Targeted lipidomics unveils complex lipids in cervicovaginal environment as diagnostic and prognostic biomarkers for endometrial cancer.

Between June 2018 and February 2020, a total of 192 women who were scheduled for a hysterectomy for benign or malignant conditions were enrolled across three clinical sites located in the Phoenix (AZ, USA) metropolitan area (Banner University Medical Center-Phoenix, Valleywise Health Medical Center, Dignity Health Chandler Regional Medical Center). Initial enrolment exclusion and inclusion criteria have been previously described by Łaniewski et al. [ 23 ]. Exclusion criteria were confirmed by a physician during a pelvic exam, self-reported survey data and medical records. Post-hysterectomy pathological reports of endometrial tissue samples were used to classify the women into four main groups based on their condition(s). The four groups were as follows: benign conditions ( n  = 108), endometrial hyperplasia ( n  = 18), less severe grade 1 or 2 endometrioid endometrial carcinoma (EEC) ( n  = 53), and more severe subtypes of EC including: grade 3 EEC, serous carcinoma, or other histological subtypes ( n  = 13). An additional 49 asymptomatic Pap-negative women from a different cohort of women previously described in Łaniewski, Sci Rep 2018 [ 24 ], were incorporated into the study to form a healthy comparison control group. Cervicovaginal lavage (CVL) samples were collected by a physician from two cohorts of women; prior to hysterectomy from 192 women with benign conditions or EC [ 23 ], and during a well-woman exam from 49 healthy, asymptomatic women [ 24 ]. A limited number of physicians followed a standardised protocol for CVL collection, ensuring consistency across participants. All 223 CVL samples were collected using a non-lubricated speculum with 10 mL of sterile 0.9% saline solution (Teknova, Hollister, CA), and immediately placed on ice before being frozen at −80 °C. Prior to downstream analyses, all frozen CVL samples were thawed on ice, centrifuged for 10 minutes (700 x g 4 °C), aliquoted to avoid freeze/thaw cycles, and stored at −80 °C. A total of 223 CVL samples were sent to Metabolon, Inc. (Durham, NC) for lipidomic analysis using the Complex Lipids Targeted Panel. Once received, samples were recorded into Metabolon’s Laboratory Information Management System (mLIMS). Prior to extraction, isotopic compounds of known quantities were added to biological samples and utilised as internal standards. Methyl tert-butyl ether was used to separate the biological compounds within the sample before the solvent was evaporated and re-dissolved in dichloromethane: methanol (1:1) with 10 mM ammonium acetate. The Complex Lipids Targeted Panel used the Shimadzu unit with nano PEEK tubing to directly deliver the lipid sample solution at 6-minute intervals into SelexION-5500 QTRAP triple quadrupole mass spectrometer (MS). For sample analysis, the spectrometer utilised multiple reaction monitoring (MRM) mode coupled with both, positive and negative electrospray ionisation (ESI) modes. The MRM mode can measure over 1,000 lipid species in one sample, repeating the measurement 17 times to obtain even replicates in up to 40 samples at any one time. Concentrations of species were reported only if the signal was detected in 12 out of the 17 replicates. Individual species were distinguished using the MRM signal by measuring the area under the peak of an ion (lipid or fragment) and referencing against the assigned internal standard on the MRM spectrum. Once identified, the concentrations were determined by multiplying the average signal of the unknown lipid to the signal of the internal standard by the known concentration of that matching internal standard added to the sample. Lipid class concentrations were determined by adding together the concentrations of all individual species within that class and reported as total class concentrations. The concentrations of individual fatty acid side chains were summed across all species in one class to produce total concentrations of each fatty acid across the 14 lipid classes. The concentrations of each fatty acid across the 14 lipid classes were further combined to produce the total concentration of a given fatty acid representative of the entire lipidome. Concentrations for all compounds were reported as nmol/mL. The origin of each lipid species was determined using the MetOrigin (2.0) web platform. Individual lipid species were mapped to known entries in MetOrigin’s database to determine their source and pathway involvement. MetOrigin integrates seven resources together to form one large database which consists of information from Kyoto Encyclopedia of Genes and Genomes (KEGG), Human Metabolome Database (HMBD), Biochemical, Genetic and Genomic knowledgebase (BIGG), Chemical Entities of Biological Interest (ChEBI), Food Database (FoodDB), DrugBank and Toxin and Toxin Target Database (T3DB). MetOrigin’s database can identify the origin of up to 191,031 metabolites, these can be grouped into six categories: host, microbiota, co-metabolism, food, drug, or environment. Lipid species data was uploaded to MetOrigin in the format of a compound concentration table, with corresponding KEGG and HMBD identification numbers for each lipid species. The results of the origin analysis were visualised as a bar plot and Venn diagram, both of which summarise the total number of lipid species from different origins. In addition, the Metabolic Pathway Enrichment Analysis (MPEA) was also performed in MetOrigin to identify metabolic pathways which were most significantly altered for each of the six origin categories. Enrichment analysis was performed in MetaboAnalyst (6.0) to identify perturbated pathways within the lipid