DRfold2

RNA tertiary structure prediction lags protein structure prediction by a wide margin, and the reason is data. The Protein Data Bank holds far fewer RNA structures, and RNA homolog searches rarely assemble alignments deep enough to expose the co-evolutionary couplings that made MSA-based protein folding work, so alignment-dependent methods do well only where families are densely populated.

DRfold2 attacks that bottleneck by extracting co-evolutionary signal from unaligned sequence data instead. Developed in Yang Zhang's group at the National University of Singapore — the lab behind I-TASSER and the original DRfold — it pairs a pretrained RNA Composite Language Model (RCLM) with a denoising end-to-end structure module, folding an RNA from a single sequence with no alignment, template, or precomputed secondary structure. It appeared as a preprint in March 2025 and was published in PLOS Biology in February 2026.

The key idea is in how the language model is trained. Conventional RNA language models maximize a masked, position-independent likelihood, discarding exactly the pairwise dependency that base pairing consists of. RCLM instead maximizes a composite likelihood summing nucleotide-wise masked prediction and pairwise joint prediction, so higher-order inter-nucleotide interactions enter the objective directly.

#Key Features

• Single-sequence folding: The pipeline needs only a FASTA sequence — no MSA construction, no homolog search, no secondary structure preprocessing — removing the dependency on alignment depth that limits MSA-based predictors.
• Composite likelihood pretraining: A pairwise joint nucleotide term added to the masked-token objective lets the language model represent dependencies between positions rather than treating each independently.
• Denoising structure module: Training perturbs the initial nucleotide frames with random rotations and Gaussian translation noise, and the module learns to correct them.
• Decoy pool and refinement: Inference draws conformations from 80 saved checkpoints, then selects and optimizes among them using the end-to-end and geometry potentials, with optional OpenMM relaxation.
• Complementarity as a design property: Sharing no components with diffusion-based all-atom predictors, its errors are largely uncorrelated with theirs, so the two combine in a joint potential.

#Technical Details

RCLM has 47.50 million parameters and was trained on approximately 30 million RNAcentral sequences over 67,000 batches of size 128 — about 15 days on a single NVIDIA A40. It emits 512-dimensional sequence and 128-dimensional pair representations through 18 RNA Transformer Blocks, with triangle attention omitted for efficiency; 16 further blocks reduce both to 64 dimensions before the Denoising RNA Structure Module applies invariant point attention. Four structure models were trained independently, the last 20 checkpoints of each forming the decoy pool.

Evaluation used 28 RNA structures under 400 nucleotides from RNA-Puzzles, CASP15, and RNA-only PDB entries released in 2024, with training restricted to pre-2024 structures below 80% identity to any test target. At the 80% identity cutoff DRfold2 averages a TM-score of 0.351, 18.6% above the second-best method, DeepFoldRNA (0.296); at the strictest cutoff it reaches 0.269 against 0.229 for RoseTTAFoldNA. It also leads on secondary structure recovery, with Watson-Crick interaction network fidelity of 0.836 and stacking fidelity of 0.732, though non-canonical interactions remain hard for every method tested at an INF of 0.176. RCLM's unsupervised top-L contact precision at 12 Å is 47.8% versus 23.7% for RNA-FM, roughly twice its size at 99.52 million parameters. Against AlphaFold 3, DRfold2 edges ahead on average (TM-score 0.351 versus 0.345, RMSD 14.6 Å versus 16.0 Å), and a joint potential over both reaches 0.365.

#Applications

DRfold2 targets structural biologists and RNA therapeutic programs working on molecules where no close homolog has a solved structure — riboswitches, ribozymes, viral structured elements, and engineered aptamers. Needing no alignment, it is usable on synthetic and orphan sequences where MSA-based methods have nothing to work from, and single-sequence inference keeps per-target cost low enough for screening. The decoy pool is itself useful output, giving a starting point for reasoning about RNAs that populate more than one fold. A standalone package and a web server are both available.

#Impact

DRfold2 makes a specific, testable claim about RNA language models: that the masked-token objective borrowed from protein and text modeling underuses RNA sequence data, and that a composite likelihood recovers pairwise information the standard objective throws away. The evidence is the doubling of unsupervised contact precision at half the parameter count, tied to downstream accuracy by a Pearson correlation of 0.528 between unsupervised secondary structure precision and final TM-score. Absolute accuracy remains modest: TM-scores near 0.35 mean most predictions capture topology rather than atomic detail. An earlier version competed in CASP16 as "dNAfold", placing fifth by average TM-score (0.510) across 20 RNA monomer targets while posting the highest AlphaFold dissimilarity score of any group that beat the AlphaFold 3 server — evidence that methodological diversity is worth preserving where one dominant architecture could otherwise homogenize errors. Code is on GitHub with the full MIT permission text stated in the README rather than in a license file; weights, roughly 1.3 GB, are fetched by the installation script. The repository also ships the evaluation inputs — data/train.fasta with 10,432 training chains and data/test08.fasta with the 28 benchmark sequences — so the reported comparisons can be reproduced on the same targets.