Auto-branch multi-task learning for simultaneous prediction of multiple correlated traits associated with Alzheimer's disease

Jiaqi Liang^1,2Zhao Xue^1,2Wenchao Zhou^1,2Guo Xiang-jie^1,3,4Yalu Wen^1,2,5*

• ¹Academy of Medical Sciences, Shanxi Medical University, Taiyuan, China
• ²Department of Health Statistics, School of Public Health, Shanxi Medical University, Taiyuan, Shanxi Province, China
• ³Translational Medicine Research Center, Shanxi Medical University, Taiyuan, Shanxi Province, China
• ⁴School of Forensic Medicine, Shanxi Medical University, Jinzhong, China
• ⁵Department of Statistics, University of Auckland, Auckland Central, New Zealand

Correlated phenotypes may have both shared and unique causal factors and jointly modeling these phenotypes can enhance prediction performance by enabling efficient information transfer. We present an auto-branch multi-task learning model within the deep learning framework for the joint prediction of multiple correlated phenotypes. This model dynamically branches from a hard parameter sharing structure to avoid negative information transfer, ensuring that parameter sharing among phenotypes does not become detrimental. Through simulation studies and analysis of seven Alzheimer's disease-related phenotypes, our method consistently outperformed Multi-Lasso model, single-task learning approaches, and commonly used hard parameter sharing models with predefined shared layers. These analyses also reveal that while genetic contributions across phenotypes are similar, the relative influence of each genetic factor varies substantially among phenotypes. The Python code for our method is publicly available on GitHub (https://github.com/jiaqi69/TAB).