PrISM: precision for integrative structural models

Integrative modeling of macromolecular structures usually results in an ensemble of models that satisfy the input information. The model precision, or variability among these models is estimated globally, i.e., a single precision value is reported for the model. However, it would be useful to identify regions of high and low precision. For instance, low-precision regions can suggest where the next experiments could be performed and high-precision regions can be used for further analysis, e.g., suggesting mutations. We develop PrISM (Precision for Integrative Structural Models), using autoencoders, a type of unsupervised deep neural network, to efficiently and accurately annotate precision for integrative models. The method is benchmarked and tested on five examples of binary protein complexes and five examples of large protein assemblies. The annotated precision is shown to be consistent with localization density maps, while providing more fine-grained information. Finally, the generated networks are also interpreted by gradient-based attention analysis. Significance Statement Validation of integrative models and data is an open research challenge. This is timely due to the new worldwide Protein Data Bank archive for integrative structures (http://pdb-dev.wwpdb.org). Currently, a single precision value is reported for an integrative model. However, precision may vary for different regions of an integrative model owing to varying amounts of information available for different regions. We develop a method using unsupervised deep learning to efficiently and accurately annotate precision for regions of an integrative model. This will ultimately improve the quality and utility of deposited structures.