Distinct gradients of cortical architecture capture visual representations and behavior across the lifespan
- Beijing Key Laboratory of Applied Experimental Psychology, Faculty of Psychology, Beijing Normal University, China
- BGI-Shenzhen, China
- Department of Psychology, School of Arts and Sciences, University of Pennsylvania, United States
- Princeton Neuroscience Institute, Princeton University, United States
- State Key Laboratory of Cognitive Neuroscience and Learning & IDG/McGovern Institute for Brain Research, Beijing Normal University, China
- Psychology Department, Princeton University, United States
Figures
Figure 1 with 4 supplements
Two architectural gradients scaffold human visual cortex.
(a) Principal component analysis on the concatenated cortical thickness and myelin/neurite density maps from all participants in HCP-YA to extract architectural gradients of human visual cortex produced a collection of orthogonal principal components (PCs), consisting of spatial maps (i.e., scores) and individual weights (i.e., loadings) in pairs. (b) The explained variance ratio of the top 5 PCs. The first two PCs (i.e., gradients 1 and 2) dominate the explainable variance. (c) Contributions of the two architectural measures (thickness and density) to the two gradients. (d) Topographic patterns of the two gradients on a flattened cortical surface. Gradient 1 (PC1) displays roughly monotonic change from negative to positive scores across visual cortex, emanating from primary visual cortex V1, while gradient 2 (PC2) showed repeated representation in four localities, mirroring the four visual streams (early, dorsal, lateral, and ventral). Black dotted lines: borders where the different visual streams meet, defined using HCP MMP label boundaries. A.U. is arbitrary units. (e) The convergence and divergence between the group average map of thickness and tissue density. (f) Histogram depicting gradient scores in the four visual stream regions. Gradient 1 is a large-scale gradient increasing from early to ventral, for example. Gradient 2 is a fine-scale gradient sampled evenly within an individual visual stream. (g) The dependence of gradient value differences on geodesic distance is different for the two gradients. Gradient 1 shows larger changes across vertices separated by a long distance, whereas gradient 2 shows larger changes for short distances. (h) Geometric models of the two architectural gradients, which were constructed using the geodesic distance of each vertex of visual cortex to specific anatomical landmarks as anchors. The calcarine sulcus and eight local minima of gradient 2 were used as anchors to model gradients 1 and 2, respectively.
Figure 1—figure supplement 1
The two hemispheres have very similar architectural gradients in visual cortex.
The two architectural gradients from the principal component analysis (PCA) on the left hemisphere (left panel) and the right hemisphere (right panel). The gradients show high spatial similarity between the two hemispheres as measured with a vertex-wise Pearson correlation (gradient 1: r = 0.99; gradient 2: r = 0.94).
Figure 1—figure supplement 2
The principal component analysis (PCA) on the concatenated cortical thickness and tissue density maps can exploit the concurrent spatial changes of the two measurements and thus extract more meaningful architectural gradients than PCA on a single architectural measurement.
(a) The gradients from the PCA performed on the concatenated cortical thickness and tissue density maps. (b) The gradients from the PCA using only cortical thickness maps. (c) The gradients from the PCA using only tissue density (T1w/T2w ratio) maps. (d) The spatial pattern similarities between the gradients from the concatenated PCA and the separate PCA on the single architectural measurement, measured by Pearson correlation coefficients. While the separate PCAs on the single measurements can extract similar PC1s to the concatenated PCA, they fail to extract a meaningful PC2.
Figure 1—figure supplement 3
Principal component analysis (PCA) and the convergent and divergent patterns between group average maps of cortical thickness and tissue density.
