Flow matching is a generative modeling technique that guides random noise toward a structured protein shape over several steps. It treats structure prediction as learning a distribution of possible conformations rather than producing one average structure. section 12.4
For protein structure prediction, flow matching helps bridge an amino acid sequence and a plausible three-dimensional arrangement. The model must represent geometric constraints and long-range interactions, such as contacts between residues far apart in the sequence. section 12.2 section 12.4
A flow-matching structure-prediction workflow can be understood as:
- Encode the amino acid sequence with per-residue embeddings from a pretrained ESM-2 protein language model. section 12.2
- Use a generative objective to guide noisy representations toward all-atom three-dimensional coordinates. section 12.2 section 12.4
- Refine the prediction iteratively and report per-residue confidence alongside the resulting structure. section 12.2
SimpleFold
SimpleFold applies flow matching with a pretrained protein language model and a largely standard Transformer. Given an amino acid sequence without a multiple sequence alignment, it predicts all-atom coordinates and confidence estimates for structure-based design. section 12.2
