Nuclear Magnetic Resonance-based Metabolomics of Blood Plasma from Dairy Calves Infected with the Main Causal Agents of Bovine Respiratory Disease (BRD)

Each year, Bovine Respiratory Disease (BRD) results in signicant economic loss in the cattle sector, and novel metabolic proling and early diagnosis techniques represent a promising tool for developing effective measures for disease management. Here, proton - Nuclear Magnetic Resonance ( 1 H - NMR) spectra were used to characterize metabolites from blood plasma collected from dairy calves intentionally infected with the main BRD causal agents, bovine respiratory syncytial virus (BRSV) and Mannheimia haemolytica (MH), to generate a well-dened metabolomic prole under controlled conditions. In response to infection, 42 metabolites (BRSV = 27, MH = 24) changed in concentration compared to the Baseline (non-infected) state. Fuel substrates and products exhibited a particularly strong effect, reecting imbalances that occur during the immune response. Glucose levels decreased only during bacterial infection, suggesting that the clinical signs of bacterial BRD are more energetically taxing than those of viral BRD. Furthermore, 1 H - NMR spectra from Baseline and Infected samples were discriminated with an accuracy, sensitivity, and specicity ≥ 95% using chemometrics to model the changes associated with disease, suggesting that metabolic proles can be used for further development and validation of diagnostic tools.


Introduction
One of the most severe and costly health problems worldwide is bovine respiratory disease (BRD), a disease complex caused by numerous microbial pathogens [1][2][3][4] . Prevalent causal agents for BRD include viral (bovine herpes-virus type 1, bovine respiratory syncytial virus, bovine viral diarrhea virus, parain uenza-3 virus, and bovine coronavirus), and bacterial (Mannheimia haemolytica, Pasteurella multocida, Haemophilus somnus, Mycoplasma bovis) pathogens 1,5 . Bovine respiratory syncytial virus (BRSV) is a major cause of respiratory disease in young calves (≤ 1 year). This viral infection can be asymptomatic and can involve the upper and lower respiratory tracts 6,7 . BRSV typically initiates infection in response to physiological and environmental stressors, suppressing the host's defense mechanisms and predisposing the replication, inhalation, and colonization of the lungs by M. haemolytica, a microorganism found in the normal ora of the upper respiratory system in ruminants 1,8,9 . In general, viral and bacterial agents activate innate immunity, which is comprised of nonspeci c defense mechanisms that are triggered shortly after the appearance of the antigen 10 . This occurs as a result of the production of a motif of molecules expressed by the pathogen known as pathogen -associated molecular patterns (PAMPs) 11,12 . In the case of BRSV the PAMPs are known components consisting of glycoprotein G, fusion protein F, and single -stranded RNA 6 . For M. haemolytica, PAMPs include agellin, lipopolysaccharide (LPS) complex, and leukotoxin (LKT) 13 . These elicitors are identi ed by pattern recognition receptors (PRRs) for rapid detection of the threat from a potential pathogen 11 . Surface-bound and intracellular PRRs, such as Toll-like receptors (TLRs), nucleotide-binding and oligomerization domain (NOD) -like receptors, and RNA helicases, are expressed by bovine respiratory tract cells 12 . The engagement of PAMPs by PRRs results in the production of damage-associated molecular patterns (DAMPs), initiating ATP -dependent signaling cascades. Activated transcription factors induce the production of in ammatory cytokines and chemokines for release into the body, which attracts neutrophils, macrophages, and lymphocytes to the respiratory tract, resulting in respiratory disease 6,9,11 .
The standard in -eld method for BRD detection is the visual -clinical diagnosis (VCD) of cattle appearance and behavior following clinical signs such as high temperature, respiratory discomfort, and other factors considered problematic when determining the condition of infected calves 14 . This diagnostic methodology has less than 65% sensitivity, meaning that around 35% of infections are not detected, and 65% speci city, such that 35% of healthy cattle are unnecessarily treated 15,16 . Thus, alternative methods to VCD are needed to pro le the infection and de nitively detect BRD before its manifestation, which will enable farmers to respond with the proper prophylactic measures 17 . A comprehensive phenotypic assessment for BRD using a speci c "omics'' platform could be obtained from blood plasma or other bio uids. To this end, metabolomics pro ling of disease state using nuclear magnetic resonance (NMR) is a means to validate both the traditional diagnostic approach (i.e., VCD) and the biochemical processes occurring throughout infection, which may be detectable using more timely and e cient diagnostic techniques such as Near Infrared Spectroscopy (NIRS) [17][18][19] .
