RoseTTAFoldNA

A transcription factor bends the DNA it binds, and an RNA-binding protein often folds its target as it grips it. The standard route — fold the protein, build the nucleic acid separately, then dock the two rigid bodies — cannot represent that, because the shape of each partner is a consequence of the binding it is meant to predict. RoseTTAFoldNA (RFNA) removes the docking step: protein and nucleic acid sequences go in, the assembled complex comes out in a single pass, and the DNA distortion and the protein conformation are decided together.

The model extends the RoseTTAFold three-track architecture — which refines a 1D sequence representation, a 2D residue-pair representation and 3D coordinates against one another — to carry nucleic acids in all three tracks. It comes from Minkyung Baek, Ryan McHugh, David Baker, Frank DiMaio and colleagues at the Institute for Protein Design, released with code in September 2022 and published in Nature Methods in November 2023. It ships as the repository RoseTTAFold2NA, weights and pipeline named RF2NA, trained on the protein data used for RoseTTAFold2 with RNA and protein-nucleic acid structures added.

The premise was not obviously sound: the Protein Data Bank holds an order of magnitude fewer nucleic acid structures than protein ones, and the finding is that this suffices to learn general interaction geometry rather than memorized homologs.

#Key Features

• Nucleic acids in every track: Ten tokens are added to the 1D track for the four DNA and four RNA nucleotides plus unknown bases; the 2D track models base-base and base-amino-acid interactions; the 3D track gives each nucleotide a frame on its phosphate group and ten torsions covering backbone, ribose pucker and glycosidic rotation.
• One network across three problems: The same weights predict protein-DNA complexes, protein-RNA complexes and RNA tertiary structure, including assemblies where several protein subunits share a duplex.
• Calibrated interface confidence: Alongside per-residue pLDDT, an interface predicted aligned error scores whether a modeled contact surface is real, which makes the output usable as a filter rather than a single guess.
• Physically grounded refinement: Frame-aligned point error is generalized to nucleotide frames, and a fine-tuning stage adds Lennard-Jones and hydrogen-bonding terms so base pairs and backbone geometry come out chemically reasonable.

#Technical Details

RFNA stacks 36 three-track layers followed by four structure refinement layers built on an SE(3)-equivariant transformer, totaling 67 million parameters. Training used PDB entries published on or before 30 April 2020: 7,396 RNA chains and 23,583 complexes, clustered to 1,632 non-redundant RNA clusters and 1,556 protein-nucleic acid clusters against 26,128 protein clusters, sampled 60/40 protein-only to nucleic-acid-containing over roughly four weeks on 64 GPUs. Protein alignments come from HHblits over UniRef30 and BFD, RNA alignments from RNAcentral, Rfam and nt. Because crystallized protein-DNA constructs usually place the binding motif in the middle of a short duplex, training pads inputs with 0-6 random nucleotides and includes negative examples with mutated base-pairing contacts, so position cannot serve as a shortcut for specificity.

On 224 monomeric protein-nucleic acid complexes (116 clusters) held out for validation, predictions average 0.73 lDDT, with 29% above 0.8 and roughly 45% recovering more than half the native interface contacts. Confidence tracks accuracy: 38% of complexes come out with mean interface PAE below 10, and 81% of those are acceptable or better under CAPRI criteria. Across 33 clusters with no detectable similarity to training complexes, accuracy holds at 0.68 lDDT. Protein components alone reach 0.87 average TM-score against 0.88 for AlphaFold 2, while an AlphaFold-plus-docking workflow placed the correct complex in its top three models in only 1 of 14 test cases.

#Applications

The model suits questions where protein and nucleic acid must be modeled as one object: transcription factor specificity, enzymes on their recognition sites, and ribonucleoprotein assemblies such as telomerase or tmRNA-bound SmpB. Because the confidence estimate is informative, a common workflow predicts many candidate pairs and keeps only the high-confidence interfaces for experimental follow-up; the authors also frame it as a design tool for sequence-specific DNA- and RNA-binding proteins. Running it is a cluster task, not a laptop one: inference needs a GPU plus UniRef30, BFD and nucleotide databases totaling several hundred gigabytes.

#Impact

RFNA moved protein-nucleic acid complexes from a docking problem to a prediction problem, and became a standard baseline for that work: later methods routinely benchmark against it or vendor its code. The paper states its limits plainly — large RNAs above roughly 100 nucleotides, very short single-stranded segments and large multidomain proteins are the dominant failure modes, and de novo RNA folding averages 0.62 lDDT against 0.64 for DeepFoldRNA. The lab's line continued through RoseTTAFold All-Atom, which generalizes the same idea to ligands, metals and covalent modifications. Code is MIT-licensed, but the checkpoint ships as a tarball from a lab file server with no separate license statement, and has not been updated since April 2023.