Double-check: validation of diagnostic statistics for PLS-DA models in metabolomics studies

Partial Least Squares-Discriminant Analysis (PLS-DA) is a PLS regression method with a special binary ‘dummy’ y-variable and it is commonly used for classification purposes and biomarker selection in metabolomics studies. Several statistical approaches are currently in use to validate outcomes of PLS-DA analyses e.g. double cross validation procedures or permutation testing. However, there is a great inconsistency in the optimization and the assessment of performance of PLS-DA models due to many different diagnostic statistics currently employed in metabolomics data analyses. In this paper, properties of four diagnostic statistics of PLS-DA, namely the number of misclassifications (NMC), the Area Under the Receiver Operating Characteristic (AUROC), Q ² and Discriminant Q ² (DQ ²) are discussed. All four diagnostic statistics are used in the optimization and the performance assessment of PLS-DA models of three different-size metabolomics data sets obtained with two different types of analytical platforms and with different levels of known differences between two groups: control and case groups. Statistical significance of obtained PLS-DA models was evaluated with permutation testing. PLS-DA models obtained with NMC and AUROC are more powerful in detecting very small differences between groups than models obtained with Q ² and Discriminant Q ² (DQ ²). Reproducibility of obtained PLS-DA models outcomes, models complexity and permutation test distributions are also investigated to explain this phenomenon. DQ ² and Q ² (in contrary to NMC and AUROC) prefer PLS-DA models with lower complexity and require higher number of permutation tests and submodels to accurately estimate statistical significance of the model performance. NMC and AUROC seem more efficient and more reliable diagnostic statistics and should be recommended in two group discrimination metabolomic studies.     

Discriminant Q2 (DQ2) for improved discrimination in PLSDA models

In this paper we introduce discriminant Q² (DQ²) as an improvement for the Q² value used in the validation of PLSDA models. DQ² does not penalize class predictions beyond the class label value. With rigorous Monte Carlo simulations we show that when DQ² is used, a smaller effect can be found statistically significant than when the standard Q² is used.     

Comparative analysis of creatinine and osmolality as urine normalization strategies in targeted metabolomics for the differential diagnosis of asthma and COPD

Introduction

Urine is an ideal matrix for metabolomics investigation due to its non-invasive nature of collection and its rich metabolite content. Despite the advancements in mass spectrometry and ¹H-NMR platforms in urine metabolomics, the statistical analysis of the generated data is challenged with the need to adjust for the hydration status of the person. Normalization to creatinine or osmolality values are the most adopted strategies, however, each technique has its challenges that can hinder its wider application. We have been developing targeted urine metabolomic methods to differentiate two important respiratory diseases, namely asthma and chronic obstructive pulmonary disease (COPD).

Objective

To assess whether the statistical model of separation of diseases using targeted metabolomic data would be improved by normalization to osmolality instead of creatinine.

Methods

The concentration of 32 metabolites was previously measured by two liquid chromatography-tandem mass spectrometry methods in 51 human urine samples with either asthma (n?=?25) or COPD (n?=?26). The data was normalized to creatinine or osmolality. Statistical analysis of the normalized values in each disease was performed using partial least square discriminant analysis (PLS-DA). Models of separation of diseases were compared.

Results

We found that normalization to creatinine or osmolality did not significantly change the PLS-DA models of separation (R²Q²?=?0.919, 0.705 vs R²Q²?=?0.929, 0.671, respectively). The metabolites of importance in the models remained similar for both normalization methods.

Conclusion

Our findings suggest that targeted urine metabolomic data can be normalized for hydration using creatinine or osmolality with no significant impact on the diagnostic accuracy of the model.

     

An NMR- and MS-based metabonomic investigation of saliva metabolic changes in feline odontoclastic resorptive lesions (FORL)-diseased cats