species dataset. Despite being referred to as an “enrichment” analysis, this method reports whether a pathway is altered as a whole, acknowledging that different lipid species within the same pathway may be simultaneously enriched and depleted. The analysis is completed by comparing lipid species data to the Small Molecule Pathway Database (SMPD) and employing a hypergeometric test to determine if the magnitude of altered lipid species associated with a given pathway is over- or underrepresented. Statistical significance for participant demographics and other characteristics was determined using GraphPad Prism version 10.0. For continuous variables, such as BMI and age, p values were calculated with unpaired t-test to determine statistical differences between two groups or one way analysis of variance (ANOVA) to determine differences between three or more groups. Fisher’s exact test or chi-squared test were performed on the categorical variables. One-way ANOVA with Tukey’s post-hoc test was used to identify differences in concentrations of group means between EC, benign conditions, and healthy controls in each of the fourteen lipid classes, using GraphPad Prism (10.0). To compare differences in concentrations of lipid species and fatty acids side chains between women with EC, women with benign conditions, and healthy women we used MetaboAnalyst (6.0). All data analysed in MetaboAnalyst was initially uploaded in the format of a compound concentration table, log10-transformed, and auto-scaled (mean-centred and divided by standard deviation of each compound). To visualise the magnitude of enrichment and/or depletion and the level of significance between two group means, we utilised volcano plots, which integrate both fold change and t-test analyses together. The fold change analysis used raw data to calculate the absolute value of change between the mean concentrations of two groups before applying the log2 function to transform the data. A t-test with false discovery rate (FDR) correction was then applied to the log2 transformed data to evaluate whether the difference was statistically significant and reported as a q value. To identify individual complex lipid species capable of discriminating between two conditions we utilised the univariable receiver operating characteristic (ROC) analysis. The analysis generated ROC curves by plotting the true positive values against false positives at different thresholds for a given feature. The area under the curve (AUC) was used to score the discriminative ability of a single biomarker where an AUC of ≤0.5 reflected a random classifier. Species with an AUC of ≥0.7 were considered suitable candidates as diagnostic biomarkers; those with an AUC ≥ 0.8 were good biomarkers, and those with an AUC ≥ 0.9 possessed excellent discriminatory ability. Prior to uploading the data onto MetaboAnalyst, samples within each disease group were split randomly using an online data splitting tool; 2/3 of samples were used for exploratory and training purposes and the remainder 1/3 samples were reserved for subsequent model validation. To assist with feature selection for multivariable biomarker models, we utilised the multivariable exploratory ROC analysis function within MetaboAnalyst and the training dataset (2/3 of samples). Exploratory multivariable models were developed using random forest algorithm and evaluated using Monte-Carlo cross validation (MCCV) to determine the most discriminative lipid combination between two tested conditions. The performance of each multivariable model was reported as an average based on 100 MCCVs. The top ten features with the highest mean importance scores were selected for the subsequent model building and testing using the multivariable ROC tester function. The 10-feature model was trained on of 2/3 samples using random forest prior to internal validation on the held-out set consisting of 1/3 samples. Key performance metrics included AUC of multivariable ROC, confusion matrix of predicted probabilities of each sample and predicative accuracy. Cohen’s kappa (K) was used to measure the extent of agreement between predicted samples and actual samples for a dataset. Youden’s index (J) combined the sum of the true positive values and true negative values derived from the confusion matrix into a single metric and was used to assess how well the model predicted conditions of interest. Information regarding the key pathological characteristics of EC tumours previously determined from histopathological assessment of endometrial tissues was analysed for differences in lipid concentrations. We grouped participants with EC based on histological grade (other EC vs grade 1 and 2 EEC) of tumour, tumour size (>2 cm vs ≤2 cm), depth of myometrial invasion (present vs absent) and mismatch repair (MMR) protein status (MMR-deficient EC vs MMR-proficient EC). Differences in concentrations of lipid species between EC groups were analysed using t-test and visualised as volcano plots. The ability of lipid species to discriminate between EC stratified based on MMR status or histological grade from women with benign conditions and healthy women was evaluated using univariable and multivariable ROC analyses. Details of all methods were described in the statistical test section.

