Comparison of HPLC and NMR for quantification of the main volatile fatty acids in rumen digesta

To reveal the real metabolomic changes caused by specific biological events, it is critical to use a suitable analytical pipeline that can accurately and precisely detect the true concentration differences of individual metabolites. Metabolomic analysis entails several steps including pre-analytical work (i.e., biofluid sample collection and storage), experimental work (i.e., sample analysis) and data analysis (i.e., pre-process, quantification, and pre-treatment)³⁰. Concentrations of rumen VFA measured using HPLC and NMR were compared for these samples, and the reliability of results was assessed using the intraclass correlation coefficient (ICC). Then we summarized the analytical workflows of HPLC (Fig. 3) and NMR (Fig. 4). Based on the results of this research in the pre-processing and quantification of NMR spectroscopy, some practical suggestions are given. PCA, PLS-DA and correlation analysis were used to compare the concentrations collected by the two methods after different pre-treatment measures. Relevant suggestions were provided consequently.

Comparison of the quantitative rumen VFA data determined by HPLC and NMR

Table 1 summarizes the VFA concentration range from literature using both methods as a basis for comparison with current values. The concentration of volatile fatty acids in the rumen will vary a lot depending on the diets and individuals. For HPLC and NMR, there are potential issues about which units the VFA are reported as—concentrations or molar proportion (Fig. S2). Molar proportions of VFA in the rumen are commonly reported because they are more reliable to represent either VFA production, absorption, or both³⁴. The balance of generation and removal or interconversion from the pool determine the concentration of each VFA in the rumen. Furthermore, concentrations might fluctuate depending on the time and site of sampling. VFA can be diluted in the rumen by water and saliva without changing their relative proportions³⁴. It has also to be considered that in the experiment, rumen fluid was always diluted into differing liquid amounts; VFA mol or VFA molar percentages are unaffected by digesta liquid amount. Hall et al. have reported that it is more biologically meaningful when ruminal data were evaluated as moles of VFA (VFA/mol) than concentrations (mM)³⁵.

Evaluation of the consistency of HPLC and NMR data

ICC is widely used in inter-rater reliability analyses to test whether different raters have the consistent evaluation for the same subject. An ICC score in the range of 0.50 to 0.71 represent moderate agreement and from 0.71 to 0.90 represent strong agreement³⁶. In the ICC results, HPLC and NMR showed strong agreements for propionate, butyrate, isobutyrate and isovalerate. There was a moderate degree of agreement for acetate and valerate (Table 2).

Comparison of HPLC and NMR results based on the molar proportions of VFA

To compare the data intuitively between HPLC and NMR, the concentrations of VFA were transformed into molar proportions. The molar proportion derived from HPLC and NMR was based on the computation of the relative concentration of each VFA as shown in Fig. 2. Figure 2 represents the individual measurements of VFA (in molar proportions) determined in rumen samples from 33 animals using both HPLC and NMR. Generally, the individual measurements over all samples indicate that the HPLC and NMR data between samples tend to be consistent. Table 3 shows the correlations between VFA (in molar proportion) using HPLC and NMR. The two sets of data (using HPLC and NMR methods) were highly correlated among VFA.

Comparison of HPLC and NMR experimental procedures

The differences between the HPLC and NMR experimental procedures for obtaining quantitative concentrations are discussed in this session. The experimental procedures of HPLC are relatively simple mainly including sample preparation, selection of chromatographic conditions (selection of chromatographic column, determination of mobile phase, flow rate, detector, and column temperature, etc.), standard curve preparation, sample determination, sample concentration calculation, and recovery experiment as illustrated in Fig. 3. This differential washout or elution of compounds is the basis for the HPLC separation¹⁰. Although HPLC is simple to set up and operate, the operator must have a thorough understanding of the system, its columns, and the chemistry of the compounds being separated to achieve optimal separation.

