AutoRNA

RNA tertiary structure determines how a transcript is recognized, regulated, and drugged, but the computational tools that transformed protein folding transfer poorly. The Protein Data Bank holds roughly 4,500 RNA structures against more than 200,000 protein structures, and the coevolutionary signal that AlphaFold 2 and AlphaFold 3 exploit is both weaker and more fragmented in RNA, where tRNA, miRNA, and siRNA follow separate evolutionary paths across prokaryotes, eukaryotes, and viruses. Methods that depend on deep alignments or elaborate empirical potentials therefore run into a data wall.

AutoRNA, from Quantori, takes the opposite tack: given the small training set, use a small generative model and no alignment at all. A fully connected variational autoencoder maps a one-hot nucleotide sequence plus a noise vector to the pairwise distance matrix between nucleotide backbone centers of mass. That matrix is converted to 3D coordinates by multidimensional scaling, atoms are filled in from a library of PDB-derived nucleotide quadruplet templates, and molecular dynamics polishes the geometry — the same broad refinement strategy AlphaFold 2 applies at its final stage.

Two points of identification matter. The preprint's first version was titled "RNAfold" before being renamed AutoRNA, so older citations may carry the earlier title; the model is unrelated to the ViennaRNA package's long-established RNAfold, which predicts secondary structure by thermodynamic folding rather than tertiary coordinates. Later revisions of the preprint add comparisons against deep-learning RNA structure predictors including RhoFold+, AlphaFold 3 and Boltz-2, on which AutoRNA trails the strongest co-folding models on RNA-Puzzles targets.

#Key Features

• Alignment-free prediction: The model consumes sequence and nothing else — no multiple sequence alignment, no coevolution statistics, no handcrafted energy terms — which is what makes it trainable on a few hundred structures.
• Distance-matrix formulation: Predicting a full pairwise distance map lets dense layers model non-local tertiary contacts directly, without the local receptive-field assumptions a convolutional network would impose.
• Homology-controlled evaluation: Sequences were hierarchically clustered by pairwise homology at a 60% cutoff into 100 clusters before splitting, so near-duplicates do not leak between training and test.
• Physics-based refinement stage: Template assembly followed by Langevin molecular dynamics in explicit solvent corrects covalent bonds and angles and removes atomic clashes, yielding a physically plausible PDB file.
• Released inference path: The repository ships a standalone predict.py that runs FASTA to structure using downloadable pretrained weights, separate from the training pipeline.

#Technical Details

The network has 64,532,480 trainable parameters. Its input concatenates a flattened one-hot sequence encoding (256 values, for sequences padded to 64 nucleotides) with a 4,096-value noise vector matching the 64x64 distance matrix. The encoder is two fully connected layers of 4,096 and 2,048 units, the bottleneck a single 1,024-unit layer, and the decoder three layers of 2,048, 4,096, and 4,096 units emitting the distance matrix, with batch normalization and dropout at 0.5 between layers. The loss combines a mean-absolute-error reconstruction term with a KL divergence weighted at 1.1. Training data came from PDB entries containing RNA only, excluding protein and DNA complexes, restricted to 4-64 nucleotides, with distances clipped at 100 A and normalized to [0, 1]; 842 sequences were split roughly 80/15/5 into training, validation, and test subsets and trained for 2,000 epochs with Adam at batch size 16.

On the test subset the model reached an RMSE of 4.49 A and MAE of 3.13 A on nucleotide positions, with a GDT score of 32.13 and a TM-score of 0.19. Inference through the VAE takes about 0.001 seconds per sequence in batch execution; the downstream reconstruction uses 256 template catalogs comprising 46,199 files, and refinement runs 1 million Langevin steps in OpenMM with the AMBER14 force field.

#Applications

AutoRNA suits exploratory structural work on short RNAs — generating functional hypotheses, pre-screening candidate conformations, and producing coarse models where no experimental structure exists. The authors report that predicted miRNA sequences form canonical hairpin architectures and that tRNA sequences adopt conformations compatible with cloverleaf organization, so class-characteristic topology is recovered even where atomic detail is not.

#Impact

The paper's chief contribution is a demonstration that a data-efficient, alignment-free generative model can capture coarse RNA geometry, and its chief virtue is candor about what that buys. The authors state directly that TM-scores below 0.25 and GDT values near 30 do not indicate correct tertiary structures and that the outputs should not be read as near-native models; performance also degrades substantially beyond 64 nucleotides, with RMSE above 6 A and GDT below 15. The work remains a bioRxiv preprint across four versions and has not been peer reviewed. Code, an inference script, and downloadable weights are public, but the repository carries no license file, leaving reuse terms undefined; the preprint text itself is CC-BY-ND.