Women who enrolled into this study ( n  = 223) were classified into three distinct groups: healthy controls ( n  = 49), women with benign conditions ( n  = 108), and women with EC ( n  = 66). Demographic and clinical variables, including comorbidities of women in the three groups are presented in Table  1 . The participants were primarily Caucasian ( n  = 169; 76%) and non-Hispanic ( n  = 145; 65%), 47 years old on average, with majority of participants being categorised with a body mass index (BMI) of ≥35 ( n  = 83; 37%), with overall mean BMI of 33 kg/m 2 . Both, overall and pairwise differences in variables among the groups were assessed. Women with EC were significantly ( p  < 0.0001) older (mean age: 59 years) and had a significantly ( p  < 0.0001) higher BMI (mean: 39 kg/m 2 ; median: 38 kg/m 2 ) when compared to benign conditions (mean age: 45 years; mean BMI: 30 kg/m 2 ; median BMI: 28 kg/m 2 ) and healthy controls (mean age: 38 years; mean BMI: 30 kg/m 2 ; median BMI: 28 30 kg/m 2 ). Significant p  < 0.17 differences in race between all three groups were observed; Caucasian women formed the majority of participants within the healthy control group ( n  = 43; 87%), while the EC group had a larger proportion of other or unspecified race ( n  = 16; 24%). We created an additional comparison combining women with benign conditions and healthy women ( n  = 175). A descriptive statistical analysis of this grouping is shown in Additional File 1: Table  S1 . Age and BMI remained significantly ( p  < 0.0001) higher in the EC group (mean age 59 years; mean BMI 39 kg/m 2 ) relative to the combined control group (mean age 43 years; mean BMI 30 kg/m 2 ). Race and ethnicity also remained statistically different ( p  = 0.0249); Caucasian women made up the majority of both, EC ( n  = 48; 73%) and combined control group ( n  = 121 (77%)). Overall, statistically significant variables such as age, BMI, race, and ethnicity were consistent with known risk factors for EC. Table 1 Patient demographics All Benign conditions (1) Healthy women (2) Endometrial cancer (3) P value ( n  = 223) ( n  = 108) ( n  = 49) ( n  = 66) overall paired Age (mean (SD)) ( n  =  223) 47.99 (12.72) 45.55 (10.01) 38.35 (8.79) 59.14 (11.16) <0.0001 (1) vs (2) 0.0002 (1) vs (3) <0.0001 (2) vs (3) <0.0001 Race (n (%)) ( n  =  221)  White/Caucasian 169 (76.47) 78 (72.90) 43 (87.76) 48 (73.85) 0.017 (1) vs (2) 0.107  Black/African American 15 (6.79) 11 (10.28) 3 (6.12) 1 (1.54) (1) vs (3) 0.585  Other/not specified 37 (16.74) 18 (16.82) 3 (6.12) 16 (24.62) (2) vs (3) 0.0176 Ethnicity (n (%)) ( n  =  222)  Non-Hispanic 145 (65.32) 76 (70.37) 18 (36.74) 51(78.46) <0.0001 (1) vs (2) <0.0001  Hispanic 77 (34.69) 32 (29.63) 31 (63.26) 14 (21.54) (1) vs (3) 0.2883 (2) vs (3) <0.0001 BMI (mean (SD)) ( n  =  222) 33.19 (10.07) 30.63 (7.54) 30.00 (9.10) 39.69 (11.40) <0.0001 (1) vs (2) 0.9161 (1) vs (3) <0.0001 (2) vs (3) <0.001 BMI (median (SD)) (n  =  222) 30.97 (10.07) 28.66 (7.54) 28.00 (9.10) 38.46 (11.40) <0.0001 (1) vs (2) 0.9108 (1) vs (3) <0.0001 (2) vs (3) <0.0001 BMI (n (%)) ( n  =  222)  <25 45 (20.27) 23 (21.30) 16 (33.33) 6 (9.09) <0.0001 (1) vs (2) 0.4176  25–29 60 (27.03) 38 (35.19) 14 (29.17) 8 (12.12) (1) vs (3) <0.0001  30–34 34 (15.32) 19 (17.59) 6 (12.50) 9 (13.64) (2) vs (3) <0.0001  ≥35 83 (37.39) 28 (25.93) 12 (25.00) 43(65.15) Education (n (%)) ( n  =  193)  High school diploma/GED or less 53 (28.21) 23 (21.90) 13 (44.12) 17 (30.36) 0.0182 (1) vs (2) 0.0048  Some college 43 (22.05) 24 (22.86) 5 (14.70) 14 (25.00) (1) vs (3) 0.4582  Association degree or technical certification 34 (17.44) 22 (20.95) 0 (0.00) 12 (21.43) (2) vs (3) 0.0083  Bachelor’s degree or higher 63 (32.31) 36 (34.29) 14 (41.18) 13 (23.21) Marital status (n (%)) ( n  =  222)  Single/divorced/ widowed 105 (47.30) 41 (37.96) 28 (57.14) 36 (55.39) 0.0003 (1) vs (2) 0.0004  Married/cohabitating 108 (48.65) 65 (60.19) 15 (30.61) 28 (43.08) (1) vs (3) 0.082  Other/ not specified 9 (4.05) 2 (1.85) 6 (12.25) 1 (1.54) (2) vs (3) 0.0411 Gravidity (n (%)) ( n  =  218)  0–1 75 (34.40) 31 (28.97) 15 (32.61) 29 (44.62) 0.1393 (1) vs (2) 0.3636  2 45 (20.64) 23 (21.50) 10 (21.74) 12 (18.46) (1) vs (3) 0.1532  3 48 (22.02) 27 (25.23) 6 (13.04) 15 (23.08) (2) vs (3) 0.0737  4+ 50 (22.94) 26 (24.30) 15 (32.61) 9 (13.85) Parity (n (%)) ( n  =  219)  0–1 79 (36.07) 31 (28.97) 19 (41.30) 29 (43.94) 0.3879 (1) vs (2) 0.3855  2 45 (20.55) 23 (21.50) 10 (21.74) 12 (18.18) (1) vs (3) 0.2031  3 49 (22.37) 27 (25.23) 7 (15.22) 15 (22.73) (2) vs (3) 0.6468  4+ 46 (21.01) 26 (24.30) 10 (21.74) 10 (15.15) Smoker (n (%) ( n  =  213)  Yes 26 (12.21) 14 (13.46) 8 (17.02) 4 (6.45) 0.0457 (1) vs (2) 0.3013  No 166 (77.93) 81 (77.89) 38 (80.85) 47(75.81) (1) vs (3) 0.1083  Quit 21 (9.86 9 (8.65) 1 (2.13) 11(17.74) (2) vs (3) 0.0128 History of PCOS (n (%)) ( n  =  186)  Yes 14 (7.53) 10 (11.36) 0 (0.00) 4 (7.84) 0.0357 (1) vs (2) 0.0148  No 172 (92.47) 78 (88.64) 47 (100.00) 47 (92.16) (1) vs (3) 0.5733 (2) vs (3) 0.1186 P -values were calculated using unpaired t-tests for age and BMI. Fisher’s exact and Chi-squared tests were used for analysing categorical variables BMI body mass index, PCOS polycystic ovary syndrome, SD standard deviation Patient demographics P -values were calculated using unpaired t-tests for age and BMI. Fisher’s exact and Chi-squared tests were used for analysing categorical variables BMI body mass index, PCOS polycystic ovary syndrome, SD standard deviation Metabolon’s Complex Lipid Panel identified a total of 914 individual complex lipid species across 223 CVL samples collected from women with EC, benign conditions, and healthy controls. Lipid species were classified into one of 14 classes, with each class containing between 12 to 518 individual species (Fig.  1 A). Triacylglycerides (TAG) accounted for the largest number of detected complex lipid classes ( n  = 518), followed by phosphatidylethanolamines (PE; n  = 114), phosphatidylcholines (PC; n  = 68), and diacylglycerols (DAG; n  = 59). Sphingomyelins (SM), lactosylceramides (LCER), hexosylceramides (HCER), and ceramides (CER), were among the least frequently detected classes, with only 12 lipid species per class. Lipid classes were further