Introduction

As one of the most well-known endangered species, the giant panda (Ailuropoda melanoleuca) has been a symbol for wildlife conservation for several decades. Their successful conservation program has benefitted greatly from the multitude of studies published on their geographical distribution1,2nutrition3,4behavior5,6genetics7,8and reproductive biology9,10,11. Scarce knowledge of the latter was particularly hampering ex situ conservation strategies early on, as both timing of ovulation and proof of conception were difficult to determine. Although extensive hormonal and behavioral research has greatly assisted in defining the fertile period6,9,12,13reliable pregnancy diagnosis is still only possible during the final weeks before birth. This is due to the occurrence of embryonic diapause after conception and because giant pandas experience obligatory pseudopregnancy if no conception occurs after ovulation, even without the act of mating14,15. Remarkably, pregnant and pseudopregnant females show highly similar physiological and endocrinological changes during the whole (pseudo)pregnancy, making an early differentiation between the two physiological states challenging16.

Patterns of urinary progesterone in giant pandas are seemingly identical between pregnancy and pseudopregnancy, with progesterone increasing immediately before ovulation after which it remains slightly elevated for a variable period of 2–3 months, corresponding to the corpus luteum dormancy (CLD) phase13,17. Recently, a triphasic pattern was described for this primary progesterone rise, subdivided in CLD1 (variable length), CLD2 (± 80 days prior to birth/end of the cycle), and CLD3 (± 60 days prior to birth/end of the cycle) (Supplementary Fig. S1)18. During the CLD phase, the activity of the corpora lutea is still relatively low and the embryo is believed to pause its development. Embryonic development is expected to resume at the start of a secondary, more drastic progesterone rise. Based on this biphasic progesterone profile, giant pandas are believed to experience delayed implantation since the endocrinology corresponds with other species in which embryonic diapause has been established16,19,20. Additionally, correlations between the timing of fetal detection with ultrasound and start of the second progesterone peak provided additional evidence that the implantation probably only occurs after the start of the secondary progesterone rise15. This secondary rise in progesterone is strongly pronounced and marks the start of the active luteal (AL) phase, lasting for 42 days. In pregnant individuals, this defines the period between blastocyst reactivation and birth of a full-grown, but altricial cub. During the AL phase, estrogens and prostaglandins, more specifically 13,14-dihydro-15-keto-PGF2α (PGFM), have shown diagnostic potential10,11. Around the presumed time of implantation, PGFM peaks for the first time, approximately three weeks prior to birth10. This peak divides the AL into an early (EAL, pre-PGFM peak) and late (LAL, post-PGFM peak) period (Supplementary Fig. S1)17. In pregnant individuals, both estrogens and PGFM will increase from the final two weeks of the LAL onwards, with PGFM climbing to extraordinary high numbers (x20) on average 24 h prior to birth10,11,17. Furthermore, glucocorticoids and even progestogens have recently been detected in higher urinary concentrations in pregnant individuals during the (L)AL phase17. In contrast, although pseudopregnant females exhibit a similar CL reactivation and AL phase, it is unclear what physiological processes are initiated or maintained during this period, causing these specific hormonal patterns17.

Despite recent advances in reproductive monitoring of (pseudo)pregnancy in giant pandas, available data on physiological and metabolic alterations during (pseudo)pregnancy remain scarce. Better understanding of these changes would both facilitate the use of artificial reproduction techniques and assist in increasing the success rate of pregnancies21. This will be essential to maintain and further improve the genetic diversity within the giant panda population under human care, as certain individual pandas with higher breeding difficulties will be genetically under-represented in the population compared to the pandas that breed more easily. Additionally, understanding hormonal and physiological changes during normal pregnancy can provide the first step in detecting and explaining pregnancy losses, which are believed to occur relatively frequently in this species but remain difficult to detect22.

This study aimed to obtain a more comprehensive view of the urinary molecular composition of pregnant and pseudopregnant giant pandas over their whole reproductive cycle, starting from the early corpus luteum dormancy phase (CLD1) until the end of the (pseudo)pregnancy (LAL). To this extent, a total of 238 urine samples from six pregnant and seven pseudopregnant cycles were analyzed using ultra-high performance liquid chromatography coupled to high-resolution mass spectrometry (UHPLC-HRMS), combining two complementary approaches: polar metabolomics and steroidomics (Supplementary Fig. S2). By comparing the urinary molecular spectra across different phases of the reproductive cycle, we evaluated the potential of MS-based profiling to discover distinct differences in metabolic changes between pregnancy and pseudopregnancy. The results provide a methodological starting point for future research into physiological differences between the two states and allow us to present novel early pregnancy biomarker (panel) candidates that require validation in larger and independent sample sets.

