Protein conformation and dynamics generation from MD data, sampling trajectories, independent ensembles, and interpolations between two known states.
No providers recorded yet. Browse all providers
Deep generative models trained on molecular dynamics data have split into two camps. Ensemble generators such as AlphaFlow and ConfDiff sample conformations that approximate a protein's equilibrium distribution but say nothing about the order in which those states are visited. Trajectory models capture temporal structure but cannot produce independent samples on demand. Neither formulation covers both, and the physical reality — a protein's accessible states and the paths between them — is a single object.
ConfRover, from ByteDance Seed, models conformation and dynamics together by treating an MD trajectory as an autoregressive sequence. Each frame is generated conditioned on the frames before it, using the causal transformer formulation familiar from language modeling. Because the conditioning context can be empty, a single frame, or a start-and-end pair, the same trained model supports forward simulation, time-independent sampling, and interpolation between two conformational states — the first model to do all three in one framework.
Trained on the ATLAS dataset with a test split by protein identity, ConfRover is evaluated on proteins whose structures and dynamics it has never seen. It was presented at NeurIPS 2025, with code and weights released under Apache-2.0.
The encoding layer draws single and pair representations from the Evoformer stack of a pretrained OpenFold model with frozen weights, after three recycling iterations; the diffusion decoder is initialized from ConfDiff and the remaining modules are trained from scratch. The released models carry 19.6M trainable parameters and are trained on the ATLAS set of roughly 1,300 proteins with triplicate 100 ns trajectories each, split by protein identity. A second checkpoint, ConfRover-interp, continues training the base model with a 1:1:1 mix of trajectory, single-frame, and interpolation objectives.
On a multi-start benchmark of about 2,700 generation conditions drawn from 82 held-out proteins, ConfRover's Pearson correlation with reference trajectory conformational change reaches 0.75 at the trajectory level and 0.63 at the frame level, against 0.56 and 0.47 for MDGen; in projected 2D PCA space the gap widens to 0.73 versus 0.18. On 100 ns simulation, state recovery improves to 0.51 Jensen-Shannon distance and 0.58 F1 against MDGen's 0.56 and 0.42, with an MD oracle at 0.31 and 0.79, and tICA-derived principal dynamic modes track the oracle closely. For time-independent sampling, ConfRover is comparable to AlphaFlow and ConfDiff, outperforming at least one of them on five of eight ensemble metrics despite not being specialized for that task.
ConfRover offers a fast proxy for short MD runs in workflows that need conformational context rather than a single structure: identifying cryptic pockets and alternative binding-competent states for structure-based drug design, characterizing loop and domain flexibility, and generating starting ensembles for downstream docking or free-energy calculation. The interpolation mode targets a distinct question — sketching plausible paths between two known conformational states, such as apo and holo forms — where no comparable model for large proteins existed.
By casting MD as next-frame prediction, ConfRover imports the training efficiency and conditioning flexibility of autoregressive sequence models into molecular simulation, and shows a general-purpose model can match specialists at their own task. The authors are direct about what remains: current checkpoints handle only single-chain proteins; ATLAS's 100 ns trajectories bias the learned dynamics toward short-timescale, local motions rather than large functional transitions; diffusion on backbone SE(3) space with reconstructed torsions loses accuracy for large rotamer changes; and the triangular updates in the structure module make scaling to larger proteins and longer trajectories computationally costly. A gap to MD-oracle fidelity in state recovery also persists.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.