Genetic variance of metabolomic features and their relationship with body weight and body weight gain in Holstein cattle

Abstract

In recent years, metabolomics has been used to clarify the biology underlying biological samples. In the field of animal breeding, investigating the magnitude of genetic control on the metabolomic profiles of animals and their relationships with quantitative traits adds valuable information to animal improvement schemes. In this study, we analyzed metabolomic features (MFs) extracted from the metabolomic profiles of 843 male Holstein calves. The metabolomic profiles were obtained using nuclear magnetic resonance (NMR) spectroscopy. We investigated 2 alternative methods to control for peak shifts in the NMR spectra, binning and aligning, to determine which approach was the most efficient for assessing genetic variance. Series of univariate analyses were implemented to elucidate the heritability of each MF. Furthermore, records on BW and ADG from 154 to 294 d of age (ADG_154–294), 294 to 336 d of age (ADG_294–336), and 154 to 336 d of age (ADG_154–336) were used in a series of bivariate analyses to establish the genetic and phenotypic correlations with MFs. Bivariate analyses were only performed for MFs that had a heritability significantly different from zero. The heritabilities obtained in the univariate analyses for the MFs in the binned data set were low (<0.2). In contrast, in the aligned data set, we obtained moderate heritability (0.2 to 0.5) for 3.5% of MFs and high heritability (more than 0.5) for 1% of MFs. The bivariate analyses showed that ~12%, ~3%, ~9%, and ~9% of MFs had significant additive genetic correlations with BW, ADG_154–294, ADG_294–336, and ADG_154–336, respectively. In all of the bivariate analyses, the percentage of significant additive genetic correlations was higher than the percentage of significant phenotypic correlations of the corresponding trait. Our results provided insights into the influence of the underlying genetic mechanisms on MFs. Further investigations in this field are needed for better understanding of the genetic relationship among the MFs and quantitative traits.

Introduction

Metabolomics is an approach that characterizes all metabolites (the metabolome) in a biological sample (Bovo et al., 2016). Nuclear magnetic resonance (NMR) spectroscopy is one of the key technologies used in metabolomics to identify and quantify metabolites and to profile metabolites in a biological sample. The price of an NMR instrument is high compared with other metabolomic analytical systems, such as gas chromatography and liquid chromatography. However, NMR is a faster, more reproducible, and less labor intensive per sample. This approach is also advantageous because it needs minimal preparation before measurement, has a nondestructive measuring process, and produces data in a quantitative manner (Holzgrabe et al., 2005; Emwas et al., 2016). Because metabolism is highly conserved across biology, with the metabolome usually having a close relationship with the phenotypic variation of an organism, analytical approaches of metabolomics are transferable across different biological systems (Fiehn, 2002).

In the field of animal breeding, it is of great interest to detect the contribution of genetic factor in biological observations and to use this contribution as a tool for improving economical traits. Some experimental studies have attempted to investigate the genetic background of metabolomic profiles in animal samples (Nicholson et al., 2002; German et al., 2005). For instance, Kühn et al. (2014) highlighted the advantages of combining genetics and metabolomics in a study of mammalian growth and tissue deposition. The authors noticed the importance of genetic variance and appropriate analytical methods to integrate metabolomics into studies on the genetic and physiological background of complex traits (e.g., mammalian growth). The same study reported a significant positive phenotypic correlation between plasma arginine at day 240 and average daily gain in a beef × dairy cattle cross population.

Despite the long history of using NMR technology, the proportion of genetic variance in data sets from biological samples remains unclear, with only a few studies addressing this issue (Wittenburg et al., 2013; Matros et al., 2017). Further investigation in this area could provide valuable insights on how we could benefit from genetic variation in intermediate phenotypes in animal breeding applications.