NMR -based metabolomics provides a window into metabolic mechanisms by combining highthroughput analytical chemistry with multivariate data analysis (MVA) 20 to identify and quantify changes in metabolic products of a biological system 21,22 . Proton NMR ( 1 H -NMR) spectra arise from active nuclei absorbing electromagnetic energy at the frequencies speci c to the 1 H nucleus, resulting in resonance within a two -level quantum system 23 . This resonance frequency, along with the signal intensity, is speci c to the local covalent -bonding structure and chemical environment and is re ected in the manifold of information -rich signals (chemical shifts) in NMR spectra 23,24 . Blood plasma is the most frequently used bio uid for NMR studies 17,18,25 , and consists of the protein -rich fraction of blood in which white blood cells (WBC), red blood cells (RBC), platelets (PTL), and coagulation factors are suspended before blood fractionation with an anticoagulant 26 . Bovine blood plasma is composed of water (92%); albumin and globulins (3%) required to maintain colloidal osmotic pressure; immunoglobulins (4%) used for signal transduction and response to antigens; coagulants and brinogen (0.4%) that aid in blood clotting; minerals (0.5%) required to maintain blood pH, and lipids (0.07%) associated with hormone content and nutrition. Blood plasma is commonly used to diagnose viral or bacterial infections by detecting antigens or pathogen-speci c antibodies using ELISA (Enzyme-Linked Immuno-Sorbent Assay) 26 . Recently, blood plasma was shown to be a suitable medium for detecting M. haemolytica infection using NIRS 19 .
In cattle, NMR has been used to conduct metabolic pro ling and diagnosis of both reproductive and nutritional disorders [27][28][29][30][31][32] . In one case, seven plasma metabolites (alanine, arginine, choline, isoleucine, leucine, phosphatidyl choline, and valine) were shown to signi cantly decrease in dairy cows during estrous compared to cows in anestrous. These changes were related to glucose, triglyceride, and amino acid metabolic pathways associated with postpartum anestrus 27 . Similarly, changes in the concentration of metabolites in blood plasma were observed in Holstein cows during postpartum and lactation periods, revealing that glucose is rerouted to synthesize lactose and fats in milk, causing the lactating cow to produce ketone bodies as an alternative energy source to maintain homeostasis 28-31 . Metabolic pro les related to fatty liver disease in lactating cows were correlated with increases in β-hydroxybutyric acid, acetone, citrulline, glycine, isobutyrate, trimethylamine-N-oxide, and valine, and decreases in γaminobutyric acid glycerol, alanine, asparagine, creatinine, and glucose, suggesting this metabolic disorder alters the concentration of metabolites related to energy imbalance pathways 32 . In contrast, NMR analysis revealed calves with bronchopneumonia detected by VCD exhibited increases in 2-methyl glutarate, phenylalanine, phosphatidylcholine, but decreases in acetate, allantoin, cholesterol, dimethyl sulfone, ethanol, propionate, and free cholesterol in the blood plasma, suggesting alteration of a different set of metabolic pathways 33 . Recently, feedlot cattle that were deemed to have BRD through VCD inspection were shown to have signi cant alterations in the concentration of α-glucose chains, hydroxybutyrate, and phenylalanine by NMR analysis of blood plasma 34 . Results from both of these studies 33,34 , combined with recent work in which NIRS pro ling of blood plasma of calves with induced M. haemolytica infection 19 , indicate that characteristic shifts in the metabolome of blood plasma may be indicative of BRD infection and perhaps pathogenic speci city.
Here, we conduct bovine infection trials, known as challenge studies, in which dairy calves were intentionally infected with the main BRD causal agents, BRSV and M. haemolytica (MH), in order to generate a well -de ned metabolomic pro le under controlled conditions. 1 H -NMR analysis of the collected blood plasma was used to (1) identify metabolites associated with infection by the two different pathogens, (2) assess concentration changes of those metabolites between the healthy and infected stages and in response to each pathogen, (3) generate a model for discriminating 1 H -NMR spectra, and the metabolites involved in the differentiation of healthy and infected calves for each causal agent, and (4) provide new biochemical information to the current blood plasma NMR -BRD metabolome, which was acquired in non-controlled eld conditions 33,34 , in order to provide insight into quanti able differences between causal agents that might be targeted for the development and validation of novel BRD management strategies.

