Accounting for biological variation with linear mixed-effects modelling improves the quality of clinical metabolomics data

Abstract

Metabolite profiles from biological samples suffer from both technical variations and subject-specific variants. To improve the quality of metabolomics data, conventional data processing methods can be employed to remove technical variations. These methods do not consider sources of subject variation as separate factors from biological factors of interest. This can be a significant issue when performing quantitative metabolomics in clinical trials or screening for a potential biomarker in early-stage disease, because changes in metabolism or a desired-metabolite signal are small compared to the total metabolite signals. As a result, inter-individual variability can interfere subsequent statistical analyses. Here, we propose an additional data processing step using linear mixed-effects modelling to readjust an individual metabolite signal prior to multivariate analyses. Published clinical metabolomics data was used to demonstrate and evaluate the proposed method. We observed a substantial reduction in variation of each metabolite signal after model fitting. A comparison with other strategies showed that our proposed method contributed to improved classification accuracy, precision, sensitivity and specificity. Moreover, we highlight the importance of patient metadata as it contains rich information of subject characteristics, which can be used to model and normalize metabolite abundances. The proposed method is available as an R package lmm2met.

Metabolomic differences between critically Ill women and men

“…Importantly, by adjusting for the absolute change in 25(OH)D level at day 3, we mitigate the effect of the trial intervention on the observed sex-specific metabolomic changes which allow for study of the entire trial cohort increasing sample size and study power 60 , 61 . Further, our use of clinical trial data allows for modelling and normalization of metabolite abundance via adjustment for subject characteristics 62 . To account for multiple comparisons we utilized a conservative Bonferroni corrected P -value < 8.65 × 10 –5 63 .…”

Deep metabolome: Applications of deep learning in metabolomics

“…Literally, each chromatogram, consisting of mass spectra or fingerprint of the detected metabolite represents the abundance of an ionized molecule [42] . Raw data files are subjected to a series of data processing steps and information extraction into an expression matrix (containing retention times, accurate mass spectrum and intensity values) of the measured metabolites for subsequent analyses [43] .…”

“…However, even with data augmentation and multiple data sets integration, compared to other domains, metabolomics data still lacks the sheer number of samples used in standard machine learning applications [103] . Even some of the standard techniques used for model validation such as k -fold cross-validation [43] , [104] may not be applicable for HDLSS metabolomics data. For example, some studies cautioned that random data splitting techniques like k -fold cross-validation might yield overfit and unstable models [79] from HDLSS data.…”

Use of Large and Diverse Datasets for 1H NMR Serum Metabolic Profiling of Early Lactation Dairy Cows

“…Furthermore, when corrected spectra were used, y-values also had to be corrected, making interpretation of phenotypic values difficult. Interestingly, Wanichthanarak, et al [20] found that "readjusting" mass spectroscopy metabolite signals using patient metadata and linear mixed models improved the sensitivity and specificity of classification of human tissue samples with and without colorectal cancer. Conversely, Posma et al [41] found that adjusting NMR data for confounding factors lead to a loss of predictive power for cardiovascular risk in a large-scale human NMR metabolomic dataset.…”

“…Frequently used fixed effects include stage of lactation, parity, and herd-year-season. Similar approaches have recently been applied to metabolomic data, for example Wanichthanarak et al [20], who used linear mixed-effects models and patient metadata to account for biological variation in metabolomics data, and Laine et al [21], who used linear models to study the effect of pregnancy on mid-infrared spectral data derived from cows' milk.…”