The feline odontoclastic resorptive lesion (FORL) is a common oral problem in cats. The disease has increased steadily since the domestication of cats and etiology of this disease has not been fully determined although several theories have been proposed. Feeding practices, vaccination, and neutering programs have all been suspected to be associated with FORL. The aim of the current study is to assess the feasibility of metabonomics to detect at an early stage the onset of the disease. The diagnostic biomarkers could then be used as “efficacy markers” for nutritional intervention in preventing and/or slowing the progression of FORL. ¹H-NMR- and LC/MS-based metabonomic analysis of saliva samples obtained from a group of 21 cats (11 healthy and 10 FORL diseased) showed clear differences in the metabolic composition of saliva from healthy and FORL-diseased cats. To identify biomarkers, the spectroscopic data was processed using partial least-squares discriminant analysis (PLS-DA) and validated by leave-one-subject-out cross validation. The PLS-DA model predicted FORL- diseased cats with over 60% accuracy. The maximum value of Q² of the random permutation sets was less than 0.3. The diseased cats showed increased levels of many organic and amino acids, such as acetate, lactate, propionate, isovalerate, tryptamine, and phenylalanine suggesting changes in oral microflora in the disease situation. This study is preliminary and a larger study with more samples to further validate the biomarker profile predictive of an early FORL pathophysiological status is in progress.     

Carbohydrate fed state alters the metabolomic response to hemorrhagic shock and resuscitation in liver

Hemorrhagic shock with injury results in alterations of the metabolic state of an organism, which contribute to organ dysfunction and death. Previous investigations have explored the effects of carbohydrate prefeed in murine models but few in clinically relevant large animal models. We performed carbohydrate prefeed in pigs undergoing simulated polytrauma and hemorrhagic shock with resuscitation to determine if carbohydrate prefeeding if the metabolic response to shock is dependent on fed state. Sixty-four Yorkshire pigs were divided into two experimental groups: fasted (32) and prefed (32). Experimental animals were subjected to a standardized hemorrhagic shock protocol, including pulmonary contusion and liver crush injury. To determine molecular alterations in response to trauma as a result of prefeeding, liver biopsies were obtained at set timepoints throughout the procedure. Fifty-one metabolites were profiled for each sample via proton nuclear magnetic resonance spectroscopy (¹H NMR). Partial-Least Squared Discriminant Analysis (PLS-DA) was used to examine clustering of the data with respect to fed state. Cross-validated models separated the fed from fasted animals. Metabolites contributing to the separation have known relationships to alternate carbon energy sources, amino acid metabolism, oxidative stress response, and membrane maintenance. In conclusion, metabolomic techniques allowed identification of an alternate response to shock and resuscitation, dependent upon fed state, through the use of metabolomics.     

Metabolomic profile in hyperthyroid patients before and after antithyroid drug treatment: Correlation with thyroid hormone and TSH concentration

Hyperthyroidism (HT) is characterized by an intense metabolic impact which affects the lipid, carbohydrate and amino acids metabolism, with increased resting energy expenditure and thermogenesis. Metabolomics is a new comprehensive technique that allows to capture an instant metabolic picture of an organism, reflecting peculiar molecular and pathophysiological states. The aim of the present prospective study was to identify a distinct metabolomic profile in HT patients using ¹H NMR spectroscopy before and after antithyroid drug treatment. This prospective study included 15 patients (10 female, 5 male) who were newly diagnosed hyperthyroidism. A nuclear magnetic resonance (¹H NMR) based analysis was performed on plasma samples from the same patients at diagnosis (HypT₀) and when they achieved euthyroidism (HypT₁). The case groups were compared with a control group of 26 healthy volunteers (C). Multivariate statistical analysis was performed with Partial Least Squares-Discriminant Analysis (PLS-DA). PLS-DA identified a distinct metabolic profile between C and untreated hyperthyroid patients (R²X 0.638, R²Y 0.932, Q² 0.783). Interestingly, a significant difference was also found between C and euthyroid patients after treatment (R²X 0.510, R²Y 0.838, Q² 0.607), while similar cluster emerged comparing HypT₀ vs HypT₁ patients. This study shows that metabolomic profile is deeply influenced by hyperthyroidism and this alteration persists after normalization of thyrotropin (TSH) and free thyroid hormone (FT3, FT4) concentration. This suggests that TSH, FT3 and FT4 assays may not be insufficient to detect long lasting peripheral effects of the thyroid hormones action. Further studies are needed to clarify whether and to what extent the evaluation of metabolomics profile may provide relevant information in the clinical management of hyperthyroidism.     

