Revisiting homology-aware cross validations for RNA secondary structure prediction

This repository contains the data and source code for the manuscript Revisiting homology-aware cross validations for RNA secondary structure prediction, by L.A. Bugnon, G. Kulemeyer, M. Gerard, L. Di Persia, G. Stegmayer and D.H. Milone, 2025 (under review). Research Institute for Signals, Systems and Computational Intelligence, sinc(i).

In this work we revise existing cross-validation strategies for RNA secondary structure prediction: random k-fold, clustering fold and family fold.

We explain and analyze in detail the advantages and disadvantages of each one of them, additionally proposing two novel evaluation methodologies for testing: human-learned fold and similarity fold.

All validation strategies are applied to state-of-the-art methods for RNA secondary structure prediction and comparative results are analyzed.

Reproducing experimental results

Distribution of structural distances

This notebook reproduces the analysis of the distributions of testing to training structural distances for the different cross-validation strategies analyzed.

Performance comparison on RNA folding

This notebook shows the performance comparison among different cross-validation strategies that can be found in literature

Performance comparison among existing cross-validation strategies. A) random $k$-fold; B) clustering fold; C) fam-fold.

Performance comparison of mean performance with 95% confidence interval in A) human learned fold; B) similarity fold.

This notebook reproduces the comparison of distribution of canonical connections (GC, AU and GU) in 3 folds of the family fold.