Computational & AI Journal of Chemical Information and Modeling

A solvent-aware neural network potential corrects ligand strain estimates, especially for charged molecules that gas-phase references systematically misjudge

  • A neural potential trained on quantum-chemical data with an implicit solvent model handles neutral and charged ligands uniformly, and GPU batching makes solvated global-minimum searches feasible for whole libraries.
  • Applied retrospectively to a large set of crystal-structure ligands, strain grows with charge; the most highly charged anions are least reliable and are treated as upper bounds.
  • Gas-phase force fields match neutral cases but overstate charged-ligand strain by collapsing the unbound reference; adding implicit solvent largely fixes this, and strain proves pose-specific rather than a ligand property.
Points Based on the abstract

Retrospective strain on crystal ligands has no experimental ground truth, so agreement between methods only shows consistency, not correctness. Before trusting it, I'd check the neural potential against DFT with the same implicit solvent on our own chemotypes, since training-set coverage of unusual charged groups is the usual weak spot. Then I'd rank-order a congeneric series by strain, and treat large values as a prompt to inspect the pose or the ligand's protonation state before blaming the ligand.