Genomic and Transcriptomic Predictors of Response to Immune Checkpoint Inhibitors in Melanoma Patients: A Machine Learning Approach

Simple Summary Our work provides novel transcriptomic biomarkers that can accurately predict immune checkpoint inhibitors (ICIs) response in melanoma patients. Using a bioinformatics analysis and supervised machine learning approach, we developed four random-forest classifiers based on clinical, genomic, transcriptomic and survival data. The results of these models can enable further insight into the potential role of these genes in immunotherapy. In addition, our findings were based on a supervised approach, in which melanoma patients treated with ICI were used to retrieve response-associated biomarkers, unlike several studies that used an unsupervised approach based on drug targets to predict ICI response in non-ICI-treated melanoma patients. Apart from ICI response, we also investigated the effect of these biomarkers on overall survival and patients' prognosis, which also revealed a high association with survival, marking these biomarkers as powerful in both ICI response and patients' prognosis. Thus, our work demonstrates a cornerstone in precision oncology and further evaluates these biomarkers in clinical practice using personalized medicine for a better prognosis and response outcomes. Immune checkpoint inhibitors (ICIs) became one of the most revolutionary cancer treatments, especially in melanoma. While they have been proven to prolong survival with lesser side effects compared to chemotherapy, the accurate prediction of response remains to be an unmet gap. Thus, we aim to identify accurate clinical and transcriptomic biomarkers for ICI response in melanoma. We also provide mechanistic insight into how high-performing markers impose their effect on the tumor microenvironment (TME). Clinical and transcriptomic data were retrieved from melanoma studies administering ICIs from cBioportal and GEO databases. Four machine learning models were developed using random-forest classification (RFC) entailing clinical and genomic features (RFC7), differentially expressed genes (DEGs, RFC-Seq), survival-related DEGs (RFC-Surv) and a combination model. The xCELL algorithm was used to investigate the TME. A total of 212 ICI-treated melanoma patients were identified. All models achieved a high area under the curve (AUC) and bootstrap estimate (RFC7: 0.71, 0.74; RFC-Seq: 0.87, 0.75; RFC-Surv: 0.76, 0.76, respectively). Tumor mutation burden, GSTA3, and VNN2 were the highest contributing features. Tumor infiltration analyses revealed a direct correlation between upregulated genes and CD8+, CD4+ T cells, and B cells and inversely correlated with myeloid-derived suppressor cells. Our findings confirmed the accuracy of several genomic, clinical, and transcriptomic-based RFC models, that could further support the use of TMB in predicting response to ICIs. Novel genes (GSTA3 and VNN2) were identified through RFC-seq and RFC-surv models that could serve as genomic biomarkers after robust validation.