Carbonara: A Rapid Method for SAXS-Based Refinement of Protein Structures

Generative machine learning models for protein structure prediction are primarily trained on X-ray crystallography data, which captures proteins in crystal lattices that can deviate significantly from their native solution conformations. Biological small angle X-ray scattering (bioSAXS) offers valuable solution-state insights, but creating atomic models that rationalise this data remains challenging. Here we present Carbonara, a rapid computational pipeline that combines coarse-grained sampling with experimental constraints to efficiently identify solution-state conformations from an initial atomic model. We demonstrate Carbonara's effectiveness by refining an AlphaFold-predicted model of the DNA repair helicase, SMARCAL1, and a crystallographically determined structure of the antigen binding domains of the anti-hCD40 monoclonal antibody, ChiLob7/4, a clinically relevant immunostimulatory antibody. In both cases, Carbonara identifies physiologically relevant solution-state conformations separated from crystal-like predictions by large energy barriers, achieving in minutes what traditional MD simulations might not accomplish in weeks.