Fig. 5
From: Dense neural networks for predicting chromatin conformation
Gradient analysis of probabilities of chromatin states. Gradient analysis of probabilities of chromatin states. a At the top, a contact map and inner sequence of a genomic window used as an example to evaluate gradients (chromosome 2L 9.95–10.75 Mbp). At the bottom, the gradient of the chromatin states of three different locations σk of the sequence (marked as red) with respect to the probabilities of contact (in the same genomic window as above). The heat maps indicate how the probabilities of contact P(cij) would need to change in order to increase value of the chromatin state σk (make it more active), or equivalently, how σk would change when increasing P(cij). b Genome-average of gradients for subsets of sites where the central chromatin state is either inactive (σk≈0) or active (σk≈1), and genome-average gradients of the central chromatin state. On average, when σk is an inactive state, an increase of contacts between the sites surrounding σk makes σk more inactive (negative gradient). For the active state, an increase of contact probabilities between sites surrounding σk tends to make the sites more active (positive gradient). The average gradient square highlights that the contacts that are most informative about the chromatin state at a given location are the contacts between the sites that flank the location