Most metabolites have unique NMR chemical fingerprints consisting of numerous signals³⁷. By comparing the integral of the reference peak, which reflects the know concentration of a standard, to the integrals of the sample peaks, it is possible to quantify the metabolites³⁷. Careful preparation of spectra for quantitative analysis consists of several steps, including phasing of the spectra, baseline correction and potentially alignment of the signals (Fig. 4). Metabolites are identified and quantified by comparing the bio-NMR sample’s spectrum to a set of legitimate standards or a spectral reference library derived from authentic standards. This can involve the acquisition of additional (2D) NMR experiments to identify individual metabolites or comparison with metabolites contained in the Bovine Metabolome Database (BMDB)^33,38, where currently the details of more than 200 metabolites found in bovine ruminal fluids can be found. An alternative approach to the quantitative analysis of NMR data is an untargeted approach using chemometric analysis (Fig. 4). Chemometric profiling is fundamentally different from quantitative metabolomics (or targeted metabolic profiling) (Fig. 4). Only the spectral patterns and intensities of the chemicals are recorded, statistically compared, and utilised to determine the key spectral characteristics that separate sample classes.

Summary of NMR spectrum pre-processing and quantitative practice

The use of sophisticated curve-fitting software and specifically created databases of NMR spectra of pure metabolites obtained at appropriate pH levels and spectrometer frequencies is required for spectrum analysis (100–900 MHz) (Fig. 4). In this study, we used Chenomx software to perform spectral pre-processing and subsequent quantitative analysis (targeted metabolomics). The spectrum pre-processing steps and order in this study follow Chenomx’s tutorial as shown in Fig. 4. When using Chenomx components to quantify compounds, we found that in the pre-processing steps of spectroscopy, as stated in the official tutorial³⁹, phase correction and baseline correction are essential steps, and largely determine the accuracy of quantitative concentration. However, it is worth noting that line broadening and shim correction can make the curve smooth, but if there is no strong evidence that the above steps should be taken, they should be used with caution. Because it was found in the study that the above two steps are likely to significantly reduce the quantified concentration of the overall metabolites. Another point worth emphasizing is as stated in the official tutorial³⁹, when determining the true concentrations of compounds with multiple ¹H signals, some are more reliable in determining the true concentration of the compound than others. Less overlap, undistorted, and more intense signals, representing more protons that are split by fewer J couplings, are all signs of a more reliable signal. This study identified signals that are more reliable in the quantification of VFA compounds with multiple ¹H signals. These are referred to as “handle signals” and presented in Table 4. Quantitative analysis of VFA signals in ¹H NMR spectra was performed by Chenomx. The ¹H signals of most compounds such as isobutyrate, isovalerate and valerate, overlap, which makes their quantification difficult. In reference to Table 4, spectra of individual compounds (Fig. S1) are colour coded and triangles indicate their individual ¹H signals.

Comparison of HPLC and NMR quantitative data analysis

The purpose of data pre-treatment is to reduce systematic variation and to separate biological variation from variation introduced during the quantification of metabolites to improve the performance of the downstream statistical analysis. Experimental heterogeneity, such as sample inhomogeneity, variations in sample preparation and ion suppression, account for the unexplained variance in data⁴⁰. Data normalisation is used to eliminate systematic bias within a data set and increase overall data accuracy to make valid biological comparisons⁴¹. In this section, we discuss several methods recently reported in the literature and test their suitability for VFA analysis.

Summary of HPLC and NMR data pre-treatment and analysis method

Data pre-treatment is a crucial ‘key’ step in the metabolomics analysis pipeline, among all the other stages listed above⁴². Figure 5 illustrated several data pre-treatment, downstream analysis and popular functional platforms recommended in the literature²¹. There are three types of normalization: sample/variable normalization, data transformation and data scaling.

1. 1.

   The normalisation phase is done to each sample’s data and consists of methods for making data from all samples directly comparable. One of its most popular applications is to eliminate or reduce the impacts of changing sample dilution⁴¹. Dilution is described as a process in which the concentrations of all metabolites, and therefore all peak intensities of the associated spectrum, are influenced by the same factor (coefficient), also known as unspecific metabolite alterations. There were three methods of normalising: row-wise, column-wise, and combine normalisation. Normalization by row tries to make each sample (row) comparable to the others (i.e., samples with different dilution effects) (Figs. S3 and S4). Examples include normalization by sum, normalization by the median, and sample-specific normalization.

