AI-powered CRISPR technology turbocharges gene therapy development

Stanford researchers and their collaborators have revealed a new device that could change the way scientists conduct gene-editing experiments. The device, CRISPR-GPT, is an artificial intelligence lab aide. Instead of scientists going through the long and complicated process of gene editing alone, the system gives step-by-step directions, answers queries, and even mistakes in the process.

The idea behind CRISPR-GPT is simple: gene editing is potent, but it is overwhelming. Choosing the right method, designing guide RNAs, delivering them into cells, running experiments, and reading results requires expertise in multiple areas. For new students and researchers, that hill is steep. Even for expert scientists, experimentation takes months or years. CRISPR-GPT hopes to flatten that curve by becoming what its creators call an "AI copilot" for gene editing.

Le Cong, an assistant professor of pathology and genetics at Stanford, spearheaded the work. His dream is to bring gene editing into broader use and speed up the process of creating new treatments. "The hope is that CRISPR-GPT will allow us to construct new drugs in months, not years," Cong said. "Having an AI agent that speeds up experiments could also potentially save lives one day."

Essentially, CRISPR-GPT is a multi-agent system. It's four specialist colleagues working together: a planner that plans out the tasks, an executor that does them in order, a user agent that communicates with scientists, and a tool provider that communicates with databases and design tools. Putting all of these things together, the system can walk you through the entire process of designing a gene-editing experiment.

It also boasts three modes of interaction. There is a guided "Meta mode" for new users who need step-by-step guidance. "Auto mode" gives veterans more freedom, with the AI building workflows automatically. And there's a Q&A mode where scientists can ask questions directly and get answers based on decades of published literature and expert discussion.

The system has learned from 11 years of CRISPR experiments and online forums where scientists have shared their success and failure. That knowledge enables it to "think" like a human scientist, not just spew information. If you ask it how to activate a gene in lung cells, for example, it won't give you just a protocol. It will explain to you why each step matters, like having an experienced lab partner explain it to you.

To see if CRISPR-GPT truly works, the scientists tested it against general large language models like GPT-3.5 and GPT-4. The scientists constructed a dataset of nearly 300 scenarios through experiment planning, guide RNA design, delivery method selection, and Q&A detailed. Altogether across all tasks, CRISPR-GPT outperformed general models in terms of accuracy, reasoning, and reliability.

In a BRD4 cancer gene test case, CRISPR-GPT correctly identified the most crucial parts of the gene to edit. Competing models were off target. The system also excelled at predicting methods for difficult-to-edit cell types, a common laboratory issue.

Eight gene-editing experts also tested CRISPR-GPT. They praised it for generating correct, complete, and precise instructions and avoiding the typical "hallucinations" that afflict most AI systems.

Despite all its clever design, the question was whether or not CRISPR-GPT could assist actual researchers in conducting experiments successfully. To find out, the researchers handed the tool over to junior researchers with minimal gene-editing background.

In one experiment, a student used CRISPR-GPT to knockout four genes in lung cancer cells. The platform provided the right editing enzyme, designed the guide RNAs, chose the delivery vehicle, and even interpreted the sequencing data later on. The edits worked, with an average efficiency of approximately 80 percent. The results withstood biological proofing, confirming that the genes had been switched off as intended.

In one instance, an undergraduate from a visit to Tsinghua University employed CRISPR-GPT to enable two genes in melanoma cells. In his first attempt, he had activation rates as high as 90 percent—something that could traditionally take months of experimentation and guesswork. "It felt less like a tool and more like an ever-present lab partner," he said.

What stands out most about CRISPR-GPT is how it lowers the threshold to entry. Gene editing has been the domain of highly skilled specialists in the past, but with this system, new entrants are successful right from the start. Cong sees that as a way of bringing more people into the field, from students in training labs to scientists in fields where CRISPR is not yet widely applied.

“Trial and error is often the central theme of training in science,” Cong said. “But what if it could just be trial and done?”

The system also makes collaboration easier. Because it can explain its reasoning step by step, CRISPR-GPT allows teams at different institutions to share experiments more clearly. Rather than piecing together details from journal articles, collaborators can review the AI’s transparent workflow.

Of course, the potential of AI-directed gene editing is unsettling. Gene editing or manipulating human embryos or disease agents is risky and in most places illegal. In response, Cong's lab built into CRISPR-GPT checks. When a user tries to request something immoral, the program halts and notifies the user. It also does not retain long stretches of DNA sequence data, in order to protect genetic privacy.

Nonetheless, the scientists concede that the system is not foolproof. It works best in human and mouse studies and can fail on very atypical or very hard cases. As with all wonderful tools, control and responsibility are still the most important.

CRISPR-GPT might go beyond experiment design in the future. The scientists envision its potential to be integrated with robotic lab tools and used to design and carry out experiments programmatically. They are also eager to transfer it to other areas of biology, including stem cell research or heart disease studies.

The group has already launched the Agent4Genomics site, where CRISPR-GPT and other software of this sort are available for scientists to experiment with. Collaborators at Google DeepMind, Princeton, and UC Berkeley have also signed on to the project, which implies a growing list of AI-based biology research.

CRISPR-GPT could speed the development of new drugs by reducing the months or even years usually needed to attempt and adjust experiments. By opening doors for novice researchers, it may also potentially broaden the base of people engaged in gene-editing research in medicine, agriculture, and biotech.

The tool's controls and ability to explain its reasoning may also set a standard for how responsibly AI can be used in labs.

If it is as effective as promised, this kind of AI-human collaboration may get lifesaving treatments to patients more rapidly and at a lower cost.

Research findings are available online in the journal Nature Biomedical Engineering.

Like these kind of feel good stories? Get The Brighter Side of News' newsletter.