Zero-shot design of drug-binding proteins via neural iterative selection−expansion
5 affiliations
- 1Department of Biological Chemistry and Molecular Pharmacology, Harvard Medical School, Boston, MA, USA
- 2Department of Cancer Biology, Dana-Farber Cancer Institute, Boston, MA, USA
- 3Harvard Graduate Program in Biophysics, Harvard University, Boston, MA, USA
- 4Department of Biological Chemistry and Molecular Pharmacology, Harvard Medical School, Boston, MA, USA. [email protected]
- 5Department of Cancer Biology, Dana-Farber Cancer Institute, Boston, MA, USA. [email protected]
Abstract
Computational design of molecular recognition remains challenging despite advances in deep learning 1–3 . The design of proteins that bind to small molecules has been particularly difficult because it requires simultaneous optimization of protein sequence, protein structure, and ligand conformation 1–7 . Despite their promise, current deep-learning algorithms have struggled to navigate this landscape, precluding the zero- or few-shot design of binders. Here we show that the combination of two neural networks in an iterative design algorithm can create small-molecule binding proteins from scratch with high accuracy. To optimize a design in the joint distribution of sequence, structure, and ligand conformation, we use a pair of neural networks that were trained on reciprocal tasks. We train and use a graph neural network, LASErMPNN, to design protein sequence given protein−ligand co-structure, and we use RoseTTAFold-All Atom 8 (RFAA) to predict protein−ligand co-structure given protein sequence. We iteratively apply these two networks to design proteins that bind the drug, exatecan, a topoisomerase I inhibitor that is prone to inactivation by hydrolysis 9 . Each of four experimentally tested designs bound the drug, with the lowest dissociation constant ( K d ) near 100 nM. The hit rate and highest affinity design each surpassed the current state-of-the-art method by 5- and 70-fold, respectively. We further show that LASErMPNN can improve upon its own designs in a manner resembling chain-of-thought reasoning. Without experimental input, LASErMPNN suggested two mutations that increased affinity by over two orders of magnitude ( K d = 1.2 ± 0.2 nM). Designs were selective, structurally accurate, and achieved their intended purpose to protect the drug from hydrolysis. Our work describes a recipe for using neural networks to automate the design of high affinity small-molecule binding proteins, which should have wide application in the creation of novel drug-delivery vehicles, antidotes, sensors, and enzymes.
Related
- Site-specific RNA photocrosslinking of CRISPR-associated and viral RNA-binding proteins using genetic code expansion
- De novo design of orthogonal far-red, orange, and green fluorophore-binding proteins for multiplexed imaging
- De novo design of ligand binding proteins using large language models alone
- SOX2 can associate with chromatin directly by binding to DNA or indirectly via association with other chromatin-bound proteins
- Binding without Docking: Small-Molecule Binding Paths on Disordered Proteins.
BibTeX
@article{fry2026zeroshot,
title = {{Zero-shot design of drug-binding proteins via neural iterative selection−expansion}},
author = {Benjamin N. Fry and Kaia Slaw and Nicholas F. Polizzi},
journal = {Nature},
year = {2026},
month = jun,
doi = {10.1038/s41586-026-10670-w},
url = {https://doi.org/10.1038/s41586-026-10670-w},
note = {CC BY},
}
RIS
TY - JOUR
TI - Zero-shot design of drug-binding proteins via neural iterative selection−expansion
AU - Benjamin N. Fry
AU - Kaia Slaw
AU - Nicholas F. Polizzi
PY - 2026
DA - 2026/06/24
JO - Nature
DO - 10.1038/s41586-026-10670-w
UR - https://doi.org/10.1038/s41586-026-10670-w
AB - Computational design of molecular recognition remains challenging despite advances in deep learning 1–3 . The design of proteins that bind to small molecules has been particularly difficult because it requires simultaneous optimization of protein sequence, protein structure, and ligand conformation 1–7 . Despite their promise, current deep-learning algorithms have struggled to navigate this landscape, precluding the zero- or few-shot design of binders. Here we show that the combination of two neural networks in an iterative design algorithm can create small-molecule binding proteins from scratch with high accuracy. To optimize a design in the joint distribution of sequence, structure, and ligand conformation, we use a pair of neural networks that were trained on reciprocal tasks. We train and use a graph neural network, LASErMPNN, to design protein sequence given protein−ligand co-structure, and we use RoseTTAFold-All Atom 8 (RFAA) to predict protein−ligand co-structure given protein sequence. We iteratively apply these two networks to design proteins that bind the drug, exatecan, a topoisomerase I inhibitor that is prone to inactivation by hydrolysis 9 . Each of four experimentally tested designs bound the drug, with the lowest dissociation constant ( K d ) near 100 nM. The hit rate and highest affinity design each surpassed the current state-of-the-art method by 5- and 70-fold, respectively. We further show that LASErMPNN can improve upon its own designs in a manner resembling chain-of-thought reasoning. Without experimental input, LASErMPNN suggested two mutations that increased affinity by over two orders of magnitude ( K d = 1.2 ± 0.2 nM). Designs were selective, structurally accurate, and achieved their intended purpose to protect the drug from hydrolysis. Our work describes a recipe for using neural networks to automate the design of high affinity small-molecule binding proteins, which should have wide application in the creation of novel drug-delivery vehicles, antidotes, sensors, and enzymes.
SN - 0028-0836
SN - 1476-4687
ER -