2. 2.

   When the data distribution is skewed or asymmetric, it will bring limitations and challenges to the application of statistical analysis. In this scenario, a suitable transformation might be required to transform variable distribution close to the normal or Gaussian distribution⁴³. Transformation is also applied to correct for heteroscedasticity. There are two widely used data transformation methods: log transformation and cube-root transformation. Among them, log transformation is the more often used method⁴⁴.

3. 3.

   Many metabolomics data characteristics (such as chemical concentration or ion abundance) have a large dynamic range. The ranges of variables can be vastly diverse, causing modelling and interpretation issues. The goal of data scaling is to reduce the fold differences between metabolites of various concentration levels so that they may be compared⁴⁵. When the variables are of significantly different orders of magnitude, this approach is beneficial (some metabolites are at micromolar levels while other metabolites are at millimolar levels). Auto-scaling, Pareto scaling, and range scaling are all techniques of scaling (column-wise operation). The standard deviation is used in auto-scaling, whereas the square root of the standard deviation is used in Pareto scaling. For correlation analysis, autoscaling is used more commonly, and the data processed using Pareto scaling is more similar to the original data structure⁴⁰. Column-wise technique, in contrast to row-wise process, aims to make each variable (column) equivalent to the others (Figs. S3 and S4)^46,47

Biological interpretability analysis of HPLC and NMR

In this section, VFA data derived from HPLC and NMR after applying different data pre-treatment were compared and analysed. The two sets of data are expected to show the same biological interpretability in the correlation, PCA analysis and PLS-DA analysis results.

The overall sample quantity or metabolite concentration might change considerably among samples when using NMR-based values (Fig. 6). It’s critical to minimise or eliminate the impact of total sample variance on individual metabolite measurement. However, we must consider the degree to which they are affected in various analyses, as well as the extent to which data pre-treatment will mitigate these effects. The experimentally obtained HPLC data are in the type of molar proportion, which means the total VFA content are all approximately equal to 1. In this case, there is no comparison of the total rumen volatile fatty acids.

Pearson correlation analysis

Pearson correlation is one of the basic measures for the downstream analysis (i.e., network analysis), so this study investigated the impact of different data pre-treatment methods on the results of correlation analysis. For correlation analysis, HPLC data showed results consistent with biological knowledge, that is, acetate and butyrate were positively correlated, and propionate was negatively correlated, and the hierarchical clustering results also showed that acetate and butyrate were classified into one category based on correlation (Fig. 7). The data processing of scaling (MC, AS, PS) and NS did not have any effect on the results of the correlation matrix (Fig. 7).

For correlation analysis, the batch effects of NMR technical artefacts together with effects from standardization may exhibit inflated variation between samples resulted in large positive correlations⁴⁶. As illustrated in Fig. 7, NMR data under pre-treatments including None, MC, AS, PS, and LT showed unusual positive correlations. This is also inconsistent with biological knowledge, for example, propionate and acetate are usually negatively correlated. In a previous study based on NMR data, the large positive correlations in the non-normalized data have also been observed⁴². In general, both negative and positive correlations should be present, rather than such large positive correlations of NMR data which obviously cannot reflect reality. The problem might be caused by batch effects or the calibration standard. The requirement for metabolomics data normalisation is critical in this case.

In the research results, NS and Combine pre-treatment the NMR data showed positive and negative correlations consistent with HPLC, and the hierarchical clustering results also divided acetate and butyrate into one category, and other VFA into another category, which also agrees with the biological knowledge (Fig. 7). To eliminate variance caused by causes other than homeostatic changes, normalisation is required. Both the HPLC and NMR data showed improved concordance after normalisation, and there were no significant discrepancies in calculated correlations. The inclusion of a random sample effect in the normalisation is most likely to eliminate the impact of the calibration standard. It was found that after scaling (MC, AS, PS) was applied, the correlation matrix of HPLC and NMR did not change. Log conversion adjusted correlation coefficients but did not eliminate the large positive correlations in the NMR data. Positive correlations were stronger in the standardized data, which is concerning.

