Improved prediction of the response duration to MAPK inhibitors in patients with advanced melanoma using baseline genomic data and machine learning algorithms

PurposeBaseline genomic data have not demonstrated significant value for predicting the response to MAPK inhibitors (MAPKi) in patients with BRAFV600-mutated melanoma. We used machine learning algorithms and pre-processed genomic data to test whether they could contain useful information to improve the progression-free survival (PFS) prediction. Experimental designThis exploratory analysis compared the predictive performance of a dataset that contained clinical features alone and supplemented with baseline genomic data. Whole and partial exon sequencing data from four cohorts of patients with BRAFV600-mutated melanoma treated with MAPKi were used: two cohorts as training/evaluation set (n = 111) and two as validation set (n = 73). Genomic data were pre-processed using three strategies to generate eight different genomic datasets. Several machine learning algorithms and one statistical algorithm were employed to predict PFS. The performance of these survival models was assessed using the concordance index, time-dependent receiver operating characteristic (ROC) curve and Brier score. ResultsThe cross-validated model performance improved when pre-processed genomic data, such as mutation rates, were added to the clinical features. In the validation dataset, the best model with genomic data outperformed the best model with clinical features alone. The trend towards improved prediction with baseline genomic data was maintained when data were censored according to the two clinical setting scenarios (duration of clinical benefit and progression before 12 months). ConclusionIn our models, baseline genomic data improved the prediction of response duration and could be incorporated into the development of predictive models of response pattern to MAPKi in melanoma. Translational RelevanceCurrently, biomarkers are lacking to robustly predict the response to therapy targeting the MAPK pathway in advanced melanoma. Therefore, in the clinic, a trial-and-error approach is often used. Baseline genomic mutation profiles represent a comparably stable biological readout that is easily accessible and measurable in clinical routine. Therefore, they might represent candidate predictive biomarker signatures. However, previous studies could not show a clear predictive signal for the response to MAPK inhibitors (MAPKi) in patients with BRAFV600-mutated melanoma. Here, our exploratory machine learning-based analysis highlighted an improved prediction of progression-free survival when clinical and genomic data were combined, even when using only partial exome sequencing data. This suggests that baseline genomic data could be incorporated in the development of predictive models of the response to MAPKi in advanced melanoma by leveraging the results of current routine partial exome sequencing. Interest statementThe authors declare no potential conflicts of interest.