Materials And Methods
Animals and controlled challenges.
Ten non -immunized Holstein steers were subjected to two controlled challenge studies, each with a different infectious agent. The rst group of dairy calves (n = 5) was challenged with M. haemolytica (isolate D153) via bronchoalveolar lavage catheterization during the summer of 2019. The second group (n = 5) was challenged with BRSV (GA -1, P5) delivered by a nebulizer (DeVilbiss Pulmo -Neb) through a custom -made face mask during the fall of the same year. The detailed procedures of pathogen preparation and challenge administration are described in the supplementary information (Supplementary Methods S1 online). The calves were sheltered at Mississippi State University (MSU), and the experiments were carried out with the approval of the MSU-Institutional Animal Care and Use Committee. All methods were performed following MSU-IACUC guidelines and regulations (IACUC-19-037) and reported in compliance with the ARRIVE (Animal Research: Reporting In Vivo Experiments) guidelines 2.0. VCD and complete blood counts (CBC) were measured following published procedures for 27 days during the bacterial challenge and for 34 days during the viral challenge 19,34 . For the BRSV challenge study, three more days were assessed prior to inducing infection while awaiting serological results to con rm the absence of antibodies against this pathogen. In general, blood samples and VCD data were collected prior to induced infection for four days, after the infection for 11 continuous days, and then every other day from days 12 -23 post -challenge.

Blood acquisition.
Blood samples (n = 202) were drawn via jugular venipuncture and immediately placed on ice in two collection tubes containing the anticoagulant EDTA (ethylenediaminetetraacetic acid). The rst tube was centrifuged at 4000 rpm for 20 minutes to separate plasma, and duplicates of 1 mL were stored at -80°C until NMR analysis. The second tube was used for CBC, where RBC, hematocrit (HTC), hemoglobin (HGB), WBC, and PTL contents were acquired using a veterinary hematology analyzer. In addition, microscopic differential counts of WBC were performed to assess the variability of neutrophils, eosinophils, basophils, monocytes, and lymphocytes. The viral and bacterial challenge studies yielded a total of 97 and 105 blood samples, respectively. Based on VCD as the reference method, blood samples were classi ed as Baseline, Asymptomatic, Infected, Treated, or Recovered. To avoid interference of antibiotics or the recovery processes in the interpretation of the NMR metabolomic pro les, only data from the samples designated as Baseline (n = 55) and Infected (n = 47) were used in the univariate and multivariate analyses (Table 1). Preparation of blood plasma for 1 H -NMR analysis.

H -NMR spectra collection.
For each pathogen challenge, an even number of Baseline and Infected blood plasma samples (n = 70) were chosen for NMR analysis to ensure homogeneity of the variance and weight of each data set for the statistical analyses (Table 1); the detailed information from these samples can be found in the supplementary material (Supplementary Table S1 online). A Bruker Avance III HD 500 MHz spectrometer out tted with a 5 -mm BBFO probe (Bruker, Massachusetts, USA) was used for NMR spectroscopy.
Samples were run in automation mode with a SampleJet, with all samples refrigerated at 4°C until just prior to loading. A perfect -echo WATERGATE sequence (PE -WATERGATE, parameter set ZGESGPPE) was applied to collect the data 36 . Data were collected at 298 K for 128 scans with a 1 s inter -scan delay and a 3 s per -scan acquisition time. The total acquisition time for each sample, including 3D -shimming, was about 15 minutes. Topspin 4.0.8 (Bruker, Massachusetts, USA) was used for spectral processing following the acquisition. All spectra were zero -lled to 128k points, and a 1 Hz line broadening was used. Automatic phasing and baseline correction were performed, 1 H -NMR spectra were segmented into successive non -overlapping regions of 0.0001 -ppm chemical shifts between 0.0 and 10.5 ppm, and the water region was truncated between 4.30 to 5.10 ppm. The processed spectra were then imported into the Chenomx NMR Suite 8.6 (Chenomx, Edmonton, Canada) to identify individual metabolites using a reference library containing 338 metabolites for 500 MHz spectrometers. Because the Chenomx library was acquired using a NOESY -based 1D pulse sequence, the reported metabolite concentrations (mM) are likely to differ slightly from the actual concentrations. However, relative differences between samples will be preserved, and absolute concentration differences for the majority of compounds are expected 37 .

