Sold on the new machine

In December, Roche announced a deal with Recursion, an artificial intelligence technology developer with its own drug discovery program. Roche will access Recursion’s AI-driven drug discovery platform to identify and develop up to 40 new medications for cancer and neurological conditions. Hailed in the business press as a multibillion-dollar AI deal—a perspective heavily influenced by potential milestone payments— the engagement will begin with the payment of $150 million to Recursion.

Within weeks, Sanofi announced a similar deal with Exscientia, another drug-hunting AI technology developer. Focused on as many as 15 novel small-molecule drug candidates in oncology and immunology, the deal carries an up-front payment of $100 million and the potential for $5.2 billion more if all future milestones are hit.

By any measure, these deals are notable for their scope and the level of investment on the part of major pharmaceutical companies in AI-enabled drug discovery and development. Several other deals, smaller in scale, were announced in the first 2 weeks of 2022, seeming to make good on claims by AI platform developers that Big Pharma is ready to go big on AI.

There is no question that major drug companies are making bets on the future by teaming up with drug discovery firms that have used AI from their inception. But AI deployment is already underway at the big companies and has been for several years via in-house technology development as well as lower-stakes partnerships, most targeting specific projects or functions.

Drug companies are working toward achieving end-to-end AI coverage with applications stretching from the discovery of disease-causing proteins to the clinical trials that test drugs on people. While some technology platform developers claim they have succeeded in initial experiments with comprehensive AI for discovery and development, the consensus among drug companies and most other technology firms is that fully functional end-to-end AI is still a ways off.

Under the radar

AI has yet to produce a new drug from scratch, but the technology is hard at work behind the scenes in labs across the pharmaceutical sector. One recent success is the expedited approval of Pfizer’s Paxlovid, the oral-dose COVID-19 antiviral that won an emergency use authorization from the US Food and Drug Administration in December.

Pfizer used artificial intelligence and virtual screening to speed the design of nirmatrelvir, the antiviral component of Paxlovid, according to Charlotte Allerton, Pfizer’s head of medicine design. Machine learning, a variety of AI in which programming models evolve as they process data, was used to predict oral drug properties. “These models helped us triage virtual compounds so that we could prioritize making the ones most likely to achieve good oral drug properties as quickly as possible,” Allerton says.

Pfizer developed the models in-house and in partnerships, Allerton says. “In my world, designing small-molecule therapeutics, something Pfizer has a long legacy of doing, we have very extensive data points, millions of data points,” she says. “We feel we are in good shape in having built state-of-the-art AI models in order to enable us to more efficiently design oral drugs.”

Roche is in a similar position, according to Barbara Lueckel, the firm’s head of research technologies partnering. “Roche has been using machine learning and advanced analytics for nearly 10 years with a multitude of applications, internally and also with partnerships,” she says.

Both executives acknowledge that AI met with initial skepticism in the pharmaceutical R&D lab—scientists are inherently skeptical, Allerton notes. But there is little question today that AI will play an essential role in all drug discovery and development going forward. “I think it’s fair to say we see machine learning as something that can no longer be disconnected from how we think about the continuum from target discovery to clinical development,” Lueckel says.

Beginning of the end-to-end

Executives at AI-specialist companies say their stars are aligning. In addition to deals with big drug companies and clinical advances for their own AI-derived drug candidates, they are hitting significant technological milestones.

Exscientia and another AI firm, Insilico Medicine, both say they have established an R&D-wide AI system featuring a data-sharing loop that runs from early-stage drug target selection to the clinic and back again.

Both companies claim advances allowing them to combine automation with machine learning algorithms in biology and chemistry. Each, though, comes at these comprehensive platforms from a different starting position.

Exscientia got its start with a focus on drug design and chemistry. CEO Andrew Hopkins says the company has since added capabilities in biology—namely, for target identification—and clinical trials through acquisition, partnership, and in-house technology development.

