Inter-domain distance prediction based on deep learning for domain assembly.

Area of Science:

• Computational biology
• Structural biology
• Deep learning

Background:

• AlphaFold2 excels at single-domain protein structure prediction but struggles with multi-domain proteins due to inaccurate inter-domain interactions.
• Accurate prediction of multi-domain protein structures is crucial for understanding protein function and biological processes.

Purpose of the Study:

• To develop an advanced inter-domain distance prediction method to improve multi-domain protein structure prediction.
• To enhance the accuracy of protein domain assembly using deep learning.

Main Methods:

• Developed DeepIDDP, a neural network incorporating attention mechanisms and novel inter-domain features for predicting distances between protein domains.
• Introduced DPMSA, a data augmentation strategy to address the lack of co-evolutionary information.
• Integrated DeepIDDP with the SADA domain assembly method, creating SADA-DeepIDDP.

Main Results:

• SADA-DeepIDDP achieved 11.3% and 21.6% higher inter-domain distance prediction accuracy compared to trRosettaX and trRosetta, respectively.
• The SADA-DeepIDDP domain assembly model showed a 2.5% improvement over the original SADA method.
• Reassembled 68 human multi-domain protein models, improving the average TM-score by 11.8%.

Conclusions:

• DeepIDDP significantly improves inter-domain distance prediction and protein domain assembly accuracy for multi-domain proteins.
• The developed method offers a valuable tool for enhancing the quality of protein structure models, particularly those generated by AlphaFold2.
• The SADA-DeepIDDP server is available online for researchers to improve multi-domain protein structure predictions.