Partial Charges

CHEESE uses Geometric Transformers to predict partial atomic charges, replacing expensive DFT (Density Functional Theory) calculations. It predicts two AI-derived charge types and computes two classical baselines.

Charge types

Training

• 70,000 DFT calculations on diverse drug-like chemotypes
• QM level: ωB97X-D/def2-svp (dispersion-corrected DFT with split-valence basis)
• Validated on an out-of-distribution scaffold-split test set (Morgan Tanimoto < 0.3)

Model architecture

Two parallel Geometric Transformers (one for ESP, one for RESP), each consisting of:

• transformer.pt (~41 MB) — a geometry-aware transformer that takes atomic features + 3D coordinates and outputs learned representations
• projection.pt (~2.7 KB) — a final linear layer mapping features to one charge value per atom

Built on the en-transformer package (equivariant/geometric transformer).

Input features

• Atomic numbers (integer per atom)
• 3D Cartesian coordinates (from MMFF-optimized conformer)
• Bond matrix (symmetric adjacency matrix with integer-encoded bond types, 13 states)

Inference pipeline

Key references

• Bayly, C. I.; Cieplak, P.; Cornell, W.; Kollman, P. A. A well-behaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model. J. Phys. Chem. 1993. Link
• Alenaizan, A.; Burns, L. A.; Sherrill, C. D. Python implementation of the restrained electrostatic potential charge model. Int. J. Quantum Chem. 2020. Link
• Gasteiger, J.; Marsili, M. Iterative partial equalization of orbital electronegativity — a rapid access to atomic charges. Tetrahedron Lett. 1978. Link
• Tosco, P.; Stiefl, N.; Landrum, G. Bringing the MMFF force field to the RDKit: implementation and validation. J. Cheminform. 2014, 6, 37. Link