“What we want to do is reinvent the drug creation process,” Hopkins says. “We have to go upstream and bring in AI for biology, but also downstream to find out how AI can be used to better understand patient responses.”

Exscientia first took control of a drug creation program in a partnership with Celgene in 2019, a year after it had added structural biophysics capabilities with the acquisition of Kinetic Discovery. Exscientia began developing target identification technology in-house, and last year it acquired Allcyte, an Austrian specialist in AI-based personalized medicine discovery.

Hopkins points to two recent deployments of Exscientia’s end-to-end approach: the nomination of an anticancer molecule in a partnership with GT Apeiron Therapeutics, an oncology drug discovery firm in China, and the launch of an in-house clinical trial in which AI used individual tissue samples to predict which therapy would be most effective for people with late-stage hematologic cancer.

Insilico, on the other hand, launched on the basis of biology-based AI, according to CEO Alex Zhavoronkov. “We originally started as a target discovery company, primarily relying on multiomics data analysis and protein expression,” he says. “However, in 2015 we realized that just being a target identification player is not enough, that the pharmaceutical companies are actually not willing to pay a lot for that part. We decided to develop our own chemistry engine.”

Also determined to front its own drug discovery efforts, Insilico developed tools to predict the performance of compounds in clinical trials, Zhavoronkov says. By last year, the company had amassed what it considers a full end-to-end AI platform, comprising separate software products for target discovery, molecule design, and clinical trial design.

Zhavoronkov has high expectations for what AI can do in the pharmaceutical industry. “End-to-end discovery,” he says, “means that you start with nothing, identify a novel mechanism of a disease, translate that into individual targets, prioritize, and zoom into specific targets for which you design or discover novel molecules.”

Core focus

Other AI technology companies are maintaining their focus on core businesses in chemistry or biology. Indeed, executives at some of these firms question the wisdom of offering end-to-end machine learning—and even whether such platforms can function as described at different steps in the R&D continuum.

“At Atomwise, we are firmly in the AI-for-chemistry camp,” CEO Abraham Heifets says. “The way we think about it, there is a lot of biology that is not contentious,” he says, whereas many biological targets prove difficult to act on with drugs. “Chemistry is the limitation.”

A chemistry specialist like Atomwise is in high demand in areas such as neuro-oncology, in which chemistry is the key to crossing the blood-brain barrier, Heifets says. Drug companies are also looking for chemistry-focused AI technology partners to help them with intractable protein targets and to optimize and amplify their stores of data.

“With partners, we’ve built a proprietary database of 3 trillion molecules,” Heifets says. “The larger and more diversified the chemistry space, the better the chance of finding a molecule that works with a protein.”

Heifets acknowledges that firms like Exscientia can configure AI into an end-to-end learning loop, “but it depends on what you want to learn.” A signal from clinical data can advance early biology work, he says, but if a target is known, signals from the clinic aren’t much use until the chemistry is worked out.

Insitro is an AI technology developer and drug discovery firm rooted in biology. CEO Daphne Koller sees an active niche for the company, which, as its name implies, seeks to meld in silico and in vitro drug discovery techniques.

“The core of the company, where we started, is creating cell-based models of human disease that are derived from human genetics,” Koller says. “We are relying on machine learning to understand what disease looks like at the cellular level and, with that, go back and understand what it correlates to for a human clinical outcome.”

The company did, however, take a step in the direction of chemistry when it acquired Haystack Sciences, a creator of DNA-encoded chemical libraries, in 2020. “It is enabling us to use machine learning in the chemistry space to much more rapidly identify novel ligands for a particular target,” Koller says.

Insitro has contracts with several pharmaceutical companies, including Bristol Myers Squibb. In 2019, Insitro collaborated with Gilead Sciences, applying machine learning to target discovery. Insitro created disease models for research on nonalcoholic steatohepatitis that incorporate data from clinical trials of potential drugs for the liver disease.

