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

AI summary

Overview

Research area: Machine learning for high-dimensional biomedical data (graph neural networks applied to omics data, specifically gene expression analysis and biomarker discovery).

Technical level: Intermediate. The paper assumes familiarity with graph neural networks, tree-ensemble models (XGBoost, random forest), and standard classification metrics, though the architecture itself is described in accessible terms.

Scope: The paper proposes and evaluates a dual-graph neural network framework that combines an external biological network with a data-driven generated graph for disease classification and feature selection on omics data.

What This Paper Is About

Omics datasets such as transcriptomics and proteomics measure thousands of features but often only a few hundred samples, making reliable disease classification and biomarker discovery difficult. Graph neural networks can incorporate prior biological knowledge by treating features as nodes in a network, but existing methods rely on only one source of graph information, either an externally curated biological network or a graph constructed from the data itself. The authors build a model that uses both simultaneously, aiming to improve predictive accuracy while still producing interpretable feature importance scores.

Key Contributions

1. A dual-graph architecture (engGNN) that processes an external undirected feature graph built from curated biological network databases alongside a data-driven directed feature graph generated by XGBoost, using a separate graph-embedded deep feedforward network (GEDFN) for each and concatenating their embeddings.

2. A combined feature importance scoring method that extends the connection weight method and the graph connection weight method by summing importance scores derived from both the external graph pathway and the generated graph pathway for each feature.

3. A systematic evaluation across nine simulation scenarios varying the sample-to-feature ratio (p_n = 0.05, 0.1, 0.2) and the proportion of true features (p_t = 0.05, 0.1, 0.2), with n = 5000 samples and feature counts p of 250, 500, and 1000, comparing against ablation variants (GEDFN_xgb, GEDFN_rf, GEDFN_e) and classical methods (DFN, XGBoost, RF).

4. Real-data validation on Alzheimer's disease gene expression data (Nachun et al., 2019a), including KEGG over-representation analysis of the top 1000 prioritized genes.

Main Findings

• Simulated classification performance: engGNN achieved the best or tied-best accuracy in 4 of 9 scenarios, led in ROC-AUC in 6 of 9 scenarios, and was best or tied on F1-score in 5 of 9 scenarios. Performance generally improved as p_n and p_t increased.

• One scenario where engGNN did not lead: At p_n = 0.05 and p_t = 0.1, engGNN reached accuracy 0.870, ROC-AUC 0.946, and F1-score 0.838, slightly below GEDFN_e, which reached 0.886, 0.956, and 0.847 respectively.

• Baseline comparison in simulations: GEDFN_xgb was the most competitive alternative with 4 ROC-AUC wins but no advantage in accuracy or F1-score; GEDFN_rf won twice each in accuracy and ROC-AUC; GEDFN_e won 3 times on F1-score but underperformed on ROC-AUC; DFN, XGBoost, and RF rarely reached top performance.

• Feature selection in simulations: engGNN consistently produced the highest ROC-AUC and PR-AUC. At p_t = 0.1 and p_n = 0.2 it reached ROC-AUC 0.654 and PR-AUC 0.693, versus 0.653 and 0.694 for GEDFN_e, while other baselines stayed near 0.50, approximately random performance.

• High-signal feature selection gap: At p_t = 0.2 and p_n = 0.2, engGNN reached PR-AUC 0.860, slightly above GEDFN_e at 0.859, while GEDFN_xgb and RF were far behind at 0.741 and 0.737.

• Alzheimer's disease real data: engGNN achieved accuracy 0.788 (SD 0.0007), ROC-AUC 0.855 (SD 0.0007), and F1-score 0.579 (SD 0.0054) across 20 replications. The next best model, GEDFN_xgb, reached 0.777, 0.852, and 0.571. All baseline differences were marked statistically significant (p-value < 0.001) using Welch's t-test for unequal variances.

• Pathway enrichment: KEGG over-representation analysis of the top 1000 engGNN-prioritized genes identified calcium signaling (hsa04020, p.adjust 0.00013), MAPK signaling (hsa04010, p.adjust 0.00033), signaling pathways regulating pluripotency of stem cells (hsa04550, p.adjust 0.00033), aldosterone synthesis and secretion, cAMP signaling, EGFR tyrosine kinase inhibitor resistance, Rap1 signaling, oocyte meiosis, insulin secretion, and endocrine resistance among the top 10 pathways.