Statistics.
The mean and standard deviation (SD) for VCD and CBC parameters collected from the Infected stage of the BRSV challenge (n = 21), the M. haemolytica study (n = 26), and the combined Baseline data points from both challenge studies (n = 55) were calculated using univariate statistics. The same parameters were calculated for the concentration (mM) of metabolites chosen by comparing those found in this investigation to those previously reported for blood plasma in bovines and other mammals 17,25,38 . Metabolites related to the referencing solution, diet ingredients, or only detected in less than two samples per disease stage (Baseline or Infected) were excluded from the analyses. ANOVA and pairwise mean comparison (Baseline vs. BRSV, Baseline vs. MH, BRSV vs. MH) using Tukey -Kramer HSD (honestly signi cant difference) test with alpha = 0.05 were used to assess for signi cance in parameter response for VCD, CBC, and metabolite concentration between the Baseline and Infected categories from both challenges. The results for these post -hoc tests were reported with connecting letters, with different letters indicating signi cant differences between the Baseline, BRSV, and MH categories (JMP® 14.0 SAS Institute Inc., NC. USA). In addition, a database of metabolite concentrations representing a general state of infection was composed of data collected from all Infection samples, regardless of pathogenic agent (n = 47), and was evaluated by a pairwise mean comparison (Baseline vs. Infected) using Student's ttest with alpha = 0.05.

Multivariate Analysis (MVA).
The VCD and CBC parameters, and the metabolite concentrations, were subjected to Principal Component Analysis (PCA) using full cross-validation and algorithm -SVD (Singular Value Decomposition) to obtain correlation loadings plots to determine the magnitude and direction of a particular constituent's contribution (In uence) to the models created for the Baseline and Infected stages (Unscrambler® v. 11, Aspen Technology Inc., Massachusetts, USA). The processed 1 H -NMR spectra contained spectral peaks ranging from 0.5 to 9.0 ppm that were chosen for the chemometrics -based MVA; peaks from free EDTA at 3.2 ppm and Ca 2+ -EDTA at 3.6 ppm were removed prior to analysis 39,40 . Three balanced datasets of spectra were created: the rst named Infected (n = 70) was created by combining information from both studies; the second corresponded to the BRSV challenge (n = 40), and the third to the M. haemolytica challenge (n = 30). SIMCA software -omics skin v.15.0 (Umetrics AB, Ume, Sweden) was used to apply pattern recognition methods. 1 H -NMR spectra were subjected to PCA, in which the scale data conversion with mean-center scaling re ects the total metabolic differences between the two groups (Baseline and Infected) as well as the degree of variability within each group 41 .
The 1 H -NMR spectra of blood plasma were analyzed using orthogonal partial least-squares discriminant analysis (OPLS -DA). The OPLS -DA models were built with t[1]P and t [2]O, which stand for the rst principal component and the second orthogonal component, respectively 41 . The OPLS -DA models were used to maximize the covariance between the measured data (X variable, peak intensities in 1 H -NMR spectra) and the response variable (Y variable, predictive classi cations). The quality of each model was assessed using R 2 X, R 2 , R 2 Y, and Q 2 , where R 2 X denotes the degree of optimization of the analytical model, R 2 symbolizes the coe cient of determination, R 2 Y denotes the percentage of variance explained by the model, and Q 2 describes the model's cumulative prediction 41,42 . In the OPLS -DA models, a 7segment cross -validation was used to estimate the optimal component number of each model to avoid over -tting. In addition, the percentage of accuracy, sensitivity, and speci city were calculated to test the ability of each model to identify true positive and true negative samples correctly 43 . Permutation analysis (n = 200) was used to validate the OPLS -DA models and to assess model reliability. In permutation analysis a model is considered reliable if R 2 > 0.4 and the intercept Q 2 < 0 41,42,44 .

Results
Clinical and hematological results.
