Inferring Protein Sequence-Function Relationships with Large-Scale Positive-Unlabeled Learning.

Area of Science:

• Computational biology
• Protein engineering
• Machine learning

Background:

• Machine learning (ML) offers a powerful way to link protein sequence to function without deep mechanistic insight.
• Supervised ML struggles with large-scale deep mutational scanning (DMS) data due to high dimensionality, correlated variables, experimental noise, missing data, and a lack of negative examples.
• Accurately inferring sequence-function relationships is crucial for protein design and engineering.

Purpose of the Study:

• To develop a robust machine learning framework for inferring sequence-function relationships from large-scale, complex DMS data.
• To address the challenge of missing negative sequence data in typical DMS experiments.
• To enable the prediction of key residues governing protein structure and function.

Main Methods:

• Developed a novel positive-unlabeled (PU) learning framework tailored for sequence-function inference.
• Applied the PU learning method to ten diverse large-scale DMS datasets covering various protein types and library designs.
• Validated the predictive performance and ability to identify critical sequence determinants.

Main Results:

• The PU learning framework demonstrated excellent predictive performance across multiple datasets.
• The model successfully identified key residues that significantly influence protein structure and function.
• The approach proved effective even with the inherent challenges of DMS data, such as missing negative examples.

Conclusions:

• Positive-unlabeled (PU) learning provides an effective solution for analyzing large-scale deep mutational scanning (DMS) data.
• The developed statistical sequence-function model can accurately pinpoint critical residues for protein engineering.
• This methodology facilitates the design of novel proteins with enhanced stability and function, as demonstrated by the creation of highly stabilized enzymes.