TF-DWGNet: A Directed Weighted Graph Neural Network with Tensor Fusion for Multi-Omics Cancer Subtype Classification

Technical Summary of TF-DWGNet: A Directed Weighted Graph Neural Network with Tensor Fusion for Multi-Omics Cancer Subtype Classification

Problem Statement
Accurate cancer subtype classification is central to precision medicine but remains challenging due to the biological heterogeneity of tumors and the complexity of multi-omics data. Multi-omics datasets (e.g., DNA methylation, mRNA, and miRNA expression) are inherently high-dimensional, heterogeneous, and exhibit complex intra- and inter-modality dependencies. While Graph Neural Networks (GNNs) offer a principled framework for modeling relational structures, existing approaches often rely on predefined, undirected, or unweighted similarity networks. These methods frequently fail to capture task-specific directionality, interaction strengths, and higher-order interactions (beyond pairwise). Furthermore, interpretability at both the feature and modality levels often relies on computationally intensive post-hoc analyses that may not reflect the model's true decision-making process.

Methodology: TF-DWGNet Framework
The authors propose TF-DWGNet, a fully supervised, end-to-end framework designed to integrate multi-omics data through four coordinated modules:

1. Supervised Directed Weighted Graph Construction:
   Instead of using correlation-based similarity networks, the framework leverages the tree-splitting structure of XGBoost to construct task-specific graphs for each omics modality.

   • Mechanism: An ensemble of $M$ boosted trees is trained on each modality. Each tree is treated as a directed graph where an edge $a \to b$ indicates that feature $a$ appears before feature $b$ along a root-to-leaf path.
   • Aggregation: Feature co-occurrence across all trees is aggregated to form a single directed, weighted graph ($G^*$). Edge weights represent the frequency of the dependency across trees.
   • Outcome: This process performs data-driven feature reduction and encodes both feature importance and directional conditional dependencies (where a feature partitions patients into subgroups for which subsequent features provide discriminative power).

2. Graph Neural Network (GNN) Encoding:
   Each modality's reduced feature matrix and its corresponding directed weighted graph are processed using a modified Graph-Embedded Deep Feedforward Network (GEDFN).

   • Architecture: The adjacency structure is embedded directly into the first hidden layer by restricting connections between input and hidden nodes based on the graph topology.
   • Normalization: The input is processed using a normalized adjacency matrix with self-loops ($\tilde{D}^{-1/2}\tilde{A}\tilde{D}^{-1/2}$) to prevent highly connected nodes from dominating the representation.
   • Output: This module learns latent unimodal embeddings ($Z_1, Z_2, Z_3$) for each omics type.

3. Tensor Fusion and Low-Rank Decomposition:
   To capture higher-order interactions (unimodal, bimodal, and trimodal) without the computational intractability of full tensor expansion, the framework employs a Tensor Fusion Network (TFN) followed by Canonical Polyadic (CP) decomposition.

   • Augmentation: Latent representations are augmented with a constant feature (1) to facilitate interaction terms involving fewer than three modalities.
   • Fusion: A third-order tensor is constructed via the outer product of the augmented embeddings.
   • Decomposition: The tensor is approximated using low-rank CP decomposition ($T \approx \sum \lambda_r p_{1,r} \otimes p_{2,r} \otimes p_{3,r}$). This reduces computational complexity from exponential to linear relative to the rank $R$, while retaining the ability to model synergistic mechanisms across all omics layers.

4. Residual Classification Network:
   The compact, fused low-rank embedding is passed to a Deep Residual Network (DFN) with residual blocks for final multiclass subtype prediction. The use of residual connections stabilizes optimization and enables deeper feature transformations.

Key Contributions
The paper identifies three primary contributions:

1. Supervised Directed Weighted Graph Construction: The use of XGBoost to generate biologically meaningful, variable-sized, directed graphs that encode task-specific directionality and interaction strength, replacing static, undirected similarity networks.
2. Tensor Fusion for Higher-Order Interactions: A low-rank tensor fusion module that efficiently captures unimodal, bimodal, and trimodal interactions, offering richer representational capacity than simple concatenation or pairwise modeling.
3. Built-in Interpretability: The framework inherently produces feature-importance rankings and modality contribution scores derived from connection weights and graph structures, eliminating the need for post-hoc analysis.

Experimental Results
The authors evaluated TF-DWGNet on three real-world cancer datasets from TCGA: BRCA (Breast Carcinoma), UCEC (Uterine Corpus Endometrial Carcinoma), and KIPAN (Pan-Kidney cohort).

• Performance: TF-DWGNet consistently outperformed state-of-the-art baselines (including RF, XGBoost, DFN, GCN, GEDFN, and ablation variants) across Accuracy, F1-weighted, and F1-macro metrics.
  • On the BRCA dataset, the model achieved an accuracy of $0.821 \pm 0.023$, significantly outperforming the next best baseline (GEDFN at $0.790$) and showing substantial gains in F1-macro ($0.785$ vs. $0.748$).
  • Statistical significance was confirmed via Welch's $t$-tests over 20 independent stratified splits, with most improvements on BRCA being highly significant ($p < 0.001$).
• Ablation Studies: Removing key components degraded performance. Replacing XGBoost with Random Forest for graph construction ($TF\text{-}DWGNet_{rf}$) and removing directionality ($TF\text{-}DWGNet_{undir}$) resulted in notable performance drops, validating the importance of supervised, directed graph construction.
• Interpretability: The model successfully identified biologically relevant biomarkers (e.g., FAM134B and FOXC1 in BRCA) and provided modality-level contribution scores. In BRCA, all three omics types contributed comparably, whereas in UCEC and KIPAN, DNA methylation emerged as the dominant predictive signal.

Significance and Claims
The paper claims that TF-DWGNet represents the first framework to jointly integrate supervised directed weighted graph construction, comprehensive low-rank tensor fusion, modality-flexible GNN encoding, and built-in interpretability for multi-omics cancer subtype classification. The authors assert that the model provides a robust, interpretable solution that captures the hierarchical and directional nature of biological processes. By avoiding predefined similarity networks and capturing higher-order interactions, TF-DWGNet offers a principled approach to multi-omics integration that yields biologically meaningful insights and supports precision oncology without requiring manual feature engineering or post-hoc interpretability methods.