Overall, all the calves showed similar patterns of health and disease during the controlled studies (Fig. 1). The typical activation of the innate immunity or nonspeci c defense mechanisms following the bacterial challenge were observed after 24 hours of infection; the increase in rectal temperature (Fever) and WBC ( Fig. 1a,b) is likely due to the release and recognition of the LPS complex and other pyrogenic features of M. haemolytica in the respiratory tract cells 45 . On Day 3, after inducing infection, severe signs of disease were detected in four dairy calves (ID 6, 8, 9, and 10) which were then treated over ve days with a broadspectrum antibiotic (Ceftiofur). The treatment caused the recovery of the calves and the change in the patterns for rectal temperature (TEMP) and WBC (Fig. 1a,b). Mild signs of infection from Day 1 until Day 19 were observed in calf 7; after D19, his signs intensi ed, and antibiotics were given. By contrast, dairy calves challenged with BRSV exhibited increased TEMP and WBC on Day 7 post-infection (Fig. 1a,b), which corresponds to the end of the asymptomatic phase, characteristic of the incubation period of this virus 6 . In this challenge study, calf 2 died on Day 8 due to disease complications.
The results of the univariate and multivariate tests for the VCD and CBC values for both challenges are shown in Fig. 2. In terms of the univariate analysis during the BRSV challenge, nine parameters (TEMP, RR, RBC, HCT, WBC, neutrophils, lymphocytes, eosinophils, and basophils) were statistically different (p < 0.05) during Infection compared to the Baseline, re ective of the innate immune response to virus replication in the respiratory cells 6 . Throughout the M. haemolytica infection, four of the evaluated variables (TEMP, RR, HCT, and PTL) were statistically different (p < 0.05) from the Baseline. These ndings can be associated with the hemorrhage and edema known to be caused by the LPS complex produced by this bacterium 46 .
The PCA correlation loading plots for VCD and CBC (Fig. 3) show the relationships between the potential explanatory variables (X -matrix) within each of the databases (Baseline, Infected, BRSV, and MH) for the development of models 47 . The inner ellipse represents 50%, while the outer ellipse represents 100% of the explained variance for the individual variables. Thus, the area between the two ellipses explains 50 to 100% of the variance, implying that the centralized parameters (inside the inner circle) have an unimportant effect on the differentiation of each X -matrix. In contrast, those inside the outer circle (shaded) present a strong in uence or signi cant impact on the differentiation of each model 47,48 . In addition, when the variables are placed in the positive or negative direction of the rst principal component (PC -1), this in uence can be described as positively or negatively correlated within each Xmatrix, meaning the variables in those directions increase or decrease together to generate each model's characteristic patterns 47,48 .
The distribution of VCD and CBC variables for the Baseline model can be seen in Fig. 3a; this pattern was compared with those obtained in the Infected stage of each challenge separated and together. When evaluating the general infection database (Fig. 3b), a similar pattern was observed in comparison to the Baseline, with HGB negatively correlated, differing from the non-infected stage when it was found to have no in uence in the model. For the BRSV infection model (Fig. 3c), TEMP, HGB, and HCT were positively correlated in comparison to the Baseline. The model for the evaluated variables during the M. haemolytica infection (Fig. 3d) differed from the Baseline in which TEMP, RBC, HGB, HCT, and PLT were positively correlated.
The representative 500 MHz 1 H -NMR spectra of blood plasma collected with EDTA from dairy calves before and after the controlled infections are shown in Fig. 4. A total of 179 metabolites were identi ed using the batch pro ler option in the Chenomx library. From these, a total of 72 metabolites were selected for the statistical analyses following the criteria explained in the methodology. After the selection, the presence of these metabolites was reliably identi ed in the samples manually using Chenomx. These compounds reproducibly appeared as well -resolved signals in the 1 H -NMR spectra, and overlapped signals were con rmed using at least two peak groups before tting in the Chenomx pro ler. After the manual con rmation of these metabolites, on average 43 ± 8 of the selected metabolites were detected in each sample (Supplementary Table S1  The ndings from the univariate and multivariate analyses are shown in Fig. 5 and Fig. 6. Results for the general Infection database, which included data points from both viral and bacterial challenges, can be seen in Fig. 5. Regardless of the causative agent, 11 metabolites were found to change signi cantly (p < 0.05) compared to Baseline levels, where increases in 2-hydroxybutyrate, acetone, 3-hydroxyisobutyrate, and dimethyl sulfone, and decreases in succinylacetone, isobutyrate, 2-hydroxyvalerate, O-acetylcholine, and isoleucine, allantoin, and ethanol, were observed. Examining the response to speci c pathogenic agents revealed the concentrations of a subset of seven metabolites (BRSV = 2, MH = 5) changed signi cantly (p < 0.05) compared to the Baseline (Fig. 6). Infection with BRSV resulted in a signi cant (p < 0.05) increase in guanidoacetate and a decrease in ethanol. On the other hand, infection with M. haemolytica showed signi cant (p < 0.05) increases in concentrations of 2-hydroxybutyrate, acetone, 3hydroxyisobutyrate, and π-methylhistidine, and a decrease in isobutyrate.