Here, we investigated the underlying genetic variance of signal intensities as an indicator of blood metabolites and considered their use as potential traits (metabolomic features [MF]). Rather than focusing on a few special metabolite compounds, we used the untargeted approach. Many small organic molecules are suspended in the body fluids of an animal (such as urine or blood; Lindon et al., 2000; Beckonert et al., 2007). They consist of metabolites steaming from the metabolic activity of the animal itself and small molecules that have entered the bloodstream from the environment, for example, through the digestive tract. If a molecule contains one or more ¹H atoms, it can be detected by ¹H NMR. The ¹H nucleus has a magnetic spin when a strong uniform external magnetic field is applied (as it is in an NMR instrument) the direction of the spin will be aligned in parallel or antiparallel to the external magnetic field (Berger, 2002). The parallel state contains less energy than the antiparallel and will be the dominating state in the sample. The ¹H nucleus can be pushed into the exited antiparallel state by the energy in an electromagnetic pulse—when the nucleus relaxes back to the parallel state the surplus energy will be emitted as a photon at a specific resonance frequency or chemical shift (Kwan et al., 2011). The resonance frequency is among other things dependents on how much the ¹H nucleus is shielded from the external magnetic field by the surroundings, for example, other atoms in the molecule (Kwan et al., 2011). Thus each ¹H nucleus of a molecule has a specific resonance frequency. This resonance frequency can differ from other ¹H nuclei in the same molecule and coincide with ¹H nucleus from different molecules. Therefore, there is no one to one relationship between a peak in an NMR spectrum from a complex biological sample and a specific metabolite. Metabolites often will have several peaks and peaks will often stem from several metabolites. Some of these metabolites have previously been assigned one or several chemical shifts in the metabolomic literature, and some of them are undescribed. Great care is therefore needed to accurately translate the NMR spectra into relative concentrations of specific metabolites, and often additional information is needed, such as ¹³C NMR spectra of the same samples. These 2D data sets can separate the signal from metabolites that have identical signatures in ¹H NMR, but not in ¹³C NMR (Markley et al., 2017). In this study, we instead chose to investigate the intensities of the ¹H NMR spectra data directly. We extracted MFs from normalized processed NMR data by binning or aligning. Binning and aligning represent 2 different methods for controlling peak shifts in metabolomic data. To the best of our knowledge, the genetic parameters of MFs in blood samples have not yet been reported, with this study presenting the first exhaustive investigation of the genetic background and relationship between metabolomic features and growth performance in cattle. Thus, this study aimed to 1) investigate whether MFs could be used as a new tool for phenotypic evaluating of animals and 2) measure the magnitude of the additive genetic effect on MFs. Specifically, we reported the heritability of MF and genetic correlations with BW and ADF for 3 different age intervals of growing young bulls.

MATERIALS AND METHODS

Data

This study was conducted on 843 male Holstein calves that were born in several private herds (468 herds that contributed from 1 to 37 calves). Calves were transferred to the experimental stations following obligatory health checks. The animals were housed at 3 experimental stations at which they arrived, ideally, before 3 mo of age. However, calves that arrived that were up to 5 mo of age were still accepted for testing. A diet based on dried grass pellets supplemented with barley straw was fed ad libitum. Pellets were offered from feed troughs, and water was freely available throughout the experimental period. With a few exceptions, animals were weighed at 154, 294, and 336 d of age, and these BW were used to compute ADG_154–294, ADG_294–336, and ADG_154–336. To acquire metabolomic information blood samples were collected from animals at 270 ± 11 d of age in 96 batches using a standardized protocol (Hayhurst et al., 2007). To avoid physiological responses to feed intake, the blood samples were taken when animals were fed at maintenance level in a single meal on the day before sampling. The heparinized plasma was frozen and stored at −20 °C and was transferred to −80 °C shortly before NMR data acquisition. Based on the protocol, BW of animals were measured 6 to 9 d before collecting the blood samples under normal feeding conditions. Figure 1 shows the workflow of the standardized protocol that was used for blood sampling.

Figure 1.

Workflow of the standardized protocol

Ancestry was traced back at least 3 generations to construct a pedigree file with 14,140 animals (Table 1).

Table 1.

Descriptive statistics for BW and ADG traits used in the analyses

Trait1	No. of observations	Average, kg	Maximum, kg	Minimum, kg	SD
BW	612	330.33	453	230	36.08
ADG154–294	767	1.36	1.90	0.59	0.16
ADG294–336	700	1.31	3.85	–1.59	0.49
ADG154–336	659	1.35	2.00	0.71	0.16

1BW = body weight at blood sampling (6 to 9 d before the blood sampling day); ADG_154–294 = average daily gain from 154 to 294 d of age; ADG_294–336 = average daily gain from 294 to 336 d of age; ADG_154–336 = average daily gain from 154 to 336 d of age.

