Evolutionary Probability and Stacked Regressions Enable Data-Driven Protein Engineering with Minimized Experimental...

Area of Science:

• Biochemistry
• Computational Biology
• Protein Science

Background:

• Protein engineering optimizes protein properties for industrial and academic applications using directed evolution and rational design.
• Limited screening throughput and vast variant possibilities hinder efficient protein engineering.
• Data-driven strategies model protein fitness landscapes computationally but require substantial training data.

Purpose of the Study:

• To introduce MERGE, a novel method for data-driven protein engineering applicable with limited available data.
• To enable reliable modeling of protein fitness landscapes even with small datasets (50-500 labeled sequences).
• To improve the prediction of protein fitness values and rankings based on sequence information.

Main Methods:

• MERGE combines direct coupling analysis (DCA) with machine learning (ML) approaches.
• The method is designed to function effectively with limited training data.
• Performance was evaluated across diverse proteins and properties for fitness prediction.

Main Results:

• MERGE demonstrates strong performance in predicting protein fitness values and sequence rankings.
• The method significantly outperforms existing state-of-the-art approaches when trained on small datasets.
• MERGE requires fewer computational resources compared to other methods.

Conclusions:

• MERGE facilitates data-driven protein engineering, especially when experimental data is scarce.
• This approach is highly promising for protein engineers working with limited datasets.
• The method offers an efficient and computationally less intensive alternative for fitness landscape modeling.