SPDesign: protein sequence designer based on structural sequence profile using ultrafast shape recognition.

Area of Science:

• Computational Biology
• Protein Engineering
• Bioinformatics

Background:

• Current deep learning methods for protein sequence design often overlook crucial protein-specific physicochemical features.
• Existing approaches primarily focus on optimizing network architectures, limiting their effectiveness.

Purpose of the Study:

• To introduce SPDesign, a new protein sequence design method that incorporates structural sequence profiles.
• To enhance protein sequence design by integrating structural information with deep learning.

Main Methods:

• SPDesign utilizes ultrafast shape recognition to identify similar protein structures in a database.
• Sequence profiles are extracted via structure alignment and combined with pre-trained structural knowledge and geometric features.
• An enhanced graph neural network is employed for sequence prediction.

Main Results:

• SPDesign achieved significant accuracy gains in sequence recovery rate compared to state-of-the-art methods (ProteinMPNN, Pifold, LM-Design).
• The method demonstrated strong performance on benchmarks with limited homologous sequences (orphan and de novo).
• SPDesign successfully reconstructed sequences for proteins with similar structures but different sequences and ensured designed sequences fold into native structures.

Conclusions:

• SPDesign represents a significant advancement in protein sequence design by effectively utilizing structural information.
• The method's superior performance and ability to handle diverse protein types suggest broad applicability in biopharmaceutical research and therapeutic development.