The model builds on a tool called CRISPR, a powerful gene-editing technology used to edit genomes and develop therapies for genetic diseases. But training on the tool to design an experiment is complicated and time-consuming — even for seasoned scientists. CRISPR-GPT speeds that process along, automating much of the experimental design and refinement. The goal, said Le Cong, PhD, assistant professor of pathology and genetics, who led the technology’s development, is to help scientists produce lifesaving drugs faster.
“The hope is that CRISPR-GPT will help us develop new drugs in months, instead of years,” Cong said. “In addition to helping students, trainees and scientists work together, having an AI agent that speeds up experiments could also eventually help save lives.”
In addition, it could expand the pool of scientists who can effectively use gene editing technology — no experience required. For instance, a student in Cong’s lab used CRISPR-GPT to successfully guide an experiment that turned off a handful of genes in lung cancer cells on his first attempt. That kind of feat usually requires a prolonged period of trial and error. But the AI tool’s ability to flatten CRISPR’s steep learning curve seems like a promising way to open access to gene editing throughout the biotechnology, agriculture and medical industries, Cong said.
“Trial and error is often the central theme of training in science,” Cong said. “But what if it could just be trial and done?”
Cong is the senior author of the study published July 30 in Nature Biomedical Engineering. The lead authors are Yuanhao Qu, a graduate student in cancer biology, and Kaixuan Huang, a collaborating graduate student at Princeton University.
AI that thinks like a human
CRISPR can target sections of DNA and snip out problematic mutations — the conceptual basis for many genetic disease therapies such as sickle cell anemia. But it can take months of guess-and-check work for researchers to evaluate whether the suspected segment of DNA is indeed the culprit they need to excise. (CRISPR can sometimes accidentally edit the wrong gene sequence, leading to unwanted genetic effects.)
CRISPR-GPT uses years of published data to hone the experimental design into something likely to be successful. It can also predict off-target edits and their likelihood of causing damage, allowing experts to choose the best path forward.
Cong’s team trained their model with 11 years’ worth of expert discussions, captured online, from CRISPR experiments and information published in scientific papers. The result? An AI model that “thinks” like a scientist.
When using CRISPR-GPT, the researcher initiates a conversation with the AI agent through a text chat box, providing experimental goals, context and relevant gene sequences. Then, CRISPR-GPT creates a plan that suggests experimental approaches and identifies problems that have occurred in similar experiments to help the researcher — novice or expert — avoid them.
Yilong Zhou, a visiting undergraduate student from Tsinghua University, used CRISPR-GPT to successfully active genes in A375 melanoma cancer cells as part of his research into better understanding why cancer immunotherapy sometimes fails.
Zhou typed his question into CRISPR-GPT’s text box: “I plan to do a CRISPR activate in a culture of human lung cells, what method should I use?”
CRISPR-GPT responded like an experienced lab mate advising a new researcher. It drafted an experimental design and, at each step, explained its “thought” process, describing why the various steps were important.
“I could simply ask questions when I didn’t understand something, and it would explain or adjust the design to help me understand,” Zhou said. “Using CRISPR-GPT felt less like a tool and more like an ever-available lab partner.”
As an early-career scientist, Zhou had designed only a handful of CRISPR experiments prior to using CRISPR-GPT. In this experiment, it took him one attempt to get it right — a rarity for most scientists.
In the past, Zhou was constantly worrying about making mistakes and double-checking his designs.
Reducing error and increasing accessibility
CRISPR-GPT can toggle between three modes: beginner, expert and Q&A. The beginner mode functions as a tool and a teacher, providing an answer and explanation for each recommendation. Expert mode is more of an equal partner, working with advanced scientists to tackle complex problems without providing additional context. Any researcher can use the Q&A function to directly address specific questions.
It’s also useful for sharing knowledge and collaborating with other labs, Cong said. CRISPR-GPT provides a more detailed and holistic response than what’s generally gleaned from a scientific manuscript and responds to repetitive inquires in a snap.
CRISPR-GPT can also check researchers’ work and apply experimental frameworks to new diseases the researchers may not be thinking about.
“People in my lab have been finding this tool very helpful,” Cong said. “The decisions are ultimately made by human scientists, but it just makes that whole process — from experiment design to execution — super simple.”
Editing responsibly and future expansion
While the technology is promising for accelerating therapeutic research, there are still some safety concerns to address before pushing CRISPR-GPT more broadly.
Cong and his team have already incorporated safeguards to protect the AI tool from irresponsible uses. For instance, if the AI receives a request to assist with an unethical activity, such as editing a virus or human embryo, CRISPR-GPT will issue a warning to the user and respond with an error message, effectively halting the interaction. Cong also plans to bring the technology to government agencies, such as the National Institute of Standards and Technology, to ensure ethical use and sound biosecurity.
In the future, the tool may serve as a blueprint for training AI to execute specific biological tasks outside of gene editing. From developing new lines of stem cells as experimental models, to deciphering molecular pathways involved in heart diseases, Cong hopes to expand the technology to other disciplines building a range of AI agents to aid in genomic discovery. To that end, he and his team developed the Agent4Genomics website, where they host a range of related AI tools for scientists to use and explore.
Researchers at Google DeepMind, Princeton University and the University of California, Berkeley contributed to this study.
Funding for this research came from the National Institute of Health (grants 1R35HG011316 and 1R01GM1416), the Donald and Delia Baxter Foundation Faculty Scholar Award, the Weintz Family Foundation, and the National Science Foundation.
# # #
About Stanford Medicine
Stanford Medicine is an integrated academic health system comprising the Stanford School of Medicine and adult and pediatric health care delivery systems. Together, they harness the full potential of biomedicine through collaborative research, education and clinical care for patients. For more information, please visit med.stanford.edu.
END