Loading…
Zero-shot protein conformation sampling that perturbs an input structure and anneals it back with a score network trained only on crystal structures.
No providers recorded yet. Browse all providers
A structure predictor answers "what does this protein look like?" with one model, but a protein that binds, gates, or signals moves between states, and which states it visits is the part that explains the biology. Molecular dynamics answers that from physics and pays in wall-clock: barrier crossings are rare events, so an honest ensemble can mean microseconds to milliseconds of simulated time. Learning-based samplers are far faster, but most are trained on trajectories of the very protein being studied, which reintroduces the cost they were meant to avoid and leaves them with nothing to say about a new target.
Str2Str removes that dependence. It is a score-based structure-to-structure translation framework: given one input conformation — typically an ESMFold or AlphaFold 2 prediction, or a deposited crystal structure — it runs a forward–backward process borrowed from simulated annealing. A forward diffusion perturbs the backbone geometry up to a transition time, and a reverse diffusion anneals it back under a learned score function, yielding a new conformer of the same protein. Because the perturbation stops short of full noise, the identity of the input survives, so the process explores around the starting structure rather than generating a backbone from scratch. The score network is trained once on general crystal structures and sees no simulation data at training or inference, which is what makes sampling zero-shot.
The work is by Jiarui Lu, Bozitao Zhong, Zuobai Zhang, and Jian Tang at Mila, Université de Montréal, and HEC Montréal, and was published at ICLR 2024. It sits in the conformational-ensemble family alongside AlphaFlow, BioEmu-1, Distributional Graphormer, and EBA.
The score model, DenoisingIPA, is a variant of the AlphaFold 2 structure module: four Invariant Point Attention layers interleaved with two multi-head attention layers, a 256-channel single representation and a 128-channel pair representation updated by edge transition layers. Training uses an amortized denoising score matching objective — the data distribution is the PDB rather than any one protein's ensemble — complemented by backbone coordinate and distogram losses. The training set is 26,459 monomeric mmCIF entries from a 9 June 2023 PDB snapshot, filtered to better than 5 Å resolution and 10–512 residues, with sequence-identity matches to the test proteins removed; training took roughly 30 GPU-days on a V100.
Evaluation covers 12 fast-folding proteins with reference all-atom MD from Lindorff-Larsen et al. (2011), sampling 1,000 conformations per target. With SDE sampling, Str2Str reaches 0.977 clash validity and 0.982 bond validity, Jensen-Shannon divergences against full MD of 0.348 on pairwise distances, 0.400 on TICA components and 0.365 on radius of gyration, and diversity errors of 0.133 (TM-score) and 0.184 nm (RMSD) — ahead of MSA subsampling, EigenFold and idpGAN. On the WW domain it matched the fidelity of a 100 µs explicit-solvent simulation in 510 GPU-seconds against more than 160 GPU-days. A BPTI case study recovered four of the five kinetic clusters known from millisecond-scale simulation.
The practical use is getting an ensemble for a protein nobody has simulated: screening candidates for flexibility or cryptic-pocket formation before committing GPU budget, generating diverse starting structures to seed selective MD or docking runs, checking whether a designed sequence has one stable fold or several, and interpreting observables such as B-factors or NMR order parameters. Because a target costs seconds rather than days, the model fits as a triage layer in front of simulation.
Str2Str established that useful conformational sampling can be learned from static crystallographic data alone, and its forward–backward formulation is architecture-agnostic enough to carry over to newer diffusion and flow-matching backbones. The authors are candid about the limits: the isotropic perturbation kernel is not steered toward any collective variable, the sampler still trails long MD on distributional metrics in general, evaluation is entirely in-silico against MD references, and reaching unbiased Boltzmann sampling would need system-specific fine-tuning. The code is MIT-licensed, with the pretrained checkpoint distributed through a Google Drive link rather than a versioned release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.