Nature Biomedical Engineering has introduced a chatbot specifically designed to help beginners with their first experiments and to support experienced researchers in their work.

Since it was first described in Science in 2012, in the landmark paper by Emmanuelle Charpentier and Jennifer Doudna, the success of the CRISPR technique has been summed up with a handful of adjectives: cheap, precise, easy to use. But since everything is relative, it’s worth asking: how easy, and compared to what? When measured against previous genetic editing platforms, CRISPR is far simpler to apply. Whereas only a few highly specialized centers could once perform these experiments, with CRISPR a standard lab, the basic skills of an ordinary biologist, and solid familiarity with bioinformatics may be enough. Still, novices need guidance, and even seasoned researchers can run into problems.

The gap between theory and practice was illustrated in 2016 by Science journalist Jon Cohen, who personally attempted to use the technology. The result was a memorable article titled “One of Our Reporters Tried to Use CRISPR and Failed Miserably”. One wonders if his adventure might have gone differently had he been assisted step by step by an artificial intelligence trained to design editing experiments and correct potential mistakes.

An international group of researchers from Stanford, Princeton, Berkeley, and Google DeepMind has developed just such a system. Based on large language models (LLMs), it can automate the design and analysis of editing experiments. It guides users not only with the classic CRISPR “molecular scissors” but also with more advanced tools such as base editing, prime editing, and epi-editing. Instead of cutting the double helix, these approaches correct individual DNA letters, rewrite selected segments, or leave the sequence intact while adjusting its expression levels. This multimodal AI agent, designed to support both expert and novice researchers throughout all stages of experimentation, has been named CRISPR-GPT, in honor of the renowned chatbot, and was presented in Nature Biomedical Engineering on July 30.

CRISPR-GPT is described by its developers as an “intelligent copilot.” Unlike other models that are limited to protein design, this system can understand the entire experimental workflow: from selecting the most suitable CRISPR system for a given goal to choosing the delivery method for introducing molecular machinery into target cells; from designing RNA guides that locate the DNA site to be edited, to analyzing data and troubleshooting in real time. All of this is done through natural language interaction, no programming or complex software required.

At its core is a four-agent architecture: an interface for user dialogue, a planner that breaks tasks down into subtasks, an executor that provides detailed instructions, and a module for accessing external sources, such as databases and scientific publications. CRISPR-GPT was trained on scientific literature, public databases, and over a decade of online forum discussions among scientists, capturing not only theoretical knowledge but also the practical reasoning of those who work daily with genome editing.

Put to the test in controlled experiments, the system achieved promising results: undergraduate students with no prior experience in the field managed to reach over 90% editing efficiency on their first attempt under AI guidance. The interface offers three modes: “Meta mode” for beginners, “Auto mode” for advanced users, and “Q&A mode” for specific questions. The system was tested with 288 cases or scenarios and outperformed OpenAI’s GPT-4o in every task.

“It’s a way to democratize access to genome correction,” said Yuanhao Qu, a Stanford PhD student and first author of the study, in an interview with GEN. Yet the democratization of genetic editing doesn’t come only from research. Another initiative, launched last year by the same university, developed a $2 CRISPR kit designed for schools. It’s a safe, simplified, cell-free system that shows CRISPR’s effect on a harmless pigment. While it doesn’t enable real research, it teaches the basic principles of editing. The goal is educational: helping students become familiar with the technology that has revolutionized contemporary biomedicine. This approach is very different from the notorious “Odin” kit marketed by biohacker Josiah Zayner, which drew harsh criticism in the past for its lack of safety precautions.

There are no recent surveys on the spread of DIY Bio, the do-it-yourself biology movement which once raised concerns even within the FBI. For now, amateur editing doesn’t appear to have produced significant impacts or concrete threats. The question remains whether artificial intelligence will also give new momentum to DIY editing, making accessible to many what once required years of training, and whether experts will deem additional oversight and safeguards necessary.

As for CRISPR-GPT, the study’s corresponding author, Le Cong, told GEN that the system was designed with built-in safety mechanisms to prevent misuse, such as the design of pathogenic agents.

[translated and adapted from Osservatorio Terapie Avanzate]