Ultra high performance liquid chromatography–high resolution mass spectrometry plasma lipidomics can distinguish between canine breeds despite uncontrolled environmental variability and non-standardized diets

The full demographics of the study dogs are shown in electronic supplementary Table S1. The mean (±SEM) age of the 96 dogs was 56 months (±3) and had a range of 8–120 months. Using a 9 point body condition score (Laflamme 1997), the mean was 3.0 with a range of 2.0–4.5. The weight varied considerably between breeds (1.8–44.2 kg) with a mean (±SEM) of 20.6 (±0.9 kg). Due to logistical constraints and the focus on the study being breed, recruitment did not result in balanced age and gender within each breed. More female dogs (n = 58) participated in the investigation than males (n = 37), and there was one unrecorded gender; however, there was representation of both females and males in each dog breed class.

Previously, we used non-targeted nominal mass fingerprinting to characterize the main drivers of variance in the plasma metabolomes of dogs (Lloyd et al. 2016). The primary fingerprint data consisted of approximately 800 signals at 1 amu in each ionisation mode. In this current work, after pre-processing, data resulted in approximately 2500 signals per profile (with high sensitivity of 1–5 ppm identification level). These data were binned at 0.01 amu for multivariate modelling and classification to allow direct comparison of all plasma profiles. Principal Component-Linear Discriminant Analysis (PC-LDA) indicated that the main source of variance (PC- Discriminant Function 1, PC-DF1, Tw 4.39, Fig. 1a) was between three breeds [Group 1: Ch, Da and Gh (on the negative side of PC-DF1 in Fig. 1a)] and five of the other six breeds (Group 2: Be, CS, GR, GS and Ma; on the positive side of PC-DF1 in Fig. 1a). LR samples were distributed across the first dimension showing intra-breed variation. Separation of breeds into two clusters was also observed in PC-DF2 in which Ch, Da, Ma and GR were distinct from the other five breeds (Tw 2.55, Fig. 1a). The Tw values calculated for the first two dimensions of variance indicated clearly that the plasma lipidome profile of dogs in a fasting state was strongly influenced by breed. As with previous observations using non-targeted metabolome fingerprinting (Lloyd et al. 2016) both breed and breed size appeared significant factors in PC-DF1.

Open image in new window

Prior to investigating possible unique inter-breed plasma lipidome differences, we elected to further characterise the signals causing breed cluster discrimination on PC-DF1 and PC-DF2 seen in Fig. 1a. PCA and PC-LDA models were determined with new meta-classes explaining PC-DF1 and PC-DF2. The first meta-class structure was to explain the variance on PC-DF1 [Group 1 (Ch, Da and Gh) vs Group 2 (Be, CS, GR, GS and Ma)], with LR data omitted due to the lack of distinctive clustering with either of the two groups. The second meta-class designations were structured to elucidate the differences on PC-DF2: Ch, Da, Ma and GR were placed into one meta-class and the other breeds including LR were placed in the comparison meta-class. RF margins, AUC values and ACC modelling scores for these models are shown in electronic supplementary Table S2.

Previously, to characterize the main drivers of variance in the plasma metabolomes of dogs, we used metabolite fingerprinting with targeted high accurate mass analysis, with information acquired on an LTQ-FT-Ultra for metabolite identification, followed by FIE-MSⁿ for structual confirmation. In this current work, m/z data were generated at high resolution and sufficient accuracy to separate signals contained previously in the same 1 amu mass bin, as well as enabling direct identification of metabolites directly in the first pass profile without the need for further targetted accurate mass analysis, only FIE-MSⁿ for structual confirmation. Coupling of MS to LC additionally allowed retention times to be associated with every accurate m/z enabling structural isomers, adducts and isobaric compounds to be distinguished and ionisation adducts to be confirmed.

The explanatory signals associated with the discrimination between the meta-classes in Fig. 1a (along PC-DF1 and PC-DF2) were investigated further as the lipidome differences were above adequate thresholds [i.e. RFIS > 0.003 and false discovery rate (FDR) corrected P-values < 0.05; (Enot et al. 2008)]. The top ranked cluster of correlated (|r| > 0.8) signals driving separation on PC-DF1 between Ch, Da and Gh and the rest of the breeds is displayed in Table 1. The remainder of features above the implemented thresholds primarily responsible for the discrimination on PC-DF1 are shown in electronic supplementary Table S2, where correlated signals (|r| >0.8) are highlighted in different colours to show association. These signals were identified as a range of triglycerides, diglycerides and phosphatidylcholines which appeared much reduced in intensity in the Ch, Da and Gh breeds compared with the other breeds (Fig. 2a). Triacylglycerides are the most common and efficient form of stored energy in mammals and higher plasma levels can reflect differences in dietary intake but also hepatic health (Ginsberg 1998; Rifai et al. 1999).

