Substrate Prediction for RiPP Biosynthetic Enzymes via Masked Language Modeling and Transfer Learning

Area of Science:

• Biochemistry
• Computational Biology
• Synthetic Biology

Background:

• Ribosomally synthesized and post-translationally modified peptides (RiPPs) are a diverse class of natural products.
• Enzyme substrate promiscuity in RiPP biosynthesis lacks simple predictive rules.
• Large language models (LLMs) show potential for predicting peptide fitness landscapes.

Approach:

• Applied masked language modeling to profile substrate preferences of LazBF and LazDEF enzymes from the lactazole pathway.
• Utilized transfer learning by applying LLM embeddings trained on one enzyme's substrates to predict the other's.
• Fine-tuned models on specific datasets to gain interpretable insights.

Key Points:

• LLM embeddings improved classification models for both LazBF and LazDEF substrates, demonstrating transferability.
• Transfer learning enhanced performance and data efficiency, especially in data-scarce scenarios.
• Fine-tuned models offered interpretable insights for designing compatible substrate libraries.

Conclusions:

• LLMs can learn transferable functional forms for enzymes within the same biosynthetic pathway.
• This approach facilitates substrate library design for targeted RiPP biosynthesis.
• The study highlights the utility of LLMs in understanding and engineering complex enzymatic processes.