StemCell-GPT: A Specialized AI Agent For Human Stem Cell Engineering

Jingwen Hui, Freja Kjellaug Amalia Ekman, Hana Yousef Ghanim, Sridhar Selvaraj, Yuanhao Qu, Matthew Porteus, Le Cong

Track: AI for Science

Keywords: AI Agent, LLM, RAG, Genetic Engineering, CRISPR

TL;DR: StemCell-GPT automates CRISPR guide-RNA design for stem cells and provides domain-specific guidance, achieving high accuracy (MAE = 4.94 %, R² =0.685, Spearman Correlation = 0.85) across four therapeutic gene targets.

Abstract: CRISPR technology has revolutionized genetic medicine, enabling programmable genome modifications. Central to its success is high-performance guide RNA (gRNA) that directs Cas9 nuclease to desired genomic targets. Although the first CRISPR therapy was approved by FDA in 2023 to edit hematopoietic stem cells, it remains the only proven gene-editing treatment to date. While numerous gRNA design tools exist, they are primarily geared toward cancer cells. As a result, designing high-efficiency, context-aware gRNAs for precise stem cell editing remains a critical bottleneck. This work introduces StemCell-GPT, a specialized AI agent designed to automate and enhance stem cell editing through multi-objective CRISPR gRNA design and by solving stem cell–specific engineering queries. The StemCell-GPT pipeline first generates candidate gRNAs in a target region and refines their on-target scores using DNA language model embeddings calibrated by a Random Forest regressor (achieving MAE = 4.94\%, \(R^2 = 0.685\)). It then fuses calibrated scores with insertion-deletion mutation profiles, using a feed-forward network to optimally weight each component based on real-world stem cell editing data. The entire workflow is managed by a large language model–based agentic process, and we benchmarked its performance on four therapeutic genes (CCR5, HBB, STING1, CFTR), achieving an average Spearman correlation of 0.85 between the predicted and experimental gRNA design rankings. These results highlight the value of agentic exploration in navigating a vast bio-design space. By streamlining multi-objective optimization and providing context-aware engineering support, StemCell-GPT accelerates precision gene editing in stem cell research and paves the way for robust, high-throughput clinical applications.

Serve As Reviewer: ~Le_Cong2, ~Yuanhao_Qu1

Submission Number: 96