This July, when an AI-enabled chess-playing robot grabbed and broke the finger of a 7-year-old player during a tournament, the president of the Moscow Chess Federation summarized the situation succinctly: “This is, of course, bad.”

The same might have been said of IBM Watson’s foray into the realm of healthcare. Some eight months after the company sold its data assets to a PE firm, it’s already a cautionary tale: Following a buying spree that netted it a number of health data companies, Watson Health couldn’t deliver on its promise to revolutionize cancer care and drug discovery.

Fortunately for the industry — and, one might argue, for humanity at large — any number of organizations are succeeding where Watson failed. The platform’s implosion, it seems, has done little to mute the enthusiasm for using AI techniques to solve the impossible riddle that is drug discovery.

“No AI-designed drug has made it to market. In fact, no AI-designed drug has gotten beyond Phase I,” notes Richard Law, chief business officer of Exscientia. “But the excitement is very fair, given that there are now multiple drugs in clinical trials that were designed with AI.”

The stakes are huge: Precedence Research reports that the AI drug development market will reach $11.9 billion by 2030, up from an estimated $1.2 billion in 2021. That’s why it is probably worth taking a moment to understand what AI is — and what it isn’t — in the context of drug discovery.

“This is not comparing human and computational chess players,” Law stresses. “This is not teaching a machine how to drive a car. This is enabling a data analysis-driven system that is able to do something that humans cannot do.”

Founded in 2012, Exscientia focuses on identifying and designing small molecules for preclinical candidates as well as clinical trial design. Small molecules are an area of particular focus of AI-enabled discovery efforts and a good match with the strengths of machine learning. The chemical structures of small molecules can be described in ways that computers can easily process and there is an abundance of data on, for example, their interactions with protein targets.

AI algorithms can track some 2,500 or so properties of a molecule and model their interactions, a task that far exceeds human capacity.

“Every single one of those 2,500 things is a reason a drug might fail,” Law says. “You could get 2,499 things right but if you get one thing wrong, that drug fails.”

Especially when compared with the early bluster of Watson Health, some uses of AI in healthcare may seem modest in scope. But those applications, their boosters say, have the potential to dramatically reduce the time required to get new drugs into clinical testing and lower the incredibly high failure rate (around 96%) that has characterized drug development. It’s a financial play in addition to a scientific one: AI promises to lower the cost of developing a new drug, which averages around $5 billion.

However, all of this remains in its earliest stages. A Pharmaceutical Technology analysis of GlobalData’s Drug Database found more than 250 unique active drugs with the terms “artificial intelligence” or “machine learning” in their descriptions, with 91% of them in discovery or preclinical stages.

Exscientia is far from alone in championing AI’s potential to speed the drug discovery process. Insitro, Valo Health and Recursion Pharmaceuticals are among the organizations that have emerged as pioneers in the field — and, in the process, attracted the interest of the VC community and big pharma.

To that latter point, Exscientia and Sanofi are collaborating on the development of 15 oncology and immunology drug candidates. Then there’s the company’s partnership with Bristol Myers Squibb, which dates back to 2019 and focuses on developing small-molecule treatments for cancer and auto-immune conditions.

While Exscientia and its competitors have focused on the discovery applications of AI, others are focusing on its utility in the recruitment of clinical trial participants and in the marketing of approved drugs.

Swoop president Ron Elwell, whose company helps pharma entities navigate privacy minefields, describes the task at hand as “building an audience based on demographic information — age, kids, where you live and other factors that resemble as closely as possible the audience that has a condition.”

However, Elwell notes that “things get complicated when you are trying to reach an audience with particular subvariants,” and stresses that all models must comply with HIPAA and state laws that safeguard patient privacy. “A fundamental axiom is that you can’t target someone based on a previous diagnosis.”

The challenge becomes even more difficult with specialty drugs, where it may take a patient seven years before accurate diagnosis.

“What we are doing is coming up with new techniques that, looking at the health data, lets us determine how likely it is that someone has a specific condition,” Elwell explains. “What we can then do is tie those high-scoring patients to their physicians, and go to those physicians with a message: ‘We think you have seven clients that have this rare or undiagnosed condition.’ It’s an exciting way to take the next big leap.”

By way of example, Elwell points to a Swoop/Abbott collaboration around mitral regurgitation (also known as leaky heart valve syndrome). By tapping more than 3,700 consumer attributes and 65 billion consumer records — and, of course, employing AI techniques that allowed them to process the huge volume of information — the organizations created a better understanding of the target patient population and then focused its efforts on specific patient niches. This application of AI resulted in tangible results, including a 29% conversion rate.

The takeaway here, for marketers and technologists and all other interested parties, is that those who focus relentlessly on the data will likely be rewarded. Klick Health SVP, insights and analytics Eamon Boyle believes that the industry’s data scientists are up to the task — and that its marketers are ready to follow them down that road.

“There is a COVID effect that has impacted healthcare marketing in terms of how people view illness, especially mental health. People share more about themselves, their conditions and how they engage with healthcare providers and healthcare-related information,” he explains. “Healthcare information is now found across all areas of social networks, from Meta to TikTok. So once the data has been made HIPAA-, GDRP- and CCPA [California Consumer Privacy Act]-compliant, AI methodologies are applied to mine the data to find the appropriate signals.”

While Law focuses on AI’s ability to simultaneously consider some 2,500 attributes of a molecule — looking less for a needle in a haystack than a needle somewhere on a farm, to paraphrase Exscientia CEO Andrew Hopkins — Boyle notes that AI can perform a similar function in identifying the patients with rare conditions who could be helped by a new drug.

“AI is incredibly important at finding signals around rarer diseases, as the incidences are smaller, but there is a need to search through larger samples of data to find these incidences,” he explains. “Moreover, AI is good at combining factors together to find signals. It might, for instance, start from larger samples around OTC drugs that may be the first signal of an underlying condition.”

While AI boosters look forward to its applications in healthcare, the path forward is being shaped by the realities of trial-and-error inherent in the scientific method. While Watson Health is generally considered a disappointment, some might view its setbacks as an essential part of the learning process.

“In certain areas, AI and machine learning have had a steeper learning curve than we expected,” Boyle notes.

Elwell points to the inherent biases of a large percentage of health-related data, and specifically the greater availability of information comparing wealthier patients and lower-income ones. It is a reality for which marketers need to take into account, and then correct for it.

Despite these obstacles, there are plenty of reasons to remain optimistic. Law notes that even the smallest advances can have huge payoffs.

“You actually need only a tiny increase on your probability of success to get an enormous increase on return on investment,” he says. “If you have a 5% increase in the probability of success, you’d get like a 500% increase in return on investment. If it works, it’s so much better that it’s ridiculous.”