LUAD-LUSC Tumor Classification: A Bioinformatics Practice Combining Graph Neural Networks with Clinical Features

This project conducts lung cancer subtype (LUAD/LUSC) classification research based on TCGA data. Key innovations include multi-omics data fusion, graph neural network (GNN) modeling of gene relationships, and integration of clinical features to enhance classification performance, providing a reference implementation for AI applications in precision medicine.

Lung cancer is one of the malignant tumors with the highest incidence and mortality rates globally. LUAD and LUSC are the most common subtypes, with significant differences in pathogenesis, treatment, and prognosis. Accurate differentiation is crucial for personalized treatment. Traditional pathology relies on empirical judgment, while modern bioinformatics attempts to extract features from genomic data to build automated models. This project presents a complete practical case combining GNN and clinical features.

Data sources are public TCGA data, including CNV, RNA expression, methylation, and clinical data of over 700 patients. The preprocessing pipeline includes: 1. File extraction and mapping (integrate scattered data and establish patient mapping); 2. STRING database integration (download protein-protein interaction data to build gene relationship networks); 3. Methylation data preprocessing (map probes to genes using Illumina 450K chip manifest); 4. Clinical feature encoding (convert categorical variables to numerical form).

Construct a personalized gene relationship graph for each patient: nodes are genes/proteins (5-dimensional features from multi-omics data), edges are protein-protein interactions (weights based on STRING confidence, 3-dimensional features). The model architectures include three types: GAT (pure GNN, using only genetic data), MLP (baseline model, using only clinical features), and MultiModalGNN (core model, fusing graph data and clinical features). Key hyperparameters include num_node_features=5, clinical_input_dim=53, etc.

Data partitioning uses a training/validation/test strategy to ensure class balance. The training script graph_classification.py implements graph data loading, model initialization, training loop (cross-entropy loss + Adam optimizer), and early stopping mechanism. Evaluation metrics include classification accuracy, AUC-ROC, and clinical feature contribution (comparing performance differences with and without clinical features).

Clinical significance: Guide treatment plan selection (LUAD is sensitive to targeted therapy, LUSC relies on chemotherapy/immunotherapy), prognosis assessment, and clinical trial stratification. Limitations: Limited sample size, class imbalance, lack of external validation, and insufficient interpretability of GNN. Improvement directions: Transfer learning, attention visualization, multi-center validation, and expansion of survival prediction tasks.

Educational value: Suitable for learning bioinformatics data processing, GNN applications, multi-modal learning, and end-to-end project practice. Entry-level recommendations: First understand the clinical differences between LUAD/LUSC → explore TCGA data format → run preprocessing scripts → study graph construction logic → modify model architecture to observe performance changes.