Rotamer-free protein sequence design based on deep learning and self-consistency.

Area of Science:

• Computational biology
• Protein engineering
• Artificial intelligence in science

Background:

• Deep learning methods for protein sequence design show promise computationally but lag behind conventional methods in experimental validation.
• Existing methods often require complex reconstruction and optimization of sidechain structures, limiting efficiency.

Purpose of the Study:

• To introduce ABACUS-R, a new deep learning approach for designing amino acid sequences that fold into predefined protein backbones.
• To overcome the limitations of previous computational methods in experimental settings.

Main Methods:

• ABACUS-R employs an encoder-decoder network utilizing multitask learning.
• The model predicts central residue sidechain types based on their 3D local environment, excluding sidechain conformations.
• Iterative application of the encoder-decoder generates self-consistent sequences for a target backbone.

Main Results:

• Experimental results, including five X-ray crystallography structures, demonstrate ABACUS-R's superior performance.
• ABACUS-R surpasses state-of-the-art energy function-based methods in both success rate and design precision.
• The simplified design process eliminates the need for explicit sidechain structure optimization.

Conclusions:

• ABACUS-R represents a significant advancement in computational protein design.
• The method offers a more efficient and accurate approach to designing functional protein sequences for specific backbones.
• This deep learning strategy holds potential for accelerating protein engineering and discovery.