Inductive Bio on a Winning Streak With ADMET Predictions
By Deborah Borfitz
April 2, 2026 | Following on the heels of a similar victory in 2025, Inductive Bio earlier this year took home gold in an all-comers prediction competition focused on the properties of chemical compounds critical to drug development. It would have been a David vs. Goliath moment had not some of the big pharma companies it outperformed already been familiar with its Beacon technology or established drug discovery partners.
In the latest OpenADMET-ExpansionRx blind challenge, Inductive Bio placed first among more than 370 submissions in what has become the “Olympics of ADMET prediction,” according to Josh Haimson, co-founder and CEO of the AI-driven company specializing in building virtual chemistry labs for designing high-quality molecules at top speed. ADMET—short for absorption, distribution, metabolism, excretion and toxicity—determines if a compound is safe and effective in the human body, and it’s data regulatory agencies like the Food and Drug Administration (FDA) require for drug approvals as it represents the key, measurable processes that define a drug’s pharmacokinetics.
In the initial Polaris ADMET challenge last year, there were only 39 competitors, and the target was predicting ADMET for a coronavirus main protease program, he says. The latest predictive modeling exercise was aimed at ADMET for myotonic dystrophy, amyotrophic lateral sclerosis, and dementia.
The therapeutic arenas were tied to the availability of blinded data donated by, respectively, the nonprofit ASAP Discovery Consortium and Expansion Therapeutics. Both competitive events were organized by the Open Molecular Software Foundation.
“These competitions really serve as validation points for the platform we’re building,” says Haimson, adding that the company’s competitive advantage is “the combination of the team, the data, and the feedback loops that we’ve built here.” The digital simulations Inductive Bio is building “mirror the way science is done in wet labs” but uses AI to do it faster and better.
The traditional process of designing a drug that can proceed into human testing can take three to four years of trial and error, he says. By building virtual models of those experiments, Inductive Bio can compress those timelines to between nine and 12 months.
The company’s AI chemistry assistants “use those models to explore millions more hypotheses than any human can, and in doing so surfaces the best molecules to move into the wet lab and ultimately the clinic,” says Haimson. “We’ve now been able to do that across dozens of active development programs.”
But the “real question” everyone is asking is what’s real and what’s fluff with AI, and “for this specific problem of predicting and optimizing the ADMET and pharmacokinetics of molecules, it’s ... not hype,” Haimson says. “We’ve now proven it in these back-to-back blinded competitions where there is no way to cheat.”
Flatiron Roots
Both Haimson and his cofounder Ben Birnbaum, Inductive Bio’s chief technology officer, hail from Flatiron Health, where they were leading machine learning and data teams heavily focused on clinical research in oncology. Flatiron’s acquisition by Roche in 2018 kicked off the journey into earlier stage drug discovery.
There was, at the time, many inflated expectations surrounding AI’s capabilities, Haimson adds, making him both an AI practitioner and skeptic. He consulted with hundreds of scientists within and outside Roche in search of truth, finally landing on the molecular optimization area as a place where state-of-the-art AI was starting to get “good enough ... to move the needle” in real-world deployment.
“If you spend enough time looking at real-world oncology data, it’s very humbling and it becomes very clear that a lot of patients still don’t have good treatments,” says Haimson, which led him and Birnbaum to the conclusion that they could do more impactful work further upstream in the discovery process. They launched Inductive Bio in 2021 with the mission of breaking down the drug property dataset silos of different pharma and biotech companies—much as Flatiron was doing to collapse the data silos between cancer clinics—by providing a secure mechanism for data-sharing and common learning.
Industry has made thousands of molecules and measured properties about how those molecules interact with the body, and Inductive Bio created the Pre-competitive ADMET Consortium to unlock that treasure trove for shared exploration. “Our partners are able to contribute data to train our models in a secure and IP-protected way, which then allows every member of that consortium to benefit from the collective information being shared across the industry,” Haimson says.
The other distinguishing feature of the Inductive Bio approach is that its models are being actively used in drug discovery programs, allowing the company’s AI engineers and drug discovery scientists to continually teach the algorithms using feedback on the predictions they got right and wrong. Among the company’s publicly disclosed biotech partners are Nested Therapeutics, Rapport Therapeutics, Arrakis Therapeutics, Architect Therapeutics, Aleksia Therapeutics, Nexo Therapeutics, Belharra Therapeutics, and Tenvie.
Customers are working across therapeutic areas, including oncology, inflammation, neuroscience, and rare diseases, says Haimson. In all cases, the focus is on ADMET, “one of the core bottlenecks in developing a drug and getting it to the clinic no matter what disease you’re trying to treat.”
