This study proposes a multi-omics integration model, EGWO-GAT, based on graph attention network (GAT) and enhanced-grey wolf optimizer (EGWO), for staging prediction of bladder urothelial carcinoma (BLCA) samples.

A total of 404 BLCA samples from The Cancer Genome Atlas (TCGA) were collected from the University of California, Santa Cruz Xena (UCSC Xena) functional genomics explorer, including mRNA, DNA methylation, and microRNA (miRNA) data. After separate preprocessing and differential analysis of the 3 omics datasets, node features and edge features were extracted. Using GAT as the backbone framework, EGWO was employed for hyperparameter optimization, and multilayer perceptron (MLP) was applied for cancer staging prediction. Through 5-fold cross-validation, the performance of EGWO-GAT was compared with classical machine learning models, along with omics contribution analysis and performance conparison of models retaining different numbers of similar edges. Accuracy, precision, recall, F1 score, and area under the curve (AUC) were used as core evaluation metrics.

Differential feature screening identified 534 differentially expressed genes in mRNA, 3 108 differential probes in DNA methylation, and 114 differential miRNAs. Model comparisons showed that EGWO-GAT achieved optimal performance when integrating all omics data and retaining edges from the top 3 similar patients, with an AUC value of 0.744, an accuracy of 0.711, a precision of 0.792, a recall of 0.782, and an F1 score of 0.785. Its performance significantly surpassed other classical methods and outperformed the GS-GAT model across all metrics. Omics contribution analysis confirmed that full integration (mRNA+DNA methylation+miRNA) outperformed all other combinations. Result of the similarity edge number performance comparison demonstrated that retaining the top 3 similar edges yielded the highest AUC, accuracy, precision, and F1 score compared to the top 5 or 7 edges.

The EGWO-GAT model exhibits excellent performance in BLCA staging, providing reliable technical support for precise diagnosis. It addresses clinical challenges caused by sample heterogeneity and holds significant potential for guiding individualized treatment and improving patient prognosis.