DNA Methylation Biomarkers Discovery Approach Using Weighted Causal Multi-Outcome Double Machine Learning Estimation

In this paper a biomarker discovery approach is proposed, where a causal double machine learning method is utilized for identifying DNA methylation biomarkers. Based on multi-outcome causal estimate, biomarkers are assessed with respect to their integrated target expressed genes, the selected biomarkers are then validated for tumor/ normal classification. The paper proposes various multi-outcome weighted functions to calculate the average multi-outcome causal estimate for a given CpG site; correlation, fold change, discriminant analysis weights, and feature importance. The proposed approach is applied on CPTAC-3 Head&Neck cancer in comparison to the commonly LIMMA approach. The proposed approach provided better classification accuracy, however with minimum possible features, the highest reported accuracy 98.79% with 11 features resulted from causal feature importance weight function, while the LIMMA accuracy is 97.67% with 92,815 features.