Methods

Animals

Urine samples of seven female giant pandas were provided by seven zoological institutions. For every sample, 1–15 mL urine was collected and transferred to Eppendorf tubes into 200–1000 µL aliquots during routine hormonal analysis with EIA. In total, urine samples from 13 cycles were included, of which six originated from pregnancies and seven from pseudopregnancies (non-bred cycles) (Table 1). Each cycle was subdivided into six different reproductive phases as described in Wauters et al.18i.e., anestrus (start of observations until cross-over between progestogens and estrogens), CLD1 (variable length, starting 7 days after day of peak estrogens), CLD2 (± 80–60 days prior to birth/end of the cycle), CLD3 (± 60 days prior to birth/end of the cycle until early onset of consistent increase in progestogens), EAL (early onset of consistent increase in progestogens until first PGFM peak), and LAL (post first PGFM peak) (Supplementary Fig. S1). Multiple samples were collected longitudinally from the same individuals on different days within the same cycle during every reproductive phase. From these repeated samples, three representative ones were selected per reproductive phase per cycle with sampling times matched, ensuring that each sample corresponded to a similar time point across cycles (Supplementary Table S1). As sample volumes could be insufficient for multiple analyses, selected samples were not always identical for the different performed analyses, i.e., polar metabolomics and steroidomics. The giant pandas were accommodated based on best practice guidelines for animal husbandry as well as recommendations given by the respective supporting Chinese giant panda experts. Because this study only included analysis of urine as a waste product, animal welfare was not affected. Therefore, no specific ethical approval was needed as all animal-related work was conducted according to relevant national and international guidelines. All experimental protocols were permitted by the respective zoos and supporting Chinese giant panda experts.

Table 1 Summary of the included animals and cycles.
Full size table

Sample collection and preparation

All urine samples were collected either via aspiration from the concrete floor with a clean syringe in the absence of the animal or by direct urination in a small container (trained animals). After collection, the samples were transferred to a collection vial and stored at -20 °C until shipment to the Laboratory of Integrative Metabolomics (LIMET) at Ghent University (Belgium). Before analysis, urinary specific gravity (USpG) was measured according to the method described and validated in Wauters et al.23. using a handheld digital refractometer designed for cat (PAL-USG (Cat), range 1.000-1.080, Atago™, Fukaya, Japan) or human urine (Pal-10 S (Human), range 1.000-1.060, Atago™, Fukaya, Japan).

Chemicals and reagents

The polar metabolomics and steroidomics methodologies included, respectively, 70 and 56 authentic analytical standards of naturally occurring compounds (Supplementary Table S2). Analytical standards and internal standards (ISTD) were purchased at Sigma Aldrich (St. Louis, Missouri, USA), Steraloids (Newport, Rhode Island, USA), Cayman chemical (Ann Arbor, Michigan, USA), CDN Isotopes (Pointe-Claire, Canada), SERVA Biochemicals (Heidelberg, Germany), VWR Chemicals (Radnor, Pennsylvania, USA), and Santa Cruz Biotechnology (Santa Cruz, California, USA). Solvents were UPLC-MS grade and obtained from Fisher Scientific (Loughborough, UK), Merck (Darmstadt, Germany), Chem-lab (Zedelgem, Belgium), and Biosolve (Valkenswaard, The Netherlands). Ultra-pure water (UPW) was obtained via a purified-water system from Millipore (Brussels, Belgium).

Instrumentation

Metabolomics

For the urinary polar metabolome analysis, the method described by De Paepe et al.24 was used. Chromatographic separation of the analytes was performed via reversed-phase chromatography with gradient elution using an Acquity HSS T3 C18 column (1.8 mm, 150 mm x 2.1 mm) mounted on a Dionex UltiMate 3000 XRS UHPLC system (Thermo Fisher Scientific, San José, California, USA). Detection of the analytes was performed on a Q-Exactive™ quadrupole-Orbitrap HRMS (Thermo Fisher Scientific, San José, California, USA), which was preceded by heated electrospray ionization (HESI-II source) in polarity switching mode. Prior to analysis, the instrument was calibrated by infusing ready-to-use calibration mixtures (Thermo Fisher Scientific, San José, California, USA). Quality control (QC) samples (constructed by combining 50 µL of the first 20 samples selected at random with a volume higher than 300 µL) were included in the beginning (n = 10) and throughout the analysis run (n = 2 every 10 samples) to stabilize the system and to correct for potential instrumental drift between batches.

