When constructing risk scoring models (such as PRS) for hereditary tumors such as breast cancer and ovarian cancer, we found significant heterogeneity between multimodal omics data (including germline SNV, CNV, RNA expression, and methylation), especially the lack of model transferability among different ethnic groups. What statistical or deep causal modeling methods can effectively handle cross-modal heterogeneity and population heterogeneity to improve the generalization performance of the model on multi-center clinical data? How to control model drift and risk interpretability in large-scale deployment?