BindCraft Pipeline
Computational protein binder design pipeline — from target selection through structure prediction to binding affinity scoring.
A protein binder is a small, custom-built molecule designed to grab one target and hold on. Get the shape right and it can quiet a runaway disease pathway; get it wrong and it drifts past, useless. The hard part is the size of the search: the number of possible protein sequences is astronomically larger than the number of atoms in the observable universe, and almost all of them fold into nothing you’d want. Finding the rare few that grab the right target is the whole game.
This pipeline plays that game end to end — from picking a target protein, through generating binder candidates, to scoring how tightly each one should hold. Think of it as a conversation between two AI systems. One, RFdiffusion, imagines what a binder might look like: the shape that could nestle against the target. The other, ProteinMPNN, works backward to the amino acid sequence that would actually fold into that shape. ESM-2 embeddings give the models a feel for the language proteins are written in, and Rosetta’s energy functions score the candidates, so the most promising ones move forward for experimental validation.
A note on the name. BindCraft is a published binder designer from another group, and it takes a different route to the same goal: backpropagating through AlphaFold2 rather than diffusing structures with RFdiffusion. I have used it in molecular design and it is good work. What is described above is my own assembly — built, wired, and re-tuned in close coordination with the biologists and research collaborators who need the binders, which is the part of this that I actually claim.
The immediate target is BAFF — B-cell activating factor — and miniprotein binders for Sjögren’s syndrome, an autoimmune disease. The broader ambition is a disease-graph that connects proteins, pathways, and clinical endpoints across the rare disease landscape, with a growing catalog of targets and mapped structures.
None of this runs without compute, and that is exactly the point. The pipeline runs on Lattice Protocol’s federated GPU infrastructure — the same verifiable decentralized compute the rest of the stack depends on — built on one stubborn commitment: compute should never be rare for the people working on rare disease. Every candidate that clears the pipeline is one someone can actually test: one more shape, one more shot at a target that used to have none. The models it leans on are public code — every one of them linked above — the target list keeps growing, and the next binder is always the interesting one.