ROCKET

ROCKET — Refining OpenFold with Crystallographic/Cryo-EM liKElihood Targets — turns a frozen protein structure predictor into an experimental refinement engine. Structural biologists routinely start model building from a predicted structure, but converting raw density into a mechanistically correct atomic model remains a bottleneck, especially below 4–5 Å resolution where sidechains blur and conventional refinement gets stuck in local minima. ROCKET closes that gap by making the structure predictor itself the prior, and steering it with the experimental data.

The key move is where the optimization happens. Standard refinement software adjusts Cartesian coordinates under geometric restraints, which cannot cross the energetic barriers separating one loop conformation or domain orientation from another. ROCKET instead performs gradient descent in the space of multiple sequence alignment cluster profiles — the continuous representation from which AlphaFold2 derives its geometry. Because a small change in that representation can produce a large, coherent change in the predicted structure, barrier-crossing rearrangements become reachable. Crucially, no weights are updated: the pretrained network is used purely at inference time.

ROCKET was developed by Alisia Fadini, Minhuan Li and colleagues across Columbia University, the University of Cambridge, Harvard University and the New Mexico Consortium, posted as a preprint in February 2025 and published in Nature Methods in 2026. It is built on OpenFold, the trainable open reimplementation of AlphaFold2, which supplies the differentiable forward pass ROCKET needs.

#Key Features

• Refinement in embedding space: Optimization targets the MSA cluster profile rather than atomic coordinates, so refinement can reorganize secondary structure, flip a peptide, or shift a domain.
• Multiple experimental modalities: Differentiable maximum-likelihood target functions are provided for X-ray diffraction intensities and cryo-EM half maps, and the same framework accommodates cryo-ET sub-tomogram averages.
• No retraining required: The pretrained network stays frozen, so ROCKET avoids the cost and data hunger of fine-tuning a structure predictor per data type.
• MSA subsampling to escape local minima: When the initial prediction is far from the observed state, ROCKET generates alternative starting models by subsampling the input alignment and selects the one with the highest experimental likelihood before refining.
• Portable target functions: The crystallographic and cryo-EM likelihood modules are separable from the refinement loop and can augment other structure prediction methods.

#Technical Details

Each refinement iteration runs a forward pass through OpenFold, aligns the prediction to the data by molecular replacement or cryo-EM rigid-body docking, evaluates the experimental likelihood target and its derivatives, and updates the MSA cluster profile. A final local refinement with phenix.refine cleans up geometry and atomic displacement parameters. The current implementation operates on one protein chain at a time. On a benchmark of 27 single-chain crystallographic datasets solved better than 3 Å, ROCKET improved Cα RMSD to PDB-REDO reference models for all but one target, with a mean 0.47 Å drop across the ten hardest cases; real-space correlation coefficients improved for both backbone and sidechains, and R-free fell by more than 3% for ten models. On the serpin PAI-1 (1.8 Å), an AlphaFold2 prediction of the metastable active state sat 10.7 Å from the latent-state structure with a 40.4 Å reactive-centre-loop deviation; MSA subsampling plus refinement brought both to 1.9 Å. ROCKET also captured ligand-induced loop rearrangements in c-Abl kinase and PTP-1B, subunit rebuilding of GroEL from cryo-EM, and structures from a 3.82 Å HAI-1 dataset and a 9.60 Å GroEL sub-tomogram average. A control experiment optimizing pLDDT instead of the experimental target failed to improve agreement with the data, showing the gain comes from the measurements rather than the network's own confidence.

#Applications

ROCKET fits into crystallography and cryo-EM pipelines at the point where an automated model has been placed in the density but disagrees with it in functionally important regions. It is most valuable for conformational states poorly represented in training data — ligand-bound loop configurations, alternative domain orientations, serpin latent states — and for low-resolution regimes such as in situ cryo-ET, where manual rebuilding is slow and often inconclusive. The direct beneficiaries are structural biologists running high-throughput crystallography and cryo-ET groups working past the usual limits for automated model building. The code is distributed under the Apache 2.0 licence with tagged releases.

#Impact

ROCKET demonstrates that a pretrained structure predictor can serve as an implicit prior for experimental structure determination without any retraining, which reframes the relationship between machine learning models and experimental data in structural biology. Rather than treating AlphaFold2 as a fixed oracle whose output is then edited, it treats the network's latent sequence representation as the refinement variable. Practical constraints remain: memory limits confine the current implementation to a single chain at a time, and ligands are not explicitly modelled during refinement. Publication in Nature Methods and the release of separable, differentiable crystallographic and cryo-EM target functions make the components reusable by other structure prediction methods.