This study explored the robustness and uniqueness of the flexible fitting of atomic structures against cryo-electron microscopy (cryo-EM) maps using elastic network motion models. The success of flexible fitting is based on the optimistic expectation of a single optimum fit that can be reached from a wide range of start conformations. We revisited this assumption for four AlphaFold models that deviated from corresponding medium-resolution cryo-EM maps but benefitted from flexible fitting. To test the dependence of the flexible fitting performance on the start structures, we systematically generated decoys using normal modes, offering a broader sampling of the conformational space compared to a single start structure. This strategy allowed exploration of the global properties of the cross-correlation (CC) scoring function landscape. Statistical analysis using multidimensional scaling revealed that the initial decoy ensembles collapsed into multiple basins of attraction in three of the four cases. The results demonstrate that a single start structure can be trapped in the local maxima of the CC during flexible fitting (spurious fits), but the decoys increase the likelihood of finding a correct fit. More precisely, there is a "winning" cluster of closely related structures that exhibit high template modeling (TM)-scores with the known true structures. Comparison of the CC and the TM-scores showed that the winning cluster can be identified by high CC values, further demonstrating the utility of cryo-EM maps as filters for screening candidate structures.

AlphaFold 3 (AF3) has recently been shown to offer improved accuracy in predicting the structures of protein multimers. Improved models may lead to new opportunities for fitting them to cryo-electron microscopy (cryo-EM) maps with medium resolution (5-10 Å). Deriving atomic models from such cryo-EM maps is still challenging due to the lack of high-resolution features. Our case study involving four AF3 multimer models and corresponding cryo-EM maps with 7-8 Å resolution showed that the predicted multimer models were partially correct. The predicted models contained fairly accurate domains, secondary structures, and individual chains, since 9 of the 17 chains exhibit TM-scores higher than 0.8 and 16 chains had TM-scores above 0.5 compared with the official atomic structures that were deposited with the cryo-EM maps. However, some cases exhibited incorrect relative positions of individual chains or domains. We observed that the order of cross-correlation (CC) scores between the multimers and their corresponding cryo-EM maps aligned with the order of the TM-scores. This shows that if regions are masked correctly, CC scores are sensitive enough to distinguish among the multimer models. A masking of monomeric chains may not always be attainable, so we also explored the level of accuracy in secondary structure segmentation for one of the cases in greater detail. Although molecular details are not fully visible in cryo-EM maps at medium resolution, the location of major secondary structures, such as α-helices and β-sheets, were detectable using our DeepSSETracer tool. Our analysis illustrates the potential for improvements in the accuracy of AF3-predicted multimer models by combining the density map-model similarity (CC scores) and the secondary structure map-model similarity in a future approach.

Understanding the biochemical mechanisms that drive protein-protein interactions is a challenging task, traditionally requiring mutation studies and expert interpretation of protein structures. A method that can generate mechanistic explanations from the biochemical properties and contributions of interactions would enhance our ability to study protein-protein interactions. In this study, we present a novel approach to interpreting mechanistic insights from machine learning methods; we manually annotated a dataset of 1225 mutation experiments with mechanistic insights focused on electrostatic, hydrogen bonding, steric and hydrophobic interactions. To show a preliminary process for evaluating mechanism prediction models, we extracted SHAP features that are representative of protein binding mechanisms from a Gradient-Boosting Tree (GBT) model trained to predict binding affinity. We found that the SHAP values generally agreed with the annotated mechanisms from our dataset, especially when looking at electrostatic and steric features. We also found that hydrophobicity consistently played a dominant role and hydrogen bonds consistently played a secondary role, challenging conventional assumptions about the role of these interactions.