A Multi-Modal Genomic Knowledge Distillation Framework for Drug Response Prediction

Precision oncology utilizes genomic data to tailor treatment to individuals. Cancer drug sensitivity studies can predict the response levels of different drugs for the same cultured cancer cell line, which is beneficial for personalized medicine. Recent studies have demonstrated that integrating multi-modal genomic data, e.g., gene expression, mutation, copy number alteration, methylation, can provide comprehensive knowledge and improve drug response prediction. Although multimodal genomic profiles are generally available from public datasets, only gene expression data is commonly used in clinical settings. In this study, we propose a framework for privileged information knowledge distillation to transfer knowledge from a multi-modal genomic teacher network, using only gene expression for inference. Specifically, we train a teacher network by feature re-weighting based on inter-modality dependencies and align the inter-sample correlations through our proposed relation-aware differentiation distillation. Experiments on the Genomics of Drug Sensitivity in Cancer (GDSC) dataset demonstrate that our framework improves drug response prediction by about 6% compared to the baseline and outperforms state-of-the-art methods. Transferable studies performed on missing GDSC data and clinical datasets further confirm the feasibility of our model for predicting drug responses using only gene expression data.