When Yilong Zhou, an undergraduate student from Tsinghua University visiting Stanford University, wanted to run a gene editing experiment, the odds were against him. To investigate why cancer immunotherapy sometimes fails, Zhou wanted to use CRISPR to edit the genetic makeup of cancer cells — a task that typically requires years of training and extensive trial and error. But Zhou had just one summer to get results, and he had limited prior gene-editing experience.
Still, Zhou managed to pull it off. With the help of a new artificial intelligence (AI) tool called CRISPR-GPT, he switched off genes of interest in lung cancer cells on his first attempt, an experimental milestone that would have been otherwise difficult to reach for such a novice researcher.
CRISPR-GPT represents a growing trend at the intersection of AI and gene editing. The tool uses a large language model trained on years of published gene-editing experiments to guide researchers through every step of an experiment. It’s one of many AI tools emerging that help design, troubleshoot, and even invent new CRISPR systems. As these tools proliferate, their proponents say they could transform the discovery and development of new gene therapies — but questions still remain about how useful they’ll really be.
The problems with CRISPR
CRISPR, short for clustered regularly interspaced short palindromic repeats, first rose to prominence in 2012, when researchers demonstrated that the Cas9 enzyme could be harnessed to make targeted cuts in DNA , enabling precise genome editing. The system is adapted from a bacterial immune mechanism that uses CRISPR sequences and Cas enzymes to recognize and destroy invading viral DNA. In the lab, this molecular tool works like a pair of scissors, cutting DNA at specific locations so researchers can delete or insert genetic material of their choosing.
CRISPR has found applications in both basic research and clinical translation, and it’s now a mainstay of medical research. It can be used for drug screening via genome-wide editing; to correct single-nucleotide mutations, such as in sickle cell anemia; and even to engineer more potent immunotherapies in chimeric antigen receptor (CAR) T cells.
But CRISPR comes with its own set of problems. Both the Cas enzymes that CRISPR uses and the target cell’s inherent genetic repair machinery can be unpredictable, leading to edits that work in some cells but not others. Edits can also introduce errors and off-target effects. All of these issues can lead to CRISPR experiments feeling like a “black box,” said Thomas Naert, a biomedical researcher at Ghent University, with researchers resorting to taking educated guesses and crossing their fingers that these time-consuming and costly experiments will actually work.
AI as a CRISPR assistant
It can think, act, and make decisions like a real CRISPR expert.
– Le Cong, Stanford University
With these limitations in mind, Le Cong, a gene-editing researcher at Stanford University, led a group of researchers to design CRISPR-GPT. By combining a large language model trained on over a decade of data from published papers and forums with existing gene editing bioinformatics tools, Cong created a sort of “virtual lab partner” for CRISPR experiments.
“It can think, act, and make decisions like a real CRISPR expert,” Cong said. Cong imagines labs where researchers, including those as early as the undergraduate level, can interact with AI agents to plan experiments, select the best guide RNAs, optimize protocols, and run analyses.
With a chat-based interface similar to ChatGPT, CRISPR-GPT is designed to serve as a one-stop shop for designing, troubleshooting, and analyzing CRISPR experiments. The chatbot can generate guide RNAs for specific edits, plan and optimize experimental protocols, analyze results, and help troubleshoot issues such as a failed cell culture or a faulty delivery of CRISPR proteins. Like ChatGPT, it features a memory system that stores information about ongoing experiments, allowing it to provide continuity across sessions.
Cong and his co-authors introduced CRISPR-GPT in a paper published this summer in Nature Biomedical Engineering. While their tool is one of the most comprehensive AI systems available for designing gene-editing experiments, it is not the only option on the market. In another paper published this summer in Nature Biotechnology, a team led by Naert recently introduced Pythia, an AI tool designed to help researchers insert larger cassettes of DNA into the genome more reliably.
The tool was inspired by Naert’s own research with frogs, where he frequently was frustrated with the unpredictability of editing large chunks of DNA. One reason these sorts of experiments frequently go wrong, he explained, is that the cell’s natural genetic repair mechanisms chew up the DNA around the inserted region.
