An enhanced genetic mutation-based model for predicting the efficacy of immune checkpoint inhibitors in patients with melanoma.

Background: Programmed death ligand 1 (PD-L1) and tumor mutation burden (TMB) have been developed as biomarkers for the treatment of immune checkpoint inhibitors (ICIs). However, some patients who are TMB-high or PD-L1-high remained resistant to ICIs therapy. Therefore, a more clinically applicable and effective model for predicting the efficacy of ICIs is urgently needed. Methods: In this study, genomic data for 466 patients with melanoma treated with ICIs from seven independent cohorts were collected and used as training and validation cohorts (training cohort n = 300, validation cohort1 n = 61, validation cohort2 n = 105). Ten machine learning classifiers, including Random Forest classifier, Stochastic Gradient Descent (SGD) classifier and Linear Support Vector Classifier (SVC), were subsequently evaluated. Results: The Linear SVC with a 186-gene mutation-based set was screened to construct the durable clinical benefit (DCB) model. Patients predicted to have DCB (pDCB) were associated with a better response to the treatment of ICIs in the validation cohort1 (AUC=0.838) and cohort2 (AUC=0.993). Compared with TMB and other reported genetic mutation-based signatures, the DCB model showed greater predictive power. Furthermore, we explored the genomic features in determining the benefits of ICIs treatment and found that patients with pDCB were associated with higher tumor immunogenicity. Conclusion: The DCB model constructed in this study can effectively predict the efficacy of ICIs treatment in patients with melanoma, which will be helpful for clinical decision-making.