Steroidomics

Steroidomics analysis was executed according to Cools et al.25. Analytes were chromatographically separated on a Vanquish Horizon UHPLC system (Thermo Fisher Scientific, San José, California, USA) equipped with a reversed phase Acquity BEH C18 UPLC column (1.7 μm, 2.1 × 100 mm) (Waters, Millford, Massachusetts, USA). An Orbitrap Exploris 120 HRMS (Thermo Fisher Scientific, San José, California, USA) was used for analyte detection, fitted with an atmospheric-pressure chemical ionization (APCI) source. Initial instrument calibration was first achieved using a Pierce Flexmix calibration solution (Thermo Fisher Scientific, San José, California, USA), which was infused for both ionization modes. QC samples were constructed by combining 80 µL aliquots of a representative subset of 70 samples (assuring that each panda and each reproductive phase was included at least once in both groups) and analyzed identically to the metabolomics analysis described above.

Fragmentation experiments

Fragmentation of significant features was performed for both polar metabolomics and steroidomics on the Q-Exactive™ instrument. For the features selected from the metabolomics method, the same analytical method was largely applied, but this time with parallel reaction monitoring (PRM). Furthermore, the resolution and AGC target were lowered to 17,500 FWHM and 2E + 5, respectively, while the injection time was increased to 100 ms. The isolation window was 1.0 Da and 3 collision energies were applied, namely 15, 30, and 50 eV. For the steroidomics fragmentation, the same strategy was used, but flow rate and gradient were slightly altered (Supplementary Table S3). The initial steroidomics method was also adjusted by adding PRM, and altering the mass resolution, AGC target, and injection time to the same values as the metabolomics fragmentation method. The isolation window was again 1.0 Da and 25, 35, and 50 eV were used as collision energies. Only samples with known high abundances of the selected features were retained. QC samples were constructed by combining 100 µL of each sample and were again analyzed in the beginning (n = 5) and throughout (n = 2 every 10 or 11 samples) the analysis run. A standard mix containing the authentic analytical standards of the targeted approach was run in the beginning and at the end of the analysis.

Sample extraction

Prior to the analysis of the polar urinary metabolome, 50 µL of urine were diluted with UPW according to the protocol used for human urine by De Paepe et al.24. The dilution factor was optimized for giant panda urine to a value of four to obtain the highest number of detected features with good repeatability, i.e., coefficient of variation (CV) of the feature abundance below 30%. In short, samples were first centrifuged for 8 min at 1,000 g and 4 °C followed by the collection of 50 µL of the supernatant in a 1.5-mL Eppendorf tube. Then, 130 µL UPW and 20 µL ISTD solution (10 ng µL− 1 in UPW) were added. The diluted sample was vortexed for 30 s after which 100 µL was transferred into a glass HPLC vial with insert. Sample extraction prior to the steroidomics analysis was performed with 200 µL of urine according to Cools et al.25.

Data processing and analysis

Targeted data processing, i.e., only known metabolic features representing the target compounds, was executed using the in-house R package TaPEx or manually in the Quan Browser of XCalibur 4.4 (Thermo Fisher scientific, San José, California, USA) (Supplementary Tables S2)26. Untargeted processing, i.e., all detected metabolic features, was carried out with Compound Discoverer (CD) 3.2 (metabolomics) and 3.3 (steroidomics) (Thermo Fisher Scientific, San José, California, USA). Signal intensities (TaPEx) and feature abundances (Quan Browser and CD) were corrected for USpG prior to further data analysis using the adjusted formula of Wauters et al.23. Following USpG correction, the untargeted data was explored using principal component analysis (PCA) with iQC (internal quality control samples) normalization to check QC clustering, natural patterning of the samples and identify potential outliers. Potentially interesting features were selected for further data-analysis by building predictive models comparing pregnant and pseudopregnant samples per reproductive phase using orthogonal partial least squares-discriminant analysis (OPLS-DA). These models were validated via the evaluation of specific quality parameters including Q2R2(Y), permutation testing, and cross-validated ANOVA (CV-ANOVA) (p-value < 0.05). Univariate statistics comparing pregnant with pseudopregnant sample data were run on the selected untargeted features and the putatively identified targeted compounds using the in-house R pipeline. Depending on their individual normality, which was first tested with a Shapiro-Wilk test (α = 0.05), a Welch two-sample t-test (parametric, for normally distributed data) or Wilcoxon rank sum test (non-parametric, for data that was not normally distributed) (α = 0.05 for both after adjustment for multiple testing with the Benjamini-Hochberg method27 was applied. Variables were deemed significantly different between pregnant and pseudopregnant sample data in a certain reproductive phase if their univariate test resulted in an adjusted p-value < 0.05 in that phase while p-value > 0.05 in the anestrus phase (negative control, no correction of the p-value was applied here). MS/MS spectra of significantly different features were further analyzed using isotope and fragmentation pattern analysis with SIRIUS 5.5.728, spectral library searching via the Global Natural Product Social Molecular Networking (GNPS)29and CD 3.3 for further annotation according to Schymanski et al.30. These analyses correspond to Tier 3 confidence, meaning a tentative candidate identification based on computational predictions and spectral similarity was obtained. Wherever possible, the aim was to confirm the identity of discriminative features with the highest level of confidence (Tier 1) by fully matching their accurate mass (m/z), retention time, carbon isotope ratio (C12/C13), and fragmentation spectra with those of authentic reference standards analyzed under identical conditions. Finally, in addition to the multi- and univariate data analysis strategies, the predictive potential of biomarker panels was investigated via elastic net regularization (α = 0.15) followed by generalized linear modelling (GLM) per reproductive phase31,32. All features with CV > 30% in their feature abundance across QC samples were excluded prior to model building. Additionally, all features that were significantly different in the anestrus phase (p-value < 0.05 with univariate statistics) were excluded from all phases to remove inherent variation between groups unrelated to the reproductive state. Optimization of the tuning parameter λ was performed via leave-one-out cross validation (LOOCV), given the small sample set33.

