Beyond sequence: A physics-informed machine learning framework for predicting DNA mutations

This paper investigates how incorporating information from a quantum tight-binding model can enhance the predictive capability of machine learning models for identifying mutation-prone sites in mitochondrial DNA (mtDNA). We employ quantum Hamiltonian techniques and machine learning to explore mutations in mitochondrial DNA's hypervariable segment 1 (HVR1). This region is recognized for its high variability and is frequently used in genealogical DNA testing and research. Our approach considers the local energy associated with each base pair, as well as the interactions among electrons within the DNA chain. For this study, we analyze data from the Mitomap database. Our findings suggest that both the local ionization energies and the context-dependent nature of the base pairs significantly influence the locations of mutations within DNA. Specifically, our machine learning model can extract valuable insights when examining homopolymeric runs-regions where a single base pair repeats multiple times within a sequence.