Characterization of the MHC class II region in cattle. The number of DQ genes varies between haplotypes

The organization of the major histocompatibility complex (MHC) class II region in cattle was investigated by Southern blot analysis using human probes corresponding to DO, DP, DQ, and DR genes. Exon-specific probes were also employed to facilitate the assessment of the number of different bovine class II genes. The results indicated the presence of single DO and DR genes, at least three DR genes, while the number of DQ genes was found to vary between MHC haplotypes. Four DQ haplotypes, DQ ^{1 1} to DQ ^{2 4}, possessed a single DQ and a single DQ gene whereas both these genes were duplicated in eight other haplotypes, DQ ^{3 5} to DQ ^{9 12}. No firm evidence for the presence of bovine DP genes was obtained. The same human probes were also used to investigate the genetic polymorphism of bovine class II genes. DQ DQ , DR DR , and DO restriction fragment length polymorphisms (RFLPs) were resolved and in particular the DQ restriction fragment patterns were highly polymorphic. Comparison of the present result with the current knowledge of the class II region in other mammalian species suggested that the DO, DP, DQ, DR, and DZ subdivision of the class II region was established already in the ancestor of mammals. The DP genes appear to be the least conserved class II genes among mammalian species and may have been lost in cattle. The degree of polymorphism of different class II genes, as revealed by RFLP analyses, shows striking similarities between species.     

Structure and spectral characteristics of diquat-cucurbituril complexes from density functional theory

Electronic structure, ¹H NMR and infrared spectra of diquat (6,7-dihydrodipyrido[1,2-b:1′,2′-e] pyrazine-5,8-diium or DQ²⁺) encapsulated by cucurbit[n]uril (n?=?7,8) hosts are obtained using the density functional theory. Theoretical calculations have shown that both CB[7] or CB[8] host possesses strong affinity toward DQ²⁺ compared to its reduced cation or neutral species. Calculated ¹H NMR spectra reveal that H_α protons on bi-pyridinium rings of DQ²⁺@CB[8] complex are de-shielded owing to C=O?H interactions. On the other hand aromatic (H_β and H_δ) of DQ²⁺ within the CB[8] cavity exhibit significant shielding. The complexation of CB[8] with DQ²⁺ splits the carbonyl stretching vibration (1788 cm^?1) into two distinct vibrations which correspond to 1765 cm^?1 arising from hydrogen bonded carbonyls and the 1792 cm^?1 band from non-interacting ones. Further, the CN stretching vibration in DQ²⁺ exhibits a frequency blue-shift of 6 cm^?1 on its encapsulation within the CB[8] cavity. The direction of frequency shift has been explained on the basis of natural bond orbital analyses.

Metabolic profiling of cervical tubercular lymphadenitis tissues by proton HR-MAS NMR spectroscopy

Proton metabolic profiling of incisional biopsied cervical lymph node tissue specimens of 109 patients suffering from tubercular (CTBL) and non-specific (NSCLA) lymphadenitis were analyzed by high resolution magic angle spinning (HR-MAS) NMR spectroscopy. In the present study, 40 endogenous metabolites namely, myo-inositol (m-Ins), branched chain amino acids (BCAA), glutamate, serine, taurine (Tau) aromatic amino acids, choline (Cho) containing compounds and glucose were characterized. To the best of our knowledge, this is the first report on metabolic profiling of cervical tubercular lymph node tissues using HR-MAS NMR spectroscopy. The principal component analysis revealed a clear discrimination between CTBL and NSCLA tissues. Increase in the concentration of mobile poly unsaturated fatty acids, BCAA, Cho, Tau, glycine and a decrease in the concentration of lactate, phosphocholine and m-Ins was observed in CTBL cases. The partial least square discriminant analysis (PLS-DA) with R ² = 0.95 and Q ² = 0.92 provided >98 % of correct classification between the two groups. A PLS-DA training set model of 75 % (CTBL = 54, NSCLA = 27) of the subjects when subjected for prediction of 25 % cases (CTBL = 18, NSCLA = 10) as an unknown dataset provided more than 98 % of diagnostic accuracy in their respective histological categories. The receiver operator characteristic curve was generated from PLS-DA factor-1 projected an area under the curve of 0.962. The metabolic profile obtained from HR-MAS NMR spectroscopy may be used as surrogate markers in vivo MRS for differentiating between CTBL and NSCLA cases non-invasively.     

Metabolomics Reveals Dynamic Metabolic Changes Associated with Age in Early Childhood

Objectives

A detailed understanding of the metabolic processes governing rapid growth in early life is still lacking. The aim of this study was to investigate the age-related metabolic changes in healthy children throughout early childhood.

