Plasma Metabolomics Does not Improve Prostate Biopsy Prediction Over the Clinical Variables
DOI:
https://doi.org/10.26821/ijshre.14.09.2026.140902Keywords:
Plasma, Metabolomics, Prostate BiopsyAbstract
Prostate cancer detection begins with the PSA blood test, but PSA is prostate specific and not cancer specific, so many men with a raised value are sent to biopsy and turn out benign. This problem is worst in the gray zone, a PSA between 4 and 10 ng/mL, where the test carries little information. Plasma metabolomics has been proposed as a richer signal, and a recent deep learning model on this same cohort reported an AUC of 0.89, which is far above what simple methods reach. In this study I tested whether a Graph Attention Network that connects metabolites through KEGG pathways can predict biopsy outcome better than flat feature models and better than the routine clinical variables that are already available, with a focus on the gray zone where the decision actually matters. I used the Early Detection Research Network cohort of 580 men and 1,169 metabolites. Every model was evaluated with repeated five-fold cross-validation, and all data-dependent steps were fit inside the training folds only. The Graph Attention Network reached an AUROC of 0.559, which was not better than XGBoost at 0.556 and not better than a logistic regression given only age and PSA at 0.563. Removing the graph and keeping only self-loops did not lower performance and slightly improved it. The KEGG graph that the hypothesis proposed was the worst of all the graph variants. The metabolites added nothing over age and PSA. The clinical variables alone reached an AUROC of 0.72 for any cancer, 0.81 for clinically significant cancer, and 0.70 inside the gray zone, while metabolomics stayed at chance in the gray zone. I also showed that the gap with the published 0.89 is largely explained by feature selection performed before cross validation rather than inside it, which inflated the AUROC by 0.14 to 0.20 in my own data. No single metabolite survived multiple testing correction. The conclusion is a clean negative result. Plasma metabolomics does not help predict biopsy outcome in this cohort, and the usable signal is in cheap clinical data.
References
Amante, E., Cerrato, A., Alladio, E., Capriotti, A. L., Cavaliere, C., Marini, F., et al. (2022). Comprehensive biomarker profiles and chemometric filtering of urinary metabolomics for effective discrimination of prostate carcinoma from benign hyperplasia. Scientific Reports, 12(1), 4361. https://doi.org/10.1038/s41598-022-08435
Ambroise, C., and McLachlan, G. J. (2002). Selection bias in gene extraction on the basis of microarray gene-expression data. Proceedings of the National Academy of Sciences, 99(10), 6562–6566. https://doi.org/10.1073/pnas.102102699
American Cancer Society. (2026). Cancer Facts and Figures 2026. Atlanta: American Cancer Society.
Benedetti, E., Chetnik, K., Flynn, T., Barbieri, C. E., Scherr, D. S., Loda, M., and Krumsiek, J. (2023). Plasma metabolomics profiling of 580 patients from an Early Detection Research Network prostate cancer cohort. Scientific Data, 10, 830. https://doi.org/10.1038/s41597-023-02750-7
Chen, T., and Guestrin, C. (2016). XGBoost: a scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 785–794. https://doi.org/10.1145/2939672.2939785
Dieterle, F., Ross, A., Schlotterbeck, G., and Senn, H. (2006). Probabilistic quotient normalization as robust method to account for dilution of complex biological mixtures. Analytical Chemistry, 78(13), 4281–4290. https://doi.org/10.1021/ac051632c
Elmarakeby, H. A., Hwang, J., Arafeh, R., Crowdis, J., Gang, S., Liu, D., et al. (2021). Biologically informed deep neural network for prostate cancer discovery. Nature, 598, 348–352. https://doi.org/10.1038/s41586-021-03922-4
Huang, J., Mondul, A. M., Weinstein, S. J., Karoly, E. D., Sampson, J. N., and Albanes, D. (2022). Metabolomic profile of prostate cancer-specific survival among 1812 Finnish men. BMC Medicine, 20, 362. https://doi.org/10.1186/s12916-022-02561-4
Kanehisa, M., and Goto, S. (2000). KEGG: Kyoto Encyclopedia of Genes and Genomes. Nucleic Acids Research, 28(1), 27–30. https://doi.org/10.1093/nar/28.1.27
Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., and Liu, T. Y. (2017). LightGBM: a highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems, 30, 3146–3154.
