Fig. 4.

The trends in DL models for cancer. There is an upward trend in using multi-omics data (blue) compared to single-omic data (orange) (A) and in the integration of domain knowledge (DK) (orange, green, red) (B) based on recent studies for DL in cancer biology. The most frequently integrated domain knowledge are pathways (orange) and other DK (red) like functional modules with recent increase in the usage of PPI networks. C There are three main categories of DK integration as: input data pre-processing (PRE) (blue), architecture definition (ARCH) (orange) and post-hoc comparison (POST-HOC) (green). There is a trend in the use of DK in PRE step, i.e. DK is used to enrich or augment the input data, which results in a change of data representation; D In recent years, there is an increasing number of specialised DL architectures which encode the structure of biological relations. Graph Neural Networks (GNNs, and Graph Convolution Networks - GCNs) based architectures were the most prevalent used (green). There is an increase in the number of sparse DNN (red) and sparse AE/VAE (blue) models