Data representation learning for disease prediction using omics data Yeonwoo Chung College of Information and Computing Gwangju Institute of Science and Technology

Abstract

Accurate prediction of complex diseases using omics data remains challenging due to patient heterogeneity, the high dimensionality of molecular features, and the limited number of available samples. Learning effective data representations is therefore essen- tial, as it transforms raw omics measurements into latent features that more clearly capture disease-related patterns. This thesis investigates two complementary strategies for representation learning using both single-omics and multi-omics data. For single- modality settings such as mRNA expression, we propose JTSC, a joint triplet loss framework with a strict semi-hard constraint that stabilizes metric learning by filtering misleading hard samples and enforcing representation-level augmentation. JTSC re- fines the embedding space under limited and heterogeneous samples, and demonstrates improved performance on Alzheimer’s disease data and multiple cancer stage predic- tion tasks. To address challenges in multi-omics integration, we introduce MOSAIC, a similarity-based framework that aligns modality-specific embeddings and fuses them using graph convolution. By weighting cross-modal relationships according to activa- tion similarity, MOSAIC captures global inter-omics structure, identifies biomarkers associated with cancer metastasis and survival, and maintains robustness even when key omics modalities are partially missing. Overall, this work highlights the impor- tance of selecting representation learning strategies that match the characteristics of the data, showing that appropriate formation, and alignment of latent spaces can lead to improved predictive accuracy, robustness, and biological interpretability in both single-omics and multi-omics analyses. ©2026 Yeonwoo Chung ALL RIGHTS RESERVED