Flipping to the ‘Avoidome’
It is unsurprising that this year’s OpenADMET-ExpansionRx event had such high participation, as ADMET has seen growing interest and focus in drug discovery, says Haimson. “It shouldn’t have to take four years and 4,000 molecules to find that one needle in the haystack ... that can make its way through the body effectively to become a drug.”
ADMET can be thought of as “everything a drug needs to do in the body aside from hitting the target,” Haimson explains. When designing an oral medication in oncology, for example, that means not just effectively killing cancer cells but also resisting degradation by stomach and liver enzymes, making its way into the intestine and then permeating across the intestinal wall into the bloodstream where it dissolves and gets distributed throughout the body to wherever the tumors are—including crossing the blood-brain barrier if the cancer has metastasized to the brain. “All along the way, it can’t hit other things and cause a bunch of chaos that would create toxicity or tolerance issues with the drug.”
The “avoidome” is a new term coined by drug discovery experts that refers to proteins that drug candidates should avoid interacting with to prevent poor ADMET, as well as adverse pharmacological effects and safety liabilities, he shares. “For a long time, the field has focused on genomics and transcriptomics and all the omics-es that help us understand what to target in the body to cure a disease, and the avoidome is really trying to flip that ... to say it’s just as important to avoid all the things that you need to steer away from that can cause a drug to fail in humans.”
While AI technology is clearly now up to the task of assessing and predicting the ADMET properties of small molecules, there are many other areas where large, curated, pre-competitive datasets don’t exist to better understand the biology and do meaningful work, Haimson says. Open benchmarking initiatives are a “true test” of whether models can work on molecules that have never been disclosed publicly before.
The latest competition attracted submissions from across the board, including first runner-up Merck in collaboration with NVIDIA and second runner-up EMD Serono—two of the largest pharma companies and the biggest AI company in the world. This was an anonymous competition but, based on disclosures, others in the race were various academics, drug discovery AI groups, and individuals employed by biotechs.
Toxicity Modeling Project
Inductive Bio plans to compete in other upcoming competitions, says Haimson, but winning them is tangential to its main goal of helping its drug discovery partners develop and discover better drugs. The competitions are “useful proof points to show with independent validation that what we’re building actually works ... and we truly are the best in the field at [ADMET modeling].”
And that’s a big advantage in an increasingly crowded AI marketplace. As Haimson rightly points out, seemingly every week now a new AI company emerges asserting that its platform can dramatically reduce drug development costs and timelines.
“I think there are a lot of overhyped folks in the space that are claiming what isn’t real yet,” he says. Even with the Beacon models of Inductive Bio, moving from early chemical matter to a development candidate that is ready for human testing “is not because AI on its own is imagining the perfect molecule. It’s really about the AI as a tool that empowers scientists to focus their time and resources on the most promising molecules and the most likely ideas that have a higher probability of success when they go into the wet lab.”
Toxicity remains one of the grand challenges of drug discovery, says Haimson, in pivoting to Inductive Bio’s news that it was awarded up to $21 million from Advanced Research Projects Agency for Health (ARPA-H) to lead a multi-institutional team in developing next-generation drug toxicity models to advance safer therapeutics and reduce reliance on animal testing. Safety issues are hard to detect preclinically with animal safety tests, making one of the big focus areas for the current FDA and administration how to better model the safety of drugs in humans while also reducing the number of animals needed for that research.
The award is from the CATALYST (Computational ADME-Tox and Physiology Analysis for Safer Therapeutics) program of ARPA-H and will be building toxicity prediction models in conjunction with Amgen and several leading academic pioneers in human liver organoids and ex vivo human tissue systems, he reports. Together they will generate an abundance of data on two of the largest drivers of toxicities, drug-induced liver injury and cardiotoxicity, which account for about 40% of safety issues that appear in the clinic and terminate drug programs.
The first three years of the so-called DATAMAP (Digital Acceleration of Toxicity Assessment with Mechanistic and AI-driven Predictions) project will be devoted to developing and validating that the “digital liver” models reflect how a human liver will respond to various drugs, including the toxicities. Much of the same underlying technology and architecture used for the Beacon models will be leveraged, Haimson notes, although “we’re going after a whole new class of biology and data to have a shot at being able to predict these [safety signals].”
In the final two years of the DATAMAP project, Amgen is planning to work with the models in support of regulatory submission to the FDA for new molecular entities, says Haimson.