Developed by the Reciprocal Space Station consortium, which includes the University of Cambridge in the UK, Columbia University in the US, and the Flatiron Institute, and published in Nature Methods in April 2026, ROCKET (Refinement of structures using Coevolutionary Knowledge and Experimental Targets) is an innovative refinement framework that combines AI-based protein structure prediction models with physical experimental data. This system delves into the deep learning architecture of OpenFold (an open-source reimplementation of AlphaFold2) and directly integrates raw physical experimental data, such as X-ray crystallography diffraction data or Cryo-EM and Cryo-ET density maps, into the model to fine-tune the combined structure at the atomic level. By simultaneously learning from user-provided structural templates and experimental waveform signals, it acts as a next-generation precision control hub for structural biology, outputting refined models that not only predict but also align with real physical entities. Traditional structure refinement tools like Phenix or COOT use Cartesian coordinate optimization, which directly manipulates atomic coordinates in 3D space. This approach often leads to getting stuck in local minima or causing physically impossible chemical distortions, especially in low-resolution or noisy data environments. In contrast, ROCKET overcomes these limitations by modifying and biasing the representation of the internal latent space, specifically the multiple sequence alignment (MSA) cluster profiles and coevolutionary embeddings within OpenFold, rather than forcibly fitting physical coordinates. As an analogy, while the traditional approach is like correcting typos in a completed manuscript one by one, potentially ruining the context, ROCKET is like an editor who guides the author's thoughts (the latent space) with the overall structure and intent of the writing, allowing the entire text to be rewritten naturally and coherently. Thanks to this approach, ROCKET can naturally and accurately correct even macroscopic structural changes, such as large-scale domain rearrangements or challenging loop flips, without violating physical constraints or evolutionary statistical norms. In real-world biotechnology and drug discovery research, this tool proves highly effective in analyzing low-resolution or highly flexible membrane protein structures. For example, researchers with a Cryo-EM map with a resolution of 4Å or lower and a rough initial PDB model can bypass complex manual atomic modeling steps and obtain reliable protein-ligand binding structures with just a few GPU optimization sessions using ROCKET's rk.refine pipeline. Even in laboratories with limited GPU resources, the phenix.rocket remote submission feature, which uses the Phenix server, allows complex multi-complex refinements to be processed in minutes without requiring direct investment in expensive hardware infrastructure, significantly shortening the research cycle.

Developed by the Reciprocal Space Station consortium, which includes the University of Cambridge in the UK, Columbia University in the US, and the Flatiron Institute, and published in Nature Methods in April 2026, ROCKET (Refinement of structures using Coevolutionary Knowledge and Experimental Targets) is an innovative refinement framework that combines AI-based protein structure prediction models with physical experimental data. This system is based on OpenFold (an open-source version of AlphaFold2).