Methodology in Plain English

The framework has four parts. First, the authors build an undirected feature graph from publicly available biological network databases, where nodes are molecular features such as genes and edges represent known functional relationships. Second, they train an XGBoost model on the same prediction task and convert each fitted decision tree into a directed graph, where nodes are the features the tree selected and edges follow the tree's splitting order; all trees' graphs are merged into one aggregated directed graph. They chose XGBoost over random forest and gradient boosting machine because of its sparsity-aware algorithm and built-in regularization.

Each graph feeds a separate GEDFN model. In a GEDFN, the first hidden layer has the same number of neurons as there are features, and the connections between the input and that first layer are masked by the graph's adjacency matrix, so only graph-adjacent features influence each other. Because the external graph is undirected, its masked connections are symmetric; because the tree-derived graph is directed, its masked connections are asymmetric. Self-loops are added to both graphs so each feature also retains its own contribution. Later hidden layers are fully connected. The last hidden layers from the two GEDFNs are concatenated and passed through a standard deep feedforward network with a softmax output to produce the final class probabilities.

For interpretability, each feature's importance score is the sum of two pieces: the absolute connection weights involving that feature in the external-graph GEDFN and the corresponding absolute weights in the generated-graph GEDFN, restricted to actually connected node pairs by an indicator function. Input features are Z-score normalized before training so weights are comparable, and the relative ranking of features is what guides downstream biological interpretation.

Implementation used TensorFlow and the XGBoost package, the Adam optimizer, ReLU activations, and categorical cross-entropy loss with K = 2 softmax outputs for binary classification. The number of XGBoost trees was set to 0.2p. Network depth was three, with the first hidden layer width fixed at p and subsequent layers of 64 and 16 neurons; the final DFN had one hidden layer of width 16. Training used a learning rate of 0.0001, 50 epochs, dropout 0.2, batch size 16, an 80/20 train-test split, 20 repeated runs, and early stopping. Experiments ran on Boston University's Shared Computing Cluster using four CPUs per job.

Why This Matters

• Methodological impact: The paper argues that relying on a single graph source gives only a partial view, because external databases may be incomplete or misannotated, similarity-based graphs may overfit expression data, and model-generated graphs may encode task-specific noise. Combining an external undirected graph with a data-driven directed graph is the paper's distinctive move, and the results suggest this improves both classification and feature selection stability, as indicated by lower variance across runs.

• Real-world applications:

  • Disease classification from blood or tissue gene expression profiles where sample sizes are small relative to the number of measured genes.
  • Biomarker prioritization, using the model's feature importance ranking to nominate candidate genes for follow-up studies.
  • Pathway-level interpretation of disease data, as demonstrated by the KEGG enrichment analysis on Alzheimer's disease that recovered calcium signaling, MAPK signaling, and cAMP signaling pathways.
  • Extension to other omics layers the authors list, namely transcriptomics, proteomics, and metabolomics.

• Industry relevance: Pharmaceutical and diagnostics groups working with high-dimensional molecular data face the same "large p, small n" constraint described in the paper, and a method that produces both a prediction and an interpretable feature ranking is directly useful for target discovery and for regulatory contexts where model transparency matters.

Future Directions

1. Extension beyond binary classification. The implementation fixes K = 2 softmax outputs; the framework's applicability to multi-class or multiclass-with-ordinal-outcome problems is not addressed in the available content.

2. Testing on other omics modalities. The external graph and generated graph are demonstrated on gene expression data; whether the dual-graph design transfers to proteomics or metabolomics is stated as a motivation but not evaluated in the content available.

3. Comparing more graph-construction strategies. The paper compares XGBoost- and random forest-derived graphs, but the space of possible data-driven graph builders and of external network databases is not exhaustively explored.

4. Broader clinical validation. The real-data application is a single Alzheimer's disease peripheral blood gene expression dataset; the details of that dataset's size and the full set of downstream analyses are truncated in the content provided, so the generalizability across cohorts and diseases remains open.

Target Audience

This paper is most useful to computational biologists and biostatisticians working on high-dimensional omics classification and biomarker discovery, and to machine learning researchers interested in how prior knowledge graphs can be combined with data-derived graphs inside a single neural architecture. Practitioners who need interpretable predictions for small-sample, many-feature datasets will find the feature importance scoring method and the simulation design most actionable.

Authors’ 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.