Distilling Multi-Modal Genomic Knowledge for Drug Response Prediction

In clinical practice, the accurate assessment of patient response to drugs is crucial for personalized treatment. Research has demonstrated that alterations in genomic profiles can significantly influence the efficacy of cancer therapies. Studies of cancer drug sensitivity have the potential to predict heterogeneous cell line responses to various drugs, facilitate drug screening processes, and identify new biomarkers for sensitive populations[3]. Unlike traditional documentation of patient drug responses, cancer pharmacogenomics research has established a comprehensive database of in vitro cultured cell lines, encompassing genomic data across multiple modalities, including gene expression, mutation, copy number variation, and methylation patterns[1]. Recent investigations have confirmed that incorporating information from diverse genomic modalities can yield a wide range of biological insights that enhance the predictive modeling capabilities of drug responses[2]. However, access to paired genomic data remains challenging within practical clinical settings, which limits the application of high-performance multi-modal models.

In this work, we propose a multi-modal knowledge distillation framework. During the training phase, we utilize multi-modal data from an open database and subsequently transfer the acquired multi-modal genomic knowledge to an unimodal network. Only gene expression data is required for inference. This approach not only preserves the knowledge of the multi-modal ensemble, but also reduces the computational cost during the inference phase. Different from previous multi-modal networks that only focus on feature fusion within the hidden space, we emphasize the correlation modeling between multi-modal genomes during the fusion phase. During knowledge distillation, we implement inter-sample relation alignment to facilitate efficient information transfer.

Experimental results show that our unimodal inference model outperforms state-of-the-art methods and achieves comparable performance to our optimized multi-modal model. The validation on real clinical data also demonstrates that our method can achieve accurate response prediction when only gene expression is available, indicating the generalization of the model in real-world applications. At the same time, fine-tuning our model on independent cell line data that are not identically distributed can achieve faster convergence and higher inference accuracy than random initialization and baselines.