Determining the structures of protein complexes can provide insights into their functions and inform the development of vaccines and therapies. Cryo-electron microscopy (cryo-EM) provides near-atomic resolution 3D maps of large molecules, but deriving all-atom structures of protein complexes from cryo-EM data has been challenging.
To accomplish this feat, the authors developed DeepTracer—a fast, accurate, fully automated software tool that leverages a tailored deep convolutional neural network.
The authors applied this technique to a set of 476 cryo-EM maps and compared its performance to that of Phenix—the state-of-the-art cryo-EM model determination method. Deposited model structures available in EMDataResource served as the ground truth.
DeepTracer outperformed Phenix on every metric calculated. DeepTracer created more complete protein structures, achieving 77% coverage of amino acid residues, whereas Phenix correctly determined only 46% of the residues.
DeepTracer also achieved greater spatial accuracy, predicted amino acid types with greater accuracy (50% versus 12%), and connected residues better. The authors obtained similar results when comparing the performance of the two methods at deriving models from 62 coronavirus-related cryo-EM maps.
DeepTracer also modeled large protein complexes exceptionally fast, tracing approximately 60,000 residues within two hours. According to the authors, DeepTracer can help to accelerate the scientific discovery process.
Fast, fully automated software constructs accurate models of protein structure
- 394 views