Machine learning interatomic potentials in biomolecular modeling: principles, architectures, and applications

Classical force fields remain widely used in molecular modeling due to their efficiency but fail to accurately capture reactivity and complex environments. Quantum mechanical methods like DFT offer higher accuracy but are computationally prohibitive for large biomolecules. Machine learning interatomic potentials (MLIPs) bridge this gap by approximating potential energy surfaces with near-DFT precision while enabling large-scale simulations. MLIPs learn directly from quantum data and can generalize across diverse chemical environments. This review outlines the theoretical basis of MLIPs, including training on energy and force data, symmetry constraints, and common architectures-ranging from descriptor-based to graph-based and equivariant neural networks. Key applications are examined in biomolecular contexts: conformational sampling, enzymatic catalysis, and ligand binding. Integration with molecular dynamics packages like OpenMM and LAMMPS is increasingly streamlined. While challenges remain-such as generalization to out-of-distribution systems, limited interpretability, and data scarcity-ongoing advances in datasets, hybrid modeling, and infrastructure are rapidly improving practical adoption. MLIPs represent a major step forward in atomistic simulations and are poised to become central tools in structural biology, enzymology, and computational drug discovery.