Cross-Modal Transformation of Structural Similarity Networks into Functional Connectomes for Behavioral Prediction

Poster Session F - Tuesday, April 1, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom

Yae Ji Kim¹, Minchul Kim², Marvin M. Chun³, Kwangsun Yoo^1,4; ¹Sungkyunkwan University, ²Kangbuk Samsung Hospital, ³Yale University, ⁴Samsung Medical Center

Previous studies have demonstrated structural similarity networks (SSNs) derived from T1-weighted MRI reflect individual’s structural brain networks, but their ability to infer functional connectivity and predict behavior remains underexplored. This study examines whether SSNs encode functionally relevant information for predicting functional connectomes (FCs) and cognitive performance through cross-modal connectome transformation models. We used task-free and attention-task fMRI data from 92 participants from Yoo et al., 2022a. SSNs were constructed using structural similarity metrics between volumetric regions parcellated by the Shen atlas (Oskar et al., 2019; Shen et al., 2013). To predict task-free FCs from SSNs, we created cross-modal connectome transformation models by adapting connectome-to-connectome transformation modeling framework (Yoo et al., 2022b). Predicted FCs were evaluated for their similarity to empirical FCs and their ability to predict individual attention functions using connectome-based predictive modeling (CPM). Predicted FCs exhibited high spatial correlation with empirical FCs (r = 0.70; q² = 0.49). Furthermore, predicted FCs demonstrated high state specificity (task-free vs. task-related states). Behavioral predictions based on the predicted FCs (r = 0.22, p = 0.04, q2 = 0.05) outperformed those using SSNs alone (r = 0.06, p = 0.56, q2 = -0.13), indicating that SSNs encode meaningful functional-information relevant to attentional behaviors. Notably, predicted FCs exhibited behavior prediction performance comparable to empirical FCs (r = 0.38 p < 0.01, q2 = 0.12). This study highlights feasibility of leveraging SSNs and cross-modal transformation models to infer the brain’s functional architecture and predict behavior, suggesting a practical tool when direct functional data are unavailable.

Topic Area: METHODS: Neuroimaging