The most interesting shift of the last two years is not what generative models can predict, but who can prove it in the clinic.
In 2024 the Nobel Prize in Chemistry went, in part, to the team behind AlphaFold — the system that learned to predict the three-dimensional structure of proteins. It was the moment generative AI stopped being a promise made to biology and became a tool working inside it.
Two years on, the conversation has changed. The question is no longer whether a model can design a molecule. It is whether anyone can turn that design into a medicine.
That distinction is quietly reorganising an entire sector.
The first wave of companies sold software: models that proposed targets, generated chemistry or predicted structures, then handed the work to pharma. The companies attracting the most serious capital today look different. They are building what investors now call “AI-native biotech” — owning the full loop, from proprietary data generation in automated labs, to foundation models, to their own therapeutic programmes.
The reason is simple. As models become widely available, the algorithm stops being the advantage. The advantage moves to whoever controls high-quality biological data that cannot be scraped from the internet, and to whoever has the discipline to test predictions against reality. Bessemer Venture Partners has described this as the race to build “biology-native data infrastructure.” It is telling that large AI labs, Anthropic among them, are now paying to access proprietary biological datasets rather than assuming that compute alone will win.
The clinic is where the claims are settled. In December 2025 Recursion reported durable reductions in polyp burden in an early trial of an AI-discovered candidate for a rare hereditary cancer condition — one of the first AI-enabled programmes to show a real clinical signal. Isomorphic Labs, the DeepMind spin-out behind AlphaFold, raised $2.1 billion in May 2026 and is approaching its first human trials, with candidates designed using AlphaFold 3. Notably, its timeline slipped from late 2025 to late 2026 — a reminder that converting a prediction into a validated therapy remains slow, expensive and uncertain.
This is the part the enthusiasm tends to skip. Generative models compress the earliest stages of discovery. They do not shorten biology, regulation or the years a trial takes to read out. The companies that endure will be the ones that respected that gap from the start.
For early-stage investors, the lesson is a familiar one. Technological waves survive the bubbles around them, but only the founders who build real companies survive the wave. In life sciences that means teams who pair computational ambition with experimental rigour, who treat data as an asset to be generated rather than borrowed, and who plan for the clinic rather than the demo.
This is where private capital with real domain knowledge earns its place. Backing these founders is not about betting on a model. It is about judging which teams can close the distance between what an algorithm proposes and what a patient receives — and staying with them long enough to find out.
The winners will not be those who build the most powerful models, but those who consistently translate computational insight into clinical impact.
