Token Talk 26: AI in the (Protein) Fold

By: Thomas Stahura

Remember in January when Larry Ellison and Sam Altman stood in the Roosevelt Room of the White House and, as part of their massive $500 billion "Stargate" AI project, talked about using AI to cure cancer?

Such a beautiful vision.

But, six months later, the reality is building the massive data centers for Stargate is one thing; curing cancer is another. Those moonshot promises haven't materialized. But that doesn’t mean AI isn’t quietly reshaping the drug discovery landscape in ways that are far more tangible and far less attention-grabbing. The real progress is happening in the lab, where AI is being used to solve the painstaking, age-old problem of protein folding. And nowhere is this more evident than in the story of AlphaFold.

Proteins are the building block of life. They’re responsible for nearly every biological process in your body, from digesting food to repairing cells to fighting off infections. What determines a protein’s function is entirely dependent on its three-dimensional shape. And if they fold incorrectly, it can lead to diseases.

Proteins themselves are made of chains of amino acids, each with unique chemical properties, such as polarity and charge. These properties drive intricate interactions between amino acids, creating a complex web of forces (like magnets pulling and repelling each other). These interactions ultimately determine how the chain folds into its three-dimensional structure, with an astronomical number of possible configurations.

For decades, scientists have struggled to run physics simulations to predict these shapes accurately. Even with supercomputers, the sheer number of calculations required to simulate protein folding was, until recently, beyond our reach.

That is until AI models like AlphaFold cracked the protein-folding problem wide open. After its launch in 2020, AlphaFold could predict protein structures with near-experimental accuracy, and it did so in hours instead of years. By 2022, AlphaFold mapped the structures of nearly every known protein — more than 200 million of them — and made the data freely available to researchers worldwide. The breakthrough was so profound that Demis Hassabis, co-founder of DeepMind, shared the 2024 Nobel Prize in Chemistry for AlphaFold.

The latest iteration, AlphaFold 3, can model how proteins interact with other molecules, like DNA, RNA, and potential drug compounds. Researchers can now use AlphaFold 3 to identify the most promising candidates before they even pick up a pipette. AlphaFold's sister company, Isomorphic Labs, is preparing to start human clinical trials for its first AI-designed oncology drugs.

I could mention Boltz-2, a new open-source protein-folding model developed by researchers at MIT. Or the folding@home project (which I already mentioned way back TT3). But AlphaFold, although closed source, is available to anyone, for free, at this website. Check out this hemoglobin protein I folded with it!

The bottle neck now seems to be the wet lab. AI can design thousands of potential drug candidates in a day, but synthesizing and testing those candidates in a real physical lab is still a slow, expensive, and manual process. Though this is a problem I imagine robotics solving soon. Startups like Synthace, Strateos, and Automata, just to name a few, are already building fully automated biology labs that run 24/7.

But AI drug discovery isn't new. Founded in 2011, Absci promised to combine generative AI with wet labs for rapid drug discovery. Four years after going public, the Vancouver, Wash.-based company’s stock cratered more than 90% from missed revenue targets. The vision was right; the timing was wrong.

We're witnessing the disruption of big pharma and this rise of "AI-native" small pharma startups. These are companies that own the entire R&D stack, from computational discovery to clinical development.

We won’t wake up tomorrow to a headline that cancer has been “solved.” But we might see a quiet Tweet from a grad student who just cut a year off her thesis because AlphaFold gave her the missing piece.

I asked AlphaFold to produce ricin’s holotoxin (A + B chains); the result popped out in under a minute, looking suspiciously like something you’d find in a Soviet lab fridge. Actually synthesizing it is a whole other challenge (and a fast-track to a federal watch-list).

Science advances through countless tiny gains that compound over time, not dramatic breakthroughs. The future belongs to teams who understand that each 0.1% improvement matters, especially when human lives hang in the balance. So stay tuned ‘cause just as Cursor democratized computer programming, AI is about to turn every curious mind into a backyard biochemist.