GTLSTMEP: A Novel Model Based on Graph Transformer and bi-LSTM for Predicting Essential Proteins in Sampled Subgraphs

Predicting essential proteins is crucial for identifying disease-causing genes and advancing drug discovery. However, existing methods face challenges, such as insufficient capture of biological properties, class imbalance, incomplete representation of network topology, and inadequate utilization of edge features. To address these limitations, we propose a novel model, GTLSTMEP, which combines Graph Transformer (GT) and Bidirectional Long Short-Term Memory (bi-LSTM). GTLSTMEP leverages biological data-including subcellular localization, protein complexes, and gene expression profiles-to initialize node and edge features. A subgraph batch sampling approach is employed to handle class imbalance. The bi-LSTM module processes time-series gene expression data, while the GT module captures global topological features using Laplacian positional encodings and local topological features through an attention mechanism that automatically identifies neighboring nodes' characteristics. Additionally, edge features are integrated to enhance the learning of protein interactions. The node features extracted from these two modules are then fed into a Multi-Layer Perceptron (MLP) for essential protein prediction. Experimental results demonstrate that GTLSTMEP outperforms all comparative methods. For instance, on the DIP dataset, GTLSTMEP achieves an $AUC$ that is 17.64% higher than the next-best model. Ablation studies further confirm the effectiveness of each component of GTLSTMEP.