SO3LR Tauri stretches one force field to 85 elements: a single set of weights runs proteins, nucleic acids, membranes and platinated DNA

A single force field now covers proteins, nucleic acids, lipids, carbohydrates, water and organic solvents, ions, metal-containing biomolecules and reaction pathways, without separate parameter sets for each class. The earlier SO3LR was trained on 4 million structures over eight elements and reached 200k atoms on one GPU; classical empirical force fields are parameterized separately for water, ions and metal centers, so simulating several biomolecular classes together requires combining parameters that were not designed to be used together. On the coupled-cluster-referenced SAPT10k set, SO3LR-T-M reaches an interaction-energy MAE of 0.45 kcal/mol, half the 0.90 kcal/mol of SO3LR; on PLA15, a set of protein–ligand active sites, it reaches 3.81 kcal/mol, a relative error of 2.1%.

The model adds learned local features to three explicit physical terms, and evaluates long-range electrostatics and dispersion with particle-mesh Ewald summation, so neither has to be truncated in large periodic systems. SO3LR used a ZBL-type screening function fitted during training for short-range repulsion, and scaled free-atom dispersion coefficients by a single learned Hirshfeld volume ratio, a single scalar rescaling too restrictive across the periodic table. Short-range repulsion is replaced by pair-specific NLH potentials pre-tabulated from dedicated DFT calculations, and the network predicts per-atom polarizability and C6 ratios; across four independent NPT replicas at 300 K and 1 atm, water and fifteen organic solvents give a mean relative density error of 6.3%, against 2.4% for OPLS4.

The same pretrained weights run a platinum-containing DNA crosslink and an unmodified B-DNA dodecamer, with metal coordination maintained by the potential itself. Classical force fields typically derive platinum parameters one complex at a time and fix the coordination geometry with improper torsions, so neither the coordination number nor the charge distribution can respond to the environment. Across four replicas of the cisplatin 1,2-d(GpG) crosslink, Pt(II) stays four-coordinated in every frame with no ligand exchange and no bonded restraints or improper torsions imposed; Pt–N bond lengths deviate from crystallographic averages within the threshold the source states.

The demonstrated scope is static benchmarks, transition-state searches for three reactions, and dynamics of water, organic solvents, salt solutions, peptide pKa, proteins, a glycoprotein, a POPC membrane, platinated DNA and a B-DNA dodecamer, all without system-specific fitting. The authors also state that only the proline-catalyzed aldol transition states were found blindly, while the benzoin and cobalamin methyl-transfer saddles were refined from published structures.

The authors list work still to be done: predictive accuracy for static and dynamic liquid properties, reaction barriers, coverage of chemical and conformational space, and the sampling and inference speed needed for biomolecular processes. They also report a systematic overestimation of organic-liquid densities, possibly partly from nuclear quantum effects; water self-diffusion about half the experimental value with a still-elevated first oxygen–oxygen peak; a peptide pKa MAE of 0.98 units with primary amines underestimated; and an excess sugar-pucker population near O4′-endo in B-DNA.