Sample Preparation and NMR Data Acquisition

Samples were removed from the freezer to thaw at 4 °C overnight. The samples were then centrifuged at 3,000 × g for 10 min at 4 °C. One hundred microliters of each of the sample and buffer (75 mM sodium phosphate pH 7.4, 0.5 mM 3-(trimethylsilyl)-1-propanesulfonic acid-d₆ (DSS-d₆), 0.1% sodium azide in 20% D₂O) was mixed in 96-well deep well plates (Corning Costar 2 mL polypropylene, cat no 3,961) and transferred to 3-mm SampleJet NMR tube racks using a SamplePro L liquid handling robot (Bruker BioSpin). The temperature of the plates and sample racks during pipetting was maintained at 2 °C.

NMR data were acquired on an Oxford magnet operating at 799.87 MHz that was equipped with a Bruker Avance III HD console and a 3-mm TCI cryoprobe and cooled SampleJet automatic sample changer. The experiment used a 1D proton pulse sequence, including excitation sculpting for water suppression and perfect echo for J-modulation suppression, and a CPMG pulse train of 193 ms as the T2 filter. The average of 64 scans were acquired at 32 k points, using a sweep width of 20 ppm, an acquisition time of 2.03 s, and a relaxation delay of 1.3 s. The temperature was kept at 6 °C in the sample changer and at 25 °C during acquisition.

We processed the spectra using an in-house custom Matlab script (personal communication Caroline Sands and Timothy Ebbels). Specifically, we zero-filled by a factor of 1.5 and applied an exponential apodization function equivalent to 0.5 Hz line-broadening, followed by Fourier transformation. All spectra were referenced to the DSS-d₆ signal, automatically phased, and baseline corrected.

Pre-processing of NMR Data

Before statistical analysis, metabolomic data must be pre-processed by several steps to extract meaningful information from a high-dimensional and complex data structure. The preprocessing steps ensure that the variation observed in the data are representative of real biological variation. Thus, these steps are crucial for data quality and for interpreting the results (Karaman, 2017). Common preprocessing steps include noise filtering, normalization, binning, peak alignment, and scaling. The optimal combination and order of these steps depends on the structure of the data and must be evaluated with great care (Wu and Li, 2016).

Noise Filtering

To ensure that NMR spectra only contain information on metabolites, some interfering effects should be eliminated at the first step of pre-processing (Lindon et al., 2011). One of the unwanted regions is signals produced by the solvent water, which does not add information on the metabolites in the samples. We thus excluded water peak in the range of 4.7 to 4.9 ppm. In addition, the region of the added standard DSS-d6 at −0.2 to 0.2 ppm was removed.

Normalization

Sample normalization is a crucial step in the pre-processing of NMR spectral data to refine variation that could be attributed to experimental sources (e.g., pipetting, weighing tissue) and signal intensities, which are not correlated to metabolism and interfere with statistical analyses (Karaman, 2017). Therefore, to consider possible variation in the overall concentrations of samples after the noise filtering step, we normalized the raw data using the probabilistic quotient method (Dieterle et al., 2006). In this way, our samples were more comparable with each other (Karaman, 2017). In addition, in this step, 9 outliers were detected based on a plot of the first 2 principal components of a PCA on the mean centered intensities that were scaled to unit variance. After removing these outliers, a new PCA on the remaining data set did not reveal any additional outliers.

Binning

Binning is a common method for data reduction, as it allows feature peaks to be detected and reduces the influence of peak shifts in ¹H-NMR spectral data (De Meyer et al., 2008; Lindon et al., 2011; Zheng et al., 2011; Smolinska et al., 2012; Alonso et al., 2015). In the binning or bucketing technique, NMR spectra are divided into a number of segments (called bins or buckets), in which the total area inside each bin is subsequently summed to produce a single value as the area under the curve of the original spectrum. We considered a bin size that was equal to 0.04 ppm as a division unit of the whole spectrum, as previous studies proposed this bin size as being large enough to encompass variation in peak shifts (Lindon et al., 2011; Smolinska et al., 2012). Our data contained 30,916 intensities, with 262 bins being present when the whole shift axis was divided by 0.04 ppm. Based on the whole number of the intensities, each bin contained 118 spectral intensities per spectra. Therefore, integration was achieved by summing the intensities within each bin. After integration, 262 data points were obtained for each sample. This data set represented the first set of MFs.