Methods

Healthy children from a birth cohort were enrolled in this study from birth through 4 years of age. Urinary metabolites were assessed at 6 months, and 1, 2, 3, and 4 yr of age by using ¹H-nuclear magnetic resonance (NMR) spectroscopy coupled with multivariate statistical analysis including principal components analysis (PCA) and partial least-squares discriminant analysis (PLS-DA). Metabolic pathway analysis was performed using the MetPA web tool.

Results

A total of 105 urine samples from 30 healthy children were collected and analyzed. Metabolites contributing to the discrimination between age groups were identified by using supervised PLS-DA (Q² = 0.60; R² = 0.66). A significantly higher urinary trimethylamine N-oxide (TMAO) and betaine level was found in children aged 6 months. Urinary glycine and glutamine levels declined significantly after 6 months of age and there was a concomitant compensatory increase in urinary creatine and creatinine. Metabolic pathway analysis using MetPA revealed similar nitrogen metabolism associated energy production across all ages assessed. Pathways associated with amino acid metabolism were significantly different between infants aged 6 months and 1 year, whereas pathways associated with carbohydrate metabolism were significantly different between children at ages 2 and 3 years.

Conclusions

Urine metabolomics ideally represents dynamic metabolic changes across age. Urinary metabolic profiles change significantly within the first year of life, which can potentially provide crucial information about infant nutrition and growth.     

Will urinary biomarkers provide a breakthrough in diagnosing cardiac surgery-associated AKI? – A systematic review

Abstract

Introduction: Acute kidney injury following cardiac surgery is a dreaded complication contributing to early mortality. Diagnosing AKI using serum creatinine usually results in a delay. To combat this, certain kidney damage specific biomarkers were investigated to identify if they can serve as early predictors of cardiac surgery-associated AKI (CSA-AKI). This study systematically reviews three such biomarkers; NGAL, tissue inhibitor of matrix metalloproteinase-2 (TIMP-2) and insulin-like growth factor binding protein-7 (IGFBP7) to identify if they can serve as early predictors of CSA-AKI.

Methods: Systematic search was carried out on literature reporting the diagnostic ability of the three biomarkers from databases in accordance with PRISMA guidelines.

Results: We found 43 articles reporting urinary-NGAL levels (n?=?34 in adults, n?=?9 in children) and 10 studies reporting TIMP-2 and IGFBP7 levels among adults. Interestingly, NGAL showed high diagnostic value in predicting AKI in children (seven among nine studies with AUROC?>?0.8). The cell cycle arrest biomarkers, namely TIMP-2 and IGFBP7, showed high diagnostic value in predicting AKI in adults (five among ten studies with AUROC?>?0.8).

Conclusion: In predicting CSA-AKI; the diagnostic value of NGAL is high in the paediatric population while the diagnostic value of TIMP-2 and IGFBP7 is high in adults.     

Metabolic Changes in Urine during and after Pregnancy in a Large,Multiethnic Population-Based Cohort Study of Gestational Diabetes

This study aims to identify novel markers for gestational diabetes (GDM) in the biochemical profile of maternal urine using NMR metabolomics. It also catalogs the general effects of pregnancy and delivery on the urine profile. Urine samples were collected at three time points (visit V1: gestational week 8–20; V2: week 28±2; V3∶10–16 weeks post partum) from participants in the STORK Groruddalen program, a prospective, multiethnic cohort study of 823 healthy, pregnant women in Oslo, Norway, and analyzed using ¹H-NMR spectroscopy. Metabolites were identified and quantified where possible. PCA, PLS-DA and univariate statistics were applied and found substantial differences between the time points, dominated by a steady increase of urinary lactose concentrations, and an increase during pregnancy and subsequent dramatic reduction of several unidentified NMR signals between 0.5 and 1.1 ppm. Multivariate methods could not reliably identify GDM cases based on the WHO or graded criteria based on IADPSG definitions, indicating that the pattern of urinary metabolites above micromolar concentrations is not influenced strongly and consistently enough by the disease. However, univariate analysis suggests elevated mean citrate concentrations with increasing hyperglycemia. Multivariate classification with respect to ethnic background produced weak but statistically significant models. These results suggest that although NMR-based metabolomics can monitor changes in the urinary excretion profile of pregnant women, it may not be a prudent choice for the study of GDM.     

