Key Drivers of AI Drug Discovery

More than two decades ago, physicists-by-training Colin Hill and Iya Khalil, PhD, co-founded Gene Network Sciences (GNS), a biosimulation company based in Somerville, MA, on the back of the Human Genome Project and the growing excitement around genomics and multiomics. The start-up was an early entrant in the bold new era of systems biology that was being championed by the likes of George Church, PhD, and Lee Hood, MD, PhD. (Khalil moved to big pharma in 2020 and is currently vice president and head of data, AI, and genome sciences at Rahway, NJ-headquartered pharma giant Merck & Co.)

Few people were talking about artificial intelligence (AI) back then, but the roots were already being laid, as Hill noted at The State of AI in Drug Discovery, a virtual summit hosted by GEN on October 30 that drew more than 5,000 registrants. “From the beginning, we were an AI company, but one really focused on human biology, the biological mechanisms of human disease,” he recalled. “The terms that were floating around kept on evolving from bioinformatics to big data to precision medicine to AI.”

Two years ago, GNS was rebranded as Aitia. The new name, which is derived from the ancient Greek word for “cause,” refers to the company’s use of causal inference models to uncover the causal mechanisms of diseases. Hill, who today is Aitia’s CEO, noted that the new name was adopted “to indicate our move to being platform plus therapeutics.”

At the virtual summit, Hill participated in a session on The Business of AI in Drug Discovery along with Simon Barnett (a research director at Dimension, a New York, NY-based venture capital firm) and Sabrina Yang, PhD (co-founder and chief innovation officer at Watertown, MA-based Empress Therapeutics). During the session, Hill observed that although most of the buzz around AI in biotechnology is around the need to transform the discovery and design of small molecules and antibodies, bigger problems loom both upstream and downstream. He asked, “What are the right drug targets to go after? Can we start to unravel the actual underlying circuitry of human disease?”

Such questions vex us despite all we’ve learned in the two decades that have passed since the Human Genome Project’s completion. “We’re lucky if we know 5% of the genetic and molecular circuitry of disease,” Hill said. “That’s clearly at the heart of why drug discovery and clinical development has been such a trial-and-error process with persistent 80–90% failure rates in clinical development.”

From the outset, Hill’s focus has been to apply causal AI (based on the Turing Award–winning work of Judea Pearl, PhD) to reverse engineer the hidden circuitry of disease from human multiomics data and animal models. “We call the resulting models Gemini digital twins, because they are becoming more accurate replicas of human disease, more so than an animal model or cells on plastic will ever be,” Hill continued. These replicas are used to conduct computational experiments, simulating the equivalent of small interfering RNA knockdowns in silico to discover hidden drivers of disease progression, which can become drug targets once validated.

Another goal downstream of designing drugs is to improve the design of clinical trials. “It takes a village,” Hill said. “I think together we will be transforming biomedicine in a major way soon.”

New dimension

Simon Barnett joined Dimension as research director from St. Petersburg, FL-based ARK Invest in 2023, believing the $350-million “venture vehicle” was the right fit for his growing interest in the critical importance of computation on life sciences.

At the virtual summit, Barnett predicted that over the next 5–10 years, a small tranche of companies wearing the “tech bio” moniker will emerge, and that “every new company in the space, and increasingly in large pharma as well, [will] make computation a central part of what they’re doing rather than something that’s a bolt-on or secondary to the central mission of making medicine.” Barnett also shared his thoughts about the digitalization of the life sciences. For example, he said that it is about “combining high-throughput experimentation and next-generation computational techniques to tackle some of the biggest problems in drug development, diagnostics, and clinical trial design.”

Most AI-focused biotech companies are focused on making drugs, but these companies also need to consider factors such as disease indications, modalities, and strategies for reaching the clinic. “These companies are product businesses at the end of the day,” Barnett said. “They need to be aware of the types of risk that investors have a palate for. It’s a very different economic regime now than it was a couple years ago.”

Another priority is having either proprietary data or a valuable data-generating platform. These are “essential ingredients for developing a compelling story that would attract investment in this era,” compared to 5–10 years ago, says Barnett.

A key for attracting capital is having some level of validation of the technology, especially in the context of pharma collaborations. Hill’s company Aitia recently announced oncology drug discovery partnerships with Espoo, Finland-headquartered Orion Pharmaceuticals and Suresnes, France-headquartered Servier. Such collaborations are important, Hill said, as they provide evidence that respected biotechnology and pharmaceutical companies have vetted the technology and placed their bets with the company. He added that such deals also drive revenue and allows his company “to get to some very interesting places without having to raise hundreds of millions or even billions of dollars.”

