G2DBridge: A Multimodal Framework Linking Genetics to Disease through Imaging Intermediates

Abstract

Genetic-based risk prediction is becoming increasingly available for a wide range of common diseases thanks to the growth of large-scale biobanks and population-scale genetic studies. However, despite widespread availability, genetic risk prediction typically offers only modest predictive power, as complex biological processes that mediate disease manifestation are often not fully captured by genetic variation alone. In contrast, imaging-based prediction can achieve substantially higher predictive accuracy for many diseases but remains costly and unavailable for the majority of populations. Here, we present G2DBridge, a predictive machine learning framework that links genetic variation to disease risk by modeling imaging-derived phenotypes as intermediate traits. Using Alzheimers Disease Neuroimaging Initiative (ADNI) for training and Alzheimers Disease Sequencing Project (ADSP) for external validation, G2DBridge outperforms state-of-the-art polygenic risk scores in Alzheimers Disease (AD) prediction while relying on the same genetic data in the target data, providing additional predictive power without requiring new data modalities. The framework generalizes across cohorts, enables risk stratification with clear separation between cases and controls, and highlights biologically meaningful regions consistent with AD pathology. This approach offers a scalable strategy for applying genetically inferred imaging features in precision medicine.