Evaluating the Reliability of AlphaFold 2 for Unknown Complex Structures with Deep Learning

Recently released AlphaFold 2 shows a high accuracy when predicting most of the well- structured single protein chains, and subsequent works have also shown that providing pseudo-multimer inputs to the single-chain AlphaFold 2 can predict complex interactions among which the accuracy of predicted complexes can be easily determined by ground truth structures. However, for unknown complex structures without homologs, how to evaluate the reliability of the predicted structures remains a major challenge. Here, we have developed AlphaFold-Eva, a deep learning-based method that learns geometry information from complex structures to evaluate AlphaFold 2. Using different types of sub-complexes of the central apparatus and recently released PDB data, we demonstrate that the reliability of unknown complex structures predicted by AlphaFold 2 is significantly affected by surface ratio, contact surface and dimension ratio. Our findings suggest that the reliability of predicted structures can be directly learned from the intrinsic structural information itself. Therefore, AlphaFold-Eva provides a promising solution to quantitatively validate the predicted structures of unknown complexes, even without a reference.