MACHINE LEARNING IN EPIGENOMIC LANDSCAPE INTERPRETATION

Abstract

The regulation of the genome is an important part of the history in order to understand the mechanisms of disease and gene regulation.  To demonstrate how regulatory states can be characterised and predictive genomic features identified, we also present a machine learning-based platform that analyses multi-layer epigenomic data where we now integrate DNA methylation, histone modification, and chromatin accessibility profiles.  Before a wide range of models was trained in the framework of supervised methods including the series of ensemble methods and the gradient-orientation classifiers, data was pre-processed, and normalised, along with the engineered-feature extraction.  In comparison with the linear approach, cross-validation provided the evidence that ensemble algorithms with specific results (Random Forests and Gradient Boosting), performed better as predictors of the stage of damage to the car (accuracy > 92%).  SHAP-based interpretation indicated that the higher regulatory indicators were enhancer accessibility, active histone mark (H3K4me3, H3K27ac) and promoter CpG sites.  Correlation analysis validated functional significance because the up-regulation of gene expression was positively linked with histone acetylation, and the reduced level of transcript abundance was negatively related to promoter methylation.  Motif enrichment analysis identified transcription factors with related aspects of development and lineage commitment as common pieces enriched in predictive areas.  Those findings demonstrate that machine learning can present biologically grounded explanatory insights on how genes are regulated in addition to presenting high precise grouping of epigenomic states.  This integrative approach enables both the translational applications of precision medicine (with potential biomarker generation) and mechanistic studies because it provides the foundation of predictive epigenomics.