundary between Brodmann areas 3b-to-1 as compared to the interfaces between Brodmann areas 3a-to-3b, and 1-to-2, which may be better characterized by a gradual change in RSFC.

Rather than investigating the gradients’ trajectory, here we investigate the correspondence between each somatotopic ROIs’ dominant gradients and Brodmann areas using K-means clustering. In the case of two clusters (K = 2), we found that the gradients clustered well into Brodmann areas 3a & 3b, and 1 & 2 (average dice coefficient across all somatotopic regions: 0.856) as expected based on previous results. Using three and four clusters (K = 3, 4) led to a decrease in cluster performance, clustering of S1 into Brodmann areas 3a & 3b, 1, and 2, and Brodman areas 3a, 3b, 1, and 2, respectively (average dice coefficient across all somatotopic regions: 0.564 for K = 3; 0.396 for K = 4). Fig. 5a provides further details regarding the performance of clustering for each somatotopic region. Furthermore, Fig. 5b shows the fraction of participants showing cluster overlap to each Brodmann area and demonstrating that high clustering stability was observed in Brodmann areas 3a & 3b, and areas 1 & 2 (i.e., K=2 clusters). The choice of K = 2 clusters was further validated by silhouette analyses, which demonstrated highest average silhouette coefficient values at K = 2 clusters across all somatotopic regions, with incremental decreases observed with the choice of K = 3, and K = 4 clusters (Supplementary Fig. 5). Although variation in RSFC is not homogeneous in somatotopic S1 regions, here, we provided evidence for the separation of somatotopic S1 into two functional parcels that respect underlying architectonics, specifically, Brodmann areas 3a & 3b, and areas 1 & 2.

Fig. 5. Agreement between K-means clusters and Brodmann areas.

(A) Average dice similarity coefficient scores (N = 65) for K clusters and their closest corresponding Brodmann areas. For K = 2, the DSC is calculated with Brodmann areas 3, and 1 & 2. For K = 3, the DSC is calculated with Brodmann areas 3, 1, and 2. For K = 4, the DSC is calculated with Brodmann areas 3a, 3b, 1, and 2. Bootstrap 95% confidence intervals are denoted by the error-bars. (B) Clustering performance using K = 2, 3, 4 clusters into their respective Brodmann areas. Maps show the fraction of participants (N = 65) showing overlap in the region.

3.3. Thalamocortical connectivity reflects different Brodmann areas

Here, we applied the functional parcels defined by Brodmann areas 3a & 3b, and 1 & 2 to investigate thalamocortical connectivity for each somatotopic region. First, spatial maps showing areas of dominant thalamocortical connectivity between unilateral thalamus and Brodmann areas 3a & 3b, and 1 & 2 were examined to identify thalamic sites that are connected to S1. Fig. 6a shows each cortical component, and its significant areas of connectivity to the thalamus (*P < 0.05; nonparametric one-sample t-test). Qualitatively, these maps show Brodmann areas 3a & 3b have widespread connectivity to the whole thalamus across left and right, LL, T and UL. Contrary to this, thalamocortical connectivity of Brodmann areas 1 & 2, although widespread in the LL, was predominantly localized in the posterior of the thalamus. We investigated the dominant thalamocortical connectivity maps to further characterize which thalamic nuclei (i.e., using the Morel atlas [Krauth et al., 2010]) showed peak connectivity for each cortical ROI of S1. In general, we found peak thalamocortical connectivity to areas adjacent to VP nucleus (i.e., posterior nucleus [PO], medial geniculate nucleus [MGN], and anterior pulvinar [PuA]) (Fig. 6b), whereas only right Brodmann areas 1 & 2 showed peak connectivity to ventral posterior (VP) nucleus. Furthermore, left and right Brodmann areas 3a & 3b trunk showed peak connectivity to more anterior regions of the thalamus (i.e., medial dorsal [MD], and intralaminar [IL] nuclei), and the left T and UL of Brodmann areas 1 & 2 also showed peak connectivity to the lateral pulvinar (PuL). Next, we assessed whether the magnitude of RSFC differed between the functional parcels (Brodmann areas 3a & 3b, and areas 1 & 2) and each thalamic nucleus. In general, higher functional connectivity was observed between each thalamic nucleus and Brodmann areas 3a & 3b than with Brodmann areas 1 & 2. Specifically, thalamocortical connectivity was significantly higher between somatotopic Brodmann areas 3a & 3b and seven out of the 13 thalamic nuclei, compared to thalamocortical connections to areas 1 & 2: AN (left & right T, and left & right UL), MD (left LL, left & right T, and left & right UL), IL (left LL, left & right T, and right UL), VP (left T only), VL (left & right T, and left & right UL), VA (left & right T, and left & right UL), and VM (left LL,left & right T, and left & right UL) (*P < 0.05; nonparametric Wilcoxon signed-rank test, specific details documenting all of the thalamic nuclei-to-S1 functional connections can be found in Fig. 6c with box plots shown in Supplementary Fig 6). Contrary to this, thalamocortical connectivity of the MGN (right T only), and PuL (right T, and right & left LL) were higher with Brodmann areas 1 & 2 compared to connections with areas 3a & 3b (*P < 0.05; nonparametric Wilcoxon signed-rank test). Overall, most thalamic nuclei demonstrated differences in thalamocortical connectivity between functional parcels with a trend towards having stronger connectivity to Brodmann areas 3a & 3b, than to areas 1 & 2. Together, these results demonstrate that VP and VP-adjacent nuclei are primarily connected to both S1 Brodmann parcels, as supported by NHP-related anatomical studies (Nelson and Kaas, 1981; Mayner and Kaas, 1986). We further showed that different Brodmann area parcels led to spatial differences in connectivity patterns to the thalamus, in addition to some differences in magnitude of connectivity to thalamic nuclei. Taken together, this suggests the application of the proposed gradient scheme in future RSFC-related connectivity studies.

