Prediction and evaluation of the effect of pre-centrifugation sample management on the measurable untargeted LC-MS plasma metabolome

Highlights: • The effect of pre-centrifugation management on plasma sample quality was comprehensively investigated. • Metabolites responded differently to pre-centrifugation management. • Metabolite profile was rapidly affected by pre-centrifugation delay at all temperatures. • Metabolite panels for pre-centrifugation management prediction were constructed using machine learning. Optimal handling is the most important means to ensure adequate sample quality. We aimed to investigate whether pre-centrifugation delay time and temperature could be accurately predicted and to what extent variability induced by pre-centrifugation management can be adjusted for. We used untargeted liquid chromatography-mass spectrometry (LC-MS) metabolomics to predict and evaluate the influence of pre-centrifugation temperature and delayed time on plasma samples. Pre-centrifugation temperature (4, 25 and 37 °C; classification rate 87%) and time (5–210 min; Q² = 0.82) were accurately predicted using Random Forest (RF). Metabolites uniquely reflecting temperature and temperature-time interactions were discovered using a combination of RF and generalized linear models. Time-related metabolite profiles suggested a perturbed stability of the metabolome at all temperatures in the investigated time period (5–210 min), and the variation at 4 °C was observed in particular before 90 min. Fourteen and eight metabolites were selected and validated for accurate prediction of pre-centrifugation temperature (classification rate 94%) and delay time (Q² = 0.90), respectively. In summary, the metabolite profile was rapidly affected by pre-centrifugation delay at all temperatures and thus the pre-centrifugation delay should be as short as possible for metabolomics analysis. The metabolite panels provided accurate predictions of pre-centrifugation delay time and temperature in healthy individuals in a separate validation sample. Such predictions could potentially be useful for assessing legacy samples where relevant metadata is lacking. However, validation in larger populations and different phenotypes, including disease states, is needed.