Accessing artificial intelligence in pharmaceutical laboratories

Artificial intelligence raises a lot of questions and a lot of anxiety, each new advance fostering both enthusiasm and skepticism. In recent months, generative AI—computer algorithms able to produce detailed analyses, thorough reports, striking images, and passable poetry given simple prompts—has been in the news, heralding an advance for computer intelligence while stirring up a host of new concerns.

AI has proved impervious to resistance, however, having already crept into our daily lives in the form of Google searches, satellite-based navigation systems for drivers, and virtual assistant technologies that turn on our coffee makers and play music on command.

The technology has also made stealth advances in industry, including the pharmaceutical industry, in which not very long ago researchers questioned its role and expressed concerns over machines taking their jobs in drug discovery labs.

Those questions rapidly gave way to ones about how to best gain access to algorithms for machine learning, the data-driven intelligence engine of AI that is now recognized as essential to the discovery and development of new medicines.

The potential for AI-derived breakthroughs in drug research is spurring change across the pharmaceutical industry. Big companies like Sanofi and Genentech are looking for knowledgeable partners to help navigate the field and improve drug discovery. Smaller firms are acquiring businesses to widen their AI-oriented offerings—both for themselves and for Big Pharma collaborators. And companies devoted to serving the drug industry are forming partnerships with AI firms to add a new tool to their offerings.

This year’s models

Two announcements, a year apart, reflect current activity as the technology advances in drug discovery.

In May, Recursion Pharmaceuticals, an AI platform company specializing in biology, announced the purchase of two chemistry-focused AI firms, Cyclica and Valence Discovery. The announcement followed the debut last year of a partnership between Charles River Laboratories International, a leading contract research organization (CRO), and Valo Health, an AI platform company that has grown rapidly in recent years through acquisition. The partners launched Logica, an approach to drug discovery that combines Valo’s AI-driven Opal Computational Platform with Charles River’s preclinical research services.

The announcements illustrate a strategic gap between traditional CROs like Charles River and the new breed of IT platform providers typified by Recursion.

“Recursion is not a CRO,” says Imran Haque, the firm’s senior vice president of AI and digital sciences. “We are a pure-play platform provider with an internal portfolio and a number of trials underway in rare and infectious disease.”The company is also providing its platform to drugmakers. Two notable examples are an engagement with Bayer focused on fibrotic diseases and one with Roche and its Genentech unit for neuroscience and gastrointestinal oncology.

Beyond the access that the big drug companies get to AI, these partnerships are not unlike traditional deals between drug companies and smaller biotechnology firms. In collaborations between drug companies and discovery firms with innovative AI platforms, the drug companies may have an option to license or acquire drug candidates, with the platform partner retaining certain rights and royalties.

Recursion’s two acquisitions are unlikely to alter its partnership model, Haque says, but the access to chemistry platforms will add a critical dimension to the company’s digital approach to drug discovery. Many other AI platform developers, such as Exscientia and Insilico Medicine, began work with a concentration on chemistry.

“Our focus has really been on using AI to build and advance what we call phenomics, a high-dimensional assay platform,” Haque says. Recursion’s microscopy-based fluorescent screening approach applies machine learning to study perturbations of cells in relation to one another in biological systems.

“We have done knockouts of essentially the whole genome,” he says, referring to a technique of altering a cell in a manner that disables a particular gene. “That’s 17,000 out of 20,000 genes.”

Leaning into chemistry with the acquisitions, Recursion quickly put two of Cyclica’s models, pareto-optimal embedded modeling (POEM) and MatchMaker, to work. POEM now applies machine learning algorithms to an automated drug metabolism and pharmacogenetic assay module that can test 500 compounds a week on multiple assays. And Recursion used MatchMaker to predict the protein targets for its entire library of over a million small-molecule compounds.

With Valence, Haque says, Recursion has its eye on future platform development. “We saw in Valence a company that had made big strides in machine learning, especially on frontiers of deep learning in chemistry and deep learning for generative AI with a strong community of academic researchers and collaborators,” Haque says.

For Cyclica, coming on board at Recursion advances a journey that was already underway. “Cyclica started as software as a service with MatchMaker,” says Michael Palovich, the firm’s chief science officer. “We found out that’s not the best use of the platform. There was a chunk of opportunity that was just missed.” The company determined about a year ago to put the tools it developed to work on its own drug discovery projects.

Cyclica is now part of an AI-based drug discovery company with a pipeline of five clinical-stage candidates and is applying its chemistry AI model to a huge compound library. “Now we have a dataset of 1.2 million molecules in Recursion’s library, where MatchMaker has said that for every one of those molecules, here is the protein it is predicted to bind to,” Palovich says.

