Abstract
We describe a modified version of AlphaFold2 that incorporates experimental distance distributions into the network architecture for protein structure prediction. Harnessing the OpenFold platform, we fine-tune AlphaFold2 on structurally dissimilar proteins to explicitly model distance distributions between spin labels determined from Double Electron-Electron Resonance (DEER) spectroscopy. We benchmark the performance of the modified AlphaFold2, refer to as DEERFold, in switching the predicted conformations of a set of membrane transporters using experimental DEER distance distributions. Guided by sparse sets of simulated distance distributions, we showcase the generality of DEERFold in predicting conformational ensembles on a large benchmark set of water soluble and membrane proteins. We find that the intrinsic performance of AlphaFold2 substantially reduces the number of required distributions and the accuracy of their widths needed to drive conformational selection thereby increasing the experimental throughput. The blueprint of DEERFold can be generalized to other experimental methods where distance constraints can be represented by distributions.