trRosetta

trRosetta predicts de novo protein tertiary structures by using deep neural networks to infer inter-residue geometries from amino acid sequences and guiding Rosetta energy minimization to produce atomic models.

Key Features:

• Deep Learning Integration: A deep neural network predicts inter-residue geometries, including distances and orientations, from an amino acid sequence to guide structure determination.
• Structure Refinement: Predicted geometries are converted into restraints that direct energy minimization within the Rosetta framework to generate final models.
• Rapid and Accurate Predictions: For proteins of approximately 300 amino acids, trRosetta can produce a final structural model in about one hour using up to 10 CPU cores in parallel.
• Homology Template Incorporation: Homologous templates are automatically incorporated as additional inputs to combine homology modeling information with de novo predictions.
• trRosettaX Enhancements: trRosettaX introduces a Res2Net multi-scale network and an attention-based module to better predict inter-residue geometries and utilize multiple homologous templates.
• Benchmark Performance: trRosettaX achieved an average TM-score of approximately 0.8 on 161 CAMEO targets (June–September 2020) and improved contact precision by ~6% on CASP13 and ~8% on CASP14 free modeling targets, outperforming top groups.

Scientific Applications:

• Functional Annotation: Predicted structures enable interpretation of functional implications of amino acid sequences and putative active or binding sites.
• Drug Discovery: Structural models can identify potential binding sites and inform structure-based drug design and virtual screening.
• Protein Engineering: Predicted models support engineering efforts by allowing evaluation of how sequence modifications may affect structure and function.

Methodology:

A deep neural network predicts inter-residue distances and orientations from the amino acid sequence; these predicted geometries are converted into restraints that guide energy minimization within Rosetta to produce atomic models. trRosettaX uses a Res2Net multi-scale network and an attention-based module to incorporate multiple homologous templates.