Aligning

For the second MF data set, we used aligning to control peak shifts. There are many factors in play during the acquisition of multiple sample NMR data, including pH, temperature, instrumental factors, salt concentration, overall dilution of the sample, relative concentrations of specific ions, and metabolites. These factors could influence the position of the peak corresponding to the same metabolites in different samples. This issue represented a particular challenge in this data set due to the high number of samples. To address this challenge in the second data set, we used the interval–correlation–shifting (icoshift) alignment algorithm and manually selected the segments to be aligned. The icoshift is a Matlab-based open-source software package that is faster to implement, with lower computational demands, than other methods when solving aligning problems in NMR data (Savorani et al., 2010; Vu and Laukens, 2013). Icoshift can be set to run automatically, but this unsupervised alignment often fails with complex large-scale data sets. Hippuric acid is highly sensitive to chemical shifts in the physiological pH range and is an abundant metabolite in herbivores. We, therefore, organized the unaligned spectra into groups according to the part of their spectra where hippuric acid was expected to be situated (between 7.4 and 7.9 ppm). The grouping was obtained by the k-means clustering method with 10 clusters to identify the most similar spectra. Then, for each cluster, a reference spectrum was identified based on the highest match with the average spectrum within each cluster. Intervals for aligning were selected based on the visual inspection of the reference spectra. The 10 reference spectra were aligned to the reference spectra with the highest number of MFs. Then, each cluster was aligned to the aligned reference spectrum originating from the same cluster. Figure 2 shows 10 spectra, one spectrum randomly selected from each of the 10 clusters: 1) after normalization, 2) after binning in the first data set with 0.04 ppm bin size, and 3) after using alignment for correcting peak shifts.

Figure 2.

Spectrums after normalization of 10 random samples, after binning and after aligning.

In general, peak-based methods such as icoshift alignment are preferable to binning for controlling peak shifts (Alonso et al., 2015). To determine which method is more suitable for investigating the genetic background of MFs, we analyzed both the binned data without aligning (the first metabolomic data set) and the aligned data without binning (the second metabolomic data set). Of note, both data sets were based on data generated after the noise filtering and normalization steps, and the difference between these data sets is purely in the peak shifts controlling step.

Scaling

Often, the distribution of individual MFs is extremely skewed compared with a standard normal distribution, which was also the case for our data. Therefore, all of the statistical analyses were carried out on transformed data using the Johnson transformation method to approach normality and to estimate (co)variance components (Johnson, 1949). The Johnson R package was used for transformation (Santos Fernandez, 2014). In brief, the Johnson transformation operates based on 3 families of transformations and the estimation of corresponding parameters; namely, the bounded system (S_B), the lognormal system (S_L), and the unbounded system (S_U). The algorithms introduced by Slifker and Shapiro (1980) and Chou et al. (1998) are the 2 most common criteria used to select one of the 3 functions that produce the closest distributions to the standard normal distribution. The Johnson transformation method is a suitable and flexible method for normalization, with respect to the algebraic sign of data. The following equation is the general form of Johnson transformation:

$${\displaystyle \begin{array}{l}{\displaystyle z=\gamma +\delta f\{(x-\xi )/\lambda \}}\end{array}}$$

In this equation, z is the transformed data, x is original metabolic intensities, and γ, δ, ξ, and λ are the Johnson parameters.