Having an interdisciplinary team—one that includes not just experts in computational AI, but also experts in other aspects of drug discovery—will also remain a priority for investment considerations.

Hill suggested that growth will come as the industry starts to see Phase IIa readouts that are positive, especially if they reflect discoveries that would not have been made but for the data and the AI technology. “When we see the first drug candidate start to have some success against a biological target that truly no-one knew before AI found it—that’s when we’re going to say, ‘Aha, we’ve entered this new era,’” Hill declared.

This new AI era is not just about understanding protein folding or molecular design of therapeutics, but about grasping the complexity of human biology. Hill indicated that he anticipates more investment will enter the space even before drugs reach the market. “It will be fantastic when that happens and we get a drug to market,” he said. “But the game is going to be played out and won or lost even before that that endpoint.”

Barnett estimates that there are several dozen companies that would self-identify as having machine learning or another data-driven technique as core to their R&D operations. One such company is Salt Lake City, UT-based Recursion Pharmaceuticals, which began with a high-throughput imaging platform to deconvolute novel biology. Last summer, the company announced plans to merge with Oxford, UK-based Exscientia, which is, like Recursion, a high-profile, publicly traded AI company.

While companies such as New York, NY-headquartered Schrödinger are using AI to design and optimize better molecules, others are solving for biology first. According to Barnett, companies focused on biology should ask, “How can I use AI to find causally relevant mechanisms and targets from, say, very complex human data, using that to deconvolute the most exciting biological mechanisms to pursue?”

Back to nature

Sabrina Yang is a senior principal at Cambridge, MA-based Flagship Pioneering, a venture creation firm. Best known as the firm behind Moderna (also based in Cambridge), Flagship has also founded more than 100 bioplatform companies, including the aforementioned Empress Therapeutics.

According to Yang, Empress Therapeutics uses “the power of AI and genetics to find better small-molecule drug starting points that are already invented by nature inside the human body.” The company uses AI to study genetic clues to the biosynthetic pathways of small molecules and elucidate what these molecules may look like, venturing into previously unknown chemical space.

“We hope that this approach will lead to a higher chance of success in the clinic,” Yang said. “We’re starting with molecules that are already pharmacologically active. Without AI, we would not be able to efficiently find the genes that encode for the compounds that are disease- and health-relevant.” She emphasized that AI is critical to understanding how these molecules have co-evolved with us and how they can be “co-opted as drugs that may have better safety advantages.”

Yang said that Empress Therapeutics generated 15 small-molecule drug candidates in less than two years with a team of just 30 people, delivering time and cost savings of 50% in comparison with traditional approaches. Ultimately, of course, the company’s approach has to be proven in the clinic, beginning with improved chances of success in Phase I safety trials.

Bottlenecks and challenges

If you were to ask each of 10 companies for its favorite list of targets, Yang said, you would likely find a big overlap between the different companies’ lists. There is still a big need to identify novel mechanisms that will ultimately translate into increased success in the clinic.

A major bottleneck is that AI is not yet positioned to immediately improve the speed and cost of clinical development, but Hill sees that changing in the future. True judgment will be seen when filing Investigational New Drug appilications: Are companies progressing drug candidates in which there is greater confidence and mechanistic understanding in the patient population? Will extensive multiomic validation of a target and the corresponding drug enhance the probability of success?

Hill predicted there will be a tipping point when AI impacts clinical development. “Yes, speed and cost [improvements] will happen too, but the real cost is in the cost of clinical development failures—that’s the bottleneck. Until we can change that, we’ll be nibbling at the margins.”

When AI companies reach the pivotal Phase IIa stage, there are bound to be failures and, Barnett warns, a backlash. “It’s going to be unfortunate,” Hill remarked. Critics will likely find fault with AI technology and complain that it has been overhyped, which could in turn hurt funding cycles. “We’re going to have to push through because, of course, technology is inexorably improving over time.”

Another challenge is in data. “The protein databank that powered AlphaFold took decades of curation and tens of billions of dollars,” Barnett said. “We’re just not going to be able to generate datasets of that scale on the fly.” He asserted that data availability, cleanliness, access, and metadata will be critical for companies, regardless of whether they work with clinical data or preclinical/biochemical data.