Improved Protein Structure Prediction Using a New Multi-Scale Network and Homologous Templates

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning in protein science

Background:

• De novo protein structure prediction accuracy has significantly advanced due to deep learning.
• Existing methods like trRosetta provide a foundation for further improvements.

Purpose of the Study:

• To present trRosettaX, an enhanced version of trRosetta for de novo protein structure prediction.
• To improve the prediction of inter-residue geometries and overall structure accuracy.

Main Methods:

• Integration of Res2Net, a novel multi-scale network, for predicting inter-residue distances and orientations.
• Incorporation of an attention-based module to leverage multiple homologous protein templates.
• Application of deep learning techniques for enhanced protein structure modeling.

Main Results:

• trRosettaX demonstrated a 6% and 8% improvement in contact precision on CASP13 and CASP14 free modeling targets, respectively.
• Achieved an average TM-score of approximately 0.8 on CAMEO targets, surpassing top-performing groups.
• A preliminary version ranked among the top server groups in the CASP14 blind test.

Conclusions:

• The multi-scale network (Res2Net) and homologous template exploitation effectively enhance protein structure prediction accuracy.
• trRosettaX represents a significant advancement in computational protein structure prediction.
• The trRosettaX algorithm is now integrated into the trRosetta server.