Explaining decisions of graph convolutional neural networks: patient-specific molecular subnetworks responsible for metastasis prediction in breast cancer

Abstract

Background: Contemporary deep learning approaches show cutting-edge performance in a variety of complex prediction tasks. Nonetheless, the application of deep learning in healthcare remains limited since deep learning methods are often considered as non-interpretable black-box models. However, the machine learning community made recent elaborations on interpretability methods explaining data point-specific decisions of deep learning techniques. We believe that such explanations can assist the need in personalized precision medicine decisions via explaining patient-specific predictions.

Methods: Layer-wise Relevance Propagation (LRP) is a technique to explain decisions of deep learning methods. It is widely used to interpret Convolutional Neural Networks (CNNs) applied on image data. Recently, CNNs started to extend towards non-Euclidean domains like graphs. Molecular networks are commonly represented as graphs detailing interactions between molecules. Gene expression data can be assigned to the vertices of these graphs. In other words, gene expression data can be structured by utilizing molecular network information as prior knowledge. Graph-CNNs can be applied to structured gene expression data, for example, to predict metastatic events in breast cancer. Therefore, there is a need for explanations showing which part of a molecular network is relevant for predicting an event, e.g., distant metastasis in cancer, for each individual patient.

Results: We extended the procedure of LRP to make it available for Graph-CNN and tested its applicability on a large breast cancer dataset. We present Graph Layer-wise Relevance Propagation (GLRP) as a new method to explain the decisions made by Graph-CNNs. We demonstrate a sanity check of the developed GLRP on a hand-written digits dataset and then apply the method on gene expression data. We show that GLRP provides patient-specific molecular subnetworks that largely agree with clinical knowledge and identify common as well as novel, and potentially druggable, drivers of tumor progression.

Conclusions: The developed method could be potentially highly useful on interpreting classification results in the context of different omics data and prior knowledge molecular networks on the individual patient level, as for example in precision medicine approaches or a molecular tumor board.

Keywords: Classification of cancer; Deep learning; Explainable AI; Gene expression data; Molecular networks; Personalized medicine; Precision medicine; Prior knowledge.

Fig. 1

The workflow to obtain a data point-specific subnetwork. For clarity, a data point represented by a gene expression profile of a patient from the breast cancer dataset. The molecular network (HPRD PPI) structures the genes and is the same for every patient. Patient’s gene-expression values are assigned to every vertex of the HPRD PPI so that the patient is represented as a graph signal. Trained Graph-CNN performs graph convolutions and as output classifies the patient as metastatic or non-metastatic. GLRP is applied as a post hoc processing, propagating the relevance from the predicted label up to the input features (vertices of the molecular network). Top 140 highly relevant vertices constitute a molecular subnetwork. Molecular subnetworks differ from one patient to another

Fig. 2

From left to right: initial image, LRP on classical CNN and GLRP on Graph-CNN

Fig. 3

The PPI subnetworks for (1) metastatic patients a (GSM519217) and b (GSM615233) and (2) non-metastatic patients c (GSM615695) and d (GSM150990). The coloring of the node is based on gene expression levels by 25% and 75% quantiles (blue=LOW, yellow=NORMAL, red=HIGH), based on the gene expression throughout the whole patient cohort. The size of vertices corresponds to the relevance scores within one subnetwork. All the subnetworks are highly relevant compared to the rest of the PPI network. Green circles highlight targetable genes

Fig. 4

Signal transduction pathway analysis of subnetwork genes reported for 79 patients in 5 subtypes. (From left to right) Blue heatmap: 238 signaling pathways clustered according to proportion of shared subnetwork genes; Orange heatmap: Enrichment significance of pathways in subnetwork genes combined from patients of given subtype. Darker orange indicates higher significance; Purple heatmap: Median difference in matched pathway genes observed in pairwise comparisons of subnetwork gene sets from patients mapped to 33 pathways. Darker purple indicates higher tendency of pairs of subnetwork gene sets to coincide with different pathway genes; Green heatmap: Enrichment significance of pathways in subnetwork genes of 79 patients. Darker green indicates higher significance. Corresponding subtypes and metastatic status are shown by the annotation above the heatmap. A detailed version of this figure capturing pathway and sample names is provided in Additional file 1: Fig. S2