The PCA scores plots showing the trends of the 1 H -NMR spectral signals can be seen in Fig. 8. In the scores plot, each point corresponds to either a Baseline or Infected sample, and overlap between the two groups indicates similarities in the metabolite composition of the blood plasma. Outliers are samples outside the con dence ellipse based on Hotelling's T 2 (signi cance level 0.05). The database for the general infection was analyzed with 10 principal components (Fig. 8a) giving modeling parameters R 2 X = 0.83 and Q 2 = 0.56; in this case, three outliers were identi ed, which corresponded to samples 33 (D9, calf 5), 34 (D11, calf 5), and 39 (D-2, calf 7). In the database of BRSV samples (Fig. 8b), 7 principal components gave a degree of optimization R 2 X = 0.80 and a cumulative prediction Q 2 = 0.38, with two outliers: samples 22 (D4, calf 2) and 34 (D11, calf 5). In the database composed of M. haemolytica samples (Fig. 8c), 6 principal components produced modeling parameters R 2 X = 0.79 and a Q 2 = 0.48. The outlier corresponded to Baseline sample 39 (D -2, calf 7).
The OPLS -DA scores plots (Fig. 9) for the general infection (Fig. 9a), BRSV (Fig. 9c), and M. haemolytica (Fig. 6e) databases demonstrated a clear distinction in the chemistry of blood plasma from non-infected and infected cattle. This suggests that even though the chemical composition of the blood plasma is similar before and after infection, as shown in the PCA scores plots (Fig. 8), there is enough information to successfully detect and discriminate infection from baseline with an accuracy, sensitivity, and speci city higher than 95% (Table 2). More importantly, all the models showed a speci city of 100%, meaning no false -positive samples (Baseline samples) were classi ed as infected. Table 2 shows the quality parameters for the calibrations and validations of each model using only two principal components (PCs). The values of R 2 X (> 0.4), R 2 (> 0.9), R 2 Y (> 0.9), and Q 2 (> 0.4) obtained in the calibration indicated that the models are robust, reliable, and have a low risk of over tting 41,49 . In addition, R 2 values were greater than Q 2 in the permutation plots, such that a more positive slope in the regression line corresponds to a higher degree of t to the data and the reliability of each model 41,42,44 . The distance to model plots and permutation plots for the discriminant models can be found in the Supplementary material les (Supplementary Figure S1 online). Table 2 OPLS -DA model quality parameters for the classi cation of 1 H -NMR spectra from blood plasma collected before and after the controlled infections. Cal = calibration, Val = validation, R 2 X = degree of optimization, R 2 = coe cient of determination, R 2 Y = percentage of variation explained by the model, and Q 2 = cumulative prediction, t[1]P = rst principal component. The color map derived from the coe cient loading plot indicates signi cant changes in spectral signals which contribute to the trends in the rst principal component (t[1]P) of the OPLS -DA scores plot (Fig. 9) that distinguish Baseline from Infected samples. These spectral signals can be compared to known metabolite peaks, allowing them to be assigned in the loading plot. In Fig. 9

Discussion
In this study, 1 H -NMR was evaluated as a means to determine the metabolomics of BRD by detecting biomarkers in blood plasma after intentionally infecting dairy calves with the main causal agents of this disease under controlled conditions. The induced infections provided a larger and more accurate data set by including animals with limited or minor VCD signs but that were actually infected and metabolite information was veri ed with clinical physiological signs and hematological parameters recorded before and after the controlled infections. Similarly, the current blood plasma NMR -BRD metabolome, which was acquired in non-controlled eld conditions 33,34 , used VCD as the reference method to determine the infection, however the lack of identi cation of speci c causal agents restricted the discovery of specialized biomarkers that might be targeted for the development and validations of novel BRD management strategies such as NIRS-based diagnosis in the early stages of the disease. As a result, the current study adds to and expands on the existing metabolome by identifying unique biomarkers in cattle infected with BRSV and M. haemolytica independently.