Statistical Analyses

To estimate the (co)variance components and genetic parameters of MFs in the 2 data sets and for the phenotypes (BW, ADG_154–294, ADG_294–336, and ADG_154–336), 2 different models, in terms of fixed effects, were fitted in univariate and bivariate analyses:

$${\displaystyle \begin{array}{l}{\displaystyle M_{ijkl}=\mu +b\times \mathrm{S}\mathrm{T}\mathrm{S}{\mathrm{A}}_i+\mathrm{B}{\mathrm{P}}_j+\mathrm{C}{\mathrm{L}}_k+A_l+{\epsilon }_{ijkl}}\end{array}}$$ (1)

$${\displaystyle \begin{array}{l}{\displaystyle P_{ijkl}=\mu +b\times \mathrm{S}\mathrm{T}\mathrm{W}{\mathrm{D}}_i+\mathrm{S}{\mathrm{T}}_j+\mathrm{Y}\mathrm{S}{\mathrm{B}}_k+A_l+{\epsilon }_{ijkl}}\end{array}}$$ (2)

In the first model, M_ijkl = MFs (in binned or aligned data), µ = intercept, STSA_i = standardized age of animals at the sampling date, b = regression coefficient, BP_j = fixed effect of batch × pipetting date (105 levels), CL_j = fixed effect of cluster (j = 1, …, 10), A_l = random additive genetic effect (l = 1, …, 14,140), and ε_ijkl = random residual effect. Because the clustering was only used in the aligned data, the cluster effect was not included in the first model for analyzing the binned data set.

In the second model, P_ijkl = observation of the phenotypes (BW, ADG_154–294, ADG_294–336, ADG_154–336), µ = intercept, STWD_i = standardized age of animals at the weighing date, b = regression coefficient, ST_j = fixed effect of the station (i = 1, 2, and 3), YSB_k = fixed effect of year-season of birth (12 levels), A_l = random additive genetic effect, and ε_ijkl = random residual effect. Because the ADG traits were calculated between identical ages of animals, STWD_i was only fitted in the model for BW. Note that a random maternal genetic effect and a random maternal permanent environmental effect were not significant and thus were not included in the second model.

It is assumed that A_l~ N(0, Aσ²_a) and ε_ijkl~ N(0, Iσ²_a), where A is the numerator relationship matrix derived from the pedigree, σ²_a is the additive genetic variance, I denotes the identity matrix, and σ²_e is the residual variance. The variance components were estimated using the average information restricted maximum likelihood algorithm implemented in the DMUAI module in the DMU software package, release 5.2 (Madsen and Jensen, 2013).

To perform bivariate analyses between MFs and the phenotypes for BW and ADG, the MFs were filtered so that only MFs with heritability significantly different from zero in the univariate analyses were retained. Heritability was considered significant if the MFs were significantly different from zero using a 1% significance level (Z > 2.326). For this purpose, we used the Z-scores of heritability. The Z-scores were also used for the correlations obtained from bivariate analyses to determine the significance of the whole correlation. The Z-score is the ratio of the variance components or parameters to their SE (Rønning et al., 2007; Isik et al., 2017):

$${\displaystyle \begin{array}{l}{\displaystyle Z=\frac{x}{{({\sigma }^2)}^{0.5}}}\end{array}}$$

In this equation, x is the heritability estimate and σ is the associated SE obtained via the DMU software.

RESULTS

Fixed Effects in the Models

To decide which fixed factor should be fitted in the final analysis model, for each MF, a general linear model (GLM) and ANOVA F-test were first applied using R software to assess the significance of the fitted fixed effects (batch × pipetting date and cluster) in the model. The combined effect of batch × pipetting date was significant on 56% (146) of MFs in the binned data set (Table 2) and on 54% in the aligned data set at the 0.05 significance level. The cluster effect was significant on a low percentage of the aligned data set (38% at the 0.05 significance level).

Table 2.

Number and proportion (in parentheses) of significant fixed effects at 2 levels on the binned and aligned data set

Fixed effect	Data set
	Binned	Aligned
Batch × pipetting date	146 (56%)	16,655 (54%)
Cluster	—	11,711 (38%)

Univariate Analyses

Relatively high heritability was obtained for the BW (0.55). The heritability of the ADG traits showed an increasing trend with increasing age of animals, so that the heritability of ADG_154-336 was almost twice that obtained for ADG_154-294 (0.15, 0.24, and 0.32 for ADG_154-294, ADG_294-336, and ADG_154-336, respectively). The highest heritability for ADG_154–336 could be due to the fact that this trait was calculated from the longest time period and based on triplicate weight records at each age, which will result in a smaller measurement error compared with the other traits.

