How to Apply AI Strategically in Chemistry — Without Repeating Past Mistakes

Cautionary Chemistry Tales

As we enter the dawn of AI in drug discovery, we will undoubtedly also cycle through what Gartner calls the AI “hype cycle.” We can prepare for the hype, along with its “peak of inflated expectations” and “trough of disillusionment,” by recalling cautionary tales from chemistry’s past.

The Combichem Bust

In 2004, The Wall Street Journal (WSJ) published an article titled, “Drug Industry's Big Push Into Technology Falls Short.” The article echoed the sentiment spreading across the industry at the time—combichem had not just fallen short, but had perhaps played a key role in a marked decline in both new drug approvals and profits.

The allure of shiny new automation and robotics tools had made us lose our way. We started playing a numbers game, rather than a target game. To make the most of commercially-available building blocks, we simplified our chemistry to sp2-sp2 couplings and A-B-C inputs. But instead of getting endless possibilities, we got endless disappointments. We had millions of flat molecules that didn’t bind well to 3D target sites. Compounds going into trials weren’t clean, but were instead riddled with impurities, or perhaps racemic mixtures. And while we pushed more targets into study, fewer came out, oftentimes because the molecules were simply metabolized too quickly to work.

As we look back nearly two decades later, the parallels with AI are striking: huge investments in the latest innovative technologies, fundamental research paradigm shifts, strategic partnerships, and the blurring of lines between hype and reality. We mustn’t let history repeat itself with AI.

Lipinski’s Rules Run Amok

Another hopeful idea gone awry is Lipinski’s rule-of-five, a set of property rules often used in the early 2000s to quickly determine a compound’s therapeutic potential. But, as Derek Lowe summarizes in his blog, “Ruling out the rule of five,” Lipinski’s were often applied too stringently and often eliminated too many promising compounds—an opinion eventually shared by Lipinski himself.

Like the story of combichem, Lipinski’s rules illustrate that good ideas don’t always pan out in practice, especially when their steadfast adoption impedes experimental exploration. Sometimes rules need to be broken. We should keep this in mind as we adopt AI and train our models.

Applying AI Strategically in Chemistry

One quote from the WSJ article still stands out today. It reads, “…The story of chemistry technologies shows how hard it is to automate a process and keep room for serendipitous insight—which has been responsible for many great drug discoveries.”

These lessons from chemistry’s past have taught us that the disruption of disruptive technology is sometimes better off tempered. That technology advances are not always best applied unilaterally and exclusively, but rather strategically and supplementally.

Strategic Application of AI

Companies need to identify where AI can best help, which can be incredibly difficult when facing seemingly endless possibilities. Deloitte reports that AI is being used from the earliest days of discovery, into clinical trials, onto production, and through commercialization. Still, their report also acknowledges that many companies struggle to identify exactly where and how to use AI.

Following the money may provide some insight. Looking specifically at discovery, a survey of the AI start-up landscape shows a number of revealing trends:

• ​AI is most often used by companies studying small molecules.
• ​AI is being applied in diverse therapeutic areas, led by oncology, neuroscience, immunology, and infectious disease.
• ​AI is most used in structure- and ligand-based virtual screening, target identification and validation, lead discovery, and data mining.

While these trends may show where AI is being most used, they can’t yet show the success of these endeavors because most drugs driven by AI discovery efforts are still in early development.

Supplemental Role of AI

Another important lesson from our past is that we need to preserve our researchers’ freedom to explore, even as we set them up to leverage new technologies. How do we do this with AI?

First, we must provide an R&D infrastructure that facilitates the collection of AI-ready data from experimental efforts. But, beyond that, we need to let researchers merge results from AI back into the larger experimental fabric, where they can be considered alongside all the diverse chemistry, biology, formulation, and physical characterization data that teams are producing across the company’s broader research cycle.

Find Your AI Fit

See how you can fit AI into your drug discovery efforts with an infrastructure driven by AI-ready data.