For correct interpretation of the data analysis, it is important to acknowledge the limited number of individual giant panda bears (n = 7) enclosed in this study, even though the majority of mature female giant pandas resident in Europe was included, representing a significant share (~ 1%) of the total global giant panda population under human care. It must also be noted that due to the inclusion of multiple samples from the same individuals, complete independence of all sample groups could not be guaranteed, which may introduce bias and affect variation and model performances. Additionally, confounding factors such as environmental influences, maturity, diet etc., and their interactions can greatly affect metabolic and hormonal changes across the reproductive cycle. Nevertheless, as the same strategy was applied for all sample groups, the loss of inherent variation within groups is expected to be similar in all sample groups, therefore limiting potential bias. Furthermore, samples from the anestrus phase were included in the study, representing a negative control. Comparing samples from the anestrus phase allowed for correction of potential confounding factors by filtering out variation between sample groups that is likely unrelated to the reproductive state.

Results

Metabolomics analysis

Out of the 70 compounds from the targeted metabolomics list, 43 were detected across the giant panda urine samples (Supplementary Table S4). Univariate statistical tests revealed that USpG corrected peak areas of 15 of those compounds were significantly different between pregnant and pseudopregnant pandas in at least one reproductive phase (Supplementary Table S4). Seven were, however, also significantly different in the anestrus phase. The remaining eight compounds were fragmented leading to three confirmed identities after spectral matching using authentic analytical standards (Tier 1)30. Urocanic acid was significantly higher in pseudopregnant urines in CLD3, while two amino acids, L-leucine and L-phenylalanine, were both significantly higher in pregnant samples in the LAL phase (Fig. 1).

Fig. 1
figure 1

Targeted polar metabolic feature abundances in pregnant and pseudopregnant giant pandas across the reproductive phases. Feature abundances are based on peak areas of the different samples per reproductive phase in pregnant (P, orange, n = 15 per phase) and pseudopregnant (PP, blue, n = 21 per phase) cycles. An anestrus, CLD corpus luteum dormancy phase, EAL early active luteal phase, LAL late active luteal phase.

Full size image

The untargeted metabolomics analysis yielded 6,550 and 4,556 features in the positive and negative ionization modes, respectively. Unsupervised data analysis was performed for all samples and per reproductive phase comparing pregnant and pseudopregnant samples (Supplementary Fig. S3A-F). The PCA model exclusively based on samples of the LAL phase showed a better group separation compared to models of the other reproductive phases, with PC1 and PC2 accounting for 22.29% and 12.58% of the variation, respectively (Supplementary Fig. S3F). Some individual sample clustering (pandas SB741 and SB868) was observed in the PCA plot of the CLD3 phase (Supplementary Fig. S3D), which may reflect biological or environmental variation unrelated to pregnancy status (e.g., diet, management, or sample handling conditions across zoos). To further define metabolic differences over the course of the reproductive cycle, supervised OPLS-DA models were constructed for each reproductive phase. When all 11,106 features were included, the OPLS-DA model of the LAL phase showed excellent Q2 (0.831), R2(Y) (0.987), permutation testing, and a CV-ANOVA p-value < 0.001. After filtering out all features with a high variance in their feature abundance across QC samples (CV > 30%) or a low VIP score (< 1), new OPLS-DA models were built with more than 2,000 features remaining in all reproductive phases. All newly built models showed excellent Q2R2(Y), permutation testing, and CV-ANOVA p-values (Supplementary Table S5). Based on the VIP scores, Jack-knifed confidence interval, and S-plot correlation, 40, 41, 68, 15, and 310 features were retained from the OPLS-DA models constructed with all features for the CLD1, CLD2, CLD3, EAL, and LAL phase, respectively (Supplementary Table S6). After employing univariate statistics, 389 discriminative features were retained across all phases (CLD1: 30, CLD2: 28, CLD3: 40, EAL: 13, LAL: 278) (Supplementary Table S7). Three features could be identified by spectral matching of the fragmentation data with authentic analytical standards (Tier 1)30. Kynurenic acid (KYNA) and indole-3-lactic acid (ILA) were both significantly higher in pseudopregnant samples during the CLD2 and CLD3 phase (Fig. 2). In contrast, in the LAL phase these metabolites were significantly higher in the pregnant samples. Tiglylcarnitine decreased in both groups during the early CLD phase compared to anestrus, but returned to higher levels during the late CLD phase (CLD3) in pseudopregnant samples (Fig. 2). The matched spectral data of putatively annotated urinary metabolomics features (UMF) such as UMF1752 and UMF7328 (o-methoxyhippuric acid and an n-aminodecanoic acid, respectively), could not be fully confirmed with analytical standards since they were not commercially available. Preliminary identification was therefore only based on m/z ratio, expected retention time based on known similar compounds, and C12/C13 ratio of the results provided by the fragmentation pattern analysis and spectral library searching (Tier 3)30. UMF1752 and UMF7328 were significantly higher in pregnant samples during CLD1 and EAL, respectively (Fig. 2).

