Single-sequence protein structure prediction by integrating protein language models

Area of Science:

• Computational biology
• Structural biology
• Deep learning applications

Background:

• Deep learning has significantly advanced protein structure prediction.
• Current leading methods, like AlphaFold2, necessitate multiple sequence alignments (MSA), which are not biologically representative of natural protein folding.
• There is a need for MSA-free protein structure prediction methods.

Purpose of the Study:

• To develop and evaluate RaptorX-Single, a novel single-sequence-based protein structure prediction method.
• To compare the performance of RaptorX-Single against MSA-based and other MSA-free methods.
• To investigate the impact of protein language models on prediction accuracy.

Main Methods:

• Integration of multiple protein language models with a structure generation module.
• Development of RaptorX-Single, a deep learning framework for MSA-free protein structure prediction.
• Comparative analysis against AlphaFold2 and other MSA-free predictors on various protein datasets.

Main Results:

• RaptorX-Single demonstrates significantly faster computation times compared to MSA-based methods.
• The method achieves superior prediction accuracy for antibodies, proteins with few homologs, and single mutation effects.
• Performance is influenced by both the scale and training data of the underlying protein language models.
• RaptorX-Single shows competitive results even when compared to MSA-based AlphaFold2 for proteins with abundant homologs.

Conclusions:

• RaptorX-Single offers a viable and efficient alternative for protein structure prediction, especially in scenarios where MSAs are unavailable or limited.
• The study highlights the importance of protein language model characteristics for prediction performance.
• This MSA-free approach broadens the applicability of deep learning in structural biology.