Data-driven protease engineering by DNA-recording and epistasis-aware machine learning.

Area of Science:

• Biochemistry
• Molecular Biology
• Computational Biology

Background:

• Machine learning (ML) tools have advanced protein structure prediction.
• Predicting protein catalytic activity and designing sequences with desired functions remain significant challenges.
• Current limitations stem from insufficient experimental data and inefficient exploration of vast protein sequence spaces.

Purpose of the Study:

• To engineer proteases with tailored substrate specificity.
• To overcome limitations in generating large-scale sequence-activity data for ML.
• To develop a data-efficient ML model for predicting protease function.

Main Methods:

• Development of a DNA recorder for deep specificity profiling of proteases in Escherichia coli.
• Parallel testing of 29,716 candidate proteases against 134 substrates.
• Generation of ~600,000 protease-substrate pair data.
• Application of epistasis-aware training set design for ML model optimization.

Main Results:

• Identification of key sequence determinants controlling protease specificity.
• Creation of a data-efficient deep learning model capable of accurately predicting protease sequences.
• Demonstration of desired on- and off-target activities prediction.
• Validation of epistasis-aware training set design for efficient sequence space exploration.

Conclusions:

• The developed DNA recorder and ML approach enable precise engineering of protease specificity.
• Epistasis-aware training set design significantly enhances model accuracy and experimental efficiency.
• This work provides a generalizable strategy for protein engineering beyond proteases.