Fig. 2
figure 2

Untargeted polar metabolic feature abundances in pregnant and pseudopregnant giant pandas across the reproductive phases. Feature abundances are based on peak areas of the different samples per reproductive phase in pregnant (P, orange, n = 15 per phase) and pseudopregnant (PP, blue, n = 21 per phase) cycles. O-methoxyhippuric acid and n-aminodecanoic acid are putatively identified (Tier 3). An anestrus, CLD corpus luteum dormancy phase, EAL early active luteal phase, LAL late active luteal phase.

Full size image

Steroidomics analysis

Using steroidomics analysis, 12 out of 56 targeted compounds could be detected in the panda urine (Supplementary Table S8). Seven compounds showed matching fragmentation patterns with the authentic analytical standards, i.e., cortisol, cortisone, tetrahydrocortisone, 1,4-androstadienedione, 11-oxoetiocholanolone, androstenedione, and allopregnanolone sulfate (Tier 1)30. The compound putatively identified as 11β-hydroxyetiocholanolone, showed a highly similar fragmentation pattern to the analytical standard of 11β-hydroxyetiocholanolone, but the primary and some minor fragments differed with 2 Da each (Supplementary Fig. S4). This steroid is likely closely related to 11β-hydroxyetiocholanolone with the double bound shifted to another position. It is further referred to as urinary androgen 1 (UA1). UA1 and 11-oxoetiocholanolone were both significantly higher in pregnant urines during the CLD3 and EAL phase (Fig. 3). During the LAL phase, cortisol, 1,4-androstadienedione, and allopregnanolone sulfate were significantly higher in pregnant samples (Fig. 3).

Fig. 3
figure 3

Targeted steroid feature abundances in pregnant and pseudopregnant giant pandas across the reproductive phases. Feature abundances are based on peak areas of the different samples per reproductive phase in pregnant (P, orange, n = 18 per phase, except for the CLD1 phase (n = 17)) and pseudopregnant (PP, blue, n = 21 per phase, except for the anestrus and LAL phase (n = 20 and 18, respectively)) cycles. UA1 = unknown androgen 1, an analogue of 11β-hydroxyetiocholanolone. An anestrus, CLD corpus luteum dormancy phase, EAL early active luteal phase, LAL late active luteal phase.