(a) Relationship between cortical thickness and tissue density across large-scale visual ROIs. Parcellations were defined using an independent dataset (Abdollahi et al., 2014). Each colored point represents the mean values within a visual area, with error bars indicating the SD. The inset shows the same analysis using visual ROIs from the multimodal parcellation (MMP) atlas (Glasser et al., 2016a). Strong negative correlations were observed across visual ROIs in both parcellations. (b) Relationship between cortical thickness and tissue density within individual visual ROIs. Bar plots show correlation coefficients (r) computed separately for each ROI, with error bars indicating the SEM across hemispheres. Most visual areas showed significant negative correlations (*p < 0.05, Bonferroni corrected), whereas V1, V2, and VO1 displayed positive correlations, consistent with previous findings (Maingault et al., 2021; Sereno et al., 2013; Shafee et al., 2015). (c) The covariance matrix of the PCA reveals high inter-individual positive covariance within cortical thickness and tissue density, respectively, and inter-individual negative covariance between cortical thickness and tissue density. This suggests that the group-level average thickness and tissue density account for a significant portion of the variance, allowing the high-dimensional features to be approximated as two primary dimensions: group-level thickness and tissue density. (d) Schematic illustration of PC1 and PC2. As indicated by (c), the PCA can be approximated as a two-dimensional problem. The point cloud, represented by the gray ellipse, indicates individual vertices. The principal component with the greatest explained variance (PC1) is oriented along thickness + (−1) * tissue density, while the PC2 is oriented along thickness − (−1) * tissue density. (e) Alignment of PC scores with local microstructural features. Within each visual ROI, mean cortical thickness and density were plotted for vertices with the top and bottom 20% of PC scores (closed and open dots, respectively), connected by lines. For PC1 (left), vertices with higher scores generally showed greater thickness and lower density, while those with lower scores showed the opposite trend. This within-ROI relationship parallels the large-scale cross-ROI anti-correlation in (a), consistent with PC1 reflecting a large-scale cortical organizational axis. In contrast, PC2 (right) exhibited more heterogeneous profiles across ROIs, indicating that it captures region-specific microstructural variations.
Figure 1—figure supplement 4
Split-half cross-validation of PC2 maps demonstrates reliable geometric features underlying visual cortex organization.
A split-half cross-validation was conducted to mitigate concerns about circularity in defining the geodesic model’s anchor points. Anchors were defined from one half of the participants and used to predict the PC2 map in the other half. Top row shows the PC2 maps derived from half 1 (left, solid outline) and half 2 (right, dashed outline) of the sample, displaying highly similar spatial patterns (r = 1.00, p < 0.001), indicating strong reliability. Middle and bottom rows demonstrate cross-prediction: the geodesic model derived on half 2 successfully predicts the PC2 map in half 1 (left, r = 0.49), and vice versa (right, r = 0.46). These results indicate reliable geometric features underlying the spatial organization of PC2 across independent samples.
Figure 2 with 2 supplements
The functional and microstructural properties of the two architectural gradients.
(a) Gradient 1 was highly correlated with the population receptive field (pRF) size (r = 0.66), while gradient 2 was not (r = 0.03). (b) Gradient 1 was perfectly correlated with the hierarchical rank of the 10 visual areas within the ventral visual stream (Spearman rank ρ = 1.00), while gradient 2 was not. (c) The functional significance of the architectural gradients was evaluated by measuring to what extent each gradient is related to areal differentiation of the visual cortex (left) and the global functional organization measured by fractional amplitude of low-frequency fluctuation (fALFF) from resting-state fMRI (right). The combination of gradients 1 and 2 greatly improved the predictive power of classifying visual areas compared to using either gradient alone. However, gradient 1 contributes more than gradient 2 in predicting the large-scale functional organization (i.e., fALFF map). (d) Cell body density from the BigBrain dataset is quantified for each cortical layer at each vertex and correlated with each gradient map. Gradient 1 was mainly correlated with cell body density in layers III and IV, while gradient 2 was mainly correlated with cell body density in layer I. (e) The architectural gradient 2 predicts the presence of novel visual representations in the anterior temporal lobe. Left: example participant with the pRF eccentricity map displayed on the inflated cortical surface. The highlighted region (white dotted line) is the subject of the zoomed insets on the right. The black outline delineates the anatomical region from which pRF data was extracted. The putative visual representation cluster is outlined on the insets, showing a radial eccentricity representation and a perpendicular representation of polar angle traveling roughly medio-anterior to latero-posterior as indicated by the white arrow. (f) Illustration on an inflated cortical surface illustrating that the AT-cluster of retinotopic maps is located near the anterior intersection of the occipitotemporal (OTS) and collateral sulci (CoS). The AT-cluster representations demonstrate perpendicular representations of pRF eccentricity and polar angle. (g) In all 21 participants, pRF size and eccentricity from all above-threshold vertices within the anatomically defined region (black solid line from panel e) are extracted, binned by eccentricity, averaged across participants, and then lines of best fit are modeled across the averaged data. Shaded regions represent bootstrapped 68% confidence intervals. (h) Average visual field coverage maps of left and right AT voxels. Lighter colors represent higher coverage density, black outline denotes 50% pRF density compared to maximum. Value of 1 denotes all participants had maximum pRF coverage at that point in space. (i) Kernel density histogram denoting laterality index across participants. Value adjusted such that +1 denotes purely right visual field coverage across an individual’s pRF fits, −1 denotes purely left laterality, and 0.5 is bilateral coverage.