Fig. 6. Spatial analysis of dominant thalamocortical connectivity between ipsilateral thalamus and functional subdivisions of S1.

(A) Areas of the thalamus significantly correlated with corresponding cortical region (*P < .05). To qualitatively assess areas of the thalamus that display higher connectivity, the t-statistic map is overlaid on these maps obtained from the one-sample t-test permutation test. (B) Mean t-statistic values for each thalamic nucleus based on the Morel atlas (Krauth et al., 2010). Only thalamic nuclei that were significantly correlated with the corresponding cortical region are shown. Numbers 1, 2, 3 corresponds to the top three thalamic nuclei with the highest t-statistic value in descending order. (C) Difference in thalamocortical connectivity (measured by partial correlation score) between Brodmann areas 3a & 3b, and 1 & 2 to each thalamic nucleus. Thalamocortical connections with significant differences are annotated by an asterisk (*P < .05).

The 13 thalamic nuclei are as follows: anterior nucleus (AN), medial dorsal (MD), internal lamina (IL), pulvinar medial (PuM), pulvinar lateral (PuL), pulvinar anterior (PuA), lateral posterior (LP), medial geniculate nucleus (MGN), posterior nucleus (PO), ventral posterior (VP), ventral lateral (VL), ventral anterior (VA), and ventral medial (VM).

4. Discussion

In the present study, we demonstrated the use of resting-state functional connectivity to characterize mesoscopic structural organization of primary somatosensory cortex. First, a novel technique, connectopic mapping, was applied to RSFC data and revealed a RSFC gradient in S1 which serves as a proxy for anatomical hierarchy of that region. Subsequent analysis of the RSFC gradient revealed evidence for two distinct functional parcels that delineate Brodmann areas 3a & 3b from areas 1 & 2. Thalamocortical connectivity using these parcels was then applied to reveal differing connectivity patterns that are supported by anatomical studies of NHPs, and further underscores the value of using this gradient scheme for future S1-related work in humans. Collectively, these results provide evidence for an anterior-to-posterior gradient in the resting primary somatosensory cortex and suggests close associations to anatomical hierarchy, microstructure, and Brodmann boundaries. Application of this novel technique provides new insight into bridging the gap between mesoscale connectivity and microstructure along the architectonic axis of S1, and builds upon previous multimodal characterization of S1 along its somatotopic axis (Kuehn et al., 2017). Crucially, our work suggests a secondary direction of functional heterogeneity intrinsic to S1 and may be used in conjunction with structural MR measures to fully appreciate the interplay between functional connectivity and structure.

Manifold learning or ‘connectopic mapping’ was used to demonstrate that the dominant RSFC gradients of somatotopic S1 regions were able to accurately predict anatomical hierarchy, and in most cases, demonstrated stronger associations when compared to other structural MR metrics. As such, application of RSFC could be used as another surrogate for local mesoscopic hierarchy, and may be used in conjunction with structural MR metrics to investigate structure-function interplay. Evidence for using structural MR metrics as a surrogate for anatomical hierarchy is supported by close associations between patterns of feedforward/feedback innervations (Barbas and Rempel-Clower, 1997), laminar differentiation (Barbas, 1986), and cytoarchitecture (Dinse et al., 2015). In principle, structural organization revealed in this manner only accounts for intraregional characteristics and does not consider interregional connectivity. For example, it is possible that structural damage to S1-connected regions may have downstream effects on the dominant RSFC gradient which may not be evident using structural measures. This has far reaching implications for S1, as S1 has been shown to topographically map onto the cerebellum (Hahamy and Makin, 2019), primary motor cortex, supplementary motor cortex (Zeharia et al., 2012), operculum and insula (Brooks et al., 2005), and parietal cortex (Huang et al., 2012), among other regions. More critically, the notion of S1 interregional connectivity is fundamental for brain function as suggested in animals to facilitate whisker-detection tasks (Yamashita and Petersen, 2016; Kwon et al., 2016), and in humans for deep brain stimulation (Horn and Fox, 2020). Thus, the identification of a surrogate for anatomical hierarchy in humans that is grounded by the principles of RSFC may provide further insight into the relationships between structure-function, behaviour, and disease.

