Loading…
Machine learning force field for all-atom protein dynamics, trained on 40 million DFT dipeptide conformations covering backbone and side-chain space.
No providers recorded yet. Browse all providers
A machine-learned force field trained on backbone dihedral scans has watched arginine's side chain visit only three of its nine rotameric states. Drive a simulation into a fourth and the network is extrapolating: in the trajectories reported here, hydrogens on a lysine side-chain nitrogen are flung out of their bonded positions and the run collapses partway through. Side chains determine molecular recognition and ligand binding, so a potential that models them by extrapolation fails at the questions it is usually asked.
The PUD+ model is the fix Dianwei Wu and Tong Wang at Tsinghua University propose: a ViSNet-PIMA equivariant graph neural network trained on the Protein Unit Dataset Plus, a 40-million-conformation quantum-chemistry corpus of capped dipeptides in which side-chain degrees of freedom are sampled as systematically as the backbone. The side-chain half comes not from a grid scan of phi and psi but from replica-exchange molecular dynamics, clustered to representative structures and labelled with density functional theory. It covers all nine chi1/chi2 rotameric states as continuous distributions, where backbone scanning reaches three.
The trained checkpoint is a fixed potential, not a per-system fit. It runs inside the AI²BMD simulation framework, which decomposes a polypeptide into dipeptide units and queries the model for each unit's energies and forces at every step — so a network that has only ever seen capped dipeptides drives dynamics of proteins absent from its training set.
At full scale the model stacks eight ViSNet layers and one PIMA layer, extending vector-scalar interactive message passing with explicit treatment of non-local interactions. Node and edge embeddings are 128-dimensional, angular information is expanded in spherical harmonics to degree l_max = 1, and the radius graph uses a 15 Å neighbor cutoff. Energy and force terms combine in a single mean-squared-error objective weighted 1:50. The corpus pairs the 20-million-conformation backbone subset inherited from AI²BMD with 20 million new side-chain conformations, all computed at the M06-2X/6-31G* level in ORCA and costing roughly 12.5 million CPU core-hours for the new half alone; 5% of each dipeptide's data is held out as a two-million-sample test set.
At matched training-sample counts, the side-chain-enriched corpus lowers test error from 0.01414 to 0.002576 kcal/mol per atom on potential energies and from 0.1825 to 0.05984 kcal/mol/Å on atomic forces. In a 250 ps AI²BMD simulation the model reproduces the full extended → 3₁₀-helix → α-helix folding transition of Ace-Ala15-Nme, with errors against DFT references of 0.05118 kcal/mol per atom on relative potential energy and 1.7211 kcal/mol/Å on forces, against 0.9473 and 23.6177 for Amber ff19SB on the same frames.
The model targets all-atom simulation where the answer depends on side-chain behavior: rotamer changes on ligand binding, hydrophobic core packing during folding, and the flexible interfaces of signaling motifs. Its demonstrated regime is short trajectories of small polypeptides in vacuum or explicit solvent. The underlying corpus reaches a second audience: refitting the ff19SB CMAP term against it yields ff19SB-PUD+, which produces more expanded ensembles for the 526-residue disordered protein FUS and tighter contact distributions for p53 transactivation domain complexes, benefiting groups running conventional Amber simulations without a neural potential.
The work makes a testable claim about machine-learned force fields: that coverage of side-chain conformational space, not corpus size alone, sets the ceiling on how far a dipeptide-trained potential generalizes. The scaling curve and the trajectory-stability comparison give it quantitative support, and the CMAP result shows the same data improving a force field with no learned component. The limits are equally clear. Dipeptide fragments cannot represent interactions spanning several residues, which the authors identify as the case for extending sampling to tripeptides and tetrapeptides, and the corpus contains no lipids or post-translational modifications, leaving membrane proteins outside its scope. The preprint has not been peer reviewed, is licensed CC-BY-NC, and announces neither a release of the PUD+ dataset nor a checkpoint for the trained model — though the AI²BMD simulation program and the original backbone dataset are publicly available under the MIT license.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.