Explainable machine learning identifies features and thresholds predictive of immunotherapy response

The identification of reliable predictors for immunotherapy response remains a critical challenge in precision oncology. Here, we integrated patient clinical features with DNA and RNA sequencing data derived from 229 melanoma samples. We use 138 melanoma samples to develop and train machine learning models predicting patient treatment response. Feature selection included a univariate Mann-Whitney U-test screen (p <  = 0.05) prior to unique tailoring for each model using SHAP. We identify random forest as the optimal classifier, on an independent cohort of 53 melanoma patients. We test the usefulness of the model using additional test data comprising patients with stable disease (n = 15) and non-cutaneous melanoma (n = 23). Using SHAP, we identified features linked to treatment response including mutation features, immune cell-type abundance and LAG3 expression, highlighting the contributions of both tumour intrinsic and extrinsic factors. Further leveraging SHAP we infer potential numerical thresholds for certain features separating good versus poor immunotherapy response. This explainability driven methodology has potential implications for the clinical translation of biomarker discovery and guiding treatment options to enable precision oncology.