Stability of feature selection utilizing Graph Convolutional Neural Network and Layer-wise Relevance Propagation

High-throughput technologies are increasingly important in discovering prognostic molecular signatures and identifying novel drug targets. Molecular signatures can be obtained as a subset of features that are important for the decisions of a Machine Learning (ML) method applied to high-dimensional gene expression data. However, feature selection is inherently unstable in this case. Several studies have identified gene sets that provide predictive success for patient prognosis, but these sets usually have only a few genes in common. The stability of feature selection (and reproducibility of identified gene sets) can be improved by including information on molecular networks in ML methods. Graph Convolutional Neural Network (GCNN) is a contemporary deep learning approach applicable to gene expression data structured by a prior knowledge molecular network. Layer-wise Relevance Propagation (LRP) and SHapley Additive exPlanations (SHAP) are techniques to explain individual decisions of deep learning models. We used both GCNN+LRP and GCNN+SHAP techniques to explain GCNNs and to construct feature sets that are relevant to models by aggregating their individual explanations. We also applied more classical ML-based feature selection approaches and analyzed the stability, impact on the classification performance, and interpretability of selected feature sets. Availabilityhttps://gitlab.gwdg.de/UKEBpublic/graph-lrp Contacttim.beissbarth@bioinf.med.uni-goettingen.de