Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

• Identification of cancer genes is a challenging problem in cancer genomics research
• Computational methods, including deep neural networks, have been developed to address this issue
• These methods fail to exploit gene-gene interactions and provide little explanation for their predictions
• Proposed EMGNN approach leverages multiple gene-gene interaction networks and multi-omics data
• EMGNN outperforms existing approaches and provides valuable biological insights into its predictions

Paper Content

Introduction

• Understanding gene function and disease pathogenicity depends on gene properties and interactions
• High-throughput experiments enable profiling of genetic and molecular properties
• Computational methods predict gene functions by combining gene properties and network connectivity
• Predicting gene pathogenicity in disease-specific contexts is challenging
• Cancer sequencing projects generate data for identifying novel cancer genes
• EMOGI models multi-omics features of cancer genes in PPI networks to predict novel cancer genes
• EMGNN proposed to address challenge of functional properties irrelevant to cancer disease physiology
• EMGNN maximizes concordance of functional gene relationships with unknown disease physiology
• EMGNN achieves state-of-the-art performance by combining information from 6 PPI networks
• EMGNN identifies most important multi-omics features and most influential PPI networks

Datasets

• Trained proposed model with 6 PPI Networks
• Used mutation, copy number, DNA methylation and gene expression data from 29,446 samples from TCGA
• Data from 16 different cancer types

Multilayer graph neural network

• Graph neural networks (GNNs) are used to leverage both network structure and node features.
• GNNs use a message-passing scheme with two steps: aggregating representations of neighbors and updating own representation.
• Popular architectures include Graph Convolution Networks (GCNs) and Graph Attention Networks (GAT).
• Multilayer Graph Construction method is used to handle multiple networks.
• Model is trained using cross-entropy loss function and ADAM optimizer.
• Data is divided into training, testing and validation sets.
• Initial GNN has 3 layers with hidden dimension of 64, meta-GNN has 1 layer with hidden dimension of 64.

Model interpretation

• Captum is a tool for understanding and interpreting machine learning models
• It offers a range of interpretability methods to analyze predictions
• Integrated gradient (IG) module assigns importance score to each input feature
• IG interprets decisions of neural networks by estimating contribution of each input feature
• Baseline input is typically chosen to be neutral or meaningless
• Traditional IG method not applicable to graph neural networks
• Modified approach proposed to compute IG in graph neural networks
• Node feature interpretation and edge feature interpretation analysis used

Novel cancer gene discovery

• Trained EMGNN model was used to predict novel cancer genes in 14019 unlabeled genes
• Genes were ranked by their predicted cancer gene probability
• EMOGI models were applied to each unlabeled gene to predict probability of it being a cancer gene
• Results of 6591 unlabeled genes were analyzed

Gene set enrichment analysis

• Used gene set enrichment analysis to analyze functional enrichment of important gene features in cancer pathways
• Aggregated maximum feature importance of each node using Captum’s feature explanation results
• Excluded genes with zero importance from analysis
• Ranked neighboring gene nodes based on importance
• Used ranked gene list as input for GSEA
• Computed enrichment p-value and multiple testing corrected FDR against cancer hallmark gene sets

Results

Overview of emgnn framework

• Developed a graph neural network model EMGNN
• Input is a feature vector for each gene and multiple graphs
• Model updates graph representation within each graph layer
• Introduces a meta graph layer to combine layer-wise node representations
• Multi-layer perceptron takes meta node representations and performs node classification
• EMGNN generalizes single graph GNN by capitalizing on complementary information stored in multiple graphs

Multilayered graph improves emgnn performance

• EMGNN was used to predict cancer genes using a dataset of 887 labeled cancer genes, 7753 non-cancer genes and 14019 unlabeled genes.
• Six PPI networks were binarized to keep only high-confidence edges.
• Performance increased as the number of input networks increased.
• EMGNN achieved state-of-the-art performance for all test sets.
• EMGNN outperformed EMOGI by a margin of over 5% AURPC.

Evaluating the performance of different gnn architectures and graph ablations.

• Performed ablation study to assess performance of EMGNN model using different GNN architectures and input perturbations
• Found that GCN is the best-performing GNN architecture in all datasets
• Examined performance of EMGNN with GCN architecture under different types of input perturbations
• Found that EMGNN decreased in performance for both random and all-one node features
• Removal of edges slightly decreased EMGNN performance

Explaining emgnn reveals biological insights of cancer gene pathogenicity

• Explainable and trustworthy models are essential for understanding cancer genes and discovering new ones.
• EMGNN was used to analyze the relative contributions from each PPI network to cancer gene predictions.
• ANOVA test showed significant difference in contributions from different PPI networks.
• DNA methylation was found to be significantly more important for known cancer gene prediction than other omics data.

Emgnn identifies novel cancer genes by integrating multilayer graphs

• Trained EMGNN model used to predict cancer genes on unlabeled genes
• Non-trivial number of unlabeled genes with high probability of predicted novel cancer genes
• EMGNN achieved accurate and unified novel cancer gene prediction by integrating multilayer graphs
• Predictions of EMOGI models trained on individual PPI networks showed substantial divergence
• EMGNN provided more accurate, unified predictions of cancer genes
• EMGNN predictions of COL5A1 as cancer gene with high confidence
• All individual PPI networks contributed similarly to EMGNN predictions
• Gene set enrichment analysis used to illustrate potential biological mechanisms of COL5A1

Discussion

• Biomedical and biological domain contains a wealth of information represented and analyzed using graph structures
• Gene interaction and protein-protein interaction networks describe functional relationships of genes and proteins
• Graph construction and integration methods render distinct predictive powers
• Developed a new graph learning framework, EMGNN, to jointly model multilayered graphs
• EMGNN outperforms previous graph neural networks trained on single graphs
• EMGNN leverages complementary information from different graph layers and omics features to predict cancer genes
• EMGNN recovers cancer genes missed by previous state-of-the-art predictors
• EMGNN integrates homogeneous, undirected graphs
• EMGNN can be extended to various types of graphs and to perform cross-data modality integration
• EMGNN reveals molecular aberrations that may be leveraged for screening and re-purposing of drugs