[P002] Poster

1:00 PM - 1:50 PM JST(4:00 AM - 4:50 AM UTC)

[P01-006] Knowledge and model driven design of industrial strains using AI agents

○Hongwu Ma¹, Zhitao Mao¹ (1. Tianjin Institute of Industrial Biotechnology, Chinese Academy of Sciences (China))

Keywords:

Synthetic Biology,Articificial inteligence,industrial strain,cellular model,large language model

Synthetic biology has emerged as a key approach for the development of industrial strains but its power is somehow limited by the lack of reliable computational methods for learning from big data and in silico design. Recent advances in deep learning algorithms and large language models have greatly enhanced the ability to discover new knowledge from big data and embedding knowledge from a huge number of publications. This report presents the our recent research in applying artificial intelligence technologies to guide the design and construction of engineered cells. By integrating more than 50,000 literatures related to microbial metabolic engineering using retrieval-augmented generation technology, we constructed SynBioGPT, an expert system for metabolic engineering strain design. With the integration of an external professional knowledge base and advanced search techniques, SynBioGPT significantly reduces the occurrence of hallucinations and errors in question answering comparing with general large language models such as ChatGPT. The system adopts a multi-dimensional problem decomposition strategy, subdividing user queries into subproblems to deliver more comprehensive and robust strain modification schemes. Using SynBioGPT in a question-and-answer format, we reproduced the strain engineering approaches described in multiple recently published metabolic engineering papers, covering areas including efficient substrate utilization, biosynthetic pathway construction, transport system optimization, and engineering of key enzymes. The results demonstrate that SynBioGPT can generate modification strategies similar to those reported in the literature, while also proposing several innovative solutions not mentioned in the original studies. We further integrated several model based pathway and target design methods into SynBioGPT through AI agents. By combining model-driven design with knowledge-based design, SynBioGPT can provide more comprehensive industrial strain design strategies reaching or even surpassing the level of human metabolic engineering experts.