Low-N protein engineering with data-efficient deep learning.

Area of Science:

• Biotechnology
• Computational Biology
• Protein Engineering

Background:

• Protein engineering offers significant academic and industrial promise.
• High-throughput experimental assays are crucial but often limited for identifying rare, enhanced protein variants.

Purpose of the Study:

• To develop a machine learning-guided paradigm for efficient protein engineering.
• To enable high-throughput screening of protein variants using accurate virtual fitness landscapes.

Main Methods:

• A machine learning model was trained on a small set of functionally assayed mutant sequences (as few as 24).
• The model builds a virtual fitness landscape for in silico directed evolution, enabling screening of millions of sequences.
• The approach incorporates a latent representation of 'unnaturalness' derived from natural protein sequence landscapes to guide the search.

Main Results:

• Accurate virtual fitness landscapes were generated for two distinct proteins: GFP (avGFP) and TEM-1 β-lactamase.
• In silico directed evolution screened ten million sequences, identifying diverse top candidates with high activity.
• Engineered mutants achieved activity levels comparable to those from previous high-throughput methods.

Conclusions:

• The developed machine learning paradigm efficiently uses resource-intensive assays without sacrificing throughput.
• This approach accelerates the discovery and development of engineered proteins for various applications.
• The method effectively guides protein sequence optimization by learning from natural sequence data.