At Valence, CEO Daniel Cohen sees a fit between the company he cofounded and Cyclica. “Cyclica has been focused on rapidly deploying their platform within Recursion’s drug discovery portfolio,” Cohen says. “Valence has been focused on developing new machine learning tools and technologies to feed into Recursion’s platform longer term.”

And Cohen notes an immediate synergy between the two acquired firms: a scoring function developed by Valence that measures the interaction between a ligand and its target and can be deployed in MatchMaker.

Meanwhile, Charles River intends for Logica to bring an AI component to its contract research service for drug companies.

“To be clear, we are a CRO and that’s what we’ll continue to be,” says Grant Wishart, Charles River’s senior director of small-molecule drug discovery. Wishart is also scientific operations lead in the partnership with Valo. He characterizes the Logica partnership as a milestone in an ongoing digitalization of Charles River’s pharmaceutical research service business.

“What we wanted to put in place is the ability to go from protein target to candidate molecule in a process driven by and aided by AI and machine learning,” Wishart says. “What we’ve done is brought Valo’s machine learning capabilities alongside Charles River’s huge experimental capabilities and drug discovery expertise to form our Logica partnership.”

Valo’s work in AI dates to the launch of a company called Numerate in 2007. Numerate, according to Guido Lanza, Valo’s chief technology officer and general manager of the Logica partnership, saw the pharmaceutical industry as “the most data-rich, algorithm-poor industry on earth” and wanted to take advantage of that.

Numerate also wanted to dig into machine learning methods of determining the pharmacokinetics, ADME (absorption, distribution, metabolism, and excretion), toxicity, and functional effects of compounds—areas that AI platform developers hadn’t fully exploited, Lanza says. The key, he adds, was access to a large store of data on compound activity.

Valo got off the ground in 2020 and immediately acquired Numerate. In 2022, Valo purchased Tara Biosystems, a specialist in generating data using induced pluripotent stem cells to evaluate compounds targeting cardiovascular and metabolic diseases. Tara also gave Valo a translational engine for data-rich hypothesis generation, Lanza says.

Valo began in-house drug discovery with its Opal platform, going it alone rather than pursuing deals with bigger drug companies. The pact with Charles River will put Opal’s molecular design capabilities to work as a service for other firms.

That partnership marks a kind of full circle for Lanza, who had been CEO of Numerate. “Before Valo, we were really in the service mode,” says Lanza, who is optimistic about a return to services in league with a CRO.

“I think the disruption AI can bring to the CRO business is real,” he says. “There is a really interesting tension between a technology that reduces the amount of experimental data you need and an industry whose business model is built on generating more data.”

Platform evolution

Brandon Allgood, formerly the chief technology officer at Numerate and now head of the consulting firm Obsidian Scientific, describes a progression of business models among the AI platform companies that poured onto the scene a dozen years ago. Most focused on small molecules. They also had a lot to learn about what was needed in pharmaceutical research, he says.

“We learned at Numerate that we can come in and predict activity,” Allgood recalls. “The pharma companies said, ‘Our problem is not finding compounds that are active. Our problem is finding compounds that are active and bioavailable and don’t metabolize too quickly.’ ” What they required was a multidimensional, multiparameter approach addressing selectivity, toxicology, and ADME—all at the same time.

The start-ups also had to develop a business model for engaging with drug companies. “What you don’t want to end up doing is selling to the IT department or chemistry department, because they see you as competition, and they just want to use your software, expecting a software pricing model,” Allgood says.

It became clear that successfully providing AI services to the pharmaceutical industry required comprehensive engagement with the drug discovery process.

“Drug development is death by a thousand cuts,” Allgood says. “If you develop a point solution for one of those cuts, you don’t derisk the program. You really have to take a stem-to-stern kind of approach where you apply machine learning to data generation.” Access to data of adequate quantity and quality is at least equal in importance to providing an effective algorithm, he says.

“People have gotten a lot more nuanced about AI and where the opportunity exists in pharmaceuticals,” says Abraham Heifets, CEO of Atomwise, a pioneering AI platform developer that is dedicated to small-molecule drug discovery. “The challenge for AI today is demonstrating that you can do something the drug companies can’t do,” he says.

Drug industry partners are looking for AI to generate something new. “They are judging whether AI is able to take you places you couldn’t have gotten to otherwise,” Heifets says. “Our focus has been extrapolating to novel chemical space.”

Access to new data is important, “but more important is how you use and clean the data,” Heifets says. Generative AI will play a critical role in arriving at novel compounds, given the ever-increasing quantity of data being generated, he says.

