SimpleFold
lastmod 2025-10-02created 2025-10-021 edits
About
- SimpleFold is the first flow-matching-based protein folding prediction model released by Apple, and an open-source framework for protein structure prediction
- Achieves high performance using only general-purpose Transformer layers and flow-matching generative training instead of the usual complex domain-specific designs
- Trained with 3B (3 billion) parameters, making it the largest ever, and competitive with SOTA (state-of-the-art) models on public benchmarks
- Efficient because it has no complex structural modules (triangle attention, pair representations, etc.), which makes it well suited to scaling to large datasets
- The model can be tuned and retrained on custom user data, making it highly useful in practice in fields such as bioinformatics and pharmaceuticals