Improving protein tertiary structure prediction by deep learning and distance prediction in CASP14

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Artificial Intelligence in Biology

Background:

• Deep learning and residue-residue distance prediction have significantly advanced protein structure prediction since CASP13.
• The MULTICOM system has been a participant in previous Critical Assessment of protein Structure Prediction (CASP) experiments.

Purpose of the Study:

• To enhance the MULTICOM protein structure prediction system for the CASP14 experiment.
• To improve both template-free (ab initio) and template-based tertiary structure prediction accuracy.
• To develop advanced deep learning-based model quality assessment methods.

Main Methods:

• Incorporated a new deep learning-based inter-residue distance predictor for template-free prediction.
• Enhanced the template-based tertiary structure prediction methodology.
• Utilized deep learning for distance-based structural model quality assessment.

Main Results:

• MULTICOM ranked seventh in tertiary structure prediction and third in inter-domain structure prediction at CASP14.
• Deep learning-based template-free modeling accurately predicted topology for most targets, including difficult ones.
• Template-free modeling outperformed template-based modeling in many cases, highlighting its robustness.

Conclusions:

• Deep learning and accurate distance prediction are crucial for high-performance, template-free protein structure modeling.
• The enhanced MULTICOM system demonstrates significant progress in protein structure prediction accuracy.
• Further improvements in model quality assessment are needed for challenging targets with limited good model predictions.