Loading…
Protein conformational ensemble generator using SE(3) flow matching from a perturbed ESMFold prior, sampling MD-like dynamics from sequence alone.
No providers recorded yet. Browse all providers
A crystal structure of adenylate kinase shows the enzyme clamped shut around its substrate. In solution it also spends time with its lid swung open, and traffic between those two states regulates cellular energy balance. A single-structure predictor returns one snapshot and no signal that the other exists; molecular dynamics recovers the full distribution but scales badly and tends to stay in whichever basin it starts in.
P2DFlow samples protein backbone conformational ensembles directly from sequence using SE(3) flow matching. It was developed at Lingang Laboratory with the Shanghai Institute of Materia Medica, posted as a preprint in 2024 and published in the Journal of Chemical Theory and Computation the following year. Two design decisions define it: where generation starts, and what tells its outputs apart. Rather than flowing from Gaussian noise or a harmonic prior, P2DFlow flows from an ESMFold prediction perturbed by Gaussian noise on each residue's coordinates. That prior already carries plausible bond lengths and torsion angles, so the network spends its capacity on conformational change rather than on rebuilding chemistry — unlike AlphaFlow, which fine-tunes the structure predictor itself.
The second decision is a scalar the authors call "approximate energy". Each frame of a training simulation is projected onto a plane of radius of gyration against RMSD to the crystal structure, and a Gaussian kernel density over that plane is inverted through the Boltzmann relation into an energy-like label. Conditioning on it tells the model which region of the landscape a sample belongs to, discouraging it from filling the gap between minima with states no simulation visits. At sampling time it is drawn from a Boltzmann distribution and held fixed.
Residues are graph nodes, with edges from a coordinate-distance and a sequence-distance cutoff. Node and pair representations come from a frozen ESM-2 encoder; the SE(3) equivariant block predicts translation and rotation vector fields along an optimal transport interpolation path, with side-chain torsions from an MLP and coordinates reassembled through the AlphaFold frame convention. Training follows FrameFlow, with Kabsch pre-alignment of the prior against the target. The training set is the ATLAS database of roughly 1,300 all-atom MD simulations, reduced to 100 representative structures per protein spread across the approximate-energy range; the single released checkpoint took eight days on four NVIDIA A100-80G GPUs.
Evaluation used about 100 ATLAS ensembles held out from training. Against AlphaFlow and Str2Str, P2DFlow led on all ten reported metrics: steric validity 0.941 (versus 0.725 and 0.667), root mean Wasserstein distance 15.168 (versus 33.628 and 78.246), and Jaccard similarity of 0.710 on weak contacts and 0.422 on transient contacts. On a test protein, its samples covered all three cluster centres of the reference simulation in principal component space.
The direct use is a proxy for molecular dynamics when the question is which conformations a protein visits, not the kinetics of moving between them. Weak contacts that dissociate across the ensemble mark loop regions that may open into cryptic pockets — a starting point for drug discovery against targets whose crystal form shows no druggable site. Per-residue RMSF profiles track simulated flexibility, giving protein engineers a cheap first read on which regions tolerate mutation, and sampled conformations seed targeted MD runs.
P2DFlow argues that the prior matters as much as the architecture: moving the starting point from noise to a perturbed folded structure buys validity and fidelity without a larger network. Its scope is bounded by ATLAS, which covers single-chain globular proteins; no multimer or multidomain evaluation is reported. The ESMFold prior also biases sampling toward whichever state the predictor favours; the authors observe this on D-ribose binding protein, where most samples cluster near the closed form, and note the same effect in BioEmu-1. Re-seeding mitigates it but adds a manual step. Code is public and the pretrained checkpoint is released, though only through a Google Drive folder. The repository's README asserts GPL-3.0, but it ships no LICENSE file and GitHub reports it as unlicensed, so those terms are claimed rather than granted. ATLAS, both the training and the evaluation set, is licensed CC BY-NC 4.0, so the preprocessed archives distributed alongside the checkpoint inherit a non-commercial upstream.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.