An AI loop just designed drug-binding proteins from scratch and watched nearly every chosen candidate work in the lab.
Protein design has a cruel checklist.
Fold right.
Make a pocket.
Hold one tiny molecule.
Ignore the wrong ones.
Do this while water, entropy, and biology heckle from the back row.
That is why small-molecule binders are hard. You must optimize sequence, protein shape, and drug pose at once.
Like solving a Rubik's Cube that keeps sweating.
Fry, Slaw, and Polizzi's Nature paper introduces NISE, short for neural iterative selection-expansion.
It is "zero-shot" in the useful sense: no target-specific lab-training round for exatecan or apixaban first.
The system gets a drug and a starting scaffold.
Then the search begins.
Two Models, One Pocket
NISE runs a loop.
First, LASErMPNN proposes protein sequences for a backbone with a docked ligand. It is a graph neural network, so atoms and residues pass messages like a chemistry Slack channel with better moderation.
Next, a structure predictor checks the proposed protein-drug complex.
For exatecan, the paper used RoseTTAFold-All Atom. For apixaban, it used Boltz-2.
Then NISE keeps designs that agree with themselves.
The sequence should make the intended fold.
The fold should keep the ligand in the intended pocket.
The ligand should not wander off like it saw a taco truck.
Design. Predict. Select. Expand. Repeat.
Simple to describe.
Painful to make work.
The Exatecan Test
The first target was exatecan, a potent cancer-drug payload used in antibody-drug conjugates.
It has a fragile lactone ring. At physiological pH, that ring can hydrolyze.
Translation: water can ruin the good version.
Chemistry remains petty.
NISE designed four proteins.
All four bound exatecan.
The best, named EPIC, bound at about 0.12 micromolar. Then LASErMPNN suggested two amino acid substitutions.
No new experimental structure.
No giant screen.
The double mutant reached 1.2 nanomolar affinity.
About 100-fold better.
It also kept more than 99% of exatecan in the active ring-closed form for at least 50 hours in buffer.
Not a therapy.
Not yet.
But as proof of concept, it has very sharp elbows.
The Apixaban Test
Then came apixaban, a blood thinner.
Different drug.
Different fold.
Same loop.
Six designs went to the bench.
Five bound.
The best, APEX, hit 80 picomolar affinity for apixaban.
That is sticky.
Like "left your password manager logged in" sticky.
The authors report it beat previous comparison binders by nearly 10,000-fold on affinity.
Why care?
A custom apixaban binder could someday act like a drug sponge during emergency bleeding or surgery.
Someday is doing work there.
The paper did not test animals or patients.
The body is not a clean tube. It is soup with lawyers.
Why The Loop Lands
Recent protein AI has been busy.
RFdiffusion made new protein structures easier to generate.
RoseTTAFold-All Atom and AlphaFold 3 brought small molecules, ions, and other chemical guests into structure prediction.
LigandMPNN made sequence design notice non-protein atoms instead of treating them like uninvited furniture.
NISE plugs that progress into one feedback circuit.
That matters because classic scoring functions often miss the weird details of protein-ligand recognition.
A hydrogen bond here.
A buried polar atom there.
One bad side chain and the pocket becomes a tiny parking garage with no exits.
NISE does not prove the problem is solved.
It tested two drugs. It used selected scaffold families. It depends on predictors and training data. The authors also disclose a provisional patent application.
Still, the result is hard to ignore.
The code for LASErMPNN and NISE is public too, which helps everyone kick the tires instead of admiring the brochure.
If the approach generalizes, scientists could design carriers, sensors, antidote-like sponges, and catalytic pockets with fewer wet-lab lottery tickets.
Fewer candidates.
Better hits.
Less molecular shrugging.
Read that again.
References
-
Fry, B., Slaw, K., & Polizzi, N. F. Zero-shot design of drug-binding proteins via neural iterative selection-expansion. Nature (2026). DOI: 10.1038/s41586-026-10670-w. PMID: 42343133
-
Lu, L. et al. De novo design of drug-binding proteins with predictable binding energy and specificity. Science 384, 106-112 (2024). DOI: 10.1126/science.adl5364. PMCID: PMC11290694
-
Dauparas, J. et al. Atomic context-conditioned protein sequence design using LigandMPNN. Nature Methods 22, 717-723 (2025). DOI: 10.1038/s41592-025-02626-1
-
Krishna, R. et al. Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science 384, eadl2528 (2024). DOI: 10.1126/science.adl2528
-
Abramson, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493-500 (2024). DOI: 10.1038/s41586-024-07487-w
-
Watson, J. L. et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089-1100 (2023). DOI: 10.1038/s41586-023-06415-8. PMCID: PMC10468394
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.