“There’s damage on both ends of the DNA,” he explained. “It kind of puts itself back together, but it’s not perfect.”
Pythia relies on a deep learning model trained to predict DNA repair outcomes at CRISPR–Cas9-induced double-strand breaks. The goal is to design AI-predicted RNA guide arms that “click” into the genome with minimal damage, working with the cell’s natural genetic repair machinery rather than against it. The program ranks different sequences based on how well they’re predicted to work, preventing sequences around the area of interest from being damaged.
“You’re not trying to swim against the stream,” Naert said. So far, the system has been successfully deployed across a range of biological contexts, including human cell lines, developing frog embryos, and adult mouse brains.
Rethinking CRISPR altogether
Some scientists are using AI to reimagine CRISPR from the ground up, like Ali Madani, CEO of Profluent.
“Natural CRISPR systems were evolved for bacteria, not for human cells, so they can run into limits when moved into new contexts,” he said.
Profluent’s solution to this problem, called OpenCRISPR-1, is a synthetic gene-editing protein whose sequence was designed with the help of AI trained on large-scale datasets of CRISPR–Cas and related proteins, mined from microbial genomes. In company assays, Profluent reported a 95 percent reduction in off-target editing relative to traditional CRISPR enzymes in human cell line experiments.
“We see a future where scientists design gene editors as readily as they design software,” Madani said.
Are AI gene-editing tools safe?
As promising as AI-assisted gene-editing tools may be, current iterations are still nascent and generally untested beyond highly controlled lab settings. These tools don’t eliminate the inherent risks of editing DNA, nor the potential pitfalls of relying on AI.
Hallucinations are one of the biggest considerations when dealing with AI systems. While CRISPR-GPT relies on a large language model for some tasks, it also integrates established bioinformatics software to handle other queries — a hybrid approach that, according to Cong, helps reduce the risk of hallucinations.
“We haven't found substantial problems,” Cong said. The potential for the tool to be harnessed for malicious use was also a concern, and the researchers did go to great lengths to build in biosecurity and privacy protections. CRISPR-GPT includes safeguards that prevent it from editing harmful or ethically controversial targets such as pathogenic viruses or human germline cells. There are also privacy protections: sensitive genetic data is stored locally rather than being sent to external servers.
Because of all of these considerations, Cong is treating the roll-out of CRISPR-GPT like that of self-driving cars. For now, the tool is only available to a small number of other groups at Stanford University, like a self-driving car on a closed course. As the accuracy and safety of CRISPR-GPT are tested in more diverse and dynamic conditions, Cong will consider opening up the roads.
The current reality
Benjamin Oakes, CEO of Scribe Therapeutics, said that while his company and many others in the industry have benefited from proprietary machine learning models trained on their own data to assist with CRISPR gene editing, he’s skeptical about the broader usefulness of AI tools developed elsewhere.
“Everyone's problems are so unique that it's going to be difficult to build something that is universal,” he said. In his view, current AI-assisted gene-editing tools can streamline the early screening stages of research but don’t yet address the most time-consuming and expensive phases like animal studies, clinical trials, or manufacturing.
Oakes believes the real transformation will come when AI systems are linked to robotic, automated labs capable of running thousands of experiments at once, generating the kinds of massive datasets needed to truly accelerate discovery. “If you can hook up a more advanced version of a model like this to a laboratory that is mostly done by robotics,” he said, “you could start to generate data on a scale that actually becomes very useful for training even more useful models.”
Overall, Oakes is impressed by CRISPR-GPT’s potential, but he cautions against overreliance on such systems, especially for newer researchers. “I think you're better off learning what you're trying to do, and that deeper knowledge actually helps you understand the outputs a lot more,” he said.
Other scientists, however, see the role of AI tools like CRISPR-GPT expanding quickly, such as Neville Sanjana, a genome engineer at the New York Genome Center and New York University who’s used AI to predict off-target activity in CRISPR systems. “Something that even two years ago I wouldn’t have believed was possible was asking these large language models to help design or interpret experiments,” said Sanjana. “That’s going to change how labs operate.”
