NUMonomer

Three-dimensional structure governs how RNA and single-stranded DNA perform regulation, catalysis, and ligand recognition, yet nucleic acid structure prediction still trails its protein counterpart. The strongest deep-learning methods narrow that gap by feeding the network auxiliary inputs: multiple sequence alignments, predicted secondary structure, or embeddings from a pretrained language model. Each carries a cost — homology search against large databases, an extra predictive step with its own failure modes, a second network to train and serve — and accuracy remained limited even so.

NUMonomer, developed by Yunda Si, Suqi Zhang, and Luonan Chen at the University of Chinese Academy of Sciences, dispenses with the auxiliary stage entirely. It is an end-to-end framework that predicts all-atom nucleic acid structures directly from primary sequence, with no MSA, no secondary-structure annotation, and no language-model representations. Its design follows two findings the authors report: that learning the hierarchical organization of experimentally determined structures across scales — from recurring local conformations up to global fold topologies — improves generalization, and that RNA and single-stranded DNA share structural representations, so joint training on both benefits each.

That positions NUMonomer alongside single-sequence RNA predictors such as RhoFold+, which achieves sequence-only prediction by routing through a pretrained language-model backbone, and in contrast to MSA-dependent pipelines such as trRosettaRNA. It is a variant of neither: it removes the auxiliary-input stage rather than optimizing it, and it extends coverage beyond RNA to single-stranded DNA.

#Key Features

• Sequence-only prediction: A bare FASTA sequence is the complete input, eliminating homology search, secondary-structure prediction, and language-model featurization from the runtime path.
• Joint RNA and ssDNA modeling: One checkpoint handles both substrates, selected per chain by a FASTA header tag (rna or dna), rather than requiring separate specialist models.
• Long-sequence capability: Training used sequences spanning thousands of nucleotides, and the pipeline folds a 5,000-nucleotide sequence in roughly 20 seconds on an NVIDIA H100 — with accuracy gains over competing methods most pronounced on longer RNAs.
• Roughly two orders of magnitude cheaper inference: Removing the auxiliary-input pipeline and using a scalable architecture cuts inference cost about 100-fold relative to the methods evaluated in the paper.
• Per-nucleotide confidence estimates: Each prediction ships with pLDDT scores and a JSON log of timing and sequence length, letting users triage which regions of a model to trust.
• Open code with downloadable weights: The inference pipeline is MIT-licensed, and a fixed pretrained checkpoint is applied zero-shot to new targets via a --weight argument.

#Technical Details

NUMonomer is an attention-based architecture that skips the MSA-representation trunk common to structure predictors and instead moves from an input embedder straight into an atom-level structure module. The structure module stacks eight blocks operating jointly on a 384-channel atom representation with 12 attention heads and a 64-channel pair representation with 4 heads, over a 28-atom-per-nucleotide representation. Local rigid-body frames and quaternions track geometry, RMSNorm handles normalization, and coordinates are refined through eight recycling iterations. A separate four-block confidence head predicts pLDDT. Outputs are written as .cif or .pdb atomic coordinates.

On benchmarks comprising CASP16 RNA targets plus non-redundant sets of experimentally determined RNA and single-stranded DNA structures, NUMonomer matches or outperforms leading prediction methods despite requiring no auxiliary inputs, with the largest margins on longer RNAs. Distribution is local-only: there is no hosted web server, and the checkpoint lives on Google Drive rather than a model registry.

#Applications

Sequence-only, low-cost inference suits work where alignment-based methods are impractical or too slow. Researchers studying orphan non-coding RNAs, newly sequenced regulatory elements, or synthetic constructs with no natural homologs can obtain models where MSA construction would fail outright, and the cost profile supports transcriptome- or database-scale campaigns that alignment-dependent pipelines make prohibitive. Single-stranded DNA support is unusual among structure predictors and is relevant to aptamer characterization and DNA nanotechnology. Predicted models can also seed virtual screening against structured RNA targets, guide construct design for crystallography or cryo-EM, and inform mutagenesis hypotheses.

#Impact

NUMonomer turns a negative result into a design principle: the auxiliary inputs the field converged on may be substitutable for better use of structural data itself, given training that captures structural organization across scales and shares representations between related substrates. That simplifies deployment considerably, since alignment generation and language-model featurization dominate the engineering burden in most nucleic acid folding pipelines, and the joint RNA/ssDNA treatment argues these substrates are better learned together than apart. The caveats are concrete: the work is a preprint awaiting peer review, so its benchmark comparisons are self-reported, and the preprint itself carries a CC-BY-ND license. Although the code is MIT-licensed with downloadable weights, Google Drive hosting and the lack of a web server add friction to reproduction and casual use.