Metaboanalyst 5

Freeze-dried samples were ground to a powder under liquid nitrogen. Then the powder was extracted with dichloromethane for GC–MS analysis and methanol for LC–MS analysis. For more details, see the supplementary information. Raw data from both GC–MS and UPLC-QTOF-MS were processed with MS-DIAL software (v4.46) [78 (link)]. Briefly, raw MS data files were converted to ABF format. Then MS-DIAL was used for peak peaking, alignment, integration, and retention time correction according to optimized parameters (summarized in Supplementary Data Table S17). The resulting output data table of metabolites (i.e. peak areas for each RT-m/z pair in each sample) was subjected to further statistical analysis. The main volatile constituents were identified by comparing mass spectra and retention indices with reference standards, published literature, and the NIST library database. MS/MS spectra were compared with spectra from reference standards and from open databases, including METLIN, MassBank, ReSpect, GNPS, and BMDMS-NP, to identify non-volatile constituents. Multivariate analysis was performed using MetaboAnalyst 5.0, R and SIMICA 14.0 software (Umetrics), which provided the heat map, PCA, and orthogonal partial least squares discriminant analysis (OPLS-DA). The correlation heat map of genes and metabolites was made using the OmicStudio tools at https://www.omicstudio.cn/tool.

Two-way ANOVA was performed on all data by using SPSS Statistics 26.0 software (SPSS, Inc., Chicago, IL, USA). The significance of mean differences among applied treatments in soil, substrate and hydroponically grown lettuce was analyzed by Tukey’s HSD test (p < 0.05). All figures were created by using R software (R-4.1.2). The metabolites were exhaustively contrasted by adopting heat map and partial least squares discriminant analysis (PLS-DA) methods for the applied treatments in lettuce grown in soil, substrate and hydroponic systems. PCA and PLS-DA analyses were performed by SIMCA-14.1 (Umetrics, Umeå, Sweden) and the heatmap was generated in metaboAnalyst 5.0 (www.metaboanalyst.ca).

SPSS Statistics 27 (Umetrics Corporation; Umea, Sweden) was used for ANOVA analysis and significance analysis (p < 0.05), and the data were expressed as mean ± s. PCA and PLS-DA were performed on three samples using MetaboAnalyst 5.0 to analyze sample differences and look for differentially labeled compounds. Unscrambler X established a partial least squares regression (PLSR) model to determine associations between the VOCs and the sensory properties. The Wayne chart and heat map were drawn using TBtools, the E-nose data was collected and processed by built-in Winmuster software, and the radar map and correlation heat map were drawn using Origin 2021 software (Originlab Corporation, Northampton, MA, USA).

Multivariate statistics were performed using SIMCA-P+ (12.0.0.0, Umetrics, Umea, Sweden) and Metaboanalyst 5.0. Partial least squares discriminant analysis (PLS-DA) and orthogonal partial least squares discriminant analysis (OPLS-DA) were carried out to define significant differences in metabolites temporally, spatially, and between species using pairwise comparisons. Seven-fold cross-validation and 1,000 permutations were used to validate statistical models. Determination of the discriminating metabolites toward the clustering in PLS-DA models was further analyzed using regression coefficient plots with 95% jackknifed confident intervals, where metabolites with Variable Importance for Projection (VIP) values exceeding 1.0 were selected as the metabolite cut-off. Specificity and sensitivity for the putative biomarkers was tested with the use of the Receiver Operating Characteristic (ROC) curve method [56 (link), 57 (link)]. False discovery rate (FDR) test was employed to address the “multiple testing problem”. Note that the actual energy content of the liver (e.g., lipids) was not quantified.

Multivariate data analysis (MVDA) was carried out by SIMCA 14.1 (Umetrics, Umeå, Sweden) and MetaboAnalyst 5.0 (https://www.metaboanalyst.ca/; access date: 20 June 2020–31 July 2022) [25 (link)]. Calculation of the descriptive statistics such as the t-statistic values, the p values of unpaired t-test, and the multiple-sample receiver operating characteristic curve (MSROCC) modeling analysis and the area under curve (AUC) calculation were carried out using MetaboAnalyst 5.0 and GraphPad Prism 8.4.3 (GraphPad Software, San Diego, CA, USA). MSROCC model-based biomarker analyses and new sample classification were carried out using MetaboAnalysis 5.0. The accuracy, precision, sensitivity, specificity, false positive rate, and false negative rate of the lipid marker panels identified by different MVDA methods were calculated by the confusion matrix [26 (link)].

The sample data were extracted by using MassHunter Profinder software (Agilent, California, United States). The initial and final retention times were set for data collection. Data were normalized with MetaboAnalyst 5.0 and then the resultant data matrices were introduced to SIMCA-P 14.1 software (Umetrics, Umea, Sweden) for principal component analysis (PCA) and orthogonal-partial least squares discriminant analysis (OPLS-DA). These variables with VIP >1.5 and |p (corr)| ≥ 0.58 in the OPLS-DA method were considered to be further data analysis.

To establish the overall differences in the metabolic profiles between groups studied, multivariate statistical analysis, including unsupervised principal component analysis (PCA) and supervised partial least squares discriminant analysis (PLS-DA), were performed. MetaboAnalyst 5.0 (https://www.metaboanalyst.ca/) was employed for the statistical analysis of all the experiments, as well as SIMCA version 14.0 (Umetrics) software for confirming the results for the untargeted metabolomic analysis comparing a healthy versus a condition group. The quality of the established statistical models was evaluated using the R²X, R²Y, and Q² parameters. A permutation test was performed in 100 cycles to evaluate the possible overfitting in the PLS-DA models. PCA models were performed, including QC samples, to evaluate stability of the analytical system and the data quality, using as an indicator the clustering of QC samples in successive injections during the sequences. To visualize significant features up-regulated or down-regulated when two selected groups of interest were studied, univariate statistical tests were performed, and volcano plots were drawn by transforming the fold change (FC) value of each substance peak to log2 (FC) and transforming the P value (P = 0.05) of Student’s t-test to − log₁₀(P value).