CS-Fold

CS-Fold is an RNA secondary structure prediction model from the Australian National University, developed by Jiren Zhou, Jiajia Xu, Jiayu Wen, and Brian John Parker. It addresses a gap in how deep learning models consume evolutionary evidence for RNA. Base pairing leaves a distinctive fingerprint in evolution — compensatory double substitutions that preserve a Watson–Crick pair (A:U becoming G:C) and compatible single substitutions that preserve a G:U wobble — and these events are the strongest sequence-level evidence that two positions are paired.

Existing methods read that evidence only indirectly, summarizing multiple sequence alignments column by column and discarding the topology and branch lengths of the phylogenetic tree the sequences came from. The distinction matters: five aligned species carrying the same substitution may reflect five independent mutations or a single change in a shared ancestor, and only the second is real evidence of a compensatory event. CS-Fold recovers it by reconstructing ancestral sequences with PAML over a 100-way vertebrate species tree, counting compensatory events along tree branches, and calibrating them against a Monte Carlo null model of neutral evolution.

The resulting per-position significance values enter a Pairformer network — the triangular update and triangular attention block introduced in AlphaFold 3, descended from the Evoformer of AlphaFold 2 — as an explicit attention bias, alongside frozen embeddings from the RiNALMo RNA language model. The priors then reappear in the training loss and in an unrolled optimization step at inference, so the evolutionary constraint is enforced at three points rather than merely supplied as an input feature.

#Key Features

• Phylogeny-aware evolutionary priors: Ancestral sequence reconstruction plus Monte Carlo simulation under a general time-reversible substitution model produce calibrated p-values for compensatory double and compatible single substitutions at every candidate base pair.
• Biased triangular attention: Of four attention heads in each Pairformer block, two are biased by the substitution likelihoods and two retain a global receptive field, so the sparse evolutionary signal guides attention without dominating it.
• Frozen RNA language model features: RiNALMo supplies the initial sequence embeddings with all weights frozen and no fine-tuning for secondary structure, isolating the contribution of the evolutionary priors.
• Constraints in the loss and at inference: A composite objective of weighted binary cross-entropy, Tversky loss, and a hinge penalty on prior-flagged positions is paired with a primal-dual unrolled post-processing step that enforces legal pairing, minimum loop size, and the prior thresholds.
• Cross-family evaluation: Families held out from training make the reported numbers a test of generalization to unseen RNA families rather than of within-family memorization.

#Technical Details

The training corpus is new to this work. Human RNA families and structures from Rfam 14.5 were combined with the UCSC 100-way vertebrate alignments and species tree; mafFrag extraction yielded 13,962 vertebrate sequences, of which PAML successfully reconstructed 13,898, and removing sequences longer than 1,028 nucleotides left 13,778 sequences spanning 604 families. Nine families dominate the corpus and were downsampled to 200 sequences each, with five assigned to validation and test and four to training with no overlap. The network is deliberately small: two stacked Pairformer layers with a hidden dimension of 64 after outer-product concatenation, against the 48-block stack of AlphaFold. On the held-out families CS-Fold reaches 78.8 precision, 77.0 recall, and 77.9 F1, compared with 67.4 / 82.5 / 73.3 for RiNALMo, 62.2 / 73.4 / 67.3 for MXfold2, 57.7 / 83.0 / 65.5 for UFold, and 58.3 / 69.2 / 63.3 for ERNIE-RNA. The gain is concentrated in precision; recall is lower than several baselines. Removing the compensatory-substitution priors from the attention bias, loss, and post-processing together drops F1 to 57.3, a 20.6-point decrease.

#Applications

CS-Fold suits annotation of structured non-coding RNAs in vertebrate genomes, where deep alignments and a resolved species tree are available: microRNA precursors, signal recognition particle RNAs, Y RNAs, and small nuclear and ribosomal RNAs. Its input requirement is also its constraint — the model needs a vertebrate alignment and tree per query, so it does not apply to single sequences, designed RNAs, or lineages without dense homolog sampling.

#Impact

The contribution is a template as much as a predictor: a statistically calibrated evolutionary prior can substitute for scale, letting a two-layer network outperform fine-tuned RNA language models on cross-family prediction. Two caveats bound its present reach. The work is a bioRxiv preprint under a CC-BY-NC license that has not completed peer review, and the authors state that data and source code will be made publicly available upon acceptance of the manuscript — so no code, weights, model card, or data card accompany it today, and the results are not independently reproducible. Comparisons are also drawn on the authors' own vertebrate dataset rather than the bacterial-heavy benchmarks the field has standardized on, which complicates direct comparison with published numbers.