Here’s a strange way to learn how something is built: film it falling apart. Take a Lego castle, blow it up in slow motion, and watch every brick tumble away from every other. Study that explosion closely enough — the order things come apart, the paths the pieces take — and you’ve learned the castle’s structure in reverse. Once you can run the film backward in your head, you can start assembling castles you’ve never seen, from a pile of loose bricks.

That, more or less, is a diffusion model. You take structured data, add noise until it’s rubble, and train a network to reverse the process — to walk noise back into structure, one denoising step at a time. It’s how the image generators work. And a few years ago a team at MIT’s Jameel Clinic asked a lovely question: what if you pointed that idea at molecules?

For decades, if you wanted to know how a drug candidate fits into a protein — the problem called docking — you searched. Tools like AutoDock Vina take a molecule and a pocket on the protein and try orientation after orientation, scoring each one with a physics-based function that estimates how tightly it binds. It works, and Vina is still a workhorse. But notice the shape of it: it’s a search. You generally have to know where the pocket is, and grinding through all those poses is slow — the heavier classical tools can take on the order of a thousand seconds for a single molecule.

The new way is a guess — a very good one

DiffDock, the model that came out of the Jameel Clinic, doesn’t search. It generates. It treats finding a binding pose the way an image model treats drawing a face: start from noise — a molecule scattered into random positions and twists — and denoise it, step by step, into a plausible pose. Because it learned the shape of “plausible” from data instead of from a fixed scoring rule, it doesn’t need you to point out the pocket. It does blind docking, and it does it in tens of seconds instead of tens of minutes.

That reframing — docking as something you imagine rather than something you search — is one of those moments where a field turns a corner. And the same lab kept going: Boltz, their open-source answer to AlphaFold 3, brings the same learned-from-data approach to predicting whole protein–ligand structures, and gives it away for anyone to build on. The honest caveat, from the careful benchmarks, is that these neural methods generally beat the classical search on well-trodden targets and still stumble on genuinely novel ones. It isn’t magic. But the direction is unmistakable, and the people at the Jameel Clinic saw it early.

Someone has to make the magic runnable

Here’s the part I lived. A remarkable model on GitHub is not a usable tool. It’s a research artifact — a repo, some weights, a README, and a hundred quiet assumptions about the machine it was born on. Between “the paper works” and “a scientist who isn’t a machine-learning engineer can run this on their molecule” sits a surprising amount of unglamorous plumbing.

That plumbing was my job for a good stretch, and it looked the same wherever I did it. As a co-architect of DeepChem and later Head of Data Science at LabDAO, a lot of the work had one shape: take a model like DiffDock — published, remarkable, and basically unrunnable by anyone outside the lab that built it — and make it into a tool. Package it into a container, wire it into a shared compute cluster so you didn’t need your own bank of GPUs, check that the packaged version behaves like the original, benchmark what it can and can’t do. It’s unglamorous, it’s ML-ops, and it’s most of what stands between a research model and a scientist who just wants to use it.

There’s a public example I can point to: at LabDAO, my colleagues Kelvin Wallace and Niklas Rindtorff containerized DiffDock and ran it for a real molecule-design project — a longevity compound — and wrote it up in a community preprint that thanks me in the acknowledgments. The specific molecule is almost beside the point. What mattered was that it became one of DiffDock’s first outings in a community lab instead of the one that built it — the docking had left the building.

The model is only half of it

The model is the dazzling part — the diffusion trick, the exploding Lego set run backward, the tens-of-seconds blind dock. It earns the dazzle. But I’ve come to believe the other half of any leap like this is quieter and just as real: the work of making the leap usable by people who didn’t invent it. The most capable model in the world, stuck behind a wall of setup, helps almost no one. The same model, containerized and running on a shared network with a notebook a grad student can open, changes what a whole community can attempt. (It’s the same thing I keep coming back to about open-source drug discovery — infrastructure is a gift you give the next person.)

The Jameel Clinic imagined a new way to dock. The best thing my team could do was make sure the rest of us could actually use it. Both halves are the story.