Fig. 3: Chromatin-associated RNAs contribute to accurate prediction of 3D genome folding.

a The architecture of the models in the AkitaR framework. b Barplots of performance for different types of AkitaR models on the held-out test set of genomic regions. Pearson’s correlation and mean squared error (MSE) between experimental and predicted contact maps are used as performance metrics. Error bars represent the mean ± standard error of the mean for each model type independently trained five times. Two-sided Mann-Whitney U tests were used to evaluate differences between all pairs of models. Every comparison was significant (p-value < =0.05) except those labeled as not significant (ns). U statistics and p-values for the comparisons are shown in Supplementary Data 1. Individual data points are shown as dots. c Violin plot of Pearson’s correlation of insulation tracks from observed test set maps versus predicted maps (n = 413) for the best model of each type. The box represents the interquartile range (IQR), with whiskers set to 1.5 times the IQR. The dot within the box represents the median value. d Examples showing better prediction of contact maps with nascent transcripts (top panel) or trans-located RNAs (bottom panel). The 3D genome contacts with better prediction are highlighted with green rectangles. Source data are provided as a Source Data file.