Some ideas to reconcile orthology with deep learning to predict regulatory regions using convolutional and graph neural networks

Current deep learning methods, eg CNNs, for functional element prediction are aimed to be trained on one species (eg human) and to predict on the same species (eg human). This is a very strong limit of such model for annotating newly sequenced genomes. Formalizing the annotation task as a GNN connecting species allows to generalize CNNs to predict annotations across species, in a very natural way.