Figure 2—figure supplement 1
The architectural gradients derived from the principal component analysis (PCA) on the mesoscale structural measurements are more functionally relevant than the gradients from resting-state functional connectivity (RSFC).
(a) The first two gradients from the PCA on the mesoscale structural measurements (left), the results from diffusion embedding on the whole-brain RSFC matrix (middle), and the PCA on the RSFC matrix limited to include only visual cortex (right). (b, c) Architectural gradient 1 is more related to visual processing hierarchy than either of the primary gradients from the whole-brain RSFC or visual-cortex RSFC, which was evaluated based on a map’s correlation to population receptive field (pRF) size (panel b) and hierarchical rank (panel c). (d) The first two architectural gradients are more powerful in classifying visual areas than the two gradients from the whole-brain RSFC and visual-cortex RSFC.
Figure 2—figure supplement 2
Anterior temporal visual representation clusters in all participants.
Individual cortical surfaces have been inflated and zoomed to make visible the ventral surface of the anterior temporal pole. Vertices are colored by properties of the fitted population receptive field (pRF) model, showing either eccentricity (top panel in each participant) or polar angle (bottom panels). Color scales denoting eccentricity and polar angle value mappings are shown in the inset of P1. White circles outline the putative AT visual representation cluster, and dotted white lines are drawn through the shared lower visual field representation in polar angle maps, bisecting the pRF cluster into two hemifield representations. Visualized data is thresholded, showing only vertices whose time course was fitted by the pRF model with at least 10% variance explained.
Figure 3
The relevance of the architectural gradients to visual behavior, development, and degeneration across the lifespan.
(a) Canonical correlation analysis (CCA) was used to associate multiple vision-related tasks with the two weight-vectors of each architectural gradient in the HCP-YA. CCA finds the linear combination of variables that best associate measures from the two data domains across participants. (b) Both architectural gradients can significantly predict individual visual ability. However, gradient 2 showed stronger correlation with visual behaviors than gradient 1. (c) The normalized magnitude of behavioral factor weights from CCA indicates that gradient 1 was correlated more with low-level visual abilities, while gradient 2 was correlated more with higher-level visual abilities. (d) Sliding-window spatial principal component analysis (PCA) was performed across the lifespan (PCA on a given age-bin results in ‘lifespan component’, LC) to compare how the patterns of gradient 1 and 2 change with age. The top two LCs (i.e., spatial maps) extracted from each age window correlate strongly with their respective component from the HCP-YA, confirming that the architectural gradients can be observed across the lifespan. (e) Correlations between LCs and PCs reveal that gradient 2 shows more development and degeneration across childhood and aging, compared to gradient 1.
Additional files
-
Supplementary file 1
Tables A–C.
Table A: Detailed descriptions of the 44 visual areas from the HCP multimodal parcellation (MMP) atlas, related to Figure 1. Table B: Location information of anchors of the secondary gradient’s geometric model, related to Figure 1. Table C: Detail information of 15 vision-related or vision-based behavioral tasks selected from HCP-YA dataset, related to Figure 3.
- https://cdn.elifesciences.org/articles/102095/elife-102095-supp1-v1.docx
-
MDAR checklist
- https://cdn.elifesciences.org/articles/102095/elife-102095-mdarchecklist1-v1.docx
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Distinct gradients of cortical architecture capture visual representations and behavior across the lifespan
eLife 13:RP102095.
https://doi.org/10.7554/eLife.102095.3
