Fig. 3

Detecting potential discrepancies between sequential and structural datasets. a Schematic of the three sequence-structure predictions performed. On the left, the original ChIP-seq is fed to a CNN (forward model) and outputs both a 1D sequence and a predicted Hi-C. In the centre, the original Hi-C is fed to a DNN (backward model) to predict the 1D sequence found in the forward model called the backward sequence. On the right, the backward sequence is fed to the dense neural network of the forward model (without fitting it again) to generate a new Hi-C prediction (Hi-C prediction 2), based on the backward sequence derived from the original Hi-C. b Shown is the correlation between the data generated by the models in (a) along the w-wide genomic windows. Genomic regions where the backward sequence differs from the original sequence tend to coincide with regions where the predicted Hi-C and the original Hi-C differ (0.35 correlation). There is an improvement between the predicted Hi-C and the original Hi-C when using the backward predicted sequence as input