A1272
Title: Modeling interactions for precision medicine: From long range epistasis to image genomic fusion in neurological disease
Authors: Shih-Kai Chu - National Cheng Kung University (Taiwan) [presenting]
Abstract: Two complementary statistical frameworks for precision medicine in neurological Disease are introduced. The first framework constructs multilocus risk scores that incorporate long-range linkage disequilibrium and epistatic interactions to improve genetic risk prediction for mild cognitive impairment in the Alzheimer Disease Neuroimaging Initiative cohort. Variant pairs are selected using population-level linkage measures and model-based multifactor dimensionality reduction, filtered to remove redundant hub structures, and then incorporated into an adapted penalized regression model that aggregates interaction patterns into an individual-level risk score achieving an area under the curve of 0.616 in the test set. The second framework is a multimodal survival model for brain tumors that fuses histopathology images with gene expression data. A pathway-informed self-normalizing neural network extracts genomic features, a convolutional neural network processes tissue images, and tensor-product fusion with a Cox loss function yields a concordance index that exceeds unimodal image-only and genomics-only baselines. Gradient-based attribution highlights predictive image regions and pathway-level genomic signatures, supporting biological interpretability and clinical relevance in individualized prognosis modeling.
