MIT's FrameDiff uses diffusion to dream up brand-new proteins

Nature took billions of years of trial and error to evolve the proteins inside us, but newly emerging viruses and cancers won't wait that long.

Researchers at MIT CSAIL just unveiled FrameDiff, a generative AI tool that dreams up entirely new protein structures that have never existed in nature.

Lead author and MIT PhD student Jason Yim and the team will present the work at the International Conference on Machine Learning in July.

Treating protein backbones like floating bricks

Designing a custom protein starts with getting its 3D backbone right. That spine dictates the molecule's ultimate function.

The MIT researchers realized that along this backbone, every triplet of atoms shares an identical pattern of bonds and types. Using differential geometry and probability, they modeled each triplet as a rigid body with a specific 3D position and rotation, calling them "frames."

From there, they borrowed the same math behind image generators like Midjourney. They built an algorithm using stochastic calculus on Riemannian manifolds, dubbed "SE(3) diffusion."

The model takes ordered frames, scrambles them with noise into complete chaos, and then learns how to rotate and translate each frame back into place. Once it mastered existing proteins, it learned to generalize, constructing novel backbones out of thin air.

Ditching the prediction crutch

When DeepMind dropped AlphaFold2 in 2021, it revolutionized structure prediction. But engineering medicine requires two distinct halves: predicting what a sequence folds into, and generating brand-new shapes from scratch.

MIT's core SE(3) diffusion math was already tapped by the University of Washington to build RFdiffusion, combining generative diffusion with the RosettaFold2 prediction model to design targeted binders and gene-delivery proteins.

FrameDiff takes that concept a step further by stripping away the heavy prediction engine entirely.

As Harvard computational biologist Sergey Ovchinnikov pointed out, discarding a pretrained structure prediction model opens up the ability to rapidly generate exceptionally long protein structures without being bottlenecked by older architectures.

The researchers are now working on training FrameDiff with larger datasets and expanding the framework to handle DNA and small molecules.

Why it matters

Designing custom proteins has historically been a slow, expensive gamble limited to nature's existing catalog. Tools like FrameDiff give biochemists a CAD-like interface to invent bespoke binders, hyper-efficient synthetic enzymes, and targeted drug delivery vehicles on demand.

That said, don't expect instant cures tomorrow. This is still fundamental computational work; whether these algorithmically perfect long-chain backbones can fold stably and avoid toxic side effects in wet-lab tests remains to be proven.

What makes FrameDiff brilliant is its architectural minimalism. While RFdiffusion leaned heavily on massive pretrained prediction models, FrameDiff shows that pure geometric diffusion on rigid frames is enough to build molecular foundations from scratch.

We think this lightweight approach will dramatically lower the compute barrier for de novo biologic design. Instead of tweaking what evolution already gave us, synthetic biology is rapidly turning into a greenfield generative discipline.