Deep learning geometrical potential for high-accuracy ab initio protein structure prediction.

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning applications

Background:

• Accurate protein structure prediction is crucial for understanding biological function.
• Deep learning models have advanced the prediction of inter-residue contact and distance maps.
• Existing methods still face challenges in accurately predicting complex protein structures.

Purpose of the Study:

• To introduce DeepPotential, a novel deep learning model for ab initio protein structure prediction.
• To improve the accuracy of predicting geometric descriptors, including a new hydrogen-bonding potential.
• To enhance the quality of full-length protein structure construction using predicted restraints.

Main Methods:

• Development of the DeepPotential deep learning model.
• Prediction of geometric descriptors using C-alpha atom coordinates.
• Utilizing a multi-tasking network architecture and metagenome-based Multiple Sequence Alignment (MSA) collection with confidence-based selection.
• Incorporating hydrogen-bonding and inter-residue orientation predictions.

Main Results:

• DeepPotential achieved higher Top-L/5 contact accuracy (4.1%) and TM-score (6.7%) for full-length models compared to other deep learning restraint prediction approaches.
• Significant advantages were observed in both geometrical feature prediction and full-length structure construction on CASP and CAMEO targets.
• Analysis revealed that multi-tasking architecture and advanced MSA strategies contributed to the performance gains.

Conclusions:

• DeepPotential represents a significant advancement in deep learning-guided ab initio protein structure prediction.
• The model's ability to predict novel geometric descriptors, like hydrogen-bonding potential, improves structural accuracy.
• These findings pave the way for more reliable and accurate prediction of protein structures.