ProteinNPT: Improving Protein Property Prediction and Design with Non-Parametric Transformers

Area of Science:

• Computational biology
• Protein engineering
• Machine learning for science

Background:

• Protein design offers vast potential in drug discovery, materials science, and sustainability.
• Computational protein engineering faces challenges like large design spaces, sparse functional data, and limited labels, especially for multi-property optimization.

Approach:

• Introduced ProteinNPT, a non-parametric transformer variant specifically designed for protein sequences.
• ProteinNPT is optimized for label-scarce and multi-task learning scenarios in protein engineering.
• Developed robust cross-validation schemes for supervised fitness prediction.
• Reimplemented and extended existing baseline methods, integrating diverse protein engineering literature.

Key Points:

• ProteinNPT consistently outperforms established baselines across various protein property prediction tasks.
• Demonstrated robust performance assessment through tailored cross-validation strategies.
• Showcased the utility of ProteinNPT in iterative protein design via in silico Bayesian optimization and conditional sampling.

Conclusions:

• ProteinNPT offers a powerful solution for label-scarce and multi-task protein engineering challenges.
• The developed method significantly advances computational approaches for optimizing protein functions and properties.
• Enables more efficient and effective protein design for diverse scientific and industrial applications.