AQuaRef

AQuaRef — AI-enabled Quantum Refinement — replaces the stereochemical restraint libraries used in macromolecular structure refinement with energies and forces from a machine-learned interatomic potential. It was developed by the Phenix and Q|R teams at Lawrence Berkeley National Laboratory with the Isayev lab at Carnegie Mellon University, the Roitberg lab at the University of Florida, the University of Wrocław, and Pending.AI, and published in Nature Communications in 2025.

Refinement is the near-final step of structure determination, where atomic coordinates are adjusted to fit experimental data while restraints hold the chemistry sensible. The standard restraints — tabulated bond lengths, angles, planes, and chirality from libraries of known chemical entities — cover no meaningful noncovalent interactions, must be hand-annotated for novel ligands or cross-links, and at low resolution are weak enough that practitioners bolt on extra Ramachandran, secondary-structure, and rotamer terms. Quantum mechanics would solve all of this at once, but density functional theory scales as O(N³) and refinement needs whole-molecule energies and gradients thousands of times.

AQuaRef closes that gap with a neural network potential built on the AIMNet2 architecture and trained specifically for polypeptides. It is trained once and thereafter applied unchanged to any new protein structure, so quantum-level restraints become a fixed cost rather than a per-structure calculation. This is the first reported case of a machine-learned potential performing quantum refinement of an entire protein rather than a QM/MM-partitioned active site.

#Key Features

• Linear-scaling energies and forces: Compute time and GPU memory both scale as O(N) with system size. Energy and forces for a 100,000-atom system take 0.5 seconds on a single NVIDIA H100, and a model of roughly 180,000 atoms fits in 80 GB.
• No parametrization step: The potential predicts energies directly from atoms and total charge, so unusual local geometry needs no manual restraint definitions.
• Proton placement: It resolves protonation states that library restraints bias, demonstrated on the short dimer-spanning hydrogen bonds of the parkinsonism-associated protein DJ-1 and its bacterial homolog YajL.
• Practical runtimes: Quantum refinement takes roughly twice as long as standard refinement, and often less than standard refinement with added secondary-structure and rotamer restraints — under 20 minutes for about 70% of the test models, on a laptop GPU.
• Shipped in Phenix: Available from Phenix dev-5395 onward through the open-source Q|R package, not as a standalone research script.

#Technical Details

The potential keeps the AIMNet2 core — rotation-invariant learnable features, message passing for many-body interactions, and explicit handling of total system charge — but is trained from random initialization rather than adapted from released AIMNet2 weights. Explicit long-range Coulomb and dispersion terms are dropped in favor of training directly to total DFT-D4 energy, and a short-range exponential repulsion term from GFN1-xTB is added for robustness against clashes. Training data came from SMILES enumerations of mono- through tetra-peptides covering all 20 standard amino acids, 11 alternate protonation forms, seven terminal caps, and disulfide- and diselenide-linked pairs, plus complexes of two to four peptides in random orientations — never naturally occurring sequences or observed interactions, so no structural data could leak in. Query-by-committee active learning over four iterations selected roughly one million GFN-FF-sampled conformations (median 42 atoms each) for labeling at the B97M-D4/def2-QZVP level with CPCM implicit water. Training ran for 1.5 million steps at batch size 256 against DFT energies, forces, and Hirshfeld partial charges.

Evaluation covered 41 cryo-EM and 30 X-ray models, 61 of them low-resolution with a high-resolution homolog as ground truth. AQuaRef restraints gave systematically better MolProbity scores, Ramachandran Z-scores, CaBLAM statistics, and hydrogen-bond geometry than standard restraints — with or without added Ramachandran, secondary-structure, and rotamer terms — while matching or improving fit to the data, including a smaller Rfree−Rwork gap for X-ray models. Against AMBER, Rosetta, REFMAC5, and Servalcat, AQuaRef and Rosetta led on geometry, with Rosetta's sampling-based optimization landing marginally closer to the reference models.

#Applications

AQuaRef is aimed at structural biologists finishing cryo-EM or crystallographic models, particularly at low resolutions where library restraints are least trustworthy and manual annotation most burdensome. Distributed inside Phenix, it slots into existing phenix.refine and phenix.real_space_refine workflows, and its ability to settle protonation states makes it useful for enzyme mechanism and drug-design questions that hinge on where a proton sits.

#Impact

The work reframes machine-learned interatomic potentials — usually a computational chemistry tool — as infrastructure for experimental structure determination, and makes quantum-quality refinement of whole proteins routine rather than heroic. Two limitations are explicit: the potential is trained only on common amino acid residues, so AQuaRef applies to protein-only models and not nucleic acids or novel ligands, and Q|R does not yet handle static disorder modeled as alternate conformations. The Q|R code is Apache-2.0 licensed and actively maintained, but the potential itself is not covered by that grant: the TorchScript weights are downloaded at runtime from a separate model-zoo repository that carries no license of its own.