INTEGRATION OF GENOMIC AND CLINICAL DATA FOR PERSONALIZED TREATMENT PLANNING

Abstract

This means that cancer heterogeneity demands the integrative analysis of multi-omic data by default to give precision oncology accurate prognostics. High-dimensional genomic, epigenomic, and clinical data translation into clinically actionable survival predictors is discussed in the paper, where a Multi-Omic Attention Network (MOANet) is presented that combines network-based prior knowledge, cross-modal attention, and quantification of uncertainty. RNA-seq, DNA methylation, somatic mutation, and clinical data of a retrospective cohort of 1,284 breast, lung, and colorectal adenocarcinoma patients were harmonized. It comprised ComBat batch correction, Multi-Omics Factor Analysis to dimension-reduce, propagation of networks on protein-protein interaction networks, and a transformer-based attention model with Shapley additive explanations for interpretability. Relative comparison of MOANet with eight state-of-the-art models showed that MOANet outperformed prognostic characteristics of time-dependent AUC of 0.837 at 36 months and concordance index of 0.791, which represented a 16.3 percent relative improvement over clinical-only Cox-PH models. All the omic layers were crucial, as observed by ablation experiments, but transcriptomic data alone has the greatest impact. The model showed almost perfect calibration, with a strong performance with missing modalities that maintained an ideal time-dependent AUC of over 0.73 and a strong performance with federated learning performance with an overall time-dependent AUC of 0.845 with a small privacy leak. The network-derived biomarkers that were identified as having the most attributes were TP53, PIK3CA, MUC16, and FAT4. It should be noted that prior knowledge of biology may be useful. These findings confirm that multi-omic combination with explicit uncertainty estimation and federated functions involves attention-based combination, which results in a powerful, interpretable, and clinically translational model of precision oncology prognostication.