The main goal of this study was to investigate the magnitude of genetic variation in the MFs. To achieve this goal, each MF was analyzed by restricted maximum likelihood to estimate the variance components and genetic parameters. Binned and aligned data sets contained 262 and 30,914 different MFs, respectively. For each MF, a separate univariate analysis was run. The same model was used for each MF in each of the 2 data sets. The main results of the univariate analyses of MFs are summarized in Table 3.

Table 3.

Results of univariate analyses in the 2 data sets of metabolomics features

Parameter1	Data set
	Binned	Aligned
No. of analyses	262	30914
Min h2	1.38E-09	3.35E-10
Max h2	0.17	0.52
No. of h2 < 0.2	262	29,798
No. of h2 between 0.2 and 0.5	0	1105
No. of h2 > 0.5	0	11

1h² = heritability.

In the binned data set, estimated heritabilities were generally low ranging from 0.00 to 0.17. In contrast, in the aligned data set, overall heritability increased considerably and was in the range of 0.00 to 0.52. Figure 3 shows the plots of estimated heritability in the binned and aligned data sets.

Figure 3.

Estimated heritabilities of the metabolomic features in the binned and aligned data sets.

To exclude noise MFs that might be present in the metabolomic data, even after the noise filtering step in the pre-processing, a conservative significance level (0.01 instead of 0.05) was adopted to test the heritability. The Z-scores showed that none of the heritability estimates were significantly different from zero in the binned data set. In comparison, heritability estimates of 1,040 MFs were significantly different from zero in the aligned data set.

Bivariate Analyses of MF With BW and ADG Traits

Bivariate analyses were carried out to investigate the relationship of the significant MFs with BW, ADG_154–294, ADG_294–336, and ADG_154–336. Based on the significance test carried out for the results of univariate analyses, bivariate analyses were run for the 1,040 significant MFs in the aligned data set together with BW, ADG_154–294, ADG_294–336, or ADG_154–336. Histogram of the obtained additive genetic and phenotypic correlations between the MFs and 4 growth traits are presented in Figs. 4 and 5, respectively. In Table 4, we summarized the overall range of the correlations and the number of significant correlations between the MFs and the 4 growth traits. In general, the additive genetic correlations (positive or negative) were considerably higher than the phenotypic correlations. The range of the additive genetic correlations between MFs and BW was slightly lower than the range observed between MFs and the ADG traits. However, the range of the phenotypic correlations was comparable between the 4 growth traits, but with the correlations to BW tending to have a slightly higher range.

Figure 4.

Additive genetic correlations for BW and ADG traits with selected MFs. MFs = metabolomic features; BW = body weight at blood sampling (6 to 9 d before the blood sampling day); ADG_154–294 = average daily gain from 154 to 294 d of age; ADG_294–336 = average daily gain from 294 to 336 d of age; ADG_154–336 = average daily gain from 154 to 336 d of age.

Figure 5.

Phenotypic correlations for BW and ADG traits with selected metabolomic features (MFs). MFs = metabolomic features; BW = body weight at blood sampling (6 to 9 d before the blood sampling day); ADG_154–294 = average daily gain from 154 to 294 d of age; ADG_294–336 = average daily gain from 294 to 336 d of age; ADG_154–336 = average daily gain from 154 to 336 d of age.

Table 4.

Results of bivariate analyses for the selected MFs with BW, ADG_154–294, ADG_294–336, and ADG_154–336¹

Data set2	MF and BW	MF and ADG154–294	MF and ADG294–336	MF and ADG154–336
Min rA	−0.77	−0.99	−0.98	−0.99
Max rA	0.57	0.80	0.69	0.64
Min rPh	−0.16	−0.11	−0.13	−0.12
Max rPh	0.20	0.2	0.08	0.13
No. of significant rA	129	37	94	93
No. of significant rPh	48	52	13	18

¹MF = metabolomic feature; BW = body weight at blood sampling (6 to 9 d before the blood sampling day); ADG_154–294 = average daily gain from 154 to 294 d of age; ADG_294–336: average daily gain from 294 to 336 d of age; ADG_154–336 = average daily gain from 154 to 336 d of age.

² r _A = additive genetic correlation; r_Ph = phenotypic correlation.

