Semantical and geometrical protein encoding toward enhanced bioactivity and thermostability

Area of Science:

• Synthetic biology
• Computational biology
• Biochemistry

Background:

• Protein engineering modifies amino acids for novel functions, requiring accurate prediction of variant effects.
• Current deep learning methods primarily use protein sequences, limiting geometric and structural information crucial for stability and function.
• Evaluating protein thermostability prediction is vital but often overlooked in existing models.

Purpose of the Study:

• To develop a novel pre-training framework integrating sequential and geometric encoders for protein primary and tertiary structures.
• To guide protein mutation directions towards desired traits by simulating natural selection.
• To enhance the in silico assessment system for protein engineering models.

Main Methods:

• Developed a pre-training framework combining sequential and geometric encoders for protein structures.
• Simulated natural selection to guide mutation directions for specific protein traits.
• Evaluated variant effects based on functional fitness using deep mutational scanning assays.

Main Results:

• The proposed framework demonstrated exceptional prediction performance across three benchmarks (over 300 assays).
• Achieved superior results compared to other zero-shot learning methods with minimal trainable parameters.
• Showcased enhanced accuracy and comprehensiveness in predicting protein variant effects.

Conclusions:

• The novel framework improves predictions for efficient protein engineering by integrating structural and sequential data.
• The study enhances in silico assessment systems for deep learning models in protein science.
• The approach offers a cost-effective method for accurate protein variant effect prediction.