Machine learning models predict the immunotherapy response in tumors on the basis of DNA methylation.

Abstract

Background: The epigenetic control of immune responses plays a crucial role in the development and progression of cancer. The need to identify biomarkers and create new predictive models is crucial to reliably estimate response rates in patients receiving tumor immunotherapy, which are currently low.

Methods: We screened for variably methylated loci associated with immunotherapy response, focusing on relevant pathways. The expression of immunotherapy-related methylation loci was analyzed in tissues, and quantitative trait locus (QTL) features were summarized. Relationships among tumor mutational Burden (TMB), neoantigen (NeoAg), programmed death-ligand 1 (PD-L1), and immunotherapy efficacy were examined. Motif analysis identified sequence preferences linked to methylation. Six machine learning models were constructed and compared to select the optimal.Due to the lack of external validation datasets, subgroup analysis was conducted.

Results: The five CpG loci showing the strongest response to immunotherapy were cg00045061, cg00107488, cg00056433, cg00090974, and cg00072957. Differentially methylated sites enriched the ubiquitin-proteasome pathway, with most loci located in the N shore region of CpG islands. GO enrichment highlighted microvillus length modulation and CXCR4 chemokine receptor binding. The support vector machine (SVM) model exhibited optimal performance, with precision of 0.796, accuracy of 0.833, F1-score of 0.663, recall of 0.55 and AUC of 0.894.The optimal SVM model achieved an AUC of 0.85 for melanoma dataset and 0.75 for lung cancer dataset.

Conclusion: Tumor methylation sites hold promise as predictive biomarkers for immunotherapy efficacy. The SVM model is the optimal machine learning approach for predicting methylation sites in immunotherapy.