Optimized graph neural network-multilayer perceptron fusion classifier for metastatic prostate cancer detection in Western and Asian populations

Abstract

Objective

Prostate cancer (PCa) exhibits significant genomic differences between Western and Asian populations. This study aimed to design a predictive model applicable across diverse populations while selecting a limited set of genes suitable for clinical implementation.

Methods

We utilized an integrated dataset of 1360 whole-exome and whole-genome sequences from Chinese and Western PCa cohorts to develop and evaluate the model. External validation was conducted using an independent cohort of patients. A graph neural network architecture, termed the pathway-aware multi-layered hierarchical network-Western and Asian (P-NETwa), was developed and trained on combined genomic profiles from Chinese and Western cohorts. The model employed a multilayer perceptron (MLP) to identify key signature genes from multi-omics data, enabling precise prediction of PCa metastasis.

Results

The model achieved an accuracy of 0.87 and an F1-score of 0.85 on Western population datasets. The application of integrated Chinese and Western population data improved the accuracy to 0.88, achieving an F1-score of 0.75. The analysis identified 18 signature genes implicated in PCa progression, including established markers (AR and TP53) and novel candidates (MUC16, MUC4, and ASB12). For clinical adoption, the model was optimized for commercially available gene panels while maintaining high classification accuracy. Additionally, a user-friendly web interface was developed to facilitate real-time prediction of primary versus metastatic status using the pre-trained P-NETwa-MLP model.

Conclusion

The P-NETwa-MLP model integrates a query system that allows for efficient retrieval of prediction outcomes and associated genomic signatures via sample ID, enhancing its potential for seamless integration into clinical workflows.