Atomwise pursues two varieties of partnerships, Heifets says: deals involving specific assets in development and more open-platform engagements. An example of the latter is the $20 million agreement it announced last year with Sanofi in which the big French drugmaker will use Atomwise’s AtomNet platform in its work on five potential drug targets.

Customer demands

The deal is one of several that Sanofi has entered into for access to AI. In another, the company paid $180 million for an equity stake in Owkin in 2021 and formed a partnership to discover and develop treatments for four types of cancer. And last year, Sanofi signed a deal with Insilico, paying $21.5 million to advance drug candidates for up to six targets using Insilico’s Pharma.AI platform.

Earlier this year, Sanofi announced the expansion of a 2021 deal with CytoReason: it will employ CytoReason’s AI platform to identify patient subtypes and pair them with therapeutic targets in inflammatory bowel disease. The companies originally joined to work on asthma patient subtypes.

Last year Sanofi announced a $100 mil-lion collaboration with Exscientia to develop up to 15 small-molecule drug candidates in oncology and immunology using Exscientia’s AI platform. In 2019, Sanofi licensed an Exscientia bispecific small molecule aimed at two distinct targets in inflammation and immunology.

It’s not unusual for large drugmakers to establish targeted partnerships with several platform providers that are doing similar work at competing drug firms.

“That’s a good thing,” says Diane Wuest, vice president and head of digital R&D at Sanofi. “The wider the access that platform companies have to drug discovery efforts across the pharmaceutical industry, the more their platforms improve,” she says. If Sanofi wanted exclusive access to a developer’s platform, “that would lead us more toward an acquisition-type arrangement,” she says.

Wuest says Sanofi pursues partnerships primarily to access an AI developer’s algorithms and scientists. “A pharmaceutical company like us—very large, in business many years—actually has a lot of data. And we have the ability to generate data,” she says. Partnerships are “almost like bolt-on capability.”

Genentech also seeks to enhance in-house efforts through AI partnerships, says Amit Mehta, head of business development at Genentech Research & Early Development (gRED).

“We look for partners who bring a novel dimension,” he says. “Some partners bring advanced platforms that can be related to data generation or data management. There are some partners out there who have robust datasets that are of interest to us or have the capability to integrate multiple types of datasets.”

Partnerships generally involve close ties, Mehta adds. “We are not looking at just a simple transaction but actually a true collaboration,” he says. “All of these partnerships entail free exchange of information and data, working hand in hand across different areas of expertise, which can be biology or chemistry.”

Like Sanofi’s Wuest, Mehta sees an advantage in close engagement with platform providers that are collaborating with competing drugmakers. “I think it’s great for platform companies because they get to see diversity of approaches,” allowing them to strengthen their technology, he says. “I think it’s just generally great for the entire industry. Companies are careful to make sure highly confidential information is not leaked by accident.”

Acquisition is also an option. In 2021, Genentech acquired Prescient Design, an AI drug discovery specialist launched by researchers at New York University and the Flatiron Institute.

“They were working on applying novel machine learning tools to antibody design,” Mehta recalls. “By fully integrating Prescient within the gRED umbrella, we were able to leverage all the datasets that we have [generated] over the years with their capabilities in machine learning. We can now focus on being highly effective in antibody design. It’s serving as a great accelerator for our large molecules.”

While drug company research leaders foresee little change in their approach to AI partnerships, a recent announcement by SandboxAQ, a molecular simulation spin-off from Alphabet, the company that owns Google, suggests a turning of the tables in which the AI platform developer is at the center of a constellation of drug companies and academic drug discovery labs.

After what it calls “several years of stealth development,” last month SandboxAQ announced it is working with collaborators including AstraZeneca, Sanofi, and the University of California, San Francisco. The group is using SandboxAQ’s AI and quantum computing capabilities in several therapeutic areas, including cancer, Alzheimer’s disease, and Parkinson’s disease.

SandboxAQ is hardly Alphabet’s debut in AI for pharmaceutical research. Its DeepMind division’s AlphaFold program made news in 2018 and 2020, when teams using the program placed first in an annual contest for predicting protein structures. The program has been hailed as a breakthrough in AI for drug research.

Research heads at major pharmaceutical companies are enthusiastic about the pace at which AI is moving into laboratories and their options for bringing outside technology and talent on board.

“It seems like an incredible time to be developing medicines,” Mehta says. “Today we can probe human biology at a much deeper level, we can generate high-resolution data at scale—which is unprecedented—and we have access to tools such as AI to draw insights into those data. When we look at just the vast array of modalities that are available for us to act on those biological insights, it’s amazing. By bringing all these capabilities under one umbrella, we feel like we might be writing the next chapter in health care.”