A statistical framework to model the meeting-in-the-middle principle using metabolomic data: application to hepatocellular carcinoma in the EPIC study

Abstract

Metabolomics is a potentially powerful tool for identification of biomarkers associated with lifestyle exposures and risk of various diseases. This is the rationale of the 'meeting-in-the-middle' concept, for which an analytical framework was developed in this study. In a nested case-control study on hepatocellular carcinoma (HCC) within the European Prospective Investigation into Cancer and nutrition (EPIC), serum (1)H nuclear magnetic resonance (NMR) spectra (800 MHz) were acquired for 114 cases and 222 matched controls. Through partial least square (PLS) analysis, 21 lifestyle variables (the 'predictors', including information on diet, anthropometry and clinical characteristics) were linked to a set of 285 metabolic variables (the 'responses'). The three resulting scores were related to HCC risk by means of conditional logistic regressions. The first PLS factor was not associated with HCC risk. The second PLS metabolomic factor was positively associated with tyrosine and glucose, and was related to a significantly increased HCC risk with OR = 1.11 (95% CI: 1.02, 1.22, P = 0.02) for a 1SD change in the responses score, and a similar association was found for the corresponding lifestyle component of the factor. The third PLS lifestyle factor was associated with lifetime alcohol consumption, hepatitis and smoking, and had negative loadings on vegetables intake. Its metabolomic counterpart displayed positive loadings on ethanol, glutamate and phenylalanine. These factors were positively and statistically significantly associated with HCC risk, with 1.37 (1.05, 1.79, P = 0.02) and 1.22 (1.04, 1.44, P = 0.01), respectively. Evidence of mediation was found in both the second and third PLS factors, where the metabolomic signals mediated the relation between the lifestyle component and HCC outcome. This study devised a way to bridge lifestyle variables to HCC risk through NMR metabolomics data. This implementation of the 'meeting-in-the-middle' approach finds natural applications in settings characterised by high-dimensional data, increasingly frequent in the omics generation.

Figure 1.

General scheme of the analytical framework developed in the study. A PC-PR2 analysis is carried out beforehand to identify relevant sources of variation. In the PLS model, the X- and Y-sets are related to each other, and scores are computed (1). X- and Y-scores are, in turn, associated to a case–control indicator of HCC status in conditional logistic regression models (2). A mediation analysis is carried out to explore the role of metabolomics in the association between lifestyle factors and risk of HCC (3).

Figure 2.

PC-PR2 analysis results* identifying the sources of variability in the NMR data (A) and in the lifestyle data (B). * 17 and 14 components were retained to account for 80% (threshold used) of total NMR (A) and lifestyle variability (B), respectively. The R2 value represents the amount of variability in NMR/lifestyle variable explained by the ensemble of investigated predictors.

Figure 3.

Variable importance plot (VIP) displaying the variable importance for projection statistic of the predictor variables for the PLS analyses. (A) Results from the main PLS model run on all observations (N = 336, X-set = 21, Y-set = 285). (B) Results from the PLS sensitivity analysis run on a subsample (N = 271, 92 cases, 179 controls) excluding sets where cases were diagnosed within the first 2 years of follow-up (X-set = 21, Y-set = 285). The horizontal line corresponds to Wold’s criterion (0.8), the threshold used to rule if a variable has an important contribution to the construction of the Y variables (see Supplementary Appendix, available at Mutagenesis Online for further details).