classified into three distinct lipid groups, of which neutral lipids were the most frequently detected ( n  = 629), followed by phospholipids ( n  = 224) and sphingolipids ( n  = 61). Due to the large number of individual lipid species within the dataset, we initially examined lipid class concentrations to evaluate differences between all patient groupings. Of the 14 lipid classes analysed, the mean concentrations of ten classes were significantly enriched in EC, relative to benign conditions ( p -value range from 0.001 to <0.001). The mean concentrations of 12 classes were significantly enriched between EC and healthy controls ( p -value range from 0.0183 to <0.001); (Fig.  1 B). Moreover, the mean concentrations of seven classes were significantly enriched ( p -value range from 0.0015 to 0.0442) between benign conditions and heathy controls. These classes included cholesteryl esters (CE; p  = 0.0222), lysophosphatidylcholines (LPC; p  = 0.0442), PE ( p  = 0.0259), CER ( p  = 0.0032), HCER ( p  = 0.0032), lysophosphatidylethanolamines (LPE; p  = 0.0070), and SM ( p  = 0.0094). No significant differences were reported in mean concentrations of MAG or diacylglycerides (DAG) in EC relative benign condition or healthy controls. Fig. 1 Lipid classes were significantly enriched between women with endometrial cancer, benign conditions, and healthy women. A Bar chart summarizes the total number of lipid species detected across different lipid classes, in all patients. Colour-coded by lipid class. B Violin plots show the total concentration of all detected lipid classes in individual samples among disease groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001) Lipid classes were significantly enriched between women with endometrial cancer, benign conditions, and healthy women. A Bar chart summarizes the total number of lipid species detected across different lipid classes, in all patients. Colour-coded by lipid class. B Violin plots show the total concentration of all detected lipid classes in individual samples among disease groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001) In addition to the lipid classes described above, we analysed two sets of data regarding fatty acid (FA) concentration. The first dataset represented the absolute concentrations of 28 individual FA chains, which were summed across all 14 complex lipid classes, consequently providing a class-independent measure of each FA chain (e.g., total FA12:0). The second data set represented the overall concentration of a given FA chain within all individual lipids designated to a specific lipid class (e.g., CER (FA12:0)). We employed fold change and t-tests to identify potential EC-associated changes to FA metabolism. The results were visualised by volcano plots to compare the absolute concentrations of fatty acid chains in EC to benign conditions and/or healthy controls. Of the total 28 FAs, 18 were significantly enriched (FC > 2.0; q   2.0; q  < 0.05) of ten FAs relative to the combined group consisting of benign conditions and healthy controls (Fig.  2 C). Overlap of significantly enriched FAs across three independent comparisons were visualised using a Venn diagram (Fig.  2 D). Unsaturation levels and length of carbon chains differed between enriched FAs; notably, FAs with higher unsaturation levels and longer carbon chains were associated with EC. Of the 18 significantly enriched FAs between EC and healthy controls, nine were also significantly enriched in EC relative to the combined group consisting of benign and healthy controls. Only the total concentration of FA16:1 was significantly enriched (FC > 2.0; q  < 0.05) in EC, independent of the comparison grouping. We identified ten FAs to be significantly enriched in EC when comparing to healthy controls or the combined group consisting of benign and healthy controls. These included FA side chains 16:1, 18:1, 18:2, 20:1, 20:2, 20:3, 20:4, 22:0, 22:2, 22:4. When the cervicovaginal concentrations of these ten FAs were plotted in individual samples and compared among disease groups (Fig.  2 E), their mean total concentrations of FAs 18:1, 20:4, 22:2 and 22:4 were significantly enriched ( p  < 0.0001) in EC relative to benign conditions and healthy controls. Only FAs 18:1, 20:3, 20:4 and 22:2 were significantly enriched between benign conditions and healthy controls ( p -value range: 0.069–0.0443). The concentrations of the remaining 18 FAs were plotted in Additional File 2: Fig.  S1 . After adjusting for age and BMI only FA 20:0, 20:3, and 24:1 remained significantly enriched in EC relative to healthy controls ( p -value range: 0.0239 – 0.0394; Additional File 1: Table  S2 ). Fig. 2 Women with EC exhibit significant enrichment in total concentrations of fatty acid side chains. Following log2 transformation and t-test with FDR-correction, volcano plots display concentrations of 28 fatty acid side chains integrated into complex lipids in ( A ) women with EC and healthy women, ( B ) women with EC and women with benign conditions, and ( C ) women with EC and combined benign conditions and healthy women. Colour-coded by enrichment and significance. D Venn diagram compares all significantly enriched (FC > 2.0, q  < 0.05) fatty acid side chains in EC relative to benign conditions and healthy women, highlighting shared side chains. E Violin plots show the total concentrations of ten significantly enriched fatty acid side chains in EC, shared by benign and healthy control groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001) Women with EC exhibit significant enrichment in total concentrations of fatty acid side chains. Following log2 transformation and t-test with FDR-correction, volcano plots display concentrations of 28 fatty acid side chains integrated into complex lipids in ( A ) women with EC and healthy women, ( B ) women with EC and women with benign conditions, and ( C ) women with EC and combined benign conditions and healthy women. Colour-coded by enrichment and significance. D Venn diagram compares all significantly enriched (FC > 2.0, q  < 0.05) fatty acid side chains in EC relative to benign conditions and healthy women, highlighting shared side chains. E Violin plots show the total concentrations of ten significantly enriched fatty acid side chains in EC, shared by benign and healthy control groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001) The relative distribution of individual FA chains across lipid classes can be found in Additional File 2: Fig.  