Novel Graph Neural Network Method for Differential Network Analysis in Biological Data

Network-based approaches are increasingly applied to the analysis of omics datasets, where nodes represent genes or proteins and edges encode their interactions. Differential Network Analysis (DINA) enables the comparison of connectivity patterns across conditions, revealing topological changes that may underlie disease mechanisms or treatment response [3]. Several studies have shown that variables such as sex, age, or treatment can significantly affect network structure [1]. However, traditional DINA methods often rely on restrictive statistical assumptions and are not well-suited to sparse and noisy RNA-seq data [4]. Bayesian and non-parametric approaches, such as CONGA [5], partially address these limitations but remain computationally demanding and difficult to scale. To overcome these challenges, we introduce Athena (Algorithm for Topological Heterogeneity Extraction in Differential Networks via Graph Neural Architectures), a novel framework that leverages Graph Neural Networks (GNNs) to construct differential molecular networks. Athena compares group-specific Pearson correlation matrices (e.g., males vs. females) to derive a differential adjacency matrix. Each gene is annotated with biologically informative features, including group-specific means, variances, log2 fold-change, and Welch's t-test p-values. These features are standardized and encoded in a PyTorch Geometric graph object, ensuring that both molecular and topological information contribute to the learning process. The model is based on a GraphSAGE encoder [2] with two SAGEConv layers, dropout, and ReLU activations. A multi-layer perceptron (MLP) decoder predicts the probability of differential edges using concatenated node embeddings. The model is trained on positive and negative edge samples using binary cross-entropy loss, optimized with Adam, and evaluated through AUC, F1-score, precision, and recall.