GC-PGE: A novel deep learning model for tumor drug resistance prediction and core resistance gene extraction based on...

Drug resistance is the main cause of tumor treatment failure and recurrence. Existing models exhibit limitations in accurately predicting tumor drug resistance and identifying resistance-associated genes. Furthermore, they often fail to effectively integrate multidimensional omics data and pathway-level information, lacking biological interpretability.

We propose GC-PGE (Gene Correlation and Pathway Graph Encoder Network), a novel deep learning model to predict tumor drug resistance and identify core resistance genes. First, leveraging graph neural network, we constructed a gene correlation network incorporating protein interactions, gene homology, and signaling pathway data. We then integrated the tasks of drug-resistant gene prediction and tumor sample classification into a unified network architecture using a Bayesian learning approach, enabling mutual validation and enhancement between the tasks. Additionally, we developed a pathway graph encoder. This module employed graph structure to encode gene feature weights, simultaneously classifying tumor sample drug resistance and extracting core resistance genes along with their characteristic pathways.

GC-PGE significantly outperformed traditional models including linear discriminant analysis (LDA), support vector machine (SVM), Random forest (RF), and naive Bayes model (NBM) in predicting tumor drug resistance. It also demonstrated superior performance over baseline models including differential expression (DE), mutual information (MI), and random forest (RF) in identifying resistance-associated genes. Furthermore, GC-PGE elucidated key signaling pathways and core resistance genes associated with drug resistance in liver cancer, ovarian cancer, and melanoma, providing valuable mechanistic insights.

By integrating multidimensional biological data and signaling pathway information, GC-PGE provides an effective framework for predicting tumor drug resistance and pinpointing core resistance genes. The model not only enhances prediction accuracy but also improves interpretability, thus positioning it as a promising approach for advancing personalized cancer therapy and novel therapeutic target discovery.