The Z-scores of the correlations were calculated to identify the position of significant correlations among the whole range of additive genetic and phenotypic correlations. In all 4 bivariate analyses, additive genetic correlations for MFs vs. BW and ADGs tend to be significant for correlations higher than 0.5 or lower than −0.5 (Fig. 4). The percentages of significant additive genetic correlations between BW and MFs were higher compared with the observed percentages of significant additive genetic correlations between ADG traits and MFs (with lowest percentages observed for ADG_154–294; Table 4).

In all bivariate analyses, the proportion of significant phenotypic correlations was less than the proportion of corresponding additive genetic correlations (Fig. 5). The highest proportion of significant phenotypic correlations was obtained between ADG_154–294 and MFs, whereas the lowest proportion was obtained between ADG_294–336 and MFs (Table 4). The phenotypic correlations between ADG_294–336 and MFs exhibited less variation compared with the other 3 traits (Fig. 5). It also shows that the correlations between BW and MFs exhibited more variation compared with the phenotypic correlations between the 3 ADG traits and MFs.

Discussion

This study analyzed MFs using a series of univariate and bivariate analyses together with records on BW and ADG traits in Holstein cattle. As expected, some of the MFs were heritable, of which many were significantly genetically correlated to the investigated growth phenotypes. To assess the magnitude of genetic control on the MFs, univariate analyses were run for each MF, with a model including a random additive animal genetic effect. In general, the results highlighted the considerable genetic control of metabolism in male Holstein calves, as we observed 1,040 significant heritability estimates of the MFs. The zero estimates of heritability of the MFs occurred at a higher frequency compared with the frequency of nonzero heritability in both binned and aligned data sets. The zero and low heritability of the MFs was expected, as we analyzed whole metabolomic profiles, including areas with low or no biological signal from metabolites. In addition, the analysis was conducted using a range of the NMR profile that was wide enough to ensure that all biological signals were detected, regardless of the chemical shift of the samples. This range resulted in an area of ppm’s on each side of the spectrum that was seemingly empty of biological signal, so that where zero heritability should be expected. We observed a low level of heritability, less than 0.2, for all of the MFs in the binned data set (Table 3). In contrast, in the aligned data set, 3.5% (1105) of the MFs exhibited moderate heritability between 0.2 and 0.5. Furthermore, 1% MFs exhibited high heritability of more than 0.5. The large difference between estimated additive genetic variance and heritability in the binned and aligned data sets emphasized how preprocessing metabolomic data sets influences the magnitude of estimated genetic variance of MFs. In fact, the main reason for such differences is that the standard binning method increases within-class variance, which increases the ratio of within-class to between-class variation (Davis et al., 2007). Davis et al. (2007) suggested adaptive binning to overcome this problem. Furthermore, Vu and Laukens (2013) mentioned that the traditional binning method causes a drastic reduction in data resolution; thus, the authors suggested using different alignment methods as the optimum way to control peak shifts in the NMR spectroscopy data. This increase in within-class variance is shown in Fig. 2. Therefore, we used the basic binning method vs. alignment method to pre-process the metabolomic profiles, and compared the results of analyses.

The novel research design used in this study makes it difficult to compare with other studies. However, studies investigating genetic effects on MFs have reported important genetic variance in MFs. In a study on winter wheat using SNP genotyping data, Matros et al. (2017) reported low broad-sense heritability for 76 metabolites and significant variances of genotypes with heritability of >0.4 for 24 metabolite abundances. For a further 17 metabolite abundances, non-significant variances of genotypes were reported (Matros et al., 2017). Furthermore, 55 metabolites were investigated in the milk samples of Holstein cows, with additive genetic variance being detected in a range of 0 to 0.57 for narrow-sense heritability, while broad-sense heritability was detected in the range of 0 to 0.70 (Wittenburg et al., 2013). In a study using 1,056 Fourier-transform infrared (FTIR) spectra of bovine milk, a maximum heritability of 0.25 for less than 1% of the FTIR waves have been reported (Bittante and Cecchinato, 2013). In the same study, more than half (55%) of the FTIR waves showed low heritability estimates. The Fourier transform technique is a set of mathematical functions that has so far been used in metabolomics studies to unravel metabolomes from spectrum data (Breitling et al., 2006). This technique, prior to the processing steps, has been done automatically in our data. Regardless of the different types of the biological samples used in our study, the heritability values reported by Bittante and Cecchinato (2013) are rather similar to what we obtained using the binned data set. Regarding the 2 times higher maximum heritability observed for MFs in the aligned data set compared with the binned data set, the impact of the processing methods of the metabolomic data becomes clearer.

