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

Area of Science:

• Genetics
• Computational Biology
• Quantum Chemistry

Background:

• Mitochondrial DNA (mtDNA) is crucial for cellular energy production and is prone to mutations.
• The hypervariable segment 1 (HVR1) of mtDNA is frequently studied due to its high variability and use in genetic research.
• Predicting mutation hotspots in mtDNA is challenging but important for understanding disease and evolution.

Purpose of the Study:

• To enhance machine learning models for identifying mutation-prone sites in mtDNA.
• To integrate quantum tight-binding model information into predictive models.
• To investigate the influence of local energy and electron interactions on mtDNA mutations.

Main Methods:

• Employed quantum Hamiltonian techniques and machine learning algorithms.
• Analyzed mutations in mitochondrial DNA's hypervariable segment 1 (HVR1).
• Incorporated local energy of base pairs and inter-electron interactions into the model.
• Utilized data from the Mitomap database for analysis.

Main Results:

• Quantum tight-binding information significantly improved machine learning model predictions for mutation sites.
• Local ionization energies and context-dependent base pair interactions were identified as key factors influencing mutation locations.
• The model demonstrated particular effectiveness in analyzing homopolymeric runs within DNA sequences.

Conclusions:

• Integrating quantum mechanical properties enhances the predictive accuracy of machine learning models for mtDNA mutation sites.
• Understanding local energy and electronic interactions is vital for pinpointing mutation-prone regions in DNA.
• This approach offers a novel method for analyzing genetic variability and mutation dynamics in mitochondrial DNA.