Koller is skeptical of claims about AI systems that run a circuit through the entire span of pharmaceutical R&D. “To be honest, I think it’s a little bit premature to actually call any platform truly end to end,” she says.

“I think there are companies that have efforts across different stages of the value chain and maybe use some components of experiments and tools between one and the other, which I think is really incredible progress. But to really think about true integration that goes from doing novel biology all the way through drug discovery into clinical trials . . . that’s a bit of hyperbole.”

The stretch

The major drug companies would like to establish R&D enabled by end-to-end AI, but research managers point to several areas that need further development. They are also awaiting the ultimate breakthrough—the approval of a drug discovered and developed by AI.

Eli Lilly and Company opened an autonomous drug discovery lab, the Lilly Life Sciences Studio, in San Diego in partnership with Strateos, an AI research-scheduling software firm. Ramesh Durvasula, Lilly’s head of research information technology, says the 2-year-old facility has worked on several discovery-stage projects for Lilly. “It is starting to have pretty significant impact across all our workflows,” he says.

But AI technology is only beginning to prove its effectiveness in drug research, says Utpal Singh, Lilly’s head of discovery chemistry. “Right now, there is a lot of buzz about tools,” Singh says. “But I think ultimately you have to translate those tools into outcomes that impact patients’ lives. We haven’t achieved that.”

Moving molecules into the clinic should be celebrated as a win for AI, Singh says, “but I’d like that to translate into a proof of concept and later clinical studies. Also, I’d like to see some data that expert medicinal chemists would take a look at and say, ‘That is a new insight.’ ”

Singh adds that much AI work has emphasized chemistry, whereas biology is a far more complex field. “It is still expanding,” he says. Researchers can point to efficient data sets they believe can predict pharmacology, “but the reality is that clinical testing has a way of humbling us.”

Allerton says Pfizer has a goal of achieving end-to-end AI coverage in R&D. She agrees with Singh and others that there are places where AI is quite well advanced in drug research—and others where more development is needed.

The company paid $13 million up front last month in a partnership with PostEra, an AI start-up based in Boston. A smaller deal than those announced by Roche and Sanofi, the engagement will bolster Pfizer’s internal AI capabilities with machine learning–based medicinal chemistry for drug design.

The deal will also expand access to data. “AI is a key focus in the design of small-molecule therapeutics,” Allerton says. “Many models require large data sets to be used effectively, and that can be harder to do when you are moving into novel target space—when there is less data around the target, less data to build those models from.”

Pfizer is interested in amplifying its internal capabilities and increasing its access to data through partnerships. “A lot of people are working in the field,” Allerton says. “That is part of the recent announcement with PostEra.”

At GlaxoSmithKline, Kim Branson, head of AI and machine learning, says the company is intent on making the most of its huge stores of data amassed over decades. GSK is also determined to use AI to generate more data that can be fed back into the system. “We generate data for the express purpose of improving machine learning models,” he says.

Last year, Branson says, GSK generated twice the amount of data that is publicly available on primary human immune cells. “On the order of 21 billion data points. And that’s all to feed that model,” he says. “All the AI systems we build have an experimental feedback loop.”

According to Branson, these loops are generally dedicated to a narrow purview in research, such as mapping gene variants. “There isn’t an overarching, coupled thing for the whole process,” he says.

Johnson & Johnson’s Janssen Research & Development unit has developed an end-to-end R&D data access and analysis platform it calls Med.ai. First deployed in 2020, Med.ai is designed to break down functional silos in R&D, giving researchers access to disparate forms of data from biology, chemistry, and clinical trials. The platform also taps externally licensed patient data from electronic health records and other sources.

But Najat Khan, chief data science officer for Janssen R&D and the firm’s head of strategy and operations, agrees with Allerton and others that AI’s successes are not evenly distributed across the R&D spectrum. “The jury on discovery is still out,” she says. “On the development side—clinical trials and stratifying patients—we are seeing more wins, and timelines to value are a little quicker.”

