Conformation-aware ML charges make cheap MM-PBSA rank ligands better
- A machine-learning/molecular-mechanics framework replaces fixed charges in endpoint free-energy calculations with ML charges that depend on binding state and conformation, using two embedding schemes.
- It was tested retrospectively on multiple targets from the JACS and Merck benchmark sets.
- Both variants ranked affinities better than classical MM-PBSA at acceptable extra cost.
- A charge-source control and a PTP1B case study link the gain to better handling of local electrostatics.
💡 Points Based on the abstract
The charge-source control is the part to read first: it tests whether the gain comes from the polarization-aware charges rather than from just swapping in a different charge model. Still, the evaluation is retrospective and the summary names no comparison with alchemical free-energy methods, the usual lead-optimization benchmark. Try it on one of your own series with known affinities.
Incorporating Polarization into ML/MM Endpoint Calculations for Protein–Ligand Affinity Ranking ↗