RoseTTAFold2

AlphaFold 2 keeps its pairwise residue representation geometrically self-consistent using triangle attention, an operation evaluated over every triplet of residues. That is part of why the network is accurate, and also why its cost climbs steeply with chain length — the wrong scaling for large assemblies. RoseTTAFold2 (RF2) asks whether the operation is necessary at all. It carries a separate 3D structure track through the full depth of the network and uses residue-pair distances read from that track to bias ordinary biaxial attention in the pair update, getting the same inductive bias toward 3D locality from coordinates the network already maintains.

RF2 was posted as a preprint in May 2023 by Minkyung Baek, Ivan Anishchenko, Ian Humphreys, Qian Cong, David Baker and Frank DiMaio, working at the Institute for Protein Design and collaborating institutions. Rather than modify an existing network, the authors rebuilt the original RoseTTAFold three-track architecture from scratch and added AlphaFold 2 features one at a time, keeping those that measurably improved accuracy: frame-aligned point error (FAPE) as the structure loss, recycling during training, and a distillation set of AlphaFold 2 predictions.

The outcome is a single set of weights matching AlphaFold 2 on monomers and AlphaFold-Multimer on complexes, without invariant point attention or triangle attention. Since nearly every accurate folding method that followed — OpenFold and ESMFold among them — reuses the AlphaFold 2 trunk, that says something about the design space: high accuracy belongs to a broader class of architectures than assumed.

#Key Features

• One model for monomers and assemblies: The same checkpoint predicts single chains, hetero-dimers and larger complexes, with MSAs optionally paired by taxonomy ID, instead of a separate multimer network.
• Structure-biased attention in the pair track: Distances read from the current 3D model gate and bias the pair-to-pair attention, supplying 3D locality at biaxial-attention cost rather than triplet cost.
• Symmetric assembly prediction: A point group declared at inference — cyclic, dihedral, tetrahedral, octahedral or icosahedral — builds large symmetric oligomers under that constraint.
• Interface-aware confidence: Beyond per-residue pLDDT and predicted aligned error, a pBind head averages interchain pAE into one probability that a pair of chains forms a complex.
• Cyclic peptide support: A cyclized relative position encoding closes a backbone into a ring, making the network usable on macrocycles rather than only linear chains.

#Technical Details

RF2 stacks 36 three-track blocks, each updating the 1D MSA, 2D residue-pair and 3D coordinate representations from one another, followed by four shared-parameter refinement iterations of an SE(3)-equivariant transformer. The first four blocks read a wide MSA slice and the remainder a subsampled one; embedding widths are 256, 128 and 32 for the MSA, pair and state features. Coordinates come from the SE(3)-equivariant transformer rather than invariant point attention, which trained unstably here, and refinement attends to each residue's 64 nearest neighbors.

Training used all PDB structures deposited before 30 April 2020 — 280,000 chains in 20,000 clusters — plus a distillation set built by folding 12 million UniRef50 sequences with AlphaFold 2 and clustering the confident predictions over 200 residues into 1.0 million clusters, with examples drawn 50/25/25 from distillation data, PDB monomers and PDB hetero-oligomers. On 113 medium and hard CAMEO monomer targets and 140 hetero-dimers released after training, RF2 matches AlphaFold 2 on GDT-TS and AlphaFold-Multimer on oligomer TM-score. The CASP14 ablation is the informative part: over the original RoseTTAFold baseline, FAPE loss added 2.4 GDT-TS, recycling with distillation 3.1, fully connected structure updates 2.2 and a deeper three-track stack 2.0, while triangle multiplication contributed 0.2. Runtime was measured on a single A100 GPU out to 4000 residues, where the gap in favor of RF2 is widest.

#Applications

RF2 suits large, oligomeric or symmetric systems: multi-chain assemblies with paired MSAs, point-group oligomers, and triaging putative interactions with the pBind score before committing to experiments. Being architecturally independent of AlphaFold 2, it also serves as an orthogonal filter on designed proteins, where re-predicting a design with the network that generated it risks confirming its own bias — the RFpeptides macrocycle pipeline uses it that way. Inference needs UniRef30, BFD and a template database totaling several hundred gigabytes, making it a cluster tool, though a Colab notebook covers smaller targets.

#Impact

RF2's main contribution is a working counterexample: the features widely treated as essential to AlphaFold 2's accuracy are substitutable, and the gains that mattered came from the training side. Its network became the trunk for later Baker lab work — RoseTTAFold All-Atom starts from RF2 and adds an atomic graph representation for ligands and modifications, and the same architecture reached protein–nucleic acid complexes as RoseTTAFold2NA. Limitations remain: predictions depend on deep MSAs and the databases behind them, the work has not been peer reviewed, and the weights ship as a tarball from a lab file server rather than a versioned registry.