Drug Response Modeling across Cancers: Proteomics vs. Transcriptomics

Cancer cell lines are the most common in-vitro models for the evaluation of anti-cancer drug sensitivities. Past studies have been conducted to decipher and characterize the pharmacogenomic feature of cell lines based on other omics data, such as genomic mutation data and whole-genome RNA sequencing (RNA-seq) profiles. In particular, proteomic data is also an essential component for the characterization of tumours. However, different from RNA-seq datasets rich in numerous transcriptome profiles of cancer cell lines and cell viability assay of drug responses, the pharmacogenomic protein quantifications are relatively scarce. With the availability of the recently enriched proteomic dataset ProCan-DepMapSanger, we systematically evaluated the interplays among genomic mutations, transcription, and protein expressions across cancer cell lines. In general, blood cancers have higher RNA-protein correlations than those in solid cancers. The differential expression analysis on protein data helped identify more expressional and functional impact of genomic mutations of cancer genes. We also integrated the proteomic map with drug molecular chemical features to construct a bi-modal machine learning model to infer the drug sensitivities of cancer cell lines. Our results demonstrated that protein quantifications can lead to better drug response prediction performance than the model trained on transcriptome profiles. In addition, integrating protein data with drug chemical features, represented as molecular graphs and learned by Graph Neural Network, outperformed the state-of-the-art model DeepOmicNet for drug response prediction in proteomics.