Combining the advantages of multilevel and orthogonal partial least squares data analysis for longitudinal...

Kidney transplantation is one of the renal replacement options in patients suffering from end-stage renal disease (ESRD). After a transplant, patient follow-up is essential and is mostly based on immunosuppressive drug levels control, creatinine measurement and kidney biopsy in case of a rejection suspicion. The extensive analysis of metabolite levels offered by metabolomics might improve patient monitoring, help in the surveillance of the restoration of a "normal" renal function and possibly also predict rejection. The longitudinal follow-up of those patients with repeated measurements is useful to understand changes and decide whether an intervention is necessary. The time modality, therefore, constitutes a specific dimension in the data structure, requiring dedicated consideration for proper statistical analysis. The handling of specific data structures in metabolomics has received strong interest in recent years. In this work, we demonstrated the recently developed ANOVA multiblock OPLS (AMOPLS) to efficiently analyse longitudinal metabolomic data by considering the intrinsic experimental design. Indeed, AMOPLS combines the advantages of multilevel approaches and OPLS by separating between and within individual variations using dedicated predictive components, while removing most uncorrelated variations in the orthogonal component(s), thus facilitating interpretation. This modelling approach was applied to a clinical cohort study aiming to evaluate the impact of kidney transplantation over time on the plasma metabolic profile of graft patients and donor volunteers. A dataset of 266 plasma metabolites was identified using an LC-MS multiplatform analytical setup. Two separate AMOPLS models were computed: one for the recipient group and one for the donor group. The results highlighted the benefits of transplantation for recipients and the relatively low impacts on blood metabolites of donor volunteers.