S2 A. Of the 28 individual FA chains detected, FAs 16:0,18:0 and 18:1 were identified within different species distributed across all 14 lipid classes. The FA profile was most altered in EC when compared to healthy controls with significant alterations observed in a total of 107 FA composites (FC > 2.0; q  < 0.05; Additional File 2: Fig.  S2 ). Similarly to the total FA data set, we observed significant enrichment in concentrations of 43 FA side chains when comparing EC to benign conditions (Additional File 2: Fig.  S2 B). This comparison had the lowest number of significantly enriched FAs as we observed significant enrichment of 66 FAs between EC and combined benign conditions and healthy controls (Additional File 2: Fig.  S2 C). FA chains 22:4, 20:2, 24:1, 22:2, 22:1 and 20:3 within the CE class were the most enriched in EC relative to benign conditions, healthy controls and the combined benign conditions and healthy control group. To identify individual lipid species associated with EC, data was log2-transformed to determine fold change differences and analysed by t-test with FDR correction for multiple comparison to determine differences between EC from benign conditions and healthy controls. Comparing EC to healthy controls, 423 individual lipid species were significantly enriched and eight were significantly depleted (FC > 2.0, q  < 0.05; Fig.  3 A). TAG species dominated the number the significantly altered species, accounting for over 70% of both enriched and depleted species in EC compared to healthy controls. CEs contributed 9% to the enriched species, with CE(22:4) being the single most enriched lipid in EC. Other significantly enriched species belonged to PE, PC, LPC, DAG, SM, HCER, CER classes. Comparison of EC to benign conditions revealed 212 significantly enriched lipid species, of which 84% were TAGs and 9% were CEs (Fig.  3 B). Together, the number of LCER, HCER, DAG, CER and DCER species accounted for less than 10% of significantly enriched species. Log2 fold change and t-test with FDR correction was also applied to determine significant differences in species between EC and the combined group consisting of benign conditions and healthy controls. (Fig.  3 C). Only one lipid was significantly depleted while 287 species (primarily belonging to neutral lipids) were significantly enriched in EC. We utilised a Venn diagram to compare, and contrast altered species associated with EC among the three groupings (Fig.  3 D). The lipid profile was most dysregulated between EC and healthy controls, with 143 uniquely altered species. We observed fewest differences between EC and benign conditions; only five species exhibited unique changes between these conditions. When EC was compared to the combined control group, no unique species were observed. Instead, 80 species were shared with findings from the comparison of EC to healthy controls. Overall, 208 species were altered in EC irrespective of whether compared to benign condition or healthy controls. To demonstrate distribution of concentrations among all analysed samples, we plotted concentrations of all EC participants, benign conditions, and healthy controls in species most enriched in volcano plot analyses (Fig.  3 E). The mean concentrations of TAG57:4-FA20:4 were significantly enriched between EC vs benign conditions ( p  < 0.0001), EC vs healthy controls ( p  < 0.0001), and benign conditions vs healthy controls ( p  = 0.0407). Mean concentrations of CE(22:4) were also significantly enriched in EC vs benign conditions ( p  < 0.0001), EC vs healthy controls ( p  < 0.0001), but no significant differences were found between benign conditions and healthy controls. Overall, this analysis identified numerous lipid species which could potentially be applied in distinguishing EC from benign conditions and healthy controls. Fig. 3 Neutral lipids were significantly dysregulated in women with EC, relative to benign and healthy controls. Volcano plots show results of lipid species with significantly altered concentrations following FDR-corrected t-test ( q  < 0.05) coupled with log2 fold change in ( A ) women with EC and healthy women, ( B ) women with EC and women with benign conditions, and ( C ) women with EC vs combined benign conditions and healthy women. Each point represents a single sample coloured based on lipid class. Top ten most significantly enriched and depleted species in EC are labelled. D Venn diagram compares all significantly enriched (FC > 2.0, q  < 0.05) species in EC, which were shared or unique to benign conditions and healthy women. E Violin plots showing cervicovaginal concentrations of key lipid species enriched in EC, in individual samples among disease groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001) Neutral lipids were significantly dysregulated in women with EC, relative to benign and healthy controls. Volcano plots show results of lipid species with significantly altered concentrations following FDR-corrected t-test ( q  < 0.05) coupled with log2 