Single-sequence protein structure prediction using a language model and deep learning.

Area of Science:

• Computational biology
• Structural biology
• Deep learning

Background:

• Protein structure prediction is crucial for understanding biological function.
• Current methods like AlphaFold2 rely on multiple sequence alignments (MSAs), limiting their use for orphan or rapidly evolving proteins.
• Exploring designed protein structures and understanding folding rules remain challenges.

Purpose of the Study:

• To develop a novel deep learning model for protein structure prediction that overcomes limitations of MSA-dependent methods.
• To enable accurate structure prediction for proteins lacking sufficient sequence data.
• To accelerate the design and analysis of novel protein structures.

Main Methods:

• Developed an end-to-end differentiable recurrent geometric network (RGN).
• Integrated a protein language model (AminoBERT) to learn from unaligned protein sequences.
• Utilized a geometric module for invariant representation of protein backbone geometry.

Main Results:

• RGN2 demonstrates superior performance compared to AlphaFold2 and RoseTTAFold on orphan and designed proteins.
• Achieved significant computational efficiency, with up to a 10^6-fold reduction in compute time.
• Validated the effectiveness of protein language models for structure prediction beyond MSAs.

Conclusions:

• Protein language models offer a powerful alternative to MSAs for protein structure prediction.
• The developed RGN model advances the field by enabling accurate and efficient prediction for challenging protein classes.
• This work highlights the potential of deep learning in deciphering protein structure and function.