E2EDA: Protein Domain Assembly Based on End-to-End Deep Learning.

Area of Science:

• Computational biology
• Structural biology
• Deep learning applications

Background:

• Deep learning has advanced single-domain protein structure prediction.
• Predicting multi-domain protein structures, especially inter-domain orientations, remains a significant challenge.
• Accurate modeling of multi-domain proteins is crucial for structure-based drug discovery.

Purpose of the Study:

• To develop an end-to-end deep learning method for protein domain assembly.
• To improve the accuracy and efficiency of full-chain protein structure modeling.
• To provide insights into structure-based drug discovery through enhanced protein modeling.

Main Methods:

• Developed RMNet, an EfficientNetV2-based model using attention mechanisms to predict inter-domain rigid motion.
• Transformed predicted rigid motions into spatial transformations for direct full-chain model assembly.
• Designed RMscore for selecting the optimal model from multiple assembled candidates.

Main Results:

• E2EDA achieved an average TM-score of 0.827 on a benchmark set, surpassing SADA (0.792) and DEMO (0.730).
• Reassembled models using E2EDA showed a 7.0% higher TM-score compared to AlphaFold2 predictions on a custom dataset.
• E2EDA demonstrated significant efficiency gains, reducing runtime by 64.7% compared to SADA and 19.2% compared to AlphaFold2.

Conclusions:

• E2EDA offers a highly accurate and efficient end-to-end approach for multi-domain protein structure prediction.
• The method effectively captures inter-domain orientations, improving upon existing state-of-the-art models like AlphaFold2.
• E2EDA's speed and accuracy provide valuable tools for structure-based drug discovery research.