GSK says AI is reshaping drug pipeline as Nuvalent deal hits shares

GSK has said agentic AI is already affecting every asset in its pipeline, as pharmaceutical firms race to use the technology to cut the cost and risk of developing new medicines.

Speaking at London Tech Week in Tuesday, Eyal Itskovits, GSK’s director of AI and machine learning, said that the technology has officially become a live operational tool within the pharma giant’s R&D division.

“At GSK, all of our assets are currently being impacted by agentic AI, so this is not like a futuristic thing… it’s happening now,” said Itskovits.

The announcement comes on the same morning that GSK shares slipped nearly three per cent on Tuesday, following a blockbuster $10.6bn (£8.4bn) cash acquisition of Boston-based biopharma firm Nuvalent.

The market’s initial trepidation over the hefty 40 per cent premium points to the intense cost pressures on pharmaceutical giants seeking to speed up sluggish development pipelines,

For decades, the industry has suffered under the weight of Eroom’s Law – the phenomenon describing how the cost of developing new drugs spirals exponentially, with average pre-approval costs now eclipsing £1.7bn ($2.2bn).

“The term ‘precision medicine’ has been with us for a couple of decades now”, said Itskovits.

“And it traditionally meant that we can find a marker that can help us distinguish responding and non-responding patients and this way, manage to give the right patient the right treatment.”

But the rapid scale of modern healthcare data has opened up entirely new avenues to bypass traditional clinical failure rates, which historically hover around 90 per cent for molecules entering development.

According to Itskovits, the sheer depth of contemporary data repositories means developers can look far beyond a single biological indicator.

“I think nowadays, with the data we have at hand and data is being generated in an unprecedented depth and scale, both clinically and preclinically, we can do much better than this,” he added.

“I think the data we have now allows us to redefine the concept of similarity. It’s not just this one marker, but just a very profound and deep concept of similarity that spans through molecular markers.”

By using generative AI to scan huge internal and external datasets, GSK is attempting to tailor treatments to a really specific degree.

“This allows us to take a patient and scan the entire, I would say, repository of data out there,” Itskovits said. “All the swaths of data that we have externally and internally, whether it be literature or things that are just coming from the bench, from clinical trials, and to try to find the right treatment for the actual single individual patient. This is something we could not have done without generative AI.”

Yet, despite the financial upside of early-stage screening, Itskovits added that the industry must operate within strict regulatory guardrails when deploying the technology.

“Having said this, obviously, I’m side-stepping some regulatory issues here,” Itskovits added. “We cannot still allow generative AI to come up with novel treatments and like, adopt them. But there’s plenty we can do even now.”

While the broader healthtech sector grapples with the rising risks of large language models (LLMs) hallucinating information, Itskovits argued that highly structured agentic systems actually offer greater transparency than traditional, standalone models.

“I think there is a jump there from just using LLM to using agentic AI. Agentic AI has at its core a very powerful LLM, but it also is equipped with tools to operate in the world,” Itskovits said. “I think that in a way, agentic AI poses a much smaller challenge than actual LLMs – when you speak about agentic AI, it’s not just a big box that gives you an answer, but it actually operates and run tools.”

With a well-built agentic system, researchers can see which analysis was run and which data was accessed, as well as which paper was used and what intermediate results were produced.

“You can actually track provenance,” he said. “Everyone that goes over it can actually reproduce what the agent has been doing.”

That distinction is likely to become central as drugmakers and health systems decide how far AI can be trusted in the sector.

It also speaks to a broader issue facing healthcare AI, wbhere patients and clinicians may tolerate mistakes from humans, but their expectations become far higher when software is involved.