Fig. 1: Data-driven structure prediction refinement with ROCKET.

a, Experimental techniques such as X-ray crystallography and cryo-EM provide high-fidelity observations of conformational states, capturing, for example, ligand-induced changes (right; activation loop of human c-Abl kinase trapped in the inactive state by a small-molecule binder; PDB 3PYY) that may not be modeled by ML-based predictions (left; AF2 prediction of c-Abl in the untrapped state). b, ROCKET extends OpenFold by integrating crystallographic and cryo-EM likelihood targets within its differentiable prediction pipeline. It accomplishes this by learning, at inference time, multiplicative and additive adjustments to MSA cluster profiles that maximize their agreement with experimental data as computed by an experimental likelihood function. For the crystallographic target, the function (L_xtal) depends on observed crystallographic diffraction intensities and their measurement errors (I_o, σ_I) and on the structure factor amplitudes computed from the predicted model (F_c). For the cryo-EM target, the function (L_cryo) depends on the complex Fourier terms from experimental half-maps (F₁ and F₂) and the complex structure factors computed from the predicted model (F_c).