Atomic context-conditioned protein sequence design using LigandMPNN
Problem
Designing protein sequences that correctly interact with small molecules, nucleotides and metals — essential for enzymes, binders and sensors — which prior deep-learning methods could not model.
Why it matters
Proteins that bind nonprotein molecules are central to medicine, biotechnology and biosensing.
What the AI did
LigandMPNN, a deep-learning sequence design method, explicitly models all nonprotein atoms and also generates sidechain conformations for evaluating binding.
Results
Sequence recovery on ligand-contacting residues: 63.3% vs 50.4% (Rosetta) and 50.5% (ProteinMPNN) for small molecules; 50.5% vs 35.2%/34.0% for nucleotides; 77.5% vs 36.0%/40.6% for metals; over 100 experimentally validated designs with high affinity and structural accuracy.
Impact
Peer-reviewed method with concrete benchmarks and experimental validation of designed proteins.
Record
| Venue | Nature Methods |
|---|---|
| Published | 2025 |
| Work type | article |
| Authors | Justas Dauparas, Gyu Rie Lee, Robert J. Pecoraro, Linna An, Ivan V. Anishchenko, Cameron J. Glasscock, David A. Baker |
| Citations | 278 (OpenAlex, at import time) |
| References | 56 |
| Model or method | LigandMPNN |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1038/s41592-025-02626-1 |
| Open access | hybrid |
| Keywords | protein design, deep learning, LigandMPNN, sequence recovery, ligand binding, sidechains |
Importance rating
significant — thousands to millions affected
Abstract calls ligand-binding protein design critical to enzymes, binders, sensors in medicine/biotech; new capability with 100+ experimentally validated designs; no scale numbers.