On the relationship of anthranilic derivatives structure and the FXR (Farnesoid X receptor) agonist activity

Farnesoid X receptor (FXR) is a nuclear receptor related to lipid and glucose homeostasis and is considered an important molecular target to treatment of metabolic diseases as diabetes, dyslipidemia, and liver cancer. Nowadays, there are several FXR agonists reported in the literature and some of it in clinical trials for liver disorders. Herein, a compound series was employed to generate QSAR models to better understand the structural basis for FXR activation by anthranilic acid derivatives (AADs). Furthermore, here we evaluate the inclusion of the standard deviation (SD) of EC₅₀ values in QSAR models quality. Comparison between the use of experimental variance plus average values in model construction with the standard method of model generation that considers only the average values was performed. 2D and 3D QSAR models based on the AAD data set including SD values showed similar molecular interpretation maps and quality (Q²_LOO, Q²_(F2), and Q²_(F3)), when compared to models based only on average values. SD-based models revealed more accurate predictions for the set of test compounds, with lower mean absolute error indices as well as more residuals near zero. Additionally, the visual interpretation of different QSAR approaches agrees with experimental data, highlighting key elements for understanding the biological activity of AADs. The approach using standard deviation values may offer new possibilities for generating more accurate QSAR models based on available experimental data.     

Characterization of second site mutations show that fast proton transfer to QB − is restored in bacterial reaction centers of Rhodobacter sphaeroides containing the Asp-L213 → Asn lesion

The structural basis for proton coupled electron transfer to Q_B in bacterial reaction centers (RCs) was studied by investigating RCs containing second site suppressor mutations (Asn M44 Asp, Arg M233 Cys, Arg H177 His) that complement the effects of the deleterious Asp L213 Asn mutation [DN(L213)]. The suppressor RCs all showed an increased proton coupled electron transfer rate k _AB ⁽²⁾(Q_A ^–Q_B ^– + H⁺ Q_AQ_BH^–) by at least 10³ (pH 7.5) and a recombination rate k _BD (D⁺Q_AQ_B ^– DQ_AQ_B) 15–40 times larger than the value found in DN(L213) RCs. Proton transfer was studied by measuring the dependence of k _AB ⁽²⁾ on the free energy for electron transfer (G_et). k _AB ⁽²⁾ was independent of G_et in DN(L213) RCs, but dependent on G_et in native and all suppressor RCs. This shows that proton transfer limits the k _AB ⁽²⁾ reaction with a rate of 0.1s^–1 in DN(L213) RCs but is not rate limiting and at least 10⁸-fold faster in native and 10⁵-fold faster in the suppressor RCs. The increased rate of proton transfer by the suppressor mutations are proposed to be due to: (i) a reduction in the barrier to proton transfer by providing a more negative electrostatic potential near Q_B ^–; and/or (ii) structural changes that permit fast proton transfer through the network of protonatable residues and water molecules near Q_B.     

Heterogeneity inHLA-DR2-relatedDR,DQ haplotypes in eight populations of Asia-Oceania

The relative distributions of 480DR2-relatedDR, DQ haplotypes have been determined in Australian Aborigines, Papua New Guinean Highlanders, coastal Melanesians, Micronesians, Polynesians, Javanese, and Southern and Northern Chinese. Using sequence-specific oligonucleotides (SSOs) for hybridization of polymerase chain reaction (PCR) products from DRBI,DRBS,DQA1, andDQBI genes, 15 differentDR2-related haplotypes were identified. The predominantDR2 haplotype in Oceania involved a novel combination ofDRBI ^* 1502,DRB5 ^* 0101 alleles; this haplotype occurred sporadically in Java, but not in China. In Southern China, the most frequent DR2 haplotype involved the unusual arrangementDRB1 ^* 1602,DRB5 ^* 0101; alternatively,DRB1 ^* 1602 was associated with a newDRB5 SSO pattern. This study has important implications for molecular HLA-typing protocols that assume particularDRB1 DRB5 orDR,DQ linkage relationships. Further, the novelDRBI,DRB5 haplotype in Oceania suggests that the mixed lymphocyte culture (MLC) determinants Dw2 and Dw12 are discriminated by codon 86 at theDRBI locus.     