A more detailed analysis assessing the anterior-to-posterior gradients found in somatotopic regions of S1 revealed gradual changes in RSFC between boundaries of Brodmann areas 3a & 3b and 1 & 2, whereas a more pronounced change in RSFC can be identified between areas 3b & 1, and suggests a possible functional boundary. Based on NHP anatomical studies it is well understood that different architectonic regions demonstrate distinct connectivity patterns (Krubitzer and Disbrow, 2008). Specifically, Brodmann area 3b connects primarily to adjacent S1 regions, such as areas 3a, 1, 2, secondary somatosensory cortex (S2), and primary motor cortex (M1) (Krubitzer and Kaas, 1990; Jones et al., 1979; Juliano et al., 1990; Darian-Smith et al., 1993). In comparison, Brodmann area 1 demonstrates more dispersed connections to areas 3b, 2, 7b, S2, in addition to sparse connections with areas 3a, M1 and frontal cortex (Pons and Kaas, 1986; Burton and Fabri, 1995; Burton et al., 1995). These separate connectivity patterns may explain the sharp divergence in RSFC patterns between Brodmann areas 3b & 1. Furthermore, Geyer et al. (2000) conducted a transmitter binding study which characterized the Brodmann areas with a ‘neurochemical fingerprint’, and the most differences in neurotransmitter binding sites were observed at the interface between areas 3b & 1, providing further evidence of a functional boundary between these two regions.

Gradual changes in functional topography observed between areas 3a & 3b, and 1 & 2 may be due to similar, but not identical connectivity patterns displayed by each pair of regions. Although area 3b is densely connected to area 3a, area 3a has additional connections to motor and posterior parietal areas of the cortex, as found in anatomical tracer studies in marmosets (Huffman and Krubitzer, 2001), and macaques (Jones et al., 1978; Darian-Smith et al., 1993). Similar findings have been observed in electrophysiological-guided tracer studies in macaques for Brodmann areas 1 and 2 (Pons and Kaas., 1985). Slightly differing connectivity patterns in these pair of regions may explain the gradual changes in RSFC observed in our data. Interestingly, we showed that RSFC also changes gradually within each Brodmann area (see Fig. 4b). These gradual changes reflect intrinsic RSFC heterogeneity in somatotopic areas along the anterior-to-posterior axis and may be related to possible connection topography of somatotopic ROIs onto other cortical areas. Such functional topography may be important for developing a better understanding of mesoscopic hierarchical function in humans.

Although thalamic connections to S1 predominantly originate in VP, it has been shown that thalamic inputs to areas 3a & 3b vary widely, whereas inputs to 1 & 2 localize more to posterior nuclei. For example, NHP anatomical tracer literature has shown that in addition to inputs from VP, areas 3a & 3b receive collective input from pulvinar (Cusick and Gould, 1990) (Pu), ventral lateral (VL), ventral anterior (VA) and central medial thalamic nuclei (Lang et al., 1979) (CL; as part of the internal lamina (IL)), and areas 1 & 2 receive inputs from VL, and Pu (Pons and Kaas, 1985; Friedman and Jones, 1981). In the present study, connections were observed between all investigated thalamic nuclei and Brodmann areas 3a & 3b, while only some thalamic nuclei, majority of which were located in the posterior of the thalamus, showed connections to areas 1 & 2. Nuclei that were connected to both Brodmann parcels trended towards having stronger functional connectivity to Brodmann areas 3a & 3b, compared to areas 1 & 2. Perhaps the most important thalamic nucleus associated to S1 is VP, which is known to have more dense connections to Brodmann area 3 compared to area 1 (Mayner and Kaas, 1986). However, here we observed only a trend favouring stronger functional connectivity between areas 3a & 3b and VP. This inconsistency with existing NHP literature may be attributed to a reduced signal-to-noise ratio and sensitivity to blood-oxygenation-level-dependent signal in the subcortex (compared to the cortex) (Puckett et al., 2018). Despite the lack of significant differences in observed functional connectivity between VP (and VP-adjacent) nuclei and Brodmann parcels, here, we demonstrated the parcels’ ability to accurately describe well-known NHP anatomical thalamocortical connections. Together, these findings suggest that consideration of S1 as two separate architectonic ROIs, rather than the common method of using whole S1 (Woodward et al., 2017; Chen et al., 2019), may provide complimentary information regarding thalamocortical RSFC. This approach could be used in clinical neuroscience to conduct more in-depth investigations into thalamocortical connectivity in future studies.

