CRISPR-GPT Turns Novice Scientists into Gene Editing Experts

CRISPR-GPT was trained on over a decade of expert discussions and evaluated against almost 300 test cases.

CRISPR technology has revolutionized biology, largely because of its simplicity compared to previous gene editing techniques. However, it still takes weeks to learn, design, perform, and analyze CRISPR experiments; first-time CRISPR users often end up with low editing efficiencies and even experts can make costly mistakes.

In a new study, researchers from Stanford University, Princeton University, and the University of California, Berkeley, teamed up with Google DeepMind to create CRISPR-GPT, an artificial intelligence (AI) tool that can guide researchers through every aspect of CRISPR editing from start to finish in as little as one day.¹ The results, published in Nature Biomedical Engineeringdemonstrate that researchers with no previous CRISPR experience could achieve up to 90 percent efficiency in their first shot at gene editing using the tool.

CRISPR-GPT is a large language model (LLM), a type of AI model that uses text-based input data. Led by The Cong of Stanford University and Mengdi Wang of Princeton University, the team trained the model on over a decade of expert discussions, as well as established protocols and peer-reviewed literature. They designed it to cover gene knockout, base editing, prime editing, and epigenetic editing systems, and benchmarked the tool against almost 300 test questions and answers.

The tool is ‘agentic;’ it can act autonomously rather than requiring human intervention, meaning that users do not need coding experience or any complex software. Inexperienced researchers can use simple prompts, such as “I want to knock out the human TGFβR1 gene in A549 lung cancer cells,” to receive step-by-step instructions; CRISPR-GPT provides guidance on what CRISPR system and delivery method to choose and how guide RNA should be designed, then generates a custom protocol, along with real-time troubleshooting and data analysis. With three different modes based on experience level, researchers can receive customized support for their specific needs.

The team assembled a team of eight gene-editing experts to evaluate the tool’s performance in generating step-by-step instructions for gene editing and compare it to general-purpose LLMs such as GPT-4o. CRISPR-GPT outperformed the general-purpose models, providing more accurate and concise information and avoiding errors.

Next, the team wanted to put CRISPR-GPT to the test in a real-world scenario. They enlisted junior researchers who had never used gene editing before to use the tool to design and conduct CRISPR experiments from scratch. One researcher used the tool to knock out four genes involved in tumor survival and metastasis in the human A549 lung adenocarcinoma cell line, while another employed CRISPR-GPT to perform epigenetic editing in a human melanoma cell line, activating two genes involved in resistance to immunotherapy.

For first-time CRISPR users, they achieved high editing efficiencies: The researcher who performed the multigene knockout achieved roughly 80 percent editing efficiency across all four target genes, while the one who carried out epigenetic editing achieved 56.5 percent efficiency in the activation of one gene and 90.2 percent in the other.

In the discussion of the paper, the authors suggested that using CRISPR-GPT as a laboratory co-pilot can help automate and enhance gene editing research, reduce errors, and improve research quality and reproducibility. “Every researcher can tackle bigger challenges without worrying about small mistakes,” said Cong in a LinkedIn post about the study.