Fig. 6

Evaluating the benefits of smoothing in GRiNCH. Recovery of topology and significant interactions from downsampled then smoothed data. A Rand index and mutual information were used to measure the similarity between TADs from high-depth GM12878 dataset and TADs from downsampled datasets smoothed by different methods (GRiNCH, Mean Filter, Gaussian Filter, HiCNN). Directionality was used as a TAD-calling method independent of any of the smoothing methods, i.e., GRiNCH. The mean is computed across chromosomes and the error bar denotes deviation from the mean. B Area under precision-recall curve (AUPR) was used to measure the recovery of significant interactions called by Fit-Hi-C. Precision and recall were measured for significant interactions from downsampled and smoothed datasets against the “ground truth” defined by the significant interactions from the high-depth GM12878 dataset. Since the pretrained HiCNN models imputes interactions up to 2MB apart, the AUPR for interactions <2MB apart and for all interactions are shown here