Chris Bahl saw the writing on the wall for his new AI startup in the final weeks of 2022: Molecules would matter more than models.
A pair of preprints had just come out, both showing diffusion models that could generate not just images, but proteins. Bahl, then the chief scientist of his startup AI Proteins, realized that the still-young AI bio sector was about to change.
He concluded that the race to build leading AI protein models was too commoditized, especially as these cutting-edge models called RFdiffusion and Chroma became freely available as open-source technology. Bahl stopped his team’s efforts on building general-purpose protein models, shifting focus to narrower AI applications for other drug development tasks.
“I saw no path to ROI for generating a proprietary model,” Bahl, now the CEO of AI Proteins, said in an interview.
Today, much of the AI bio field has come around to Bahl’s position, turning its focus to the drug industry’s bread and butter of creating and developing new molecules. New, advanced AI models for biology are still being developed and used in that process. But there’s less interest in startups advancing cutting-edge AI tech with a hunch that a viable business will follow.
These startups need to show that the latest AI technologies can produce not just preprints, but actual drug programs. This next generation of companies, like Bahl’s, is sprinting to the clinic, with 2026 shaping up to be a reality check on whether generative AI can finally move the needle on R&D productivity.
To analyze what 2026 holds for AI in biotech,
Endpoints News
reviewed the pipelines of 11 leading AI drug startups, mostly founded between 2018 and 2021.
Many are finally moving into human testing, where the sector’s big promises will be tested: Out of the 11 companies, there are currently eight programs in human testing. By the end of next year, that number is expected to nearly triple to 21 programs in the clinic.
“There’s a perennial swing between assets and platforms,” said Elliot Hershberg, a biotech investor at Menlo Park-based Amplify Partners, which raised its first dedicated biotech fund this year. “That has clearly dialed more towards assets in this recent market,” he said in an interview.
There are many examples. This year, San Diego-based Iambic
raised $100 million
after its first clinical readout, and plans to use the cash to bring two more drugs into human testing in 2026.
Earlier this month, Generate:Biomedicines
took its lead program into Phase 3
— a milestone that the first wave of AI-focused biotechs like Recursion, Insilico Medicine, and Absci have yet to reach. The Flagship-founded company expects to soon file two more new drug applications with regulators, for cancer programs.
And while the field’s richest startups, including Isomorphic Labs and Xaira Therapeutics, have remained tight-lipped on lead programs and clinical timelines, both have said they’re also largely focused on building pipelines. Isomorphic, for example, used a chunk of the $600 million it raised this year
to establish a Boston area office
and hire up a clinical development team led by ex-Relay Therapeutics’ Ben Wolf as chief medical officer.
Others, however, seem stuck in what Hershberg calls a “weird danger zone,” between building a pipeline and selling tools or services. That includes cautionary tales like EvolutionaryScale, which raised a $142 million seed round in 2024 with plans to build massive AI models in biology. This year, with little to show for progress since its debut, it was
folded into Mark Zuckerberg’s nonprofit science project
, under undisclosed deal terms.
The sector’s evolving focus has sometimes led to major changes at the companies. After five years of “AlphaFold-for-X” startups inspired by DeepMind’s 2020 protein structure-predicting model, many of the companies are pivoting to pipelines.
London-based CHARM Therapeutics was founded in 2021 by Laksh Aithani, then a 23-year-old wunderkind who developed an AI model called DragonFold to predict protein-ligand structures.
A first-time CEO, Aithani raised a $50 million Series A in 2022 from top-tier investors and dreamed of building the next AstraZeneca off AI’s potential.
Instead, this spring, he
left as CEO
in a mutually agreed decision with the board. The company is now led by Gary Glick, its 64-year-old executive chair and a serial biotech entrepreneur, and the CEO slot is empty. Under Glick, CHARM
raised an $80 million Series B
, meant to get its lead drug into the clinic in early 2026.
While CHARM’s Series A was driven off the potential and excitement for DragonFold, the Series B was fueled by preclinical data packages and a plan of reaching clinical proof-of-concept.
“You don’t get credit for academic papers,” said Carl Hansen, the CEO of AbCellera, a computational-focused biotech he co-founded in 2012. “You don’t get credit for things on the pipeline that are preclinical. You get credit for molecules.”
Hansen pivoted AbCellera’s own business strategy in 2023 to focus on the pipeline. He expects a readout in mid-2026 for its lead candidate, an antibody targeting NK3R.
“Our belief is: ‘Show me the money,’” Hansen said in an interview. “The money is drugs in the clinic that ultimately get good data and make a difference.”
Not that getting to the clinic solves an AI startup’s challenges. It’s just the beginning of a grueling journey full of hard decisions on how to move forward. Many of the high-flying AI startups will ultimately be measured by how they do at the boring, old work of designing and running the right clinical studies, often in competitive crowded drug classes like TL1A or HER2.
“Suddenly you’re competing not on discovery. You’re competing against your ability to do clinical development and have a low cost of capital,” Hansen said. “Before you know it, you’re $400 million into that and it no longer makes any sense to develop, to invest in, the platform capabilities.”
AbCellera doesn’t publish or talk much about the intermediate steps behind its own antibody discovery process. With the market’s emphasizing assets, Hansen said the company wouldn’t — and shouldn’t — get credit for that type of work.
“Everyone should be skeptical of everything everyone is saying, including me, until I show you the evidence,” Hansen said. “The evidence has to be smart choices on drugs that are highly differentiated that are at least in the clinic.”