Discovery of a phenazine-thiol conjugase from sparse data using genome-informed machine learning

Area of Science:

• Biochemistry
• Genomics
• Machine Learning

Background:

• Machine learning models excel with large datasets but struggle with data sparsity in biological discovery.
• Identifying enzymes for understudied substrates, like phenazine natural products, is often limited by insufficient training data.

Purpose of the Study:

• To develop a machine learning framework integrating genomic data and contrastive learning for enzyme discovery.
• To identify novel enzymes involved in phenazine modification using limited known sequences.

Main Methods:

• Developed ML-CITO (Machine Learning for genomic Context-Informed Transferable discOvery), combining genome-informed data augmentation with contrastive learning in protein language space.
• Applied ML-CITO to a dataset of 14 known phenazine-modifying sequences to identify phenazine-interacting proteins.
• Utilized in silico simulations and experimental validation, including recombinant expression and biochemical characterization.

Main Results:

• Successfully identified phenazine-thiol conjugase (PTC), an enzyme catalyzing phenazine thioconjugation, a reaction previously thought to be nonenzymatic.
• Demonstrated PTC's ability to bind both phenazine and glutathione, facilitating glutathione-dependent phenazine modification with substrate-specific outcomes.
• Showcased that PTC-deficient strains did not exhibit significant fitness disadvantages, highlighting enzymes missed by phenotype-based screens.

Conclusions:

• Coupling comparative genomics with protein machine learning effectively addresses "small data" challenges in enzyme discovery.
• ML-CITO provides a powerful framework for identifying enzymes involved in modifying understudied natural products.
• The discovery of PTC expands the known enzymatic repertoire for phenazine modification and its biological implications.