PCA analysis

For PCA analysis, the first two principal components were selected to present the results. Under different pre-treatments, 74.1%-99.5% of metabolite-based phenotypic variation can be explained. For NMR data, only the pre-treatment of NS maximized the 95% confidence interval difference between the two groups of CONC and FOR diet samples (Fig. 8). And the explanation of phenotypic variation under NS treatment is also the highest among all pre-treatment results, reaching 99.5% (Fig. 8). For the HPLC data, the original data, MC, and NS pre-treatment explained 99.4% of the phenotypic variation and achieved similar characteristic values with NMR (NS), including direction and size (Fig. 8). In the phenotypic-related PCA exploratory analysis, both HPLC and NMR data were processed by NS to extract similar high-interpretation feature values. Because PCA seeks to explain as much variation as possible in as few components as feasible, whereas correlation focuses on the investigation of (dis)similarities. Using data pre-treatment to alter data attributes may thus improve the outcomes of correlation approach while blurring the feature variance of PCA analysis²².

PLS-DA analysis

For the PLS-DA analysis of HPLC and NMR data, the first two main components under each pre-treatment can explain 71.1–99.5% of the variation of diet-based samples (Fig. 9). The first two principal components extracted from the NMR data under NS processing in the PLS-DA analysis explained 99.5% of the phenotypic variation and achieved the highest R² and Q² in cross-validation (Fig. S6). Comparing all the two components model, HPLC achieved the R², Q² and accuracy greater than 0.6 in None, MC, PS, NS and Combine which indicated the prediction and fitting effects of these models are good. NMR data have the best combination of R², Q² and accuracy under NS pre-treatment. And in the PLS-DA results of NMR data, only NS has the 95% confidence interval of different diet samples based on extracted features that are well separated (Fig. 9). For HPLC data, the data under LT and Combine reduced the extracted feature values of PLS-DA and blur the differences between phenotypes. Finally, based on the NMR and HPLC data of NS, the results of PCA and PLS-DA analysis have obtained the two most explanatory principal components and can maximize the diet-based phenotypic differences, which is also consistent with biological knowledge.

As a summary in Table 5 showed, pre-treatment that has a positive impact on the analysis results should be given priority. Pre-treatment that has no impact is acceptable, but pre-treatment that has a negative impact should be carefully used. Scaling makes it possible to compare the variability between samples based on the same metabolite. This type of data processing has achieved good results in some case studies, especially for reflecting the accurate rank of metabolites²². However, in our study, it is also very likely that there are only a limited number of metabolites in the VFA data, which does not reflect the magnitude of differences between metabolites. Therefore, scaling does not play a positive role in the specific scenario with a lack of fold difference⁴². In similarity-based analysis, such as clustering or correlation, the Combine normalization approach performing predefined row- and column-wise procedures was found to work best with each other³¹. Our research showed that the Combine method does not perform well in PCA and PLS-DA analysis which indicated complicated data pre-treatments are likely to over-eliminate some important biological differences. NS can highlight the variation of the same metabolite between different samples. Normalization to total intensity is the most common method that showed a superior role by reducing the sample variability due to differences in sample concentrations.

In the downstream analysis results of NMR and HPLC, in the hierarchical clustering based on Pearson correlation, acetate and butyrate are found in one category, whilst isobutyrate, isovalerate, propionate and valerate are found in another. This is in line with the established knowledge about propionic acid production increasing under concentrate feeding and acetate production increasing under forage feeding. The consistency of the two sets of NMR and HPLC data in PCA and PLS-DA also shows that diet can explain more than 90% of the variation in the composition of VFAs. The Table S1 provides the standard deviation of the concentration of each VFA under the FOR and CONC diets, which shows that majority of data from FOR diet are less divergent. The regulation of rumen digestion and metabolism through diets varies greatly due to individual differences in cattle. The production and ratio of VFA are not only related to the feed itself, but also affected by other factors, including minerals, ionophores, animal age and feeding time, organic matter, outflow rate, enzyme preparations, health conditions, etc. The CONC diet may cause more uneven changes in volatile fatty acids in the rumen due to its faster fermentation.