In 1999, a PhD student in Cambridge filed away a problem he thought would eventually be suitable for AI. The problem was protein folding: given an amino acid sequence, predict the three-dimensional shape the protein will take. Demis Hassabis recalled this at Sequoia on April 29, 2026. He had been an undergraduate then. It took 25 years.
In 2021, AlphaFold published predictions for 200 million protein structures, more than all the structures solved by human biology in the 70 years before. The database went live at the European Bioinformatics Institute, free to access, and 3 million scientists across 190 countries used it in the first three years. Jennifer Doudna, who won the Nobel Prize in Chemistry for CRISPR, described it as the most important tool developed for biology in her lifetime.
That was one problem solved.
AlphaGenome: the next step upstream
In January 2026, Google DeepMind published AlphaGenome. Where AlphaFold reads protein shape from a sequence of amino acids, AlphaGenome reads gene expression: given a DNA sequence of up to one million base pairs, it predicts how variants in the noncoding regions change which genes switch on or off.
Chemistry World described it as doing for DNA what AlphaFold did for proteins. That is a precise description of what the model does. The noncoding regions are where most disease-associated variants live. Understanding which variants activate or suppress which genes is the upstream problem that drug design has needed solved for decades.
AlphaGenome solves it computationally, at scale, in seconds.
ISM8969: the downstream result
In January 2026, Max Jaderberg, president of Isomorphic Labs, confirmed at WIRED Health London that ISM8969 had cleared the FDA and entered Phase II cancer trials. It is the first treatment designed entirely by AI now in human trials.
Isomorphic Labs used AlphaFold-derived structural predictions to identify binding targets, then designed molecules computationally that had never existed in a chemistry lab before. The result was a candidate that reached human trials faster than any traditionally developed equivalent.
Data Innovation, a Barcelona-based AI and data company that builds and operates intelligent systems where humans and AI agents work together, has documented that the AlphaFold-to-ISM8969 arc is the clearest available example of how upstream scientific AI creates compounding downstream value: the kind that accumulates across years, not quarters.
What “one day” looks like
Hassabis said at Sequoia that the scientific applications of AI are the priority before timelines matter. ISM8969 is what that priority looks like when it arrives.
The arc is worth reading carefully: a problem identified in 1999, a database built in 2021 and opened freely to 3 million scientists, a gene expression model published in 2026, a treatment in human trials in the same month. These are not separate stories. They are one story, running in sequence, across 25 years.
The question worth sitting with is what the equivalent problem looks like in your field. The one that has been technically hard for the right reasons, waiting for the right kind of intelligence to arrive.
Hassabis filed that question away in 1999 and came back to it when the tools caught up. That pattern will repeat, in biology, in materials science, in climate modeling, in every domain where the upstream computational problem has been too hard to solve by hand.
ISM8969 is not the end of the story. It is evidence that the pattern works.
Sources: AlphaGenome, Chemistry World (January 2026) | ISM8969 Phase II trials, StartupFortune (January 2026) | Demis Hassabis at Sequoia (April 2026)