We will develop and validate a general framework for modelling complex diseases using the large-scale multimodal data. The project has two linked aims: (i) improve individual-level risk prediction through principled fusion of heterogeneous modalities; and (ii) move beyond associations by identifying and validating interpretable causal pathways.
Research questions
(1) Do multimodal fusion models outperform single-modality and conventional risk models for predicting incident complex disease outcomes?
(2) Which cross-modal interactions are most predictive and biologically plausible?
(3) Which candidate drivers are supported by causal evidence and are robust across subgroups?
Objectives
(1) Build a multimodal architecture that combines tabular transformers with imaging encoders using cross-attention and masking to handle partial-modality samples.
(2) Deliver calibrated prediction with rigorous internal validation and fairness/subgroup checks.
(3) Learn a constrained causal graph guided by genetic instruments and biological priors, then validate key links with Mendelian randomization and sensitivity analyses.
(4) Translate validated effects into interpretable risk scores and candidate biomarkers.
Scientific rationale
Complex diseases are rarely isolated and they are driven by intricate, cross-system interactions that traditional single-modality models fail to capture. This project aims to bridge this gap by establishing a computational framework for complex disease modeling. While the methodology is applicable to various complex traits, we will initially validate it on metabolic and psychiatric disorders. This approach allows us to rigorously test the model’s ability to unravel the biological mechanisms linking distinct physiological systems(e.g., the liver-brain axis), paving the way for precision prevention strategies applicable to a wide range of chronic conditions.
