Optimal preprocessing of serum and urine metabolomic data fusion for staging prostate cancer through design of experiment

Abstract

Accurate classification of cancer stages will achieve precision treatment for cancer. Metabolomics presents biological phenotypes at the metabolite level and holds a great potential for cancer classification. Since metabolomic data can be obtained from different samples or analytical techniques, data fusion has been applied to improve classification accuracy. Data preprocessing is an essential step during metabolomic data analysis. Therefore, we developed an innovative optimization method to select a proper data preprocessing strategy for metabolomic data fusion using a design of experiment approach for improving the classification of prostate cancer (PCa) stages. In this study, urine and serum samples were collected from participants at five phases of PCa and analyzed using a 1H NMR-based metabolomic approach. Partial least squares-discriminant analysis (PLS-DA) was used as a classification model and its performance was assessed by goodness of fit (R2) and predictive ability (Q2). Results show that data preprocessing significantly affect classification performance and depends on data properties. Using the fused metabolomic data from urine and serum, PLS-DA model with the optimal data preprocessing (R2 = 0.729, Q2 = 0.504, P < 0.0001) can effectively improve model performance and achieve a better classification result for PCa stages as compared with that without data preprocessing (R2 = 0.139, Q2 = 0.006, P = 0.450). Therefore, we propose that metabolomic data fusion integrated with an optimal data preprocessing strategy can significantly improve the classification of cancer stages for precision treatment.