Full size image

A total of 14,081 features were detected following untargeted steroidomics data processing, 12,474 in the positive and 1,607 in the negative ionization mode. PCA models built per reproductive phase comparing pregnant and pseudopregnant samples revealed no separation between the two groups for any of the reproductive phases (Supplementary Fig. S5). However, some degree of individual sample clustering was observed in the PCA plots potentially reflecting site-specific variation unrelated to reproductive status. Identical to the metabolomics analysis, supervised OPLS-DA models were built per reproductive phase to better investigate the differences between pregnancy and pseudopregnancy. When including all 14,081 features, the OPLS-DA model of the LAL phase again showed excellent Q2 (0.736), R2(Y) (0.981), permutation testing, and a CV-ANOVA p-value < 0.001. OPLS-DA models of the other reproductive phases could not be validated. However, when excluding all features with a CV > 30% in their abundance across QC samples and all features with a VIP score < 1, the OPLS-DA model of the CLD3 phase, built with 3,457 features remaining, showed an improved Q2 (0.562), R2(Y) (0.995), permutation testing, and a CV-ANOVA p-value of 0.01. Potentially discriminating features were again selected from the OPLS-DA models including all features for CLD1 (n = 77), CLD2 (n = 26), CLD3 (n = 30), EAL (n = 82), and LAL (n = 198) based on the VIP scores, Jack-knifed confidence interval, and S-plot correlation (Supplementary Table S9). Following univariate statistical tests, the final number of features remaining for CLD1, CLD2, CLD3, EAL, and LAL counted 42, 15, 19, 60, and 171, respectively (Supplementary Table S10). After processing with fragmentation software tools, spectral libraries and CD, the identities of arachidonic acid (AA) and indole could be confirmed based on fragmentation patterns of the corresponding authentic analytical standards (Tier 1)30. Both compounds showed significantly higher levels in pregnant compared to pseudopregnant pandas in the LAL phase (Fig. 4). Similarly to the untargeted metabolomics approach, the identity of two other urinary steroidomics features (USF) could not fully be confirmed but they could putatively be annotated as 21-dehydrocortisol (USF2198) and 6,15-diketo-13,14-dihydro-PGF1α (USF5681) based on m/z ratio, expected retention time based on known similar compounds, and C12/C13 ratio of the results provided by the fragmentation pattern analysis and spectral library searching (Tier 3)30. Both USF2198 and USF5681 were significantly higher in pregnant samples in the LAL phase (Fig. 4). Moreover, USF5681 already showed significantly higher levels in pregnant samples in both the CLD3 and EAL phase.

Fig. 4
figure 4

Untargeted steroid feature abundances in pregnant and pseudopregnant giant pandas across the reproductive phases. Feature abundances are based on peak areas of the different samples per reproductive phase in pregnant (P, orange, n = 18 per phase, except for the CLD1 phase (n = 17)) and pseudopregnant (PP, blue, n = 21 per phase, except for the anestrus and LAL phase (n = 20 and 18, respectively)) cycles. 21-deoxycortisol and 6,15-diketo-13,14-dihydro-PGF1α are putatively identified (Tier 3). An anestrus, CLD corpus luteum dormancy phase, EAL early active luteal phase, LAL late active luteal phase.

Full size image

Predictive modelling

After QC filtering and removal of the features deemed significantly different in the anestrus phase, 7,486 and 7,397 features remained from the polar metabolomics and steroidomics data, respectively. When combining the metabolomics and steroidomics data, 14,406 features remained for model building. Because all significantly different features were removed, no reliable predictive models could be built with data from the anestrus phase. As expected, the most predictive models were built with data from the LAL phase showing accuracies of 0.972, 1.000, and 1.000 with metabolomics, steroidomics, and the combined data, respectively (Fig. 5). Identical to the PCA plots built with the metabolomics data in this phase, only one sample was wrongly classified as pseudopregnant. All other models, excluding the anestrus phase, however, also showed accuracies higher than 0.760. Furthermore, the models built with only metabolomics data, and with the combination of the metabolomics and steroidomics data for the CLD3 phase offered accuracies higher than 0.940. The number of retained features to build the final models ranged from 64 to 342 (excluding models from the anestrus phase) (Table 2).

Fig. 5
figure 5

Accuracy dot plots of the GLM models with elastic net regularization per reproductive phase. (A) Models built with the metabolomics data. (B) Models built with the steroidomics data. (C) Models built with the combined metabolomics and steroidomics data. Acc = Accuracy, CI = 95% confidence interval. All models were built after QC filtering (CV > 30%) and exclusion of features significantly different in the anestrus phase (p < 0.05). Remaining features used to build the models encompassed 7486 for the metabolomics data, 7,397 for the steroidomics data, and 14,406 for the combined data. CLD corpus luteum dormancy phase, EAL early active luteal phase, LAL late active luteal phase.

Full size image
Table 2 Predictive performance parameters of the GLM models with elastic net regularization per reproductive phase.
Full size table

Discussion

In this study, we demonstrate distinct urinary molecular profiles between pregnant and pseudopregnant giant pandas by employing an MS-based discovery approach including two different molecular omics methods, i.e., polar metabolomics and steroidomics. Although our findings must be considered within the context of the relatively small sample size and exploratory nature of this work, the observed differences in metabolic and hormonal features across different reproductive phases provide intriguing insights. These insights can establish a valuable basis for future research aiming at improving pregnancy detection and understanding reproductive physiology in this endangered species. The observed metabolic and hormonal differences were expressed from the early dormancy phase onwards but gained in magnitude towards the late active luteal phase, which is the current window for pregnancy diagnosis through ultrasound and endocrine monitoring. Untargeted polar metabolomics and steroidomics analysis revealed a total of, respectively, 389 and 307 significantly different features from which respectively 98 and 76 were retained in the early luteal (CLD) phases. Furthermore, identification of these features revealed changes in various metabolic pathways including amino acid metabolism (urocanic acid, L-leucine, L-phenylalanine, ILA, and KYNA), energy metabolism (tiglylcarnitine and n-aminodecanoic acid), arachidonic acid metabolism (6,15-diketo-13,14-dihydro-PGF1α and AA), and steroid hormone biosynthesis (UA1, 11-oxoetiocholanolone, 1,4-androstadienedione, cortisol, and allopregnanolone sulphate).

