Read about an open-source all-atom biomolecular foundation model focused on co-folding and structural reasoning for AI-driven drug discovery.
Key Takeaways
- Top-ranked antibody-antigen co-folding success reached 51.0%, 70.0%, and 66.4% across three benchmarks
- Model has about 655 million parameters and required roughly 414,000 GPU-hours to train
- Training code, inference pipelines, checkpoints, and benchmarks were released under the Apache-2.0 license