Fig. 6. Twins can learn to be robust discrepancies in sequencing depth.

a Schema describing the training process a network which is trained to learn sequencing depth artefacts. High-resolution and low-resolution Hi-C maps are compared to another pair as though they are from separate conditions. b Contrastive loss and cross embedding loss terms during training of a network which follows the schema described in (a). c Twins embedding distances distribution on test chromosome for comparison between high-depth and low-depth Hi-C windows using network trained to learn sequencing depth artefacts. d Twins embedding distances across chromosome 2 for comparison between high-depth and low-depth Hi-C windows using network trained to learn sequencing depth artefacts. Data are presented as mean values ± the 95% confidence interval. e Schema describing the training process for the control network, a network which learns to be robust to sequencing depth artefacts. One high-resolution and one low-resolution replicate are compared to another pair as though they are from separate conditions. f Contrastive loss and cross embedding loss terms during training of a network which follows the schema described in (e). Loss does not decrease significantly and the network does not learn artificial differences. g As in (c) but using network trained to be robust to artefacts. h As in (d) but using network trained to be robust to artefacts.