engGNN: A Dual-Graph Neural Network for Omics-Based Disease Classification and Feature Selection

TL;DR

This work tackles prediction and biomarker discovery in high-dimensional omics data by addressing the limitation of single-graph priors in GNNs. It introduces engGNN, a dual-graph GNN that jointly leverages an external undirected biological graph and a data-driven generated directed graph through two GEDFN modules, whose embeddings are combined and fed to a DFN classifier for final predictions. The approach yields interpretable feature importance scores and enables pathway enrichment analyses, demonstrating superior performance on synthetic datasets and a real Alzheimer's disease gene expression dataset. Overall, engGNN offers a robust, flexible framework that improves predictive power in small-sample, high-dimensional settings while supporting biologically meaningful biomarker discovery.

Abstract

Omics data, such as transcriptomics, proteomics, and metabolomics, provide critical insights into disease mechanisms and clinical outcomes. However, their high dimensionality, small sample sizes, and intricate biological networks pose major challenges for reliable prediction and meaningful interpretation. Graph Neural Networks (GNNs) offer a promising way to integrate prior knowledge by encoding feature relationships as graphs. Yet, existing methods typically rely solely on either an externally curated feature graph or a data-driven generated one, which limits their ability to capture complementary information. To address this, we propose the external and generated Graph Neural Network (engGNN), a dual-graph framework that jointly leverages both external known biological networks and data-driven generated graphs. Specifically, engGNN constructs a biologically informed undirected feature graph from established network databases and complements it with a directed feature graph derived from tree-ensemble models. This dual-graph design produces more comprehensive embeddings, thereby improving predictive performance and interpretability. Through extensive simulations and real-world applications to gene expression data, engGNN consistently outperforms state-of-the-art baselines. Beyond classification, engGNN provides interpretable feature importance scores that facilitate biologically meaningful discoveries, such as pathway enrichment analysis. Taken together, these results highlight engGNN as a robust, flexible, and interpretable framework for disease classification and biomarker discovery in high-dimensional omics contexts.

Figures (14)

• Figure 1: Architecture of the proposed external and generated Graph Neural Network (engGNN). The inputs include feature matrix $X$, an external undirected feature graph $G_e$, and a generated directed feature graph $G_{g}$. Step 1: $X$ and $G_{e}$ are processed by a graph-embedded deep feedforward network ($\mathrm{GEDFN}_{\text{e}}$); Step 2: $X$ and $G_{g}$ are processed by another GEDFN ($\mathrm{GEDFN}_{\text{g}}$); Step 3: the last hidden layers of $\mathrm{GEDFN}_{\text{e}}$ and $\mathrm{GEDFN}_{\text{g}}$ are concatenated; Step 4: the concatenated representation is passed through a deep feedforward network (DFN), followed by a softmax activation to generate the final predicted class probabilities $\hat{Y}$. Overviews of $\mathrm{GEDFN}_{\text{e}}$ and $\mathrm{GEDFN}_{\text{g}}$ are detailed in Figure \ref{['GEDFN_e']} and Figure \ref{['GEDFN_g']}, respectively.
• Figure 1: Average classification performance of various models on simulated datasets across different values of sample size-to-feature size ratio ($p_{n}$) and proportion of true features ($p_{t}$). $p_{n}, p_{t} \in \{0.05, 0.1, 0.2\}$. Metrics include accuracy, ROC-AUC, and F1-score. Subplots (a) for different $p_{n}$ values; Subplots (b) for different $p_{t}$ values.
• Figure 2: Overview of $\mathrm{GEDFN}_{\text{e}}$. The left panel shows the inputs, including feature matrix $X$ and external undirected feature graph $G_e$, while the right panel shows the layer-wise structure of the graph-embedded deep feedforward network (GEDFN) model. The first hidden layer $H_1$ has the same number of neurons as the input feature size $p$. Connections from $X$ to $H_1$ reflect the adjacency structure of $G_e$ and are therefore symmetric, whereas subsequent layers (e.g., $H_1$ to $H_2$) are fully connected. The outputs are the last hidden layer $H_{out}$ of GEDFN.
• Figure 2: Average classification performance of various models on simulated datasets across nine scenarios. The sample size is $n = 5000$, and each scenario was replicated 20 times.
• Figure 3: Overview of $\mathrm{GEDFN}_{\text{g}}$, illustrated here using $\mathrm{GEDFN}_{\text{xgb}}$. The left panel illustrates the feature graph extractor using XGBoost, and the right panel presents the layer-wise structure of the graph-embedded deep feedforward network (GEDFN) model. The inputs are the feature matrix $X$ and the generated directed feature graph $G_g$. The first hidden layer $H_1$ has the same number of neurons as the input feature size $p$. Connections from $X$ to $H_1$ reflect the adjacency structure of $G_g$ and are therefore asymmetric, whereas subsequent layers (e.g., $H_1$ to $H_2$) are fully connected. The outputs are the last hidden layer $H_{out}$ of the GEDFN model.