Open image in new window

Metabolite identifications driving the separation on PC-DF2 were also assigned using the same strategy (Table 2 and electronic supplementary Table S4). These metabolites appeared reduced in intensity in the Ch, Da, Ma and GR compared with the other breeds (Fig. 2b) and represented different chemical classes to those causing the separation on PC-DF1. Correlation cluster 1 contains the sphingolipid C₁₇ sphinganine and long chain hydrocarbons with a single alkene group indicating possible lipid breakdown products. Features in the other two clusters could only be elucidated as unknown sphingomyelins or phospholipids due to the distinctive FIE-MSⁿ patterns (electronic supplementary Table S4).

By analysing the meta-data and the habitual eating patterns (see electronic supplementary Table S1) associated with each dog we identified that the discrimination seen in PC-DF1 in Fig. 1a also was impacted by diet types. This is supported by the inter-individual variation seen in LR, as those were distributed in line with whether the dogs followed a diet containing chicken meat (group1; on the negative side of PC-DF1) or red meat (group 2; on the positive side of PC-DF1) as their main protein source. We suggest that the metabolites seen at higher concentrations in group 2 (shown in Table 1 and electronic supplementary Table S3) compared with group 1 correlate with higher red meat consumption by group 2. Raw red meat including beef, lamb and mutton has been shown to contain increased total saturated and monounsaturated fatty acids when compared with white meat such as chicken (Droulez et al. 2006; Williams 2007). Additionally concentrations of total glycerophosphocholines and phosphatidylcholines have also been shown to be much higher in pork/ beef muscle meat and sausages/burgers when compared with chicken muscle meat and processed meat (Zeisel et al. 2003). In human plasma total saturated and monounsaturated fatty acid concentrations increased after a diet high in red meat when compared with a diet high in fatty fish (Wolmarans et al. 1991). Triacylglycerols, derived from glycerol and three fatty acids (both saturated and monounsaturated), have also been shown to increase with high habitual consumption of red meat (Wolmarans et al. 1991) in human plasma. It is important to note that all plasma samples were derived from fasted animals and thus as domesticated dogs may be maintained for many years on the same daily diet it is suggested that any studies on endogenous metabolism should ensure that confounding diet-derived signals are dealt with appropriately.

We have previously used a data filtering approach to allow the identification of metabolite signals associated specifically with exposure to test foods within a mix of ‘standardized food items’ in a human intervention study (Lloyd et al. 2011). The filtered list included all those explanatory signals with a RFIS > 0.002 after the consumption of the standardised food items when compared with fasting urines. A similar approach was employed here. From the breed model on PC-DF1 (Fig. 1a) the top RFIS ranked signals causing the differences between the binary comparison of the two classes (Ch, Da and Gh vs the other breeds in Table 1), were filtered and removed as they were assumed to be responsible for the confounding diet effect. We incrementally increased the filtering cut-off until the PC-LDA model (Fig. 1a) collapsed and the Ch, Da, Gh vs Be, CS, GR, GS, Ma clustering along PC-DF1 was no longer predominant. This heuristic RFIS threshold was >0.0003; additionally, features which were below the critical threshold, yet showed high correlation with the filtered top ranked signals with a |r| > 0.8, were also omitted from subsequent analyses as these have been shown to be isotopes and adducts in previous analyses. The resultant matrix contained 21.7% fewer features (539 signals out of 2496 were removed) than the complete matrix and when used to model the breed classes the main source of variance (PC-DF1) was now between four breeds (Ch, Da, Ma and GR) and the other five breeds (Tw 2.65, Fig. 1b), as observed previously in PC-DF2 in the original un-filtered model (see Fig. 1a). Notably the LR samples now formed a single distinctive cluster, suggesting the differences in diet were responsible for much of the variance in fasting plasma lipidome. Furthermore, with the exception of GR, the main axis of discrimination (PC-DF1) may be related to breed average weight (Table 1 and electronic supplementary Table S1) as observed previously (Beckmann et al. 2010) in the urine metabolome analysis. The top ranked signals discriminating the Ch, Da, Ma and GR from the other breeds of dogs were identical to those identified previously (Table 2), suggesting shared metabolome similarities in these four breeds. The sphingolipid C₁₇ sphinganine and the long chain hydrocarbons with amine groups were lower in the Ch, Da, Ma and GR breeds. Labrador was one of the breeds that had increased concentrations of these discriminatory sphingolipids and related amines when compared with Ch, Da, Ma and GR breeds. Differences in sphingolipid concentrations evident between the breeds in this current study may reflect how the dog breeds differentially transport or mobilise plasma sphingolipids, or may indicate different responses in the insulin signalling pathway between canine breeds (reviewed in Ilan 2016). Also, it has been suggested that endogenous sphinganine may inhibit cholesterol transport in Niemann-Pick Type C (NPC) disease (Roff et al. 1991) providing a suggestion that these dog breeds may have differential cholesterol esterification abilities.

