NN-PCP: Screening phenotype-related core pathways to construct a prostate cancer metastasis prediction model based on multiple types of mutation data.

BACKGROUND AND

OBJECTIVE

Prostate cancer causes >400,000 deaths annually worldwide, being the fifth leading cause of cancer-related death in men. However, accurate prediction of prostate cancer metastasis based on mutations remains a challenge.

METHODS

We proposed a novel neural network constructed with phenotype-related core pathways, called NN-PCP, to accurately predict prostate cancer metastasis. NN-PCP comprises three main modules: IORA-driven modules, IGSEA-driven modules, and a hierarchy with two differential layers. IORA-driven modules are constructed using highly variable genes and core pathways that are gained with an improved over-representation analysis. They are applied to extract pathway features from each type of mutation data. IGSEA-driven modules are constructed using the phenotype-related core pathways that are obtained with an improved gene set enrichment analysis. They are used to transform pathway features from each type of mutation data into differential features. Finally, the hierarchy with two differential layers is employed to mimic the synergy effect of differential features from different types of mutation data to gain deeply differential features.

RESULTS

NN-PCP outperformed four state-of-the-art methods, with improvements of 5.7 %, 7.8 %, 1.2 %, 0.058, 0.031, and 0.047 in accuracy, precision, recall, F1-score, AUC, and AUPR, respectively. Furthermore, NN-PCP identified some important genes and pathways in prostate cancer metastatic prediction, such as PIK3R1, PIK3CA, and the PI3K-AKT signaling pathway.

CONCLUSIONS

The proposed method outperforms other pathway-based deep learning methods, which can be used as an auxiliary tool for doctors to improve the accuracy of disease diagnosis.