Meeting of minds

Sam Heaps, a specialist in machine learning and bioinformatics at BioTeam, a health-care informatics consultancy, says deals between drug companies and AI specialists will foster multidisciplinary teams inside pharmaceutical labs. Heaps notes that drug companies have experts in computer science, statistics, information technology, and traditional biomedical research. “But the real key is to get these people together and talking to each other,” he says.

Big Pharma research managers agree that connecting with AI-focused drug discovery companies will strengthen their teams. But it’s important to find the right fit.

Roche’s Lueckel says her firm’s engagement with Recursion hinges on complementary skills and expertise as well as a shared mindset about how to go about drug discovery. And putting Recursion’s automated high-throughput imaging platform to work with Roche’s in-house database gleaned from high-resolution single-cell sequencing is potentially transformational, Lueckel says.

Recursion says its imaging system generates petabytes of data. “We perturb human cells using a wide variety of perturbants—CRISPR knockouts, cytokine and chemokine viral toxins, for example—and look at everything from size, shape, and spatial distribution to organelle structure,” says Shafique Virani, the firm’s chief business officer.

Machine learning algorithms interrogate the cells before a massive screening of small molecules to identify molecules that can return cells to a normal or healthy state. The process is enhanced by cell painting, an imaging technique that uses fluorescent dyes to label organelles.

Lueckel says that combining Recursion’s platform with Roche’s database will give researchers access to more than just phenomaps showing the correlation between cells that have been modified by genetic intervention or chemical compounds. The combined technologies will generate multimodal maps that provide more details of human cellular biology, including how changes to cells affect messenger RNA, the code that cells use to translate genes into proteins.

Virani agrees with Lueckel that Recursion and Roche make a good match in the lab. “They have a strong focus on computational biology, and from an expertise and infrastructure perspective they have certainly built up a critical mass of data with information scientists and AI scientists as well.”

Virani, who once worked at Roche as head of neuroscience, ophthalmology, and rare-disease partnering, says Recursion is also making headway on its own. The company has licensed Takeda Pharmaceutical’s TAK-733, a clinical-stage compound being developed for hereditary cancer syndrome and related diseases. And the company hopes to bring four of its own pipeline candidates up to Phase 2 or 2/3 studies this year.

Sanofi’s deal with Exscientia, a follow-on to a partnership between the companies focused on the design of bispecific small molecules, aims to deploy Exscientia’s newly expanded capabilities, according to Frank Nestle, chief scientific officer at Sanofi.

Nestle views AI in the context of remaking research—as an effort to “reimagine drug discovery and development from end to end.” The goal, he says, is the long-standing one of speeding the process and reducing the cost of bringing a drug to market, a venture that currently takes 12 years and costs $1 billion on average, he says.

“The two keys are how long to a clinical candidate and what is the likelihood of success,” Nestle says.

Like Heaps, Nestle points to an evolution toward multidisciplinary research teams as a benefit of partnerships. Sanofi appointed Arnaud Robert in 2020 as its first full-time chief digital officer charged with directing the company’s digital transformation regimen in R&D.

But Derek Lowe, a drug researcher and author of the In the Pipeline blog, shrugs at the notion of partnerships fostering interdisciplinary teams.

“I haven’t noticed it if it’s happening,” Lowe says. “The larger companies that have more data to work with have been integrating data science into their stuff for a long time now.” Instead, he sees the recent large-scale, multiyear contracts as sensible bets on the future.

“AI is getting better all the time, and we’re not,” he says. “Machine learning programs like retrosynthesis are OK right now. They are kind of clunky and give us weird answers, but they are getting better every week.”

Describing himself as a “short-term pessimist but a long-term optimist” in regard to AI, Lowe says researchers are coming around to the technology. “I think people are very interested in it,” he says. “There is a lot of skepticism, especially among people that have been around the industry for a while. But at the same time, there is a lot of interest and a lot of hope that this thing may help out.”