One of the objectives of the study was to determine if there were any unique breed-specific signals (i.e. a signal that is not part of a common series of signals discriminating several breeds). Using the diet-filtered dataset, the modelling scores (RF margins, AUC and ACC values) for each breed pair-wise comparison are shown in Table 3. Adequate discrimination modelling scores were shown by CS and LR for only four pair-wise breed comparisons: therefore, we did not pursue these two breeds further (Table 3). For the remaining seven breeds (Ma, GS, Be, Ch, GR, Gh and Da) the modelling scores for ≥6 pair-wise breed comparisons were above the adequate thresholds criteria set, and we sought to identify the explanatory signals specific to each breed. By comparing top ranked signals (RFIS > 0.002 and P-values < 0.05) for each breed comparison, none of the Ma, GS, Da nor Be plasma lipidomes showed any unique breed-specific features consistent across all comparisons. In contrast, consistent unique breed specific features were discovered for the Ch (5 decreased in signal intensity compared to other breeds, example in Fig. 2c), Golden Retriever (7 elevated, example in Fig. 2d) and Greyhound (1 decreased, Fig. 2e). The full RF-IS and P-value data for all pair-wise comparisons are shown in electronic supplementary Table S5. It was interesting to note that despite the absence of any unique breed differences the Beagle plasma lipidome still discriminated from many other breeds suggesting a broad variance in the Beagle plasma lipidome, which is consistent with previous observations in relation to the urinary metabolome of this breed (Beckmann et al. 2010).

Direct annotation of signals unique to each of the three breeds, Ch, GR and Gh using MZedDB in the first pass profile was made without the need for further accurate mass analysis (Table 4). We have recently shown that by using a non-targeted ‘global’ metabolomic approach, Ch and Gh were discriminated from other breeds in the study (Lloyd et al. 2016). In Ch the plasma metabolites that were lower than in other breeds appeared to be vitamin D3 sterol lipids and phosphatidylcholines, which is in agreement with a previous report that the size of a dog appears to influence vitamin D3 metabolism (Tryfonidou et al. 2003). It has also been reported that Ch plasma has a low concentration of total cholesterol 197.6 ± 10.07 mg/dL compared with another small dog in this study, the Maltese, (235.2 ± 19.04 mg/dL), as well as larger dog breeds GR (301.1 ± 16.76 mg/dL), LR (229.4 ± 22.45 mg/dL) and CS (228.5 ± 26.34 mg/dL) (Usui et al. 2014). Cholesterol is the main sterol in animal tissues which plays a fundamental role in central metabolic pathways including steroid hormone and vitamin D synthesis (Ginsberg 1998; Rifai et al. 1999). Phosphatidylcholines are predominantly structural lipids important in plasma membranes and lung surfactants as well as being major components in lipoproteins. Inter-breed differences in lipoproteins have been reported previously, with the Cairn Terrier having particularly low levels of Low Density Lipoproteins compared to the LR, Be, Dh and other breeds (Downs et al. 1993).

The explanatory signal that was at a lower concentration in Greyhound plasma compared with other breeds could not be identified. As Gh generally are seen to have a lower percent body fat than other breeds (Jeusette et al. 2010) and a high percent muscle mass (Drost et al. 2006), it is speculated that this single explanatory signal may be related to the unusual body conformation of the Gh and a robust marker of the Gh breed. The signals elevated in intensity in the Golden Retriever plasma appeared to be sphingolipids, sphingomyelins and phosphatidylcholines. This observation may be consistent with previous evidence of the role of sphingolipids in skin disorders in Golden Retrievers, which are prone to atopic dermatitis (Shaw et al. 2004). Healthy dogs of mixed breeds have been shown to have erythrocytes that contain increased concentrations both of saturated sphingomyelin and also some phosphatidylcholines when compared to dogs suffering with atopic dermatitis (Fuhrmann et al. 2006). Additionally, there is evidence that sphingomyelin undergoes significant interactions with cholesterol (García-Arribas et al. 2016; Ridgway 2000) and Golden Retrievers have been shown to have the highest total cholesterol concentration out of 51 dog breeds studied and a top 14 concentration of total triacylglycerides (Usui et al. 2014).