🧬 Announcing PoET-2: A breakthrough protein language model that achieves trillion-parameter performance with just 182M parameters, transforming our ability to understand proteins.
Comments
Log in with your Bluesky account to leave a comment
Most protein language models rely on massive scale - up to 100B parameters - to memorize sequences from nature. PoET-2 takes a fundamentally different approach, learning the grammar of protein evolution.
Key breakthrough: PoET-2's multimodal architecture learns to reason about sequences, structures, and evolutionary relationships simultaneously through in-context learning.
The results are remarkable:
- 500x more compute efficient than contemporary models
- 30x less experimental data needed for protein optimization
- Improved on structure understanding
- Handles insertions and deletions naturally
PoET-2 introduces a powerful prompt grammar for controlled protein generation - enabling everything from inverse folding to motif scaffolding in a single model.
In real-world testing, PoET-2 can:
- Design proteins with multiple simultaneous constraints
- Learn from just dozens of examples
- Make accurate predictions for challenging proteins
- Run fast inference on standard hardware
How does it work? PoET-2's tiered attention mechanism processes large protein families with order equivariance and long context lengths, letting it learn from evolutionary examples at inference time.
Comments
- 500x more compute efficient than contemporary models
- 30x less experimental data needed for protein optimization
- Improved on structure understanding
- Handles insertions and deletions naturally
- Design proteins with multiple simultaneous constraints
- Learn from just dozens of examples
- Make accurate predictions for challenging proteins
- Run fast inference on standard hardware