Interestingly, in Brodmann areas 3a & 3b, and 1 & 2, we saw ipsilateral thalamocortical connectivity to be higher in thalamic nuclei adjacent to VP (posterior nucleus (PO), medial geniculate nucleus (MGN), and anterior pulvinar (PuA)), compared to VP itself. It is possible that misregistration between the Morel cytoarchitectonic atlas (used to define the nuclei) may explain for these inconsistencies. For example, the Oxford thalamic connectivity probability atlas (Behrens et al., 2003) derived from diffusion tractography showed that PO, VP and PuA demonstrated similar likelihood of connectivity to the somatosensory cortex (31, 26.8, and 23.6 %, respectively), suggesting that Morel’s atlas definition of PO should be considered the primary relay nucleus of S1, which is traditionally acknowledged as VP. With these considerations in mind, our thalamocortical analyses of Brodmann areas suggests peak RSFC to VP or VP-adjacent nuclei, and further supports the VP-centric role of the thalamus for S1.

Our results are subject to several methodological limitations. In this study, the ensuing dominant gradients were generated from RSFC data which relies on inter-brain region timeseries correlation values during an at-rest paradigm. In principle, this technique cannot be used to draw inference on the functional role of a cortical region, whereas task-based multivariate fMRI analyses could be used to fill this gap (Yokoi and Diedrichsen, 2019). Nonetheless, RSFC enables an accessible way to index brain connectivity, and holds strong parallels to how anatomical hierarchy has been traditionally annotated based on anatomical connectivity information (Felleman and Van Essen, 1991). Furthermore, RSFC provides a more practical solution for acquiring data from clinical participants who may be unable to cooperate and perform task-related experimental designs.

Another limitation is that our results only investigate the primary gradient generated from each somatotopic S1 subregion while ignoring other higher order gradients. In the context of this work, constraining our investigations to only the primary gradient was sufficient for us to accurately characterize an anterior-to-posterior axis of the somatosensory cortex while demonstrating its correspondence with underlying cytoarchitectonic boundaries. It is acknowledged that multiple overlapping gradients may exist within any given ROI (for example, a retinotopic and visuotopic gradient in V1 [Haak and Beckmann, 2020]), thus future investigations could evaluate whether higher-order gradients provide further information towards the characterization of S1.

Finally, as this work uses previously defined ROIs, any limitations associated with ROI-based analyses are also shared with this study. To mitigate ROI-inaccuracies, we used Brodmann area ROIs taken from Glasser and colleagues’ atlas which takes advantage of convergent multi-modal information to reduce errors associated with mis-defining of architectonic borders (Glasser et al., 2016). Furthermore, it is suggested that the medial-superior definition of Brodmann area 2 may slightly overlap with area 5L (Scheperjans et al., 2008), which in turn, may skew embeddings of LH/RH upper limb. Nonetheless, although inaccuracies in ROI may attend this work, we believe our results demonstrate a clear anterior-to-posterior axis in somatotopic S1 that is strongly associated with microstructure and corresponds well with Brodmann areas. Moving towards finer-grained descriptions of functional topography using RSFC, we stress the importance of accurate ROI definition to reliably capture biologically meaningful gradients.

The present study uses RSFC data to demonstrate anatomical hierarchy of S1 in humans, in addition to its association with microstructure and correspondence to Brodmann areas. Such insight suggests close coupling between structure and function and offers a framework for studying structure-function interplay in humans. Beyond this, examination of S1 at the systems level could lead to improved understanding of sensorimotor behaviours, and deficits whose pathophysiology is not well understood.

Supplementary Material

Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.neuroimage.2021.118031.

Data and code availability statement

All relevant MRI data are publicly available at https://db.humanconnectome.org. The processed data including the S1, and somatotopic subarea gradients, in addition to K-means clustering analyses can be found in the following repository: https://github.com/gngo4/S1_RSFC_Gradients.

The code for connectopic mapping is available in the following repository: https://github.com/koenhaak/congrads. Note, connectopic mapping was performed to resting-state fMRI data in CIFTI format. To generate RSFC gradients, the CIFTI formatted resting-state fMRI data must first be converted to NIFTI format. For further requests please contact the corresponding author.