Dual-encoder contrastive learning accelerates enzyme discovery

Area of Science:

• Biotechnology and Synthetic Biology
• Computational Biology and Bioinformatics
• Enzyme Engineering

Background:

• Enzyme engineering is crucial for biotechnology, but finding suitable starting enzymes is a major challenge.
• Current computational methods like contrastive learning for enzyme discovery are not yet scaled or experimentally validated.
• A scalable and effective computational approach is needed to overcome the bottleneck in enzyme discovery.

Purpose of the Study:

• To present Horizyn-1, a computationally efficient deep learning framework for large-scale reaction-to-enzyme recommendation.
• To validate the framework's effectiveness through comprehensive experimental testing across various enzyme discovery scenarios.
• To demonstrate the potential of Horizyn-1 in accelerating biocatalysis and metabolic engineering.

Main Methods:

• Developed Horizyn-1, a deep learning framework combining reaction fingerprints and protein language models.
• Trained the model on millions of reaction-enzyme pairs for reaction-to-enzyme recommendation.
• Experimentally validated Horizyn-1 for identifying enzymes for orphan reactions, predicting enzyme promiscuity, and discovering enzymes for nonnatural reactions.

Main Results:

• Horizyn-1 achieved state-of-the-art performance, identifying correct enzymes within the top 100 hits for over 75% of test reactions.
• Experimental validation confirmed Horizyn-1's efficacy in diverse enzyme discovery tasks, including noncanonical amino acid synthesis.
• Fine-tuning with minimal data significantly improved performance on underrepresented reaction classes, showing logarithmic scaling with dataset size.

Conclusions:

• Horizyn-1 effectively addresses the bottleneck of sourcing initial enzymes for optimization, enabling efficient in silico screening.
• The framework's scalability and experimental validation promise to accelerate biocatalysis and metabolic engineering.
• Continued improvement is expected with larger and more diverse training datasets, highlighting the framework's future potential.