Evolutionary-scale prediction of atomic-level protein structure with a language model
Problem
Predicting the 3D structures of the exponentially growing catalog of protein sequences was too slow with existing methods, requiring a breakthrough in folding speed.
Why it matters
Atomic-level protein structures underpin biology and medicine at evolutionary scale, and slow prediction blocked characterization of billions of sequences.
What the AI did
A large protein language model (up to 15B parameters) was trained on millions of sequences and directly infers atomic-resolution structure from primary sequence.
Results
Order of magnitude speed-up in high-resolution structure prediction; structure prediction at resolution of individual atoms as models scale
Impact
The abstract claims this extends high-resolution structure prediction to the scale of hundreds of millions of cataloged proteins.
Record
| Venue | Science |
|---|---|
| Published | 2023 |
| Work type | article |
| Authors | Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil et al. |
| Citations | 5,455 (OpenAlex, at import time) |
| References | 83 |
| Model or method | Large language model of proteins, up to 15B parameters |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1126/science.ade2574 |
| Open access | hybrid |
| Keywords | protein structure, language model, ESM, deep learning, folding, genomics |
Importance rating
major — millions of lives, an industry or a field
Abstract states structure prediction extended to two hundred million cataloged proteins and billions of sequences, with 617 million structures characterized and 225 million high-confidence predictions, a foundational advance.