ChatGPT Gaining Foothold in Drug Development, Clinical Trials

Gaurav Sharma says ChatGPT is already helping in the search for new medicines, cutting the time it takes to identify molecules for potential new drugs from a month down to a week.

“I’m using it on my research and, and it does wonders,” said Sharma, a researcher at biotech firm Atom Bioworks. “You can see my happy face because of ChatGPT.”

Sharma, who co-authored a preliminary scientific manuscript on ChatGPT in drug discovery and has been using the most well-known large language model program since it launched in November, is quick to note what it can do. And what it can’t.

“Obviously it is not a very intelligent person,” said Sharma. “But it is a very, very good assistant.”

While ChatGPT has swept into popular culture in recent months with promising and sometimes cautionary results, artificial intelligence has been part of medical research and drug development for years.

But large language models—the form of artificial intelligence that ChatGPT popularized—potentially can alter the industry, perhaps most immediately by speeding up clinical trials long criticized as clunky and expensive.

And some large language models are being developed specifically with medicine in mind.

Future ‘Already Here’

Vial, a technology platform for clinical trials and early adopter of large language models, has compressed tasks that used to take two to three weeks down to a day and a half, said Simon Burns, company co-founder and chief executive.

And Google’s own medical large language model, called Med-PaLM 2, is scheduled to be released in the coming weeks, available at first for limited testing to a select group of Google Cloud customers.

The rush of change in drug research could become an issue for government regulators and industry, said Amy Abernethy, the former Food and Drug Administration official in charge of data modernization, and now chief medical officer at Verily, part of Google’s parent company Alphabet.

“There’s only so much capability available to regulate what we currently have, let alone to stop and consider and think through how we’re going to do it into the future,” said Abernethy.

“And the future is already here,” she said. “That’s the interesting challenge at the moment and time that we’re in.”

FDA Oversight

Health care offers its own challenges and opportunities for ChatGPT (which scored higher in empathy than actual physicians when comparing written responses in an online forum).

Bringing a new drug to market on average takes more than a decade and costs about $2.6 billion, according to an analysis from Tufts University.

Large language models have the potential to speed up that process by cutting the time it takes to review the scientific literature, identify new drug targets, and identify potential clinical trial volunteers.

Burns of Vial, a clinical trial management technology platform, said only a few years ago natural language processing was nowhere near able to build the types of applications his company can now create in two to three weeks.

“LLMs [large language models] are a completely new category of machine intelligence,” Burns said. “It really is kind of a complete game changer for the speed at which you’re able to build really intelligent applications for complicated tasks like large scale, image processing, and labeling.”

But using large language models in drug development and clinical studies raises questions for US regulators, who must determine whether medical products are safe and effective enough to come onto the market.

FDA already regulates software as a medical device, and will have to define when a large language model crosses the barrier from improving communication and efficiency of a product to itself becoming a medical device, Abernethy said.

“We’re still trying to think our way through and figure out what exactly that regulatory framework is,” said Abernethy, whose data modernization work at FDA considered how to adapt to the use of AI and machine learning in drug and device regulation.

“The ability to imagine what AI could do for us has proliferated really quickly,” she said. “I see it as a fast speeding train coming right at a very tight hole in terms of the size of the tunnel and to get through the mountain. And so that train is going to get stuck.”

FDA Commissioner Robert M. Califf has said repeatedly that the agency needs to stay ahead of the industry and figure out how to regulate products that use AI-assisted applications.

“If we’re not nimble in the use and regulation of large language models—otherwise known as ChatGPT in the vernacular—we’ll be swept up quickly by something that we hardly understand,” Califf said in a speech last month.

‘Response, Diagnosis, Prediction’

Kim Branson, who has spent his career at the nexus of machine learning and drug development, expects large language models to be transformative, particularly as they move to multimodal platforms that can integrate images of pathology, gene expression, and text about a patient.

“They look to be extremely powerful for understanding response, diagnosis, prediction, and things like that,” he said.

Branson is the senior vice president and global head of AI and machine learning at the GSK plc, a UK-based pharmaceutical giant developing its own custom large language model.

“A lot of our things will be a very particular narrow use case, they’re not a general use case,” Branson explained. “We will never be rolling out, for example, an AI doctor LLM. But we might have a specific LLM that might be using multimodal data in oncology.”

While language models are important, the data, and ensuring the quality of that data, will always be the key driver, he said. “It doesn’t mean that we won’t be generating data, I think it means that we will just be more efficient with our use of data.”

Clinical Trials

In clinical trials, large language models are expected to help identify what research has already been done, find prospective study volunteers, and even analyze responses from participants, such as whether someone reacted positively to a medical intervention or if they had concerns, explained Alan Karthikesalingam, research lead of Google Health.

And large language models could address longstanding challenges with informed-consent forms, ensuring that prospective volunteers understand what they’re signing and grasp a medical trial’s risks and potential benefits.

A key tool in the ethical oversight of research, consent forms are supposed to be written at an eighth-grade reading level, but have evolved into sometimes lengthy, jargon-filled legal documents.

A rewrite of the federal regulations five years ago requires that key information appears at the top of the form, but large language models could simplify the language further, tailor it to specific audiences, and present the form in multiple languages.

Google does something similar now when its Gmail auto-drafts email replies, Karthikesalingam said.

“One can imagine the same sort of sentence completion, but people who are writing a clinical trial protocol,” he said. “Those protocols can be quite laborious to write.”

Because large language models can analyze all of the medical jargon, it should speed up the process.

“The control is still with us,” Karthikesalingam said. “But it’s like having a kind of writing buddy.”