SimpleFold predicts a protein’s three-dimensional structure from its amino acid sequence. Its predictions can support structure-based design when an experimental structure is unavailable. The model produces all-atom coordinates and a confidence estimate for each residue, without requiring a multiple sequence alignment. section 12.2
A protein sequence is a chain of amino acids, but its function depends on how that chain folds. SimpleFold combines a pretrained protein language model, a largely standard Transformer, and flow matching to predict a shape from the sequence. section 12.1.1 section 12.2
The prediction workflow combines sequence representations with generative modeling:
- ESM-2 converts the amino acid sequence into an embedding for each residue. These representations encode evolutionary and structural propensities. section 12.2
- Flow matching learns to guide random noise toward a structured protein conformation over several steps. This generative approach can capture a distribution of possible conformations rather than a single average structure. section 12.4
- The resulting coordinates describe the predicted structure; per-residue confidence helps identify regions that should not be trusted. section 12.2
Evaluate predictions using structure-accuracy metrics on held-out targets, such as CAMEO or CASP sets. Pay particular attention to targets with little evolutionary signal: accuracy can drop for orphan and designed proteins and for some antibody loops. A confidence estimate is not a substitute for validation. section 12.2
