Explaining decisions of Graph Convolutional Neural Networks: patient-specific molecular subnetworks responsible for metastasis prediction in breast cancer

Abstract

MotivationContemporary 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. 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.ResultsWe 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 applied 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. As a result this method could be potentially highly useful on interpreting classification results on the individual patient level, as for example in precision medicine approaches or a molecular tumor board.Availabilityhttps://gitlab.gwdg.de/UKEBpublic/graph-lrphttps://frankkramer-lab.github.io/MetaRelSubNetVis/Contacttim.beissbarth@bioinf.med.uni-goettingen.de

Text mining-based word representations for biomedical data analysis and machine learning tasks

“…Graph-CNNs, trained with word2vec-embedding-based networks, have shown a good enough performance to classify metastatic events vs. non-metastatic events. Predictions of Graph-CNN applied to the same gene expression data used in this study with the HPRD PPI were explained in a recent study and provided patient-specific subnetworks [27]. An interesting research question brought up by this study is whether patient-specific subnetwork genes predicted using an embedding-based gene-gene network would give different insights into the tumor biology of a patient than those predicted using PPI networks.…”

“…One of the approaches for validation of the embedding networks is to analyze how the underlying molecular network influences performance of the machine learning method utilizing prior knowledge. The Graph-CNN [41] method was applied on the breast cancer dataset introduced in section 2.3 in recent studies [26][27]. We subtracted the minimal value (5.84847) of the data from each cell of the quantile normalized gene expression matrix to keep the gene expression values non-negative.…”

“…CNN models were also effective for NLP tasks such as text classification [22-23]. They have been used in bioinformatics [24], namely in drug discovery and genomics [25], and motivated further progress on graph structured prior information with promising results on the prediction of metastatic events [26-27].…”

MetastaSite: Predicting metastasis to different sites using deep learning with gene expression data

“…Table 1 provides the GEO accession numbers of the samples used in this study, along with the sample statistics. Similar to the approach used in (Chereda et al, 2019), we used the RMA probe-summary algorithm (Irizarry et al, 2003) to process each dataset, after which they were combined based on the HG-U133A array probe names, and quantile normalization was applied across all datasets. In cases where multiple probes were mapped to one gene, the probe with the highest average value was taken.…”

MixupExplainer: Generalizing Explanations for Graph Neural Networks with Data Augmentation

“…Graph Neural Networks (GNNs) [29], a powerful technology for learning knowledge from graph-structured data, are gaining increasing attention in today's world, where graph-structured data such as social networks [12,27], molecular structures [6,25], traffic flows [19,21,41,47], and knowledge graphs [32] are widely used. GNNs work by propagating and fusing messages from neighboring nodes on the graph using message-passing mechanisms.…”