AI‑Driven Quantum Refinement: AQuaRef Enables Full‑Atom Protein Model Optimization with Quantum Constraints

The article presents AQuaRef, a new AI‑driven quantum refinement approach for protein structures proposed by a joint research team from Carnegie Mellon University, the University of Wrocław, and the University of Florida. Built on the AIMNet2 machine‑learning atomic potential and specially trained for refinement tasks, AQuaRef attains computational efficiency comparable to classical force fields while reproducing quantum‑mechanical accuracy, offering a novel pathway for full‑atom quantum refinement of biomacromolecules.

Million‑Sample Training Set for Peptide Machine‑Learning Potentials

To train a robust potential, the authors constructed a dataset covering three dimensions: chemical composition, conformational space, and intermolecular interactions. Small‑peptide databases were generated from SMILES strings, encompassing 20 standard amino acids, 11 protonation states, 3 N‑terminal and 4 C‑terminal modifications. All mono‑ and di‑peptides were enumerated, with random selections of tri‑ and tetra‑peptides, plus disulfide‑linked and seleno‑analogues. OpenEye Omega performed dense torsional sampling without chiral constraints, enabling applicability to D‑, L‑, and mixed‑stereochemistry peptides.

Complexes of 2–4 peptide fragments were also built, randomizing spatial orientation to mimic intermolecular interactions. All generated structures contained ≤120 atoms (including hydrogens) to keep calculations tractable.

Initial conformations were first relaxed with the GFN‑FF force field while preserving overall geometry via Cartesian constraints. Subsequently, a query‑by‑committee active‑learning loop was employed: 500 k random samples trained an ensemble of four models; four iterative rounds selected high‑uncertainty structures for DFT (B97M‑D4/def2‑QZVP) evaluation and added them to the training pool; the final round prioritized low‑energy, high‑uncertainty boundary structures. The process yielded ~1 M samples (average 42 atoms each).

Experimental structures from RCSB and EMDB were also harvested for validation, filtered by atom count (1 000–10 000 non‑hydrogen atoms), resolution (2.5–4 Å), MolProbity clash score (<50), and geometry deviations within four times the standard limits.

Model Architecture and Training Details

AQuaRef inherits the core architecture of AIMNet2 but introduces two key modifications for refinement. First, long‑range Coulomb and dispersion terms are omitted; instead, the model is trained to reproduce DFT‑D4 total energies, justified by the screening effect of the CPCM implicit solvent and the negligible contribution of interactions beyond 5 Å to refinement forces.

Second, an explicit short‑range exponential repulsion term from GFN1‑XTB is added, improving stability when resolving steric clashes. Training targets include DFT‑derived energies, atomic forces, and Hirshfeld charges. The network starts from random weights, uses a batch size of 256, and runs for 1.5 M steps, with all other hyper‑parameters identical to the original AIMNet2 setup.

Computational Efficiency

Benchmarking on an NVIDIA H100 GPU (80 GB) shows that both energy and force evaluation times, as well as peak memory usage, scale linearly (O(N)) with the number of atoms. For a protein of ~100 k atoms, a single‑point energy‑force calculation takes ~0.5 s; the implementation can handle up to ~180 k atoms on the same hardware.

Benchmarking on Low‑Resolution Structures

The authors assembled a test set of 61 low‑resolution models (41 cryo‑EM, 20 X‑ray) each paired with a high‑resolution reference. Three refinement regimes were compared: (1) AQuaRef’s quantum‑based constraints, (2) standard geometric restraints, and (3) standard restraints plus additional hydrogen‑bond and secondary‑structure constraints.

Results indicate that quantum‑refined models achieve markedly better MolProbity scores and Ramachandran Z‑scores than traditional methods, while maintaining comparable agreement with experimental data. For X‑ray structures, the R_work–R_free gap shrinks slightly; for cryo‑EM structures, CC_mask drops modestly but EMRinger scores remain stable, suggesting reduced over‑fitting.

Comparison with Established Refinement Packages

AQuaRef was further benchmarked against AMBER, Rosetta, REFMAC5 (for X‑ray) and Servalcat (for cryo‑EM). Across the board, AQuaRef yields the lowest R_free and the least over‑fitting. Its geometric quality rivals Rosetta and surpasses REFMAC5 and Servalcat. Both AQuaRef and Rosetta generate realistic hydrogen‑bond geometries, whereas AMBER performs moderately and REFMAC5/Servalcat often fail to recover these details.

Case Studies: Short Hydrogen Bonds in DJ‑1 and YajL

To assess proton‑position handling, the authors refined symmetric doubly‑protonated DJ‑1 and its homolog YajL. Traditional restraints force bond lengths toward database‑derived non‑protonated standards, whereas AQuaRef places protons consistently with unrestrained refinements. Even when experimental data are truncated to 2 Å resolution, AQuaRef recovers structures nearly identical to the original 1.15 Å data, correctly locating the proton on D24 Oδ2 in DJ‑1. In YajL, the shared proton between E14 and D23 yields a low‑barrier hydrogen bond, reflected in a flat energy surface from AIMNet2 and corroborated by >3σ peaks in difference electron‑density maps.

Broader Context and Related Advances

The article notes parallel progress in the field, such as Oxford’s nn‑tm‑fcc neural‑network potentials achieving near‑quantum accuracy for residue‑level fragments, and a German team’s BF‑DCQO algorithm that accelerates a 12‑amino‑acid folding problem from 72 h on GPU clusters to ~4 min on an IonQ quantum processor.

Overall, the integration of quantum mechanics, machine‑learning potentials, and experimental data provides a promising new route for protein structure refinement, especially for low‑resolution models, ligand‑binding analyses, and functional site investigations.