Protein structure prediction using deep learning distance and hydrogen-bonding restraints in CASP14

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Protein 3D structure prediction is crucial for understanding protein function.
• Existing methods like I-TASSER and QUARK have limitations in accuracy.

Purpose of the Study:

• To report the 3D structure prediction results of the enhanced D-I-TASSER and D-QUARK algorithms in CASP14.
• To evaluate the impact of four newly developed components on prediction accuracy.

Main Methods:

• Integration of DeepMSA2 for multiple sequence alignment (MSA).
• Application of FUpred for protein domain boundary prediction.
• Utilization of DeepPotential (a residual convolutional neural network) for spatial restraint prediction.
• Optimization of spatial restraint energy potentials for structure assembly.

Main Results:

• D-I-TASSER and D-QUARK showed significant TM-score improvements over I-TASSER and QUARK.
• Average TM-score increases were 96% for D-I-TASSER and 112% for D-QUARK.
• All four new components contributed to the observed improvements, particularly DeepPotential and the optimized force field.

Conclusions:

• The novel components enhance protein 3D structure prediction accuracy.
• Challenges remain in modeling multi-domain proteins and protein complexes.
• Further tuning of deep learning predictors is suggested for complex modeling.