Modeling and application of alzheimer's disease complex trait prediction based on multi-task learning.

Alzheimer’s disease (AD) is a complex disorder influenced by genetic factors, and related phenotypes often share common genetic mechanisms. However, traditional genetic risk prediction models typically focus on a single phenotype, neglecting valuable cross-phenotype information. This study aims to explore the effectiveness of multi-task learning (MTL) in modeling genetic data for AD and to assess whether aggregating genetic signals at the gene level can further improve prediction accuracy. We applied two representative MTL models, including the Sluice Network and the Learning to Branch (LTB) model, and compared them with single-task learning (STL) and hard parameter sharing (HPS) models. Simulation experiments were conducted to evaluate model performance under different conditions, followed by validation using genetic and phenotypic data from the ADNI cohort. Results show that the Sluice Network outperformed all other models under various sharing conditions, benefiting from its flexible parameter-sharing mechanism. LTB and HPS performed well under high sharing but showed significant degradation when sharing was low. Furthermore, gene-level genetic variants aggregation further improved prediction performance, particularly in the Sluice Network and LTB models. These findings suggest that MTL combined with biologically-informed feature aggregation holds promise for genetic risk prediction in AD and other complex diseases. By utilizing cross-phenotype genetic information and aggregating signals at the gene level, MTL offers a more comprehensive and accurate approach to genetic risk prediction.