An Innovative Multi-Omics Model Integrating Latent Alignment and Attention Mechanism for Drug Response Prediction

By using omics, we can now examine all components of biological systems simultaneously. Deep learning-based drug prediction methods have shown promise by integrating cancer-related multi-omics data. However, the complex interaction between genes poses challenges in accurately projecting multi-omics data. In this research, we present a predictive model for drug response that incorporates diverse types of omics data, comprising genetic mutation, copy number variation, methylation, and gene expression data. This study proposes latent alignment for information mismatch in integration, which is achieved through an attention module capturing interactions among diverse types of omics data. The latent alignment and attention modules significantly improve predictions, outperforming the baseline model, with MSE = 1.1333, F1-score = 0.5342, and AUROC = 0.5776. High accuracy was achieved in predicting drug responses for piplartine and tenovin-6, while the accuracy was comparatively lower for mitomycin-C and obatoclax. The latent alignment module exclusively outperforms the baseline model, enhancing the MSE by 0.2375, the F1-score by 4.84%, and the AUROC by 6.1%. Similarly, the attention module only improves these metrics by 0.1899, 2.88%, and 2.84%, respectively. In the interpretability case study, panobinostat exhibited the most effective predicted response, with a value of −4.895. We provide reliable insights for drug selection in personalized medicine by identifying crucial genetic factors influencing drug response.