Ruminants have a specialized digestive system to degrade grass and get the necessary nutrients to maintain homeostasis, and most of the glucose comes from the gluconeogenesis of oxaloacetate obtained from the propionate produced by the microorganism species Megasphaera, Veillonella, and Selenomonas in the rumen 50,51 . Because this is a slow process, supplying the energy demands during the cell signaling cascades and immune response caused by PAMPs recognition requires metabolism reprogramming by immune cells, where alternative energy sources such as triglycerides and proteins are used for ATP production [52][53][54] . Glucose is an upstream regulator of 26 genes associated with BRD, and if glucose homeostasis is disrupted, hypoglycemia or hyperglycemia occurs 55 . Decreases in glucose as observed in the present M. haemolytica challenge were previously reported in BRD studies as a result from natural or arti cially induced infections, LPS injections, stress-related to transport, and receiving calves 34,55,56 . It has been suggested that, in addition to the metabolic changes caused by the immunological response, the decrease in glucose levels is also due to the hypoglycemic effect of BRD and the decline in diet due to the discomfort caused by the respiratory signs 34 . Increases in glucose in response to viral infection, as shown in the BRSV challenge, are uncommon and have only been documented once in blood serum from cattle arti cially infected with bovine herpesvirus type 1 (BHV -1) and M. haemolytica 57 . This suggests that most of the known biochemical pro les for BRD, where glucose is reported to decrease and metabolites involved in ketosis appear to increase, could be primarily associated with the secondary bacterial infection characteristic of this disease.
To meet the energy demands due to the recruitment of in ammatory cells and the phagocytotic processes for microbial death, immune cells such as neutrophils, monocytes, macrophages, and lymphocytes undergo aerobic glycolysis 53,54 . In this scenario, pyruvate does not enter the mitochondrion but is instead metabolized to lactate in the cytoplasm, with glycolysis rapidly providing minor amounts of ATP 52,53 . Lactate increases were detected only during the BRSV challenge. Similarly, cattle that died from a combination of BHV -1 and M. haemolytica infection had higher lactate concentrations than those that survived 57 . It has been reported that the decrease in oxygen levels during BRD due to stress, blockage of the respiratory tract with mucous secretions, and lung affection also increased lactate concentration, the likelihood of disease progression, and eventual mortality in dairy cattle 58 . Aerobic metabolism in neutrophils is associated with an increase in reactive oxygen species (ROS), an important source of bactericidal activity 53 . Increases in 2-hydroxybutyrate, which is a metabolite associated with ROS production and lipid oxidation (ß-oxidation), was detected here in the M. haemolytica infection and has been previously linked to diabetes, as well as being a predictor of insulin and glucose resistance, causing proin ammatory responses and increased oxidative stress 59,60 .
To offset the consequences of metabolic diseases involving energy imbalance, ruminants and, more speci cally, bovine species are known to use the alternative triglyceride route [27][28][29][30]32 . Triglycerides are made up of glycerol and fatty acids. Fatty acids undergo ß-oxidation in the liver, producing acetyl -CoA, which enters the Tricarboxylic acid (TCA) cycle. As a result, the reducing agents NADH and FADH 2 are produced, which feed the electron transport chain (ETC) and drive large amounts of ATP production to address energy imbalance 53,61  In this work, chemometric-based MVA successfully distinguished the 1 H -NMR spectra from bovine blood plasma collected during the Baseline and Infected stages of both challenge studies with an accuracy, sensitivity, and speci city ≥ 95 %, which is greater than the 65% found in VCD 15,16 and similar to the threshold reached by transthoracic ultrasound evaluation and serological and molecular techniques used in the diagnosis of infection [71][72][73] . These ndings indicate biochemical differences between healthy and sick animals with the main causal agents of BRD, where metabolites related to homeostasis in the baseline and energy imbalance during the infections were found to in uence the discrimination plots.
Previous research using NIRS and NMR also successfully discriminated blood plasma from cattle infected with BRD, with sensitivities and speci cities close to 90% when using VCD as reference method 19,33,34 . The ndings in the current study are also consistent with previous research that used discriminant analysis on NMR spectra of blood plasma from cattle to identify the metabolomics of animals with ketosis, ovarian quiescence, and fatty liver disorder 27,29−32 showing the potential of this technique for the detection of metabolic disorders related with nutrition, reproduction, and disease.          (Table S1).