fold change in ( A ) women with EC and healthy women, ( B ) women with EC and women with benign conditions, and ( C ) women with EC vs combined benign conditions and healthy women. Each point represents a single sample coloured based on lipid class. Top ten most significantly enriched and depleted species in EC are labelled. D Venn diagram compares all significantly enriched (FC > 2.0, q  < 0.05) species in EC, which were shared or unique to benign conditions and healthy women. E Violin plots showing cervicovaginal concentrations of key lipid species enriched in EC, in individual samples among disease groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001) Pathway enrichment tools were used to identify perturbated metabolic pathways associated with EC. We discovered five enriched pathways in EC relative to healthy women: phospholipid biosynthesis, bile acid and steroid biosynthesis, glycerolipid and arachidonic acid metabolism (Additional File 2: Fig.  S3 ). Additionally, sphingolipid metabolism was enriched when comparing EC with benign conditions and the combined group of benign conditions and healthy women. Arachidonic acid metabolism was the most enriched pathway (log10 p -value range: 0.75–0.20) in EC, and bile acid biosynthesis was the least (log10 p -value 0.00). We also performed an origin-based enrichment analysis to identify the source of species detected by lavage sampling (Additional File 2: Fig.  S4 ). The greatest number of species originated from food-related sources ( n  = 319), and the least from pharmaceutical sources ( n  = 33). The source could not be determined for 595 species in the dataset. A total of 175 species were mapped to the host, 167 of which could also originate from microbial pathways. Significantly enriched species were co-metabolites identified within steroid biosynthesis ( p  = 0.063) and glycerophospholipid metabolism ( p  = 0.063) pathways. Sphingolipid metabolism pathway was significantly enriched by species from host only and host-microbial metabolism, the latter of which was most significant ( p  = 0.02 vs p  = 0.00000497). To estimate the diagnostic potential of altered lipid species in EC, binary logistic regression was applied to each lipid species and corresponding univariable ROC curves were plotted to determine the area under the curve (AUC). Comparing univariable ROC analyses for EC vs healthy controls, EC vs benign conditions and EC vs combined benign and healthy controls, we identified a greater number of species that belonging to various classes and met the AUC ≥ 0.8 threshold for discriminating EC from healthy controls (Fig.  4 A). Of the 14 identified lipid classes, species reaching the AUC ≥ 0.8 criteria spanned ten classes in EC vs healthy controls comparison, five in EC vs combined benign and healthy controls and only two classes in EC vs benign conditions. Of these, TAGs represented the highest quantity of species that reached the AUC ≥ 0.8 threshold for EC relative to healthy controls ( n  = 125) and combined benign and healthy controls ( n  = 19). Of the seven species that exhibited an AUC ≥ 0.8 for discriminating EC from benign conditions, five were CEs. To directly compare the diagnostic potential of species for distinguishing EC from different control groups, we plotted 25 lipids with the highest AUC values for EC relative to healthy controls (Fig.  4 B ; Additional File 1: Table  S3 ). Information on candidate biomarkers for EC relative to benign conditions and EC relative to combined benign and healthy controls can be found Additional File 1: Tables  S4 and S5 . AUC values of 25 lipids ranged between 0.850 to 0.907 for EC relative to healthy controls: between 0.780 and 0.847 for EC relative to benign conditions: between 0.804 and 0.865 for EC relative to combined benign and healthy controls. Overall, TAGs and CEs were the strongest discriminators for EC regardless of the comparator groups and demonstrated highest potential for discriminating EC from healthy controls. Fig. 4 Complex lipid species discriminate EC from healthy controls most accurately relative to benign conditions. A Scatter plot of 25 lipid species identified by univariate receiver operating characteristic (ROC) with strongest potential to discriminate between EC and healthy controls, relative to their discriminative potential for EC vs benign conditions or combined benign conditions and healthy controls. B Bar graph displays total number of species with area under the curve (AUC) ≥ 0.8, indicating good predictive potential of EC when compared to benign conditions, healthy controls or combined benign conditions and healthy controls. C Univariate AUC plots for the top three potential EC biomarkers Complex lipid species discriminate EC from healthy controls most accurately relative to benign conditions. A Scatter plot of 25 lipid species identified by univariate receiver operating characteristic (ROC) with strongest potential to discriminate between EC and healthy controls, relative to their discriminative potential for EC vs benign conditions or combined benign conditions and healthy controls. B Bar graph displays total number of species with area under the curve (AUC) ≥ 0.8, indicating good predictive potential of EC when compared to benign conditions, healthy controls or combined benign conditions and healthy controls. C Univariate AUC plots for the top three potential EC biomarkers Notably, CE species were consistently ranked as the top features with the highest potential for discriminating EC from healthy controls, benign conditions, and the combined benign and healthy control group. Univariable ROC curves of three most discriminative CE species for EC relative to healthy controls are displayed in Fig.  4 C. ROC curves for EC