Machine learning has revolutionised structural biology, with tools like AlphaFold2 enabling rapid protein structure prediction from sequence alone. Yet challenges remain in modelling details such as side-chain arrangements and condition-dependent conformational changes, partly due to limited training data. Meanwhile, experimental techniques like cryo-EM and X-ray crystallography are producing structural data faster than it can be interpreted into accurate atomic models.
This paper from Alisia Fadini and the rest of the Read lab (plus international collaborators) introduces ROCKET, an extension of AlphaFold2 that directly incorporates experimental data to refine predicted structures. Rather than optimizing in physical coordinate space, ROCKET works within AlphaFold2's internal coevolutionary representation, capturing biologically meaningful structural variation, particularly in noisy, low-signal or low-resolution datasets, that neither AlphaFold2 nor existing automated tools can access. ROCKET requires no retraining and offers a scalable framework for combining machine learning with experimental structural biology.