A generalisable framework to inject distance information into Alphafold-like structure predictors

Abstract

Structure prediction methods are now highly successful at predicting three-dimensional structures from sequence. However, it is still often desirable to supplement these methods with additional external priors on pairwise distances in the structures. We present a general method for injecting prior information into AlphaFold-like structure predictors by biasing the pair representation to produce desirable features in the distogram, which are then reflected in the structures. We demonstrate this approach to: sample alternate states by selectively pushing or pulling mobile amino acid pairs; integrate NMR NOESY data with structure prediction; and improve the success of protein-protein and protein-ligand complex prediction. We demonstrate that this approach is applicable both to AlphaFold 2 and a reproduction of AlphaFold 3 (OpenFold3). resTrain is open source, available to all users on GitHub and as a Colab notebook: github.com/clami66/resTrain