Biologically interpretable deep learning to predict response to immunotherapy in advanced melanoma using mutations and copy number variations

Abstract

Only 30–40% of advanced melanoma patients respond effectively to immunotherapy in clinical practice, so it’s necessary to accurately identify the response of melanoma patients to immune therapy pre-clinically. Here we developed the KP-NET, a deep learning model whose structure is sparse by the KEGG pathways, which can accurately predict melanoma patients’ response to immunotherapy using information at the pathway level that is enriched from gene mutations and copy number variations data prior to immune therapy. The KP-NET demonstrated the best performance on predictive melanoma response to anti-CTLA-4 and anti-PD-1 treatment with corresponding AUROC values of 0.889 and 0.867, and selected some relevant biomarkers, genes such as PLCB2, FGFR4, RHEB and CCNB2, pathways such as Inflammatory mediator regulation of TRP channels, Notch signaling pathway, mTOR signaling pathway and TNF signaling pathway, etc. In conclusion, deep learning model guided by biological information enables accurate prediction of the response of melanoma patients to immunotherapy and relevant biomarkers before clinical treatment, which may be helpful in melanoma precision medicine.

Computational Methods Summarizing Mutational Patterns in Cancer: Promise and Limitations for Clinical Applications

“…In another study, mutated pathways were correlated with different DNA-damage-response mechanisms to detect tumors mainly associated with aneuploidy and those with defective DNA repair or microsatellite instability, thus identifying groups of mutated genes that predict patients’ outcomes [ 135 ]. Recent work using deep learning has used pathway information, mutations, and copy-number variation to predict patient response to immunotherapy in melanoma [ 136 ]. An important benefit of these pathway-based approaches is an emphasis on biological interpretation of predictions, which are often considered more important than model performance ( Table 2 ) [ 137 ].…”