Allelic forms of the alpha- and beta-chain genes encoding DQw1-positive heterodimers

On chromosome 6, in the HLA region, the DQ subregion is located immediately centromeric to the DR subregion. Even though only three serological specificities to date have been officially recognized (DQwl, DQw2, and DQw3), it seems likely that the phenotypical polymorphism expressed by DQ molecules is much more complex. There are reasons to believe that fixed alpha-beta combinations exist, each of them associated with a different DR allele. DQw1 is a determinant present on DQ molecules that are found associated with DRI-, DR2-, and DRw6-positive haplotypes. By restriction fragment length polymorphism analysis, we recognized three allelic DQ-alpha and three allelic DQ-beta patterns associated with DQw1 . In addition, one of these alpha/beta pairs associated with DR1, two with DR2, and a fourth with DRw6. We have obtained evidence using nucleotide sequencing that there are as many allelic forms of DQ-alpha and DQ-beta genes as there are different molecular DQ-alpha and DQ-beta patterns. The DQ-alpha and DQ-beta chains of DQwl-positive molecules each are encoded by at least three distinctly different allelic genes, and particular alpha/beta gene combinations are associated with the same DR alleles as their corresponding molecular alpha/beta pairs.     

Identification of putative biomarkers for leptomeningeal invasion in B-cell non-Hodgkin lymphoma by NMR metabolomics

Introduction

B-cell non-Hodgkin lymphoma (B-NHL) is the most common hematological malignancy and different genetic alterations are frequently detected in transformed B lymphocytes. Within this heterogeneous disease, certain aggressive subgroups have an increased risk of central nervous system (CNS) involvement at diagnosis and/or relapse, resulting in parenchymal or leptomeningeal infiltration (LI) in 5–15% of cases. The current sensitivity limitations of cerebrospinal fluid (CSF) cytology and contrast-enhanced MRI for CNS involvement, mainly at early stages, motivates the search for alternative diagnostic methods.

Objectives

Here we aim at using untargeted ¹H-NMR metabolomics to identify putative biomarkers for LI in B-NHL patients.

Methods

CSF and peripheral blood samples were obtained from B-NHL patients with a positive (n?=?7, LI group) or negative LI diagnostic (n?=?13, control group). For seven patients, CSF samples were collected during the course of intrathecal chemotherapy, making it possible to assess the patient´s response to treatment. ¹H-NMR spectra were acquired and statistical multivariate and univariate analysis were performed to identify significant alterations.

Results

Significant metabolite differences were found between LI and control groups in CSF, but not in serum. A predictive PLS-DA cross-validated model identified significant pool changes in glycine, alanine, pyruvate, acetylcarnitine, carnitine, and phenylalanine. Additionally, increments in protein signals were detected in the LI group. Significantly, the PLS-DA model predicted correctly all samples obtained from the group of patients in remission during LI treatment.

Conclusions

The results show that the CSF NMR-metabolomics approach is a promising complementary method in clinical diagnosis and treatment follow-up of LI in B-NHL patients.

     

Validation of a urine metabolome fingerprint in dog for phenotypic classification

Selective breeding of dogs over hundreds of years has inadvertently resulted in breed-specific propensities to particular diseases. Furthermore, it has likely induced more subtle affects on the physiology of certain breeds and moved them from their evolutionary optima. In the absence of obvious disease phenotypes such subtle changes could have yet unrecognised breed-specific implications for health and well-being. Here we have applied NMR metabolomics as a discovery-driven approach to identify the impact of breed on the urinary profile of dog and to determine if non-disease-related breed differences can be identified. Multiple urines were collected non-invasively over a two-week period from seven neutered male Labrador retrievers and miniature Schnauzers. Following NMR analyses by 1-dimensional ¹H and 2-dimensional ¹H J-resolved (JRES) spectroscopy, principal component analysis revealed that the metabolic variability within each individual is relatively small compared to inter-individual variability, and that some separation between breeds was evident. A supervised model, using partial least squares discriminant analysis (PLS-DA) with class based upon breed, was trained using the JRES data. The model predicted correctly the breed of seven additional urines, yielding a model sensitivity and specificity of 100%. Several significant metabolic differences between the breeds were identified. A second model was developed using PLS-DA with class based upon individual dogs, which again achieved high classification accuracy for the test set. Overall, this confirms that canine urine is information-rich and that breed is a major determinant of urinary metabolic fingerprints. In the future this may enable a more accurate development of specific nutritional care for an individual or breed.