Fig. 3

Data representation paradigms and the impact of the integration of domain knowledge. Domain knowledge (DK) can be derived from a database (blue blocks) or expert DK (yellow blocks). DK can be used in pre-processing and data augmentation before the training process. DK from databases can be represented in two ways: A as a step in the pre-processing of input data, before the training process. This first paradigm has emerged for the representation of multi-omic data, which are transformed into graphs or a network and fed into GNN or GCN. This paradigm has been applied to DL models such as: struc2vec, GLUE, several GCN and CNN models; B as inductive bias when creating the neural network architecture, defining the connections between nodes in layers. In this case, DK impacts the training process as it affects the back-propagation. This paradigm has emerged mainly for the representation of multi-omic data, which are fed into sparsely connected Deep Neural Network, where connections are defined by biological relations. This paradigm has been applied to DL models such as: VNN, PNET, KPNN, VAE, CNN