Leveraging large language models for metabolic engineering design

Abstract

Establishing efficient cell factories involves a continuous process of trial and error due to the intricate nature of metabolism. This complexity makes predicting effective engineering targets a challenging task. Therefore, it is vital to learn from the accumulated successes of previous designs for advancing future cell factory development. In this study, we developed a method based on large language models (LLMs) to extract metabolic engineering strategies from research articles on a large scale. We created a database containing over 29006 metabolic engineering entries, 1210 products and 751 organisms. Using this extracted data, we trained a hybrid model combining deep learning and mechanistic approaches to predict engineering targets. Our model outperformed traditional metabolic engineering target prediction algorithms, excelled in predicting the effects of gene modifications, and generalized well to out-of-distribution products and multiple gene combinations. Our study provides a valuable dataset, a chatbot, and an engineering target prediction model for the metabolic engineering field and exemplifies an efficient method for leveraging existing knowledge for future predictions.