SimpleFold

Nearly every accurate protein folding model since AlphaFold2 has been built around domain-specific machinery: an explicit pair representation, triangular attention updates that enforce distance-geometry consistency, equivariant frame prediction, and hand-designed structural losses. SimpleFold, released by Apple's machine learning research group in September 2025, asks whether any of that is necessary for performance. It is a folding model assembled entirely from standard transformer blocks with adaptive layers — no pair track, no triangular updates, no equivariant operations, and no multiple sequence alignment — trained with a single generative flow-matching objective supplemented by an LDDT term.

The model treats folding as conditional generation over atomic coordinates. A frozen ESM2-3B protein language model (from the ESM-2 family) encodes the amino acid sequence into per-residue conditioning embeddings, and the network integrates a learned velocity field from Gaussian noise to a full-atom structure. Because the objective is generative rather than a deterministic regression to a single reference structure, SimpleFold samples distributions of conformations by construction, instead of being retrofitted for ensembles the way AlphaFlow fine-tunes AlphaFold2.

Apple released a family of six sizes, from 100M to 3B parameters. The 3B checkpoint is trained on roughly 8.6M distilled structures alongside experimental PDB entries, and it reaches performance competitive with MSA-free baselines on standard folding benchmarks while beating MSA-based methods on multi-state conformational tasks.

#Key Features

• No folding-specific inductive bias: Atom encoder, residue trunk, and atom decoder are all the same general-purpose transformer block, with rotational symmetry learned from SO(3) data augmentation instead of being architecturally enforced.
• Generative flow matching over all atoms: Structures are produced by integrating a learned velocity field with a Langevin-style SDE, with a stochasticity parameter tau that trades single-structure accuracy against conformational diversity.
• Native ensemble prediction: The same checkpoint samples multiple conformers per sequence, covering apo/holo state changes and molecular dynamics ensembles without a separate generative fine-tune.
• Six checkpoints, 100M to 3B: SimpleFold-100M recovers roughly 90% of the 3B model's folding performance, giving a practical accuracy-versus-cost ladder.
• Efficient inference: A single sequence representation avoids the quadratic pair-tensor cost of triangle attention, and the package ships PyTorch and MLX backends so smaller models run on a laptop.
• Confidence estimates: A separately trained four-layer pLDDT head predicts per-residue confidence, correlating with true LDDT-Cα at Pearson r = 0.77.

#Technical Details

The architecture follows a fine-coarse-fine scheme. A lightweight atom encoder with local attention embeds noisy coordinates and atomic features, average-pools atoms into residue tokens, and passes them through a heavy residue trunk that carries most of the parameters; an ungrouping step and a symmetric atom decoder then emit the velocity field. Blocks use QK-normalization, SwiGLU feed-forwards, and rotary position embeddings, extended to a 4D axial variant in the atom modules. Training mixes ~160K PDB structures with a May 2020 cutoff, ~270K high-confidence AFDB SwissProt models, and ~1.9M AFESM cluster representatives — about 2M structures for models up to 1.6B — while the 3B model additionally uses an extended AFESM set of 8.6M structures.

On CAMEO22 (183 targets), SimpleFold-3B reaches a mean TM-score of 0.837 and GDT-TS of 0.802, close to ESMFold (0.853 / 0.826) and AlphaFold2 (0.863 / 0.844) without any MSA search. On the harder CASP14 subset (70 targets) it overtakes ESMFold, at TM-score 0.720 and GDT-TS 0.639 versus 0.701 and 0.622. On the apo/holo two-state benchmark it reaches a TM-ensemble score of 0.893, above MSA-based AlphaFlow at 0.864. Inference scales far better with length than recycling-based models: on an H100 at 1024 residues, SimpleFold-100M takes 5.6 seconds against 43.6 for ESMFold and 111.5 for AlphaFold2.

#Applications

SimpleFold suits workflows where flexibility matters more than a single static answer — identifying cryptic pockets, sampling apo and holo states for docking, and generating molecular-dynamics-like ensembles without running simulations. Its MSA-free design makes it useful for orphan and metagenomic sequences with shallow alignments, and the small checkpoints plus the MLX backend let structural biologists fold sequences locally on a Mac rather than queueing on a GPU cluster.

#Impact

SimpleFold's contribution is as much a negative result as a model: it shows that the pair representations and triangular updates treated as essential since AlphaFold2 are one design point rather than a requirement, and that a generic transformer with enough data and scale gets most of the way there while being cheaper at inference and better at ensembles. The scaling curves — consistent gains from 100M to 3B, larger on the harder CASP14 benchmark — argue that this design space rewards scale. Limits are real: the work remains a preprint, the models fold single chains only and do not handle complexes, ligands, or nucleic acids, absolute accuracy still trails AlphaFold2 on CASP14, and failure cases mirror ESMFold's because both inherit the same ESM2 embeddings. Code is released under MIT, but the weights carry Apple's research-only, non-commercial model license.