Toward Chemical Accuracy in Biomolecular Simulations through Data-Driven Many-Body Potentials: III. A Library of Amino Acid Building Blocks

Extending data-driven many-body potentials to peptides and proteins requires reusable building blocks that capture the chemical diversity of amino acids. Building on the many-body expansion (MBE) of the energy, we construct an intramolecular MB-nrg library for all 20 natural amino-acid residues by combining physics-based interactions with machine-learned (ML) n-body terms trained on DLPNO-CCSD(T) n-mer reference data and described by permutationally invariant polynomials. Covalently bonded functional groups define the bodies of the MBE, allowing each ML term to be reused wherever the same chemically defined nmer occurs. Without training on whole-dipeptide energies, the assembled MB-nrg potentials predict gas-phase Ramachandran potential energy surfaces with a pooled root-mean-square deviation (RMSD) of 1.26 kcal/mol relative to DLPNO-CCSD(T), compared with 6.65 and 5.43 kcal/mol for Amber ff14SB and ff19SB force fields, respectively. Because the MB-nrg potentials are modular, the remaining deviations from the DLPNO-CCSD(T) reference can be systematically analyzed within the MBE framework and addressed by expanding the library with additional nonbonded and higher-body terms, as demonstrated for asparagine. The library thus provides an initial set of reusable n-body building blocks for peptide potentials and a blueprint for their systematic refinement and expansion toward coupled-cluster accuracy in biomolecular simulations.