SimpleFold is a protein structure prediction method that exclusively relies on transformer blocks and flow matching (1). It relies on embeddings from ESM2-3B and models from the AlphaFold2 DB and ESMFold Atlas. Authors scaled it from several hundred million parameters to 3 billion parameters, observing improved performance with larger sizes. They also fine-tuned versions of all six models on ATLAS MD data.
Details
(2) noted that transformer blocks are extremely inefficient relative to other approaches, although it isn’t clear if these are the same.
See also
1.
Wang Y, Lu J, Jaitly N, Susskind J, Bautista MA. SimpleFold: Folding Proteins is Simpler than You Think. 2025; Available from: https://arxiv.org/abs/2509.18480
2.
Ouyang-Zhang J, Murugan P, Diaz DJ, Scarpellini G, Bowen RS, Gruver N, et al. Triangle Multiplication Is All You Need For Biomolecular Structure Representations. 2025; Available from: https://arxiv.org/abs/2510.18870