Although no features from the CLD1 phase could be fully annotated, UMF1752 was putatively identified as o-methoxyhippuric acid. O-methoxyhippuric acid is not well described in literature, but it could be formed via the conjugation of o-methyl-salicylic acid and glycine34. The former is the methylated form of salicylic acid, a compound known for its anti-inflammatory capacities as cyclooxygenase (COX) inhibitor35,36. The implantation of an embryo is a pro-inflammatory reaction in which prostaglandins derived via COX play an important role37. It is thus probable that a salicylic acid-like compound is increased during the early CLD phase in giant pandas as part of a mechanism to halt implantation of the blastocyst37.

In the CLD2 and CLD3 phases, two metabolites from the tryptophan pathway, i.e., ILA and KYNA, were identified with levels increased in pseudopregnant urine. In the LAL phase, on the other hand, the levels of both metabolites and a third metabolite from the tryptophan pathway, i.e., indole, were increased in the pregnant urine instead (Figs. 2 and 4). KYNA has already been described as a potentially important metabolite in giant panda reproduction by Cao et al.38 who found an increased KYNA excretion in the luteal phase of pregnant pandas compared to their corresponding anestrus phase. The latter finding is also consistent with studies on rats (Rattus spp.) and humans39,40. KYNA has been hypothesized to be present in higher levels during pregnancy because of its potential neuroprotective effects on the fetus41. Alternatively, urinary KYNA levels can be increased because of a higher availability of KYN in the umbilical cord. Indeed, during pregnancy tryptophan is catabolized to protect the fetus against the maternal immune system42. As indole and KYN are intermediate products of tryptophan degradation, this process explains the increased indole and KYNA excretion observed during pregnancy39. ILA is another metabolite from the tryptophan pathway that has some antioxidant and immunosuppressive effects43,44,45. It can play a role in regulating the microbiome of the newborn cub, which can explain the higher levels in pregnant pandas during the LAL phase43,44,45. The reasons why KYNA and ILA are both higher in pseudopregnant pandas during the CLD2 and CLD3 phases remain unclear.

As a short-chain acylcarnitine46tiglylcarnitine functions as a fatty acid transporter into the mitochondria for energy production via fatty acid oxidation47. During specific conditions with increased energy demand (e.g., exercise or pregnancy) or lower energy intake (e.g., fasting), fatty acid oxidation will increase and thus also the production of acylcarnitines48,49. In pregnant pandas, tiglylcarnitine levels indeed start to increase during the EAL phase. This is consistent with results regarding the fecal output of giant pandas published by Wauters et al.17where a steep decline was noted midway through the EAL phase, approximately 30 days before parturition and thus around the same time window as the EAL samples of this study. However, further investigations are warranted to reveal why tiglylcarnitine levels appear to increase earlier (during CLD3) in pseudopregnant pandas (Fig. 2). UMF7328 in its turn, was increased in pregnant urine samples from the EAL phase and although its identity could not be fully confirmed, it was putatively classified as an n-aminodecanoic acid, which is a medium-chain fatty acid (MCFA). MCFAs are not only readily available energy sources, they also play an important part in fetal development and specifically reduce the chances on early pregnancy loss by improving ovarian steroidogenesis and endometrial receptivity50,51,52. The EAL phase marks the implantation/attachment period of the giant panda embryo and is thus a crucial period that could benefit from these protective effects.

