Deciphering Chemical Environments in Conformational Data Sets Using an IQA Band-Matching Protocol

In this work, we introduce BMIQA, a band matching (BM) protocol that harnesses the physical information encoded in topological atomic energies to identify equivalent chemical environments across conformational data sets. By analyzing distributions of atomic energies obtained through the Interacting Quantum Atoms (IQA) formalism, BMIQA discovers chemically equivalent environments semiautomatically without relying on any structure-based representation, marking a shift toward a fully physics-driven atom-typing strategy. We validate our approach on data sets comprising seven capped amino acids, three oligopeptides, monomeric water, and water pentamers. We also extend our analysis to uncapped and sulfur-containing amino acids. Strikingly, capped aliphatic and aromatic amino acids without side-chain heteroatoms are made of only seven atomic environments: three carbon types, two hydrogen types, one oxygen type, and one nitrogen type. This compact and minimal representation persists in homopeptides, and the number of atom types increases only slightly in heteropeptides, remaining far smaller than the total atom count. BMIQA captures subtle physical equivalences. In cysteine, it reveals that the SH and CH substructures correspond to equivalent hydrogen environments, consistent with the similar electronegativities of carbon and sulfur. While isolated water exhibits two local environments, water pentamers display four, including two oxygen and two hydrogen types arising from variable hydrogen bonding. We further show that the energetic fingerprints that characterize atomic environments depend strongly on the level of theory and that Gaussian process regression models trained on IQA energies reproduce the same environments when deployed in molecular dynamics simulations. By providing the means to build a rigorous atom-typing scheme, BMIQA constitutes a robust and physically grounded framework for next-generation atom-typed machine learning force fields.