Gao Qian, Dragsted Lars O, Ebbels Timothy
Department of Nutrition, Exercise and Sports, University of Copenhagen, 1958 Frederiksberg, Denmark.
Computational and Systems Medicine, Department of Surgery and Cancer, Imperial College London, London SW7 2AZ, UK.
Metabolites. 2019 May 9;9(5):92. doi: 10.3390/metabo9050092.
Metabolomic studies with a time-series design are widely used for discovery and validation of biomarkers. In such studies, changes of metabolic profiles over time under different conditions (e.g., control and intervention) are compared, and metabolites responding differently between the conditions are identified as putative biomarkers. To incorporate time-series information into the variable (biomarker) selection in partial least squares regression (PLS) models, we created PLS models with different combinations of bilinear/trilinear and group/time response dummy . In total, five PLS models were evaluated on two real datasets, and also on simulated datasets with varying characteristics (number of subjects, number of variables, inter-individual variability, intra-individual variability and number of time points). Variables showing specific temporal patterns observed visually and determined statistically were labelled as discriminating variables. Bootstrapped-VIP scores were calculated for variable selection and the variable selection performance of five PLS models were assessed based on their capacity to correctly select the discriminating variables. The results showed that the bilinear PLS model with group × time response as dummy provided the highest recall (true positive rate) of 83-95% with high precision, independent of most characteristics of the datasets. Trilinear PLS models tend to select a small number of variables with high precision but relatively high false negative rate (lower power). They are also less affected by the noise compared to bilinear PLS models. In datasets with high inter-individual variability, bilinear PLS models tend to provide higher recall while trilinear models tend to provide higher precision. Overall, we recommend bilinear PLS with group x time response for variable selection applications in metabolomics intervention time series studies.
采用时间序列设计的代谢组学研究被广泛用于生物标志物的发现和验证。在此类研究中，会比较不同条件下（如对照和干预）代谢谱随时间的变化，并将在不同条件下有不同反应的代谢物鉴定为假定的生物标志物。为了将时间序列信息纳入偏最小二乘回归（PLS）模型的变量（生物标志物）选择中，我们创建了具有双线性/三线性和组/时间响应虚拟变量不同组合的PLS模型。总共在两个真实数据集以及具有不同特征（受试者数量、变量数量、个体间变异性、个体内变异性和时间点数）的模拟数据集上评估了五个PLS模型。视觉上观察到并经统计确定显示特定时间模式的变量被标记为区分变量。计算自展VIP分数用于变量选择，并根据五个PLS模型正确选择区分变量的能力评估其变量选择性能。结果表明，以组×时间响应作为虚拟变量的双线性PLS模型具有最高的召回率（真阳性率），为83 - 95%，且精度较高，与数据集的大多数特征无关。三线性PLS模型倾向于高精度地选择少量变量，但假阴性率相对较高（功效较低）。与双线性PLS模型相比，它们受噪声的影响也较小。在个体间变异性较高的数据集中，双线性PLS模型倾向于提供更高的召回率，而三线性模型倾向于提供更高的精度。总体而言，我们推荐在代谢组学干预时间序列研究的变量选择应用中使用具有组x时间响应的双线性PLS。