relative to benign conditions, plus the combined benign and healthy control group were also displayed on the same graph for comparison. CE(22:4) was the single species most capable of distinguishing EC from all control groups; exhibiting excellent diagnostic potential for EC relative to healthy controls (AUC: 0.907), and close to excellent diagnostic potential for EC relative to benign conditions (AUC: 0.845) and relative to combined benign and healthy controls (AUC: 0.865). Additionally, CE(22:4) reached the highest sensitivity value when discriminating EC from healthy controls with 0.864, respectfully. CE(20:3) and CE(20:2) were most sensitive for EC when discriminating against benign conditions or combined benign conditions and healthy controls. Among all three species, CE(22:4) exhibited the highest specificity (>0.8) for distinguishing EC from all control groups, reflecting its potential for identifying benign conditions and healthy controls. Overall, lipid species present as promising biomarker candidates accurately distinguishing EC from both benign conditions and healthy controls, especially CEs. Furthermore, for the top predictive lipids identified in ROC analysis, the discriminatory potential remained comparable when distinguishing women with EC from women with benign conditions among those who reported symptoms of abnormal uterine bleeding (AUB) and/or post-menopausal bleeding (PMB) (Additional File 2: Fig.  S5 A). Notably, two out of the four species, CE(20:3) and CE(20:2) even exhibited slightly higher AUC values (0.812; 0.840) in the presence of AUB/PMB symptoms relative to the absence of symptoms (0.798; 0.835). The levels of all four lipid species CE(22:4), CE(20:3), CE(20:2), TAG54:7-FA20:4 were significantly higher in EC to benign conditions within women reported AUB/PMB ( p  < 0.0001) as well as without reported AUB/PMB ( p range 0.0010 to <0.0001) (Additional File 2: Fig.  S5 B). Subsequently, while women often present with EC-associated bleeding, blood in samples did not drastically affect the ability of lipid species to discriminate between EC and benign conditions. Machine learning was used to further assess predictive capabilities of lipid species for distinguishing EC from a combined group of benign conditions and healthy controls. The combined benign and healthy control group better represents the heterogeneity of women in clinical settings. Individual species were used in different combinations and quantities to build multivariable diagnostic models with the aim of developing a diagnostic lipid panel for EC with high specificity and sensitivity. The initial exploratory phase utilised the random forest algorithm and the training dataset (consisting of 2/3 of samples) to identify an optimal combination of lipid features and construct a model to distinguish EC from combined benign and healthy controls (Fig.  5 A). The exploratory multivariable models were built by incorporating 5, 10, 15, 25, 50 or 100 lipid species together. The AUC values of the exploratory models remained consistent (0.825 and 0.820) across all configurations, except for the five-feature model. A multivariable model consisting of 10 features exhibited the highest predictive accuracy of 75.9%, along with high specificity and sensitivity; 73.6% and 79.5% (Fig.  5 B). Additionally, lipid species such as CE (22:4), CE(20:2) and TAG54:7-FA20:4, which were previously identified as top predictors for EC by univariable ROC analysis, were identified of high importance and therefore consistently incorporated within multivariable models (Fig.  5 C). Performance did not improve substantially in models which incorporated more than 10 features. Therefore, to support the development of a practical and minimally invasive assay for EC, subsequent analysis focused on a 10-feature model for downstream evaluation. Fig. 5 Exploratory multivariable ROC analysis identifies complex lipids with discriminatory potential for EC. A ROC curves of exploratory multivariable models built using random forest. Exploratory multivariable models were generated based on 2/3 of samples and reported as average performance across MCCV runs. Confusion matrix ( B ) and top 15 important features ( C ) for the exploratory 10-feature multivariable model Exploratory multivariable ROC analysis identifies complex lipids with discriminatory potential for EC. A ROC curves of exploratory multivariable models built using random forest. Exploratory multivariable models were generated based on 2/3 of samples and reported as average performance across MCCV runs. Confusion matrix ( B ) and top 15 important features ( C ) for the exploratory 10-feature multivariable model The top ten most important features for differentiating EC from a combined group of benign and healthy controls, which were pre-selected by random forest ((CE(22:4), TAG54:3-FA16:0, TAG58:7-FA18:1, CE(22:1), CE(24:1), TAG54:7-FA20:4, TAG54:4-FA20:3, TAG58:7-FA18:2, CE(20:2) and CE(22:2)) were used to build the 10-feature model and utilised for training and internal validation purposes (Fig.  6 D). AUCs of 0.865 and 0.831 for the training and validation models demonstrate comparable performance, indicating stable performance within an internally validated data set. This 10-feature model was then trained and validated to differentiate EC from healthy and benign controls. Figure  6 A depicts the ROC curves for EC vs healthy controls using a training set and validation set, both of which exhibit close to excellent discriminatory potential with AUCs of 0.889 and 0.881, respectively. The confusion matrix shows the number of samples correctly classified as EC or healthy controls in the training (Fig.  6 B) and validation sets (Fig.  6 C). The 10-feature model obtained high