NA-MPNN

As of January 2025 the Protein Data Bank held 8,961 RNA-containing entries against 235,538 protein-containing ones. Learning RNA inverse folding — recovering a sequence that folds into a given three-dimensional backbone — from RNA structures alone therefore means training in the data-poor corner of structural biology. NA-MPNN's answer is to stop treating that corner as separate: protein residues, DNA bases and RNA bases all become nodes in one biopolymer graph, tagged by a one-hot polymer-type feature, so every protein-DNA co-crystal in the PDB contributes training signal to RNA design.

NA-MPNN was developed at the Institute for Protein Design at the University of Washington, with a Rice University collaborator, and posted as a preprint in October 2025. It extends ProteinMPNN and LigandMPNN — the same message-passing encoder-decoder stack and random-order autoregressive decoder — from protein-only nodes to nucleic acids. The direction of the task is what readers most often get backwards: LigandMPNN designs a protein sequence with nucleic acids and ligands as fixed context, while NA-MPNN designs the nucleic acid sequence, holding any protein chains fixed.

One architecture then covers two problems normally served by unrelated tools, each with its own shipped checkpoint: backbone-conditioned RNA sequence design, and fixed-dock protein-DNA specificity prediction — per-position base preferences at an interface whose geometry is given.

#Key Features

• Unified biopolymer graph: Edges are built from the Cα atom of protein residues and the C1′ atom of nucleic acids, featurized as radial-basis-function embeddings of all pairwise backbone-atom distances, including a virtual first side-chain atom constructed for both polymer classes. Each node exchanges messages with its 32 nearest neighbors.
• Collapsed nucleic acid alphabet: A single token is shared between the deoxyribo- and ribo-form of each canonical base (DA/A, DC/C, DG/G, DT/U), forcing cross-learning between DNA and RNA contexts — valid precisely because the network never sees side-chain chemistry.
• Specificity head trained on binding data: That checkpoint is supervised against experimental position probability matrices rather than crystallographic sequences, and predicts a PPM by averaging 30 stochastic decodes of the DNA chain. Since only backbone coordinates enter the graph, the prediction cannot leak the target sequence, making it a stringent filter that runs before rotamer packing.
• Partial sequence fixation: The random-order decoder lets any subset of positions be held fixed, freezing a protein chain as context while only its bound RNA is redesigned.

#Technical Details

Both checkpoints use a hidden dimension of 128, three encoder and three decoder layers and 32 neighbors. The design set is 15,633 PDB entries carrying at least one nucleic acid chain at ≤ 3.5 Å resolution, clustered at 80% nucleic acid identity; the specificity set adds protein-DNA complexes distilled from RoseTTAFoldNA and RoseTTAFold All-Atom predictions paired with CIS-BP and TRANSFAC motifs, for 45,093 entries. Each model trained for 100,000 token-budgeted batches of 6,000 tokens on a single A100.

Median native sequence recovery is 57.4% for DNA-only and 60.5% for RNA-only backbones, and 58.6% and 55.4% respectively in protein context. On the RNA-monomer test set NA-MPNN recovers 58.0% against 51.7% for gRNAde and 66.7% for RhoDesign, a ranking that holds on pseudoknots (58.9% vs 47.2% and 67.7%). Structural fidelity inverts that order: pseudoknot designs score a median OpenKnot consistency of 83.2 from RibonanzaNet reactivity predictions (against 81.7 and 72.9) and fold to 9.0 Å median C1′-RMSD under AlphaFold 3 (against 11.6 Å and 22.6 Å). On 228 held-out protein-DNA complexes, median per-position mean absolute error against the experimental motif is 0.53 and cross-entropy 1.00, versus 0.86 and 1.44 for DeepPBS.

#Applications

RNA inverse folding is the sequence half of designing aptamers, ribozymes, CRISPR guide RNA scaffolds and non-native ribosomes: a structure generator proposes a fold, and the design model supplies a sequence that adopts it. NA-MPNN was tested this way in Round 6 of the community OpenKnot pseudoknot design challenge, where 10 to 15 designs for each of 12 fixed backbones were synthesized and profiled by SHAPE-Seq, reaching a median experimental OpenKnot score of 89.9 against 80.7 for gRNAde, 87.0 for the wild-type starting sequences and 89.4 for individual Eterna player submissions. The specificity checkpoint instead serves people engineering or interpreting DNA-binding proteins, reading a transcription factor's base preferences off a docked complex without side-chain sampling.

#Impact

NA-MPNN is the first model to treat RNA sequence design and protein-DNA specificity prediction as one inverse-folding problem, and the unification pays for itself: the shared alphabet and graph let abundant DNA and protein-DNA structures compensate for scarce RNA ones. The public repository ships both trained checkpoints with training, inference and evaluation code, and inference runs on a CPU in under a minute. The limits are equally clear: the work is a preprint and has not been peer reviewed, RhoDesign still recovers more native bases on RNA-only backbones, and the specificity test set is built on predicted distillation complexes rather than co-crystal structures.