We plotted all of the obtained heritability estimates and the standardized mean intensity of MFs in the aligned data set based on the maximum heritability of MFs, which was 0.52 (Fig. 6). The standardization based on the maximum heritability using the following equation: (0.52 × mean of each intensity point)/max (mean of each intensity point), made it possible to visualize both heritability estimates and intensity of MFs in the same plot. Although the highest intensity MFs did not have the highest heritability, the plot revealed a relationship between heritability and the magnitude of the MFs. This relationship was particularly clear in the area of 1 to 6 ppm, and most of the significant heritabilities appeared in this range. Our interpretation of this relationship was that the metabolomic points, which appear at a high intensity in the NMR instrument, represent biological compounds that are under considerable influence of additive genetic effects.

Figure 6.

Estimated heritabilities in the aligned data set and standardized mean intensity. The equation used to obtain the standardized mean intensity of MFs based on maximum heritability of MFs: (0.52 × mean of each intensity point)/max (mean of each intensity point).

In addition to the univariate analyses, we used a bivariate approach to investigate the relationship between the MFs and growth traits. For the bivariate analyses, we only used MFs with heritability significantly above zero. Testing the Z-scores of the additive genetic correlations showed that BW was more significantly correlated with MFs compared with ADG traits. The proportion of significant additive genetic correlations to the total number of additive genetic correlations for each trait were ~12%, ~3%, ~9%, and ~9% for BW, ADG_154–294, ADG_294–336, and ADG_154–336, respectively. The proportion of significant additive genetic correlations in the 4 bivariate analyses might be related to the magnitude of the genetic variance of BW or ADG traits. For instance, 12% significant additive genetic correlations between MF and BW decreased to 3% between MF and ADG_154–294. Furthermore, the obtained heritability for BW was equal to 0.55 in comparison to 0.15 for ADG_154-294. Alternatively, the higher proportion of significant additive genetic correlations of MFs with BW and ADG_294-336 could be due to the young animals approaching puberty, which increases hormonal activity, such as growth hormone (GH), in the period covered by ADG_294–336. The literature shows that changes occurring in the secretion of GH in young bulls during puberty contribute to ADG (Connor et al., 1999, 2000). However, the secretion of GH is associated with general changes to hormones in the body, directly or indirectly stimulating anabolic processes, such as cell division, skeletal growth, and protein synthesis (growth promoting activity), while also increasing the oxidation of fat (lipolytic activity) and inhibiting the transport of glucose into body tissues (diabetogenic activity; Husvéth, 2011). Perhaps, the simultaneous activation and expression of regulatory genes associated with the secretion of GH and metabolic processes results in their interacting, leading to higher genetic correlations of MFs and ADG_294–336.

There were no strong phenotypic correlations between the MFs and BW or ADG traits (Fig. 5). In all cases, phenotypic correlations were less significant compared with the corresponding additive genetic correlations, of which the highest percentage of significant phenotypic correlations was observed for ADG_154-294 (5%). This is highlighting the impact of the genetic factor on the interaction between the metabolism and functional traits. This issue might be beneficial to develop more efficient selection indices.

In conclusion, the results of the univariate analyses of the MFs showed that heritability of the MFs could vary from zero to more than 0.50, with some MFs being highly heritable. Estimates of heritability of the MFs were generally larger when using alignment to control for peak shifts in the preprocessing step of the metabolomic profiles instead of binning. This phenomenon is due to an increase in within-class variation when using binning compared with aligning for controlling peak shifts. Based on the results of our study and also the previous studies, we concluded that the binning method is not a suitable method to process the metabolomic data and preferably alignment method gives better input data to access to the genetic variance of the MFs. Overall, our results provided more insights into the influence of the underlying genetic mechanisms on the MFs. However, further research are needed to confirm our results and to investigate the ability of the MFs as an indicator of complex traits, such as feed efficiency, carcass quality, and meat quality characteristics.