Precision Medicine Through Multi-Omics Modeling

Abstract

Precision medicine aims to tailor therapeutic strategies to individual patients based on their unique molecular profiles, yet clinical implementation remains limited because single-omics data (genomics alone, transcriptomics alone) capture only one dimension of disease biology. Multi-omics integration -- combining genomics, transcriptomics, proteomics, metabolomics, and epigenomics -- promises comprehensive molecular characterisation that enables accurate patient stratification, treatment response prediction, and mechanism-based therapy selection. However, multi-omics datasets are high-dimensional (>100,000 combined features), heterogeneous (different data types, scales, missing patterns), and expensive to generate, creating fundamental computational and practical challenges. This study developed and benchmarked six multi-omics integration approaches -- single-omics baselines (genomics-only, transcriptomics-only), simple concatenation, Similarity Network Fusion (SNF), Multi-Omics Factor Analysis (MOFA+), autoencoder-based integration (OmiVAE), and a proposed Multi-Omics Precision Medicine Transformer (MOPMT) combining modality-specific tokenisation, cross-omics attention, missing-omics robustness, and interpretable pathway-level embeddings -- for three precision medicine tasks: cancer subtype classification (TCGA pan-cancer, n = 9,624, 33 cancer types, 5 omics layers), drug response prediction (GDSC + CCLE, n = 986 cell lines, 198 drugs), and patient survival stratification (TCGA + METABRIC breast cancer, n = 3,842). MOPMT achieved the highest accuracy across all tasks: cancer subtyping macro-F1 0.918 (vs best single-omics 0.762, concatenation 0.842, MOFA+ 0.872; p < 0.001), drug response AUROC 0.846 (vs 0.724 genomics-only, 0.812 OmiVAE; p < 0.001), and survival C-index 0.748 (vs 0.682 transcriptomics-only, 0.724 SNF; p < 0.001). Critically, MOPMT maintained 92.4% of full-omics performance when any two of five omics layers were missing -- addressing the clinical reality that complete multi-omics profiling is rarely available. Cross-omics attention analysis revealed that proteomics contributed the most incremental value for drug response prediction (24.6% attention weight), while methylation dominated survival stratification (28.4%), providing omics-prioritisation guidance for cost-constrained clinical implementation.