DAQ-score database: Deep-learning based quality estimation of cryo-EM derived protein models

Abstract

An increasing number of protein structures are determined by cryo-electron microscopy (cryo-EM) as cryo-EM has become one of the most important methods to determine structures. On the other hand, it has been noticed that errors occur in the model building process from cryo-EM maps, probably more frequently than one might think, particularly when the map resolution is not very high. Thus, establishing quality assessment methods has become a crucial and urgent task for biomolecular structure determination with cryo-EM. We have recently developed a quality assessment method to detect protein structural model outliers using machine learning techniques. Our method, called DAQ (Deep-learning-based Amino acid-wise model Quality) score, uses deep neural network to capture local density features of amino acids and atoms in proteins and assesses the likelihood that modeled residues in a structural model is correct (Terashi et al., Nature Methods, 2022). DAQ is also able to detect not only errors in conformations but also shifts in sequence assignment to otherwise correct main-chain conformations, which is often not easy to detect by checking density fitting. Here, we performed a PDB-scale model analysis by DAQ. We applied DAQ to around 10,000 protein structure models in PDB that were derived from cryo-EM maps deposited in Electron Microscopy Data Bank (EMDB). We report the tendency of common errors made in the models through the large-scale analysis. When authors deposited updated structure models to PDB over an initial model, we see clear improvement of DAQ score in the updated version of the model. A common type of errors observed include sequence shifts along alpha helices. Model assessment results with DAQ are made available in a database (https://daqdb.bio.purdue.edu/), where models can be searched by PDB IDs, EMDB IDs and keywords.