From Code to Bio: Why Every AI Lab Is Suddenly a Drug Company

Five years ago, if you'd asked the people running the world's leading AI labs what they were building toward, the answer was some version of "a system that understands and generates language, code, or images." Ask the same question today and a striking number of them answer with a molecule.

DeepMind spun out Isomorphic Labs in 2021 to apply its structure-prediction breakthroughs directly to drug design. Nvidia built an entire biology division -- BioNeMo, Clara -- on top of its GPU stack. Meta's protein language model research spun out as EvolutionaryScale, an independent company that raised over $140M to keep pushing generative protein design. Microsoft Research has quietly shipped foundation models for molecular generation and protein dynamics. And a new generation of AI-native biotechs -- Isomorphic, Xaira Therapeutics, Insilico Medicine, Recursion (now merged with Exscientia), Chai Discovery -- have collectively raised many billions of dollars on a single shared thesis: the tools that made language models work will make biology tractable in the same way.

The moment that changed the calculus

The credibility for this thesis traces to a specific, verifiable result. In 2020, DeepMind's AlphaFold2 solved a fifty-year-old grand challenge in biology -- predicting a protein's three-dimensional structure from its amino acid sequence -- with an accuracy that stunned the structural biology community at the CASP14 competition. By 2024, AlphaFold3 extended that capability to predict how proteins interact with DNA, RNA, small molecules, and each other. That same year, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry to Demis Hassabis and John Jumper for AlphaFold, alongside David Baker for his complementary work on de novo protein design. It was, in effect, the field's official confirmation that this was no longer a research curiosity -- it was a new instrument for biology, on par with X-ray crystallography or cryo-EM in the scale of what it unlocked.

The Isomorphic thesis: decades into months

Demis Hassabis has been consistent and public about what he thinks this instrument is for. In interview after interview, he has framed the ambition in stark, almost uncomfortable terms: drug discovery today can take ten to fifteen years and cost billions of dollars per approved medicine, and he believes AI-driven methods could compress critical parts of that timeline dramatically -- from years to months, and for specific computational steps, from months to weeks. Isomorphic Labs, which he chairs, has since signed collaboration and licensing deals with Eli Lilly and Novartis worth a combined figure north of $3 billion, and in 2025 raised $600 million in its first external funding round from investors including Thrive Capital and Google's own venture arm.

The claim isn't that computers will replace clinical trials. It's that the front half of the pipeline -- deciding which molecule to make and why it should work -- no longer has to take years of trial and error in a wet lab.

It's not just DeepMind

What makes this moment different from previous waves of "computational drug discovery" hype is the breadth of serious capital and serious researchers now pointed at the problem simultaneously, from multiple independent directions:

• Nvidia has built BioNeMo and Clara into a genuine platform business, supplying the GPU infrastructure and pretrained biological foundation models that dozens of biotechs now build on top of, effectively becoming the arms dealer for the entire AI-bio wave.
• Meta's protein language model research -- the ESM family, culminating in ESM3, a model that can reason across sequence, structure, and function simultaneously -- was significant enough that its creators spun it out as EvolutionaryScale rather than keeping it inside a social media company.
• OpenAI's leadership has repeatedly named biology as one of the domains where they expect frontier models to have the most consequential real-world impact, and the company has backed biotech ventures including Retro Biosciences, which is explicitly working on extending human healthspan using AI-designed biological tools.
• Xaira Therapeutics launched in 2024 with over $1 billion in funding and essentially no products -- just a thesis, deep AI/ML talent (including David Baker as a scientific co-founder), and a bet that AI-native drug design would outcompete traditional discovery from day one.
• Insilico Medicine's rentosertib, an AI-discovered and AI-designed candidate for idiopathic pulmonary fibrosis, reached mid-stage clinical trials -- one of the first drug candidates where both the biological target and the molecule itself were substantially generated by AI systems rather than found by traditional screening.

Why now, and not ten years ago

The honest answer is that none of the individual ingredients here are entirely new -- computational chemistry and structure-based drug design have existed for decades. What changed is the same thing that changed for language models: the transformer architecture turned out to transfer remarkably well from text to biological sequences, GPU compute got cheap and abundant enough to train models at a scale that actually captures the complexity of protein folding and molecular interactions, and public structural datasets like the Protein Data Bank -- accumulated over sixty years of painstaking experimental work -- turned out to be exactly the kind of large, high-quality training corpus these models needed. Biology, it turns out, obeys scaling laws too.

The part the headlines leave out

None of this means a drug can go from idea to pharmacy shelf in months. Generating a promising molecule faster is genuinely transformative for the discovery phase of drug development -- but discovery is only the first of many stages, and several of the stages that follow are bounded by biology, statistics, and regulation in ways that raw compute doesn't touch. A Phase 3 trial still needs enough patient-years of exposure to detect a rare adverse event with statistical confidence. A safety follow-up period is still measured in the number of months a human immune system needs to reveal a problem, not in GPU-hours.

That distinction -- between compressing discovery and compressing the entire pipeline -- is the subject of the next piece in this series. It's also, not coincidentally, the reason the statistical and regulatory-evidence layer of drug development is becoming more important in this new era, not less. Faster candidate generation only turns into faster approvals if the evidence generated along the way is rigorous enough to be trusted by regulators, physicians, and the patients who will eventually take these medicines.