Lung adenocarcinoma (LUAD) is the most common subtype of non-small cell lung cancer and remains one of the leading causes of cancer death worldwide. What makes it particularly difficult to treat is its profound molecular heterogeneity - different tumors look similar under a microscope but behave very differently at the molecular level.
Three distinct molecular subtypes of LUAD have been defined through genomic profiling: Terminal Respiratory Unit (TRU), Proximal-Inflammatory (PI), and Proximal-Proliferative (PP). TRU tumors tend to have stable gene activity and better outcomes, while PI and PP tumors show progressively more inflammatory and proliferative behavior, making them harder to treat.
Multi-omics data - simultaneously studying a tumor's gene expression, microRNA activity, DNA methylation, and copy number changes - offers a richer picture than any single data type alone. However, combining these four data types is computationally challenging because each has thousands of measurements and they interact in complex ways.
Existing computational methods largely treat patients as individual data points (nodes in a network) without capturing the gene-level regulatory dynamics that drive tumor behavior. This limits both predictive accuracy and the biological insights these models can provide.
MoAGNN (Multi-omics Hierarchical Graph Neural Network) is a novel artificial intelligence framework designed to make sense of the complex molecular landscape of lung adenocarcinoma. Unlike previous models that represent patients as network nodes, MoAGNN uses individual genes as nodes, allowing it to model gene-to-gene regulatory relationships directly.
The key innovation is that MoAGNN constructs a heterogeneous graph - a network with different types of connections - that captures both relationships between genes within a single data type and regulatory interactions across different omics layers. This allows the model to understand, for example, how a microRNA might suppress a gene's expression or how DNA methylation might silence it.
Self-attention mechanisms within the model allow it to focus on the most informative genes for each prediction task, essentially learning which molecular signals matter most for classifying tumor subtypes, predicting disease stage, or estimating survival risk. This produces a ranked list of important genes that can be biologically validated.
The framework was evaluated on three clinically important tasks using data from The Cancer Genome Atlas (TCGA): classifying tumors into their molecular subtypes, determining whether disease is early or late stage, and predicting whether patients face high or low survival risk.
Data preparation began with four omics datasets from TCGA for lung adenocarcinoma patients: transcriptome (gene expression), microRNA expression, DNA methylation, and copy number variation. Each data type was preprocessed and filtered using statistical methods appropriate to its characteristics, ensuring only informative features were retained.
Graph construction placed differentially expressed genes as nodes in the network. Edges between genes were drawn based on co-expression patterns calculated using WGCNA (Weighted Gene Co-expression Network Analysis), a method that identifies which genes tend to be active or inactive together. A soft-thresholding approach kept the network structured without too many spurious connections.
Cross-omics connections were incorporated by adding edges between the gene expression nodes and their corresponding microRNA, methylation, and copy number data. These inter-omics edges were assigned biological direction weights: methylation and microRNA edges received negative weights (reflecting their typical gene-silencing roles), while copy number edges received positive weights (reflecting their gene-activating tendency).
Three separate cohorts were assembled for the three prediction tasks. The subtype cohort included 185 patients labeled as TRU, PI, or PP; the stage cohort had 450 patients split into early and late stage; and the survival cohort included 441 patients divided into high-risk and low-risk groups based on a two-year survival cutoff.
Graph convolutional layers form the computational heart of MoAGNN. These layers update each gene node's representation by aggregating information from its neighbors in the network, allowing the model to learn from gene co-expression patterns rather than treating each gene in isolation.
Self-Attention Graph Pooling (SAGPooling) selectively retains the most important genes at each stage of processing while pruning less informative ones. The model scores every gene node and keeps only the top-ranked fraction, creating progressively smaller but more focused representations of the tumor's molecular state. This hierarchical pooling is applied three times, producing a compressed but rich summary of the whole gene network.
A final multi-layer perceptron (a standard neural network classifier) takes these compressed graph representations and produces predictions for the three clinical tasks. The model was trained using the Adam optimizer with careful regularization steps - including dropout and early stopping - to prevent overfitting on the relatively small patient cohorts.
Hyperparameter optimization involved tuning two key settings: the edge threshold (which gene co-expression connections to include in the network) and the pooling ratio (what fraction of genes to retain at each pooling step). The best subtype classification results were achieved with the threshold at the 75th percentile of co-expression strength and a pooling ratio of 0.35, meaning only the top 35% of genes were retained.
MoAGNN outperformed all comparison methods on the LUAD subtype classification task, achieving an accuracy of 89.1%. This beat conventional machine learning approaches including Random Forest (83.7%), Support Vector Machine (86.4%), and Multi-Layer Perceptron (78.3%), as well as two state-of-the-art graph-based models designed specifically for multi-omics integration.
The F1-score of 0.897 and MCC of 0.838 confirm that MoAGNN's advantage is genuine and not driven by class imbalance. MCC (Matthews Correlation Coefficient) is particularly informative because it accounts for all four categories of correct and incorrect predictions, giving a single reliable score even when tumor subgroup sizes differ.
The transcriptome contributed most strongly to classification accuracy across omics combinations, while microRNA had the smallest individual impact. DNA methylation and copy number variation provided meaningful complementary information, confirming that the multi-omics integration strategy adds value beyond gene expression alone.
Subtype classification worked best because molecular subtypes are inherently gene-driven - the three LUAD subtypes are defined by their gene expression profiles, making gene-level features naturally informative. PCA analysis confirmed clear molecular separation between subtypes, explaining why the gene network approach excels at this task.