In the final phase of the reproductive cycle, i.e., the LAL phase, several of the observed metabolic and hormonal differences were related to AA metabolism and steroid hormone biosynthesis. Both have been extensively investigated in giant pandas via the use of enzyme immunoassays (EIA)10,11,16,53,54. Nevertheless, confirmation of the identity of the detected hormones is absent when merely using EIA for non-invasive endocrine monitoring as the hormone metabolization process in giant pandas is yet to be unraveled. For example, prostaglandin levels diverge between pregnant and pseudopregnant pandas during the final two weeks to prepare the pregnant female for parturition and are therefore monitored via an EIA that targets PGFM, a metabolite of PGF2α10,17,55. According to our results, however, giant pandas do not excrete PGFM. Instead, another metabolite is excreted during these stages, namely 6,15-diketo-13,14-dihydro-PGF1α, which is probably picked up via EIA cross-reactivity. Remarkably, 6,15-diketo-13,14-dihydro-PGF1α already shows significant differences in earlier stages than can be seen with PGFM EIA monitoring (Fig. 4). More specific EIAs should therefore be developed as they could possibly also pick up this trend and thus provide earlier signs for pregnancy in giant pandas than the current PGFM EIA. Similar to PGFM, we were not able to detect progesterone in any of the urine samples, despite this usually being the main target of the progesterone EIA used for giant panda reproductive monitoring16,17. The detection of allopregnanolone sulfate in giant panda urine on the other hand, has not been described in literature yet, while our results suggest it is the main urinary progestogen in this species. In contrast, the presence of cortisol, which has been targeted with EIA in giant pandas, was confirmed with our analysis. It also demonstrated the expected increase during the LAL phase, priming both the fetus and the mother for the nearing parturition17,54,56,57. However, a second potential glucocorticoid, putatively identified as 21-dehydrocortisol, showed a more distinct increase than cortisol towards the end of the cycle in pregnant females which was absent in their pseudopregnant counterparts (Fig. 4). As both steroids show highly similar structures, it is possible that 21-dehydrocortisol is detected with the same EIA as cortisol. Although glucocorticoids stimulate the synthesis of estrogens during these final weeks, resulting in improved myometrial activity, this could not be observed with our method as we were unable to identify any estrogens yet. Nevertheless, estrogens are synthetized via the conversion of androgens by aromatase enzymes58. As the steroid biosynthesis pathway gradually alters to favor estrogen over progestogen production, androgen levels can also increase being an intermediate between progestogens and estrogens. This can explain the higher levels of 1,4-androstadienedione observed solely in the pregnant giant pandas in the LAL phase in this study. Remarkably, increased levels of two other androgens, UA1 and 11-oxoetiocholanolone, were also noted in earlier stages of pregnant pandas in this study, i.e., the CLD3 and EAL phase (Fig. 3). The EAL corresponds with the CL reactivation and the window for embryo implantation. Androgens play a role in the normal endometrial physiology for embryo implantation both via the conversion to estrogens as via direct binding with androgen receptors59,60. It could be hypothesized that in giant pandas, these androgens are involved in preparing the endometrium for implantation of the embryo around reactivation of the CL. As the levels are higher in pregnant females, embryonic signals probably assist in their upregulation.

Despite the successful annotation of multiple features in this study, identification of biomarkers remains a bottleneck in untargeted metabolomics analyses. Therefore, it was explored whether panels of metabolic (including steroids) features could be used to classify samples using binomial logistic regression. Given the small set of samples (n = 33–39 per phase), the high number of variables (n = 7,397 − 14,406), and the absence of a set of independent samples for model validation, a regularization penalty was applied to prevent overfitting of the models on the training data. Elastic net regularization was selected because it allows feature selection by removing unimportant variables from the model, while taking into account potential correlations between variables31,32. The latter is expected to be present as most features will be part of biological systems, which often show complex interactions. The excellent parameters of the logistic regression models, particularly those built from data of the early stages, i.e., CLD (Table 2), suggest that panels of biomarkers measured with UHPLC-HRMS could provide an alternative strategy to the use of EIA for early pregnancy diagnosis. Nevertheless, it must be noted that these models were only validated on the training data. They will need to be validated on newly collected data to fully confirm their potential as diagnostic tools.

Conclusions

To the best of our knowledge, this is the first study to employ a non-invasive metabolomics approach to investigate metabolic and hormonal differences between pregnant and non-pregnant (e.g., proven pseudopregnancies) giant pandas across different phases of their reproductive cycles. Remarkably, our results demonstrate that discriminative changes can be observed as early as the start of the embryonic diapause, i.e., CLD1, long before the current window of pregnancy diagnosis. Our findings did not only confirm differences between pregnant and pseudopregnant pandas in prostaglandin and steroid hormone biosynthesis, but also provided new insights in potential alterations of amino acid and energy metabolism during the reproductive cycle. After further elucidation and validation of the underlying mechanisms of these changes in an independent sample set, these discriminating metabolites could be explored as possible biomarkers for pregnancy monitoring in giant pandas. Alternatively, the potential of biomarker panels, without the need for individual feature annotation, was also demonstrated. Similar to the individually identified metabolites, these classification models will require additional validation in independent sample sets before implementation can be considered. Furthermore, for the first time to our knowledge, we were able to fully annotate discriminatory urinary steroids, some of which are already included in non-invasive pregnancy monitoring in giant pandas using EIA. Although our findings revealed that the excreted steroid metabolites are not necessarily the intended target compounds of those commonly intended with EIAs, they are probably still detected via cross-reactivity of the antibodies. These results are important for future optimization of employed EIAs and will as such contribute to improving the routine reproductive monitoring of giant pandas as part of their conservation efforts.