Distance-based protein folding powered by deep learning.

Area of Science:

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

Background:

• Direct coupling analysis (DCA) is effective for protein folding but struggles with proteins lacking numerous sequence homologs.
• Existing methods often require time-consuming conformation sampling, limiting their applicability.

Purpose of the Study:

• To develop a deep learning method for accurate prediction of interresidue distance distribution in proteins.
• To enable 3D protein model construction using predicted geometric constraints, bypassing extensive conformation sampling.
• To improve protein structure prediction for proteins with limited sequence homologs.

Main Methods:

• Utilized deep learning to predict interresidue distance distribution from protein sequences.
• Employed the resulting distance matrix to generate 3D protein models without intensive conformation sampling.
• Validated the method on CASP12 and CASP13 datasets, including hard targets and membrane proteins.

Main Results:

• Accurately predicted interresidue distances for proteins with as few as ~60 sequence homologs.
• Successfully folded 21 out of 37 CASP12 hard targets, outperforming DCA.
• Achieved high precision in contact prediction and successfully folded challenging membrane proteins in CAMEO.
• Demonstrated feasibility of predicting correct folds even for proteins with limited structural similarity in the Protein Data Bank.

Conclusions:

• Deep learning-based distance prediction offers a powerful alternative to DCA for protein structure prediction, especially for proteins with sparse homologs.
• The developed method significantly advances the ability to model complex protein structures efficiently and accurately.
• This approach broadens the scope of predictable protein folds, including those lacking existing structural templates.