MetaFold-RNA

Predicting which nucleotides pair with which is the foundational step in understanding what an RNA molecule does, and deep learning has pushed within-family accuracy far past classical thermodynamic folding. That progress has come with a well-documented failure mode: neural predictors trained on bpRNA and RNAStralign tend to memorize family-specific motifs and degrade sharply on RNA families absent from training, sometimes falling below the free-energy minimization tools they were meant to replace. Thermodynamic methods generalize more gracefully because they encode biophysics rather than statistics.

MetaFold-RNA, from Huazhong University of Science and Technology, attacks that generalization gap by declining to pick a paradigm. Rather than predicting base pairs from sequence alone, it seeds a learned network with base-pairing probability maps drawn from four deliberately orthogonal predictors — LinearFold (thermodynamic energy minimization), CONTRAfold (shallow probabilistic modeling), and UFold and SPOT-RNA (deep learning) — plus one handcrafted pairing-propensity channel. A compact network then reconciles these disagreeing hypotheses into a single structure, and the authors show the fused prediction beats each of its own priors on a majority of test sequences rather than merely averaging them.

Released as a bioRxiv preprint in September 2025 and not yet peer reviewed, it takes the opposite bet from single-sequence predictors such as RNAformer: where those strip away external dependencies, MetaFold-RNA treats the existing tool ecosystem as a feature source.

#Key Features

• Ensemble priors as input, not output: Contact maps from thermodynamic, probabilistic, and neural predictors enter as channels of the initial pair representation, so the network learns where each paradigm is trustworthy instead of voting over finished structures.
• Co-evolutionary dual-track architecture: MetaFormer blocks pass information both ways between a 1D sequence track and a 2D pair map — outer products push sequence context into the pair map, and pairwise-biased attention feeds the emerging structure back to condition sequence attention.
• Cross-family generalization: On bpRNA-new, built entirely from families held out of training, it reaches 0.709 F1 against 0.661 for the next-best method, a 7.3% relative gain.
• Long-range base pairs: For pairs separated by more than half the sequence length — the interactions that define global topology — it leads on TS0 (0.741 F1), bpRNA-new (0.753), CASP15, and PDB-new.
• Engineered RNA constructs: It resolves designed topologies absent from natural training data, scoring 0.950 F1 on the CASP15 Traptamer RNA origami tile and 0.911 on a broccoli-pepper aptamer FRET tile.

#Technical Details

Nucleotides are embedded into a 64-dimensional sequence representation; the pair representation is a 32-dimensional projection of one handcrafted pairing matrix plus four predicted contact maps, augmented with relative-position encodings discretized into 65 bins over ±32 positions. Three stacked co-evolution blocks alternate an outer-product update from sequence to pair, a Mobile Inverted Bottleneck Convolution block with squeeze-and-excitation for 2D refinement, and pairwise-biased multi-head attention from pair back to sequence. The final pair representation projects to a symmetrized logit matrix, which a constrained augmented-Lagrangian post-processing step — inherited from the E2Efold/UFold lineage — filters to permit only canonical A-U and C-G pairs plus G-U wobbles while forbidding sharp loops under four nucleotides. Pseudoknots are outside the model's scope.

Training merged 20,923 non-redundant RNAStralign sequences with 10,814 sampled from bpRNA-1m, deduplicated at 80% identity with cd-hit-est to 12,628 sequences, optimized with weighted binary cross-entropy (positive weight 300) and Adam at a 1e-5 learning rate. A transfer-learning stage then fine-tuned on 114 curated PDB-derived structures. Reported F1 scores are 0.725 on TS0 (precision 0.676, recall 0.815), 0.709 on bpRNA-new, 0.818 on CASP15, 0.780 on CASP16, and 0.792 on PDB-new, where the next-best tool reached 0.699. The released checkpoints are under 1 MB, reflecting how much of the model's power is carried by its priors rather than its parameters.

#Applications

The model targets researchers annotating structure for newly discovered non-coding RNAs, where the relevant family is by definition underrepresented in training data, and designers of synthetic RNA — riboswitches, aptamers, origami nanostructures, and therapeutic constructs — whose sequences fall outside natural sequence space entirely. A public web server provides predictions without local installation, which matters because local inference first requires running the four upstream predictors to generate input contact maps — a setup burden single-sequence models avoid.

#Impact

MetaFold-RNA offers evidence that the out-of-distribution brittleness of neural RNA folding is addressable by architecture rather than more data: fusing paradigms with complementary failure modes yields gains precisely where single-paradigm models collapse. The ceiling this imposes is acknowledged: accuracy stays bounded by the constituent priors, and the framework cannot recover a structure all four predictors miss. It also predicts only nested canonical pairs and accepts no experimental probing data as input. Practical reuse carries a caveat: although the preprint states the code is released under an open-source license, the repository contains no license file, leaving the terms for the code and the bundled weights unspecified.