The present study provides evidence of a direct relationship between structural connectivity and function in the human brain. Specifically, we demonstrate that the responses to faces within an individual’s right fusiform gyrus can be predicted from that individual’s patterns of structural connectivity alone. This approach further reveals which targets are most influential in predicting function. Voxels with higher responses to faces had characteristic patterns of connectivity to other brain regions that distinguished them from neighboring voxels with lower responses to faces, or higher responses to scenes.
The connectivity model outperformed the random permutation control, indicating that there exists a strong relationship between connectivity and function. Moreover, it outperformed the distance control, suggesting that spatial information alone is insufficient for predicting functional activity and that connectivity offers information above and beyond the topographic information inherently embedded within it (due to the posited small-world organization of cortical connectivity24,25
). The relationship between function and spatial information was highly variable across participants, while the connectivity data was consistent across participants in its relationship with the functional responses. When compared to the group-average benchmark, a standard method of defining face-selective ROIs in fMRI studies, connectivity was a significantly better predictor of the individual’s actual activation pattern in over seventy-percent of the participants. One reason that the group-average did not successfully predict the activation pattern could be due to the high variability of activation loci, relative to the standard template (e.g.26
While we have treated spatial metrics as potential confounds and controlled for them by using distance and group activation models as controls, future studies may build other geometric models which do predict inter-subject variability in functional activation. For example, detailed models of cortical folding patterns27
, and/or cortical thickness29
may be detectable with MRI and predictive of functional regions. Connectivity can provide a complementary source of evidence in some cases, whereas in others it may be the only gross morphological marker available.
Despite spatial variability in functional responses, the connectivity model was highly accurate across participants. We found that the spatial distribution of face- and scene-selectivity varies in tandem with connection strength to their most predictive targets. A direct analysis of subject-to-subject variability revealed that while each participant’s connectivity profile does well at predicting their own functional response, it predicts another participant’s functional responses relatively poorly. Overall, the connectivity patterns appeared highly sensitive to individual variation in function.
While the results from Group 1 are noteworthy, they could be specific to one dataset22
. The findings from Group 2 demonstrate that this is not the case: the connectivity model’s predictions from Group 1 were much more accurate than both the distance and group-average models in over seventy percent of the new group of participants. This result was especially remarkable, because the participants in Group 2 had been scanned while performing a different functional task. The two tasks differed in the type of stimuli presented (1s static images versus 3s movie-clips), type of design (event-related versus block), number of runs (1 versus 3), and scan parameters (also see Methods
for other differences). Further, the structural connectivity measures in this second group were acquired using a DWI sequence with half as many gradient directions (30 versus 60), indicating the generalizability of the connectivity model across functional tasks and diffusion sequences.
This analysis also reveals the target brain regions for which connectivity with the fusiform is most predictive of face- or scene-selective activity in the fusiform. Face-selective fusiform voxels were predicted by connectivity with regions that have been previously reported to have a role in face processing, such as the inferior and superior temporal cortices (e.g.30, 31
). Scene-selective voxels, on the other hand, were best predicted by their connectivity to key brain areas involved with scene recognition, such as the isthmuscingulate (containing the retrosplenial cortex) and the parahippocampal cortex10, 32, 33
. Unlike functional connectivity, structural connectivity models are naïve to the functional responses of the target regions. Therefore, a region need not be category selective to be connected (and predictive of) selective voxels in the fusiform. For example, unexpected predictors of face selectivity were also discovered, such as the cerebellar cortices. Even though the cerebellum is not commonly considered as part of the “core” or “extended” face processing network3, 30, 34
studies have revealed disynaptic connections with extrastriate visual cortices via pons, which tractography is able to reconstruct (see Supplementary Fig. 1,2
), and is corroborated by functional connectivity38
. Future studies may explore these relationships to further expand on the role of functional responses in components of a structural network. Novel structure-function relationships could be investigated in macaques with functional and connectivity data, and subsequently validated more directly through more invasive techniques involving tracer injections (e.g.39 ,40
The final connectivity model also provides a framework with which to evaluate the impact of the most predictive targets and their spatial distribution. The model built from only the significantly predictive targets resulted in more accurate predictions than the predictions based on all of the target regions. While some of the best predictors from this model were nearby regions, most of them were distant to the fusiform; additional analyses excluding the fusiform’s neighbors (Supplementary Materials
) revealed that while proximal targets are part of the fusiform’s network, they do not fully account for the connectivity model’s performance. Altogether, a distributed network of brain regions characterizes category-specific visual processing in the fusiform gyrus.
The connectivity fingerprint has practical applications, both for defining ROIs independently of a task, and also for exploring group differences in structural connectivity signatures. Researchers or clinicians can apply the relationships discovered here to predict functional activation at the single-subject level in populations who do not or cannot have a functional localizer, and should expect that this will be a more accurate prediction than group-based methods. The connectivity model provided here can also be directly compared to a connectivity model built from participants with specific lesions or conditions. For instance, compromised structural connectivity in congenital prosopagnosics has previously been suggested to play a role in their deficits of face-recognition, in light of their surprisingly normal functional activation in the fusiform41
. This type of analysis can shed light on which components (if any) of the fusiform connectional fingerprint are altered or compromised in individuals with congenital prosopagnosia. A similar analysis can be used to explore possible substrates of face-processing differences in autism, normal development, and aging.
Future studies can also extend the present methods to other brain regions and contrasts that are commonly used as functional localizers, such as retinotopy in visual cortices, scene-selectivity in the parahippocampal place area10
, or expression-specificity in the superior temporal sulcus. In some cases, more complex or nonlinear approaches might better capture the relationship of connectivity and function. We implemented a linear fit in order to provide more parsimonious interpretations and to establish the feasibility of modeling structure-function relationships. Since these relationships are probably not strictly linear in a complex system such as the brain (Supplementary Fig. 3
), future work can expand these findings, creating better models, and elucidating a more detailed relationship between connectivity and function. Additionally, voxel-to-voxel tractography may help to more finely characterize the structure-function relationships identified here.
These findings open a window into the coupling between structural and functional organization in the brain. The operations of a brain region are determined by both its intrinsic properties (i.e., cytoarchitecture) that likely determine the operations that it can perform, and the extrinsic connectivity that defines the input/output relations of that brain region. Neuroimaging can relate localized functions (via fMRI) to input/output patterns of cortical connectivity (via probabilistic tractography) in an individual. The present findings demonstrate that brain structure/function relations can be defined for category-selective functional activation.