Stage classification proved more challenging for MoAGNN, with performance not exceeding that of baseline methods. The likely reason is biological: tumor stage is influenced by many factors beyond gene expression, including tumor size, smoking history, age, and other clinical variables that the model did not have access to.
The co-expression networks for stage prediction were notably sparser than for subtype classification, with fewer strong gene-gene connections. This structural weakness in the input graph limited the model's ability to extract useful representations for distinguishing early from late-stage disease.
For survival prediction, MoAGNN performed comparably to other approaches overall but achieved the highest precision (83.7%) among all models. High precision in identifying high-risk patients is particularly clinically valuable - it means that when the model flags a patient as high-risk, that prediction is more likely to be correct, reducing unnecessary interventions for lower-risk patients.
External validation using the independent GSE81089 dataset confirmed that the transcriptome-derived gene features and network structure could generalize beyond the TCGA training data. Stage prediction achieved accuracy of 77.5% on this completely independent cohort, demonstrating robustness of the molecular signatures identified by the model.
SAGPooling identified 74 key genes prioritized as most informative for LUAD subtype classification. Among these, several showed expression patterns that increase progressively from the favorable TRU subtype to the more aggressive PP subtype, tracking with disease progression.
FGB (fibrinogen beta chain) was markedly elevated in the PP subtype, consistent with its known role in remodeling the tumor microenvironment to favor progression and immune evasion. Higher FGB expression correlates with poorer performance status and more advanced stage in published studies. BMP6, while generally tumor-suppressive in lung cancer, showed PP-restricted elevation, possibly reflecting activation of a pathway (TGF-beta/BMP) associated with tumors becoming more invasive.
RSPO2 decreased progressively from TRU to PP, consistent with prior reports that its loss promotes cancer cells to undergo epithelial-to-mesenchymal transition (EMT) - a process that makes tumors more invasive and prone to spread. Two newly highlighted genes, AVPR1B and CD1E, also declined from TRU to PP, with CD1E's loss potentially contributing to an immune-cold microenvironment where the immune system cannot effectively attack the tumor.
DRP2 increased along the TRU-to-PP progression axis, with higher DRP2 associated with poorer survival and experimental evidence linking it to EMT-driven invasiveness, making it a potential therapeutic target. The model's identification of these biologically plausible gene patterns validates that its learned representations capture genuine cancer biology rather than statistical artifacts.
Gene Set Enrichment Analysis (GSEA) compared each LUAD subtype against normal lung tissue to identify which biological programs are activated or suppressed in each subtype. The analysis revealed distinctly different pathway landscapes across the three subtypes, validating that TRU, PI, and PP represent genuinely different disease programs.
The TRU subtype showed broad downregulation of metabolism and signaling pathways relative to normal tissue. Pathways related to cholesterol metabolism and fat breakdown signaling were notably suppressed, suggesting that TRU tumors maintain a relatively stable metabolic state close to normal lung biology - consistent with their more favorable clinical outcomes.
The PI subtype was dominated by upregulation of cell cycle and proliferation pathways, indicating that PI tumors are actively dividing faster than normal tissue. Cellular senescence pathways were also elevated, possibly reflecting stress-induced molecular brakes that partially restrain uncontrolled growth.
The PP subtype showed the most striking pattern: strong suppression of immune pathways including T-cell differentiation and immune recognition, combined with continued proliferative activity. This immunosuppressive profile helps explain why PP tumors are the most aggressive - they simultaneously evade immune attack and grow rapidly, creating a particularly challenging tumor microenvironment.
Kaplan-Meier survival analysis was performed on 431 LUAD patients for each of the 74 key genes identified by MoAGNN. Of these, 45 genes showed statistically significant associations with overall survival, providing a set of candidate prognostic biomarkers that could potentially guide clinical decisions.
RHOV (Ras Homolog Family Member V) emerged as the strongest indicator of poor prognosis. Patients with low RHOV expression had significantly better survival outcomes, suggesting RHOV acts as an oncogene - promoting cancer progression. This is consistent with published evidence linking RHOV to tumor cell growth, metastasis, and resistance to targeted therapies in lung cancer.
TSPAN11 and NR1I2 showed the opposite pattern - patients with higher expression of these genes had better survival. TSPAN11 (a member of the tetraspanin family) has proposed roles in immune modulation and cell-cell interactions that might help contain tumor growth. These favorable associations suggest TSPAN11 and NR1I2 may function as tumor suppressors or markers of less aggressive disease.
The identification of both risk and protective biomarkers from a single computational framework demonstrates MoAGNN's utility for translational research. Validated prognostic genes like these could eventually be incorporated into clinical diagnostic panels to refine risk stratification and guide treatment intensity for individual LUAD patients.
MoAGNN advances multi-omics cancer analysis by demonstrating that gene-centric graph neural networks with biological prior knowledge can outperform both conventional machine learning and existing graph-based approaches for tumor subtype classification. The hierarchical self-attention pooling mechanism provides interpretability - a critical feature for clinical translation.
The model's key limitation is its reliance on molecular data alone. Tumor stage and survival are influenced by many non-molecular factors including age, smoking history, treatment regimens, and post-diagnosis events. Future versions that incorporate clinical variables alongside multi-omics data are expected to substantially improve stage and survival prediction.
The biological coherence of the model's outputs - identifying genes with known and plausible roles in LUAD, finding pathway signatures that match established subtype biology, and discovering prognostically meaningful biomarkers - builds confidence that the computational approach is capturing real biology rather than overfitting to training data.
The open-source release of MoAGNN's code makes this framework available for the broader research community to apply and extend. The approach is general enough to be adapted for other cancer types where multi-omics data is available, potentially accelerating biomarker discovery and subtype characterization across oncology.