PathoGAT: A Multi-scale Pathogenic Gene Prediction System Integrating Machine Learning Ensembles and Graph Attention Networks

PathoGAT is a multi-scale system for pathogenic gene prediction. Its core innovation lies in integrating five traditional machine learning models (random forest, gradient boosting tree, etc.) with Graph Attention Networks (GAT) to achieve integrated analysis of topological information from protein-protein interaction (PPI) networks and tabular genetic features. It provides high-precision consensus scores for pathogenic gene prediction and addresses the key issue that traditional methods struggle to capture functional associations in biological networks.

Pathogenic gene identification is a core component of precision medicine and genetic disease diagnosis. Traditional gene variant annotation relies on statistical indicators such as sequence conservation, making it difficult to capture functional associations of genes in biological networks. With the popularization of high-throughput sequencing, there is a lack of effective methods to determine the pathogenicity of a large number of rare variants. Although PPI networks provide new ideas (disease genes have specific topological features), traditional machine learning struggles to handle graph-structured data and easily loses topological information.

PathoGAT adopts a multi-scale fusion architecture, including two core paths: 1. Machine Learning Ensemble Module: Integrates five algorithms—random forest, gradient boosting tree, support vector machine, logistic regression, and naive Bayes—to process multi-dimensional tabular features such as gene expression profiles and functional enrichment; 2. Graph Attention Network (GAT) Module: Uses an attention mechanism to assign differentiated weights to neighbor nodes in the PPI network, learns node embedding representations, and captures network topological information.

PathoGAT fuses information through three scales: micro, meso, and macro. The micro scale evaluates the functional impact of individual gene variants; the meso scale captures the role of genes in local network modules via GAT attention weights; the macro scale reflects the systemic importance of genes in the global network. Information from these three scales is concatenated and weighted, then fed into the consensus scoring layer.

PathoGAT uses the PyTorch Geometric framework; it integrates authoritative databases such as STRING (PPI data), OMIM (disease-gene associations), ClinVar (variant annotations), and GTEx (expression profiles). For training, stratified sampling is used to balance positive and negative samples, five-fold cross-validation is employed to evaluate generalization performance, and strict gene-level partitioning is followed to avoid data leakage.

PathoGAT outperforms single methods on multiple benchmark datasets. Its application scenarios include: prioritization of candidate variants in rare disease diagnosis, drug target discovery (key nodes in disease modules), construction of polygenic risk scores, and generation of research hypotheses (revealing molecular mechanisms).

Currently, PathoGAT relies on static PPI networks and lacks consideration of tissue specificity and dynamic changes; the interpretability of the attention mechanism is limited. Future directions include: integrating single-cell transcriptome data to capture cell-type-specific networks, introducing temporal modeling to study network dynamics, and developing interactive visualization tools to explore biological patterns.