Kipf, T. N., and Welling, M. (2017). Semi-supervised classification with graph convolutional networks. International Conference on Learning Representations. https://arxiv.org/abs/1609.02907
Lee, S., Lim, S., Lee, T., Sung, I., and Kim, S. (2020). Cancer subtype classification and modeling by pathway attention and propagation. Bioinformatics, 36(12), 3818–3824. https://doi.org/10.1093/bioinformatics/btaa203
Lima, A. R., Pinto, J., Amaro, F., Bastos, M. L., Carvalho, M., and Guedes de Pinho, P. (2021). Advances and perspectives in prostate cancer biomarker discovery in the last 5 years through tissue and urine metabolomics. Metabolites, 11(3), 181. https://doi.org/10.3390/metabo11030181
Lundberg, S. M., and Lee, S. I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1705.07874
Nadeau, C., and Bengio, Y. (2003). Inference for the generalization error. Machine Learning, 52(3), 239–281. https://doi.org/10.1023/A:1024068626366
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., et al. (2011). Scikit-learn: machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830.
Penney, K. L., Tyekucheva, S., Rosenthal, J., El Fandy, H., Carelli, R., Borgstein, S., et al. (2021). Metabolomics of prostate cancer Gleason score in tumor tissue and serum. Molecular Cancer Research, 19(3), 475–484. https://doi.org/10.1158/1541-7786.MCR-20-0548
Schmidt, J. A., Fensom, G. K., Rinaldi, S., Scalbert, A., Appleby, P. N., Achaintre, D., et al. (2020). Patterns in metabolite profile are associated with risk of more aggressive prostate cancer. International Journal of Cancer, 146(3), 720–730. https://doi.org/10.1002/ijc.32314
Sreekumar, A., Poisson, L. M., Rajendiran, T. M., Khan, A. P., Cao, Q., Yu, J., et al. (2009). Metabolomic profiles delineate potential role for sarcosine in prostate cancer progression. Nature, 457(7231), 910–914. https://doi.org/10.1038/nature07762
Sud, M., Fahy, E., Cotter, D., Azam, K., Vadivelu, I., Burant, C., et al. (2016). Metabolomics Workbench: an international repository for metabolomics data and metadata. Nucleic Acids Research, 44(D1), D463–D470. https://doi.org/10.1093/nar/gkv1042
Sun, L., Fan, X., Zhao, Y., Zhang, Q., and Jiang, M. (2024). Deep learning based metabolomics data study of prostate cancer. BMC Bioinformatics, 25, 391. https://doi.org/10.1186/s12859-024-06016-w
Sun, W. (2025). The research progress on diagnostic indicators related to prostate specific antigen gray zone prostate cancer. BMC Cancer, 25, 1264. https://doi.org/10.1186/s12885-025-14505-1
Tang, C., Guo, X., Zou, Q., Wang, J., Tang, Y., Tang, H., et al. (2025). Untargeted metabolomics revealed urinary metabolic pattern for discriminating prostate cancer from benign prostatic hyperplasia in Chinese participants. Frontiers in Oncology, 15, 1604169. https://doi.org/10.3389/fonc.2025.1604169
Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., and Bengio, Y. (2018). Graph attention networks. International Conference on Learning Representations. https://arxiv.org/abs/1710.10903
Yang, B., Zhang, C., Cheng, S., Li, G., Griebel, J., and Neuhaus, J. (2021). Novel metabolic signatures of prostate cancer revealed by 1H-NMR metabolomics of urine. Diagnostics, 11(2), 149. https://doi.org/10.3390/diagnostics11020149
