Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Kirigami: large convolutional kernels improve deep learning-based RNA secondary structure prediction

Marc Harary, Chengxin Zhang

TL;DR

Kirigami tackles RNA secondary structure prediction by reframing base-pairing as a graph adjacency problem and applying a large-kernel fully convolutional network ($k=11$) with no pooling to capture long-range dependencies. A series of post-processing constraints—including symmetry, canonical pair enforcement via prime-encoding, and multiplet elimination—produce physically plausible structures, enabling pseudoknots. On the bpRNA-based TS0 benchmark of $1{,}305$ structures, Kirigami achieves a mean MCC of $0.706$ (notably higher than SOTA methods) and $0.615$ on pseudoknots, highlighting the benefit of large receptive fields for RNA topology. These results suggest neural approaches can surpass traditional thermodynamic models in many cases, while underscoring data scarcity as a key limitation and pointing toward future directions like attention mechanisms and expanded experimental structure datasets.

Abstract

We introduce a novel fully convolutional neural network (FCN) architecture for predicting the secondary structure of ribonucleic acid (RNA) molecules. Interpreting RNA structures as weighted graphs, we employ deep learning to estimate the probability of base pairing between nucleotide residues. Unique to our model are its massive 11-pixel kernels, which we argue provide a distinct advantage for FCNs on the specialized domain of RNA secondary structures. On a widely adopted, standardized test set comprised of 1,305 molecules, the accuracy of our method exceeds that of current state-of-the-art (SOTA) secondary structure prediction software, achieving a Matthews Correlation Coefficient (MCC) over 11-40% higher than that of other leading methods on overall structures and 58-400% higher on pseudoknots specifically.

Kirigami: large convolutional kernels improve deep learning-based RNA secondary structure prediction

TL;DR

Kirigami tackles RNA secondary structure prediction by reframing base-pairing as a graph adjacency problem and applying a large-kernel fully convolutional network () with no pooling to capture long-range dependencies. A series of post-processing constraints—including symmetry, canonical pair enforcement via prime-encoding, and multiplet elimination—produce physically plausible structures, enabling pseudoknots. On the bpRNA-based TS0 benchmark of structures, Kirigami achieves a mean MCC of (notably higher than SOTA methods) and on pseudoknots, highlighting the benefit of large receptive fields for RNA topology. These results suggest neural approaches can surpass traditional thermodynamic models in many cases, while underscoring data scarcity as a key limitation and pointing toward future directions like attention mechanisms and expanded experimental structure datasets.

Abstract

We introduce a novel fully convolutional neural network (FCN) architecture for predicting the secondary structure of ribonucleic acid (RNA) molecules. Interpreting RNA structures as weighted graphs, we employ deep learning to estimate the probability of base pairing between nucleotide residues. Unique to our model are its massive 11-pixel kernels, which we argue provide a distinct advantage for FCNs on the specialized domain of RNA secondary structures. On a widely adopted, standardized test set comprised of 1,305 molecules, the accuracy of our method exceeds that of current state-of-the-art (SOTA) secondary structure prediction software, achieving a Matthews Correlation Coefficient (MCC) over 11-40% higher than that of other leading methods on overall structures and 58-400% higher on pseudoknots specifically.
Paper Structure (14 sections, 2 equations, 3 figures, 3 tables)

This paper contains 14 sections, 2 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Results
  3. Datasets
  4. Overview of pipeline
  5. Comparison with existing methods
  6. Ablation studies
  7. Biophysical analysis
  8. Discussion and Conclusion
  9. Materials and methods
  10. Input features and output labels
  11. Network architecture
  12. Post-processing
  13. Training procedure
  14. Evaluation metrics

Figures (3)

  • Figure 1: Flowchart of the Kirigami pipeline. Green nodes represent input-output values, yellow miscellaneous operations, and blue trainable neural network layers (or closely associated activation functions). Dashed red boxes indicate network submodules. Constraints (I-IV) are indicated in their corresponding nodes.
  • Figure 2: Stages for post-processing using bpRNA_RFAM_35590 as an example. A. Constraint I: symmetric predicted base-pairing probabilities $\left( \mathbf{\hat{y}}_1 \right)$ in log scale. The receptive field of an kernel is indicated in the red dashed box. B. Adjacency matrix of the ground truth $\left( \mathbf y \right)$. C. Final predicted adjacency matrix $\left(\mathbf{\hat{y}}_3 \right)$. D. Constraints II: Removal of sharp angles $\left( \mathbf{\hat{y}}_2 \right)$. In this case none are present. E. Constraint III: Removal of non-canonical pairs $\left( \mathbf{\hat{y}}_3 \right)$. In this case none are present. F. Constraint IV: Removal of multiplets $\left( \mathbf{\hat{y}}_4 \right)$—in this case none are present—and thresholding $\left( \mathbf{\hat{y}}_5 \right)$. G. Ball-and-stick representation of the ground truth $\left( \mathbf y \right)$. Positions assigned using Pseudoviewer 50. H. Ball-and-stick representation of the final, thresholded structure $\left( \mathbf{\hat{y}}_5 \right)$.
  • Figure 3: A. Difference between the Gibbs free energy ($\Delta G$) of the predicted structure and that of the ground truth by model as predicted by ViennaRNA. B. MCC versus $\Delta G$ by model with the corresponding correlation coefficient.