predictive accuracy of 83.1% and 84.2% and specificity of 81.8% and 75.0% on both training and validation sets. Notably, the 10-feature model demonstrated high sensitivity of 90.9% in the validation set by correctly classifying 20 out 22 EC samples. Moreover, the 10-feature model also achieved AUC values of 0.858 and 0.813 for EC vs benign conditions in training and validation sets; demonstrating good discriminatory potential (Fig.  6 D). Predictive accuracy (79.3% and 75.9), sensitivity (81.8% and 77.3%) and specificity (77.8% and 75.0%) values were also high for EC vs benign (Fig.  6 E and F), although slightly less than compared to EC vs healthy controls. This analysis revealed that a model created from 10 complex lipids can accurately discriminate EC from benign and healthy controls, with high specificity and sensitivity of ≥75.0%. By pairing complex lipid species with machine learning we were able to effectively discriminate EC from benign and healthy controls and validate them on an internal set of one third of the samples. Statistical tests were performed to adjust for BMI and age differences for key potential biomarkers (Additional File 1: Table  S6 ). Subsequently, CE(24:1), CE(22:1), LCER(24:0), DCER(26:0), HCER(26:0) and CE(24:0) were no longer significantly associated with EC. Fig. 6 Training and validation of 10-feature model demonstrates diagnostic ability for EC. Plot of ROC curves created using pre-selected ten-feature model on training (2/3 samples) and validation (held-out 1/3 samples); computed by random forest in ( A ) EC vs healthy controls and ( B ) EC vs benign conditions. Confusion matrix demonstrating percentage of EC and healthy control samples predicted correctly/incorrectly by pre-selected 10-feature model in training ( B ) and validation sets ( C ). Confusion matrix of showing percentage of EC and benign samples predicted correctly/incorrectly by 10-feature model in training ( E ) and validation sets ( F ) Training and validation of 10-feature model demonstrates diagnostic ability for EC. Plot of ROC curves created using pre-selected ten-feature model on training (2/3 samples) and validation (held-out 1/3 samples); computed by random forest in ( A ) EC vs healthy controls and ( B ) EC vs benign conditions. Confusion matrix demonstrating percentage of EC and healthy control samples predicted correctly/incorrectly by pre-selected 10-feature model in training ( B ) and validation sets ( C ). Confusion matrix of showing percentage of EC and benign samples predicted correctly/incorrectly by 10-feature model in training ( E ) and validation sets ( F ) On the basis that complex lipid species presented robust potential as diagnostic biomarkers in distinguishing EC from benign conditions and healthy controls, we also investigated associations between complex lipid species and tumour characteristics. Using data obtained from pathology reports we were able to stratify participants diagnosed with EC into four distinct groups based on mismatch repair (MMR) status (MMR-deficient vs MMR-proficient), presence of myometrial invasion (present vs absent), tumour size (>2 cm vs ≤2 cm) and histologic grade (other EC subtypes vs grades 1 and 2 EEC). The number of women within each category is shown in Additional File 1: Table  S7 . Initially, we utilised the univariable ROC analysis to identify lipids which could discriminate EC based on histologic grade (other EC subtypes and grades 1 and 2 EEC) or MMR status (MMR-deficient and MMR-proficient) from healthy controls, benign conditions, or the combined group consisting of benign conditions and healthy controls. When comparing healthy controls to EC based on histologic grade, the AUC values for other EC subtypes were substantially higher. The AUC values of top 25 species for other EC subtypes ranged from 0.930 to 0.980 (Fig.  7 A), in comparison to the AUC values of top 25 species for discriminating grades 1 and 2 EEC from healthy controls (AUC range from 0.840 to 0.890). Similarly, all the top 25 species used to discriminate between healthy women and women with MMR-deficient EC surpassed the AUC > 0.9 threshold (AUC range: 0.930–0.980; Fig.  7 C). While the univariable ROC analysis revealed no lipid species with AUC values of 0.9 for women with MMR-proficient EC relative to healthy women, top species did exhibit close to excellent discriminatory potential for this comparison (AUC range: 0.850–0.880). Notably, CE(22:4) exhibited the greatest discriminatory potential, irrespective of the tumour characteristic used for stratification. The cervicovaginal concentrations of CE(22:4) were compared among disease groups. Significant enrichment was observed in other EC subtypes, relative to benign conditions ( p  < 0.0001) and healthy controls ( p  < 0.0001), and in grades 1 and 2 EEC relative to benign conditions ( p  < 0.0001) and healthy controls ( p  < 0.0001); (Fig.  7 B). Significantly higher concentrations of CE(22:4) were also observed in MMR-deficient and MMR-proficient EC relative to benign conditions and healthy controls ( p  < 0.0001); (Fig.  7 D). Building on this, we aimed to determine alterations in lipid species associated with each of the four tumour characteristics. Volcano plots were used to visualise enriched and/or depleted lipid species between tumour characteristic subgroups. We identified the highest number of altered species when comparing women with MMR-deficient EC to women with MMR-proficient EC, with 23 species enriched and one depleted (Additional File 2: Fig.  S6 A). For histological grade, six species were enriched and two depleted (Additional File 2: Fig.  S6 B). For myometrial invasion, six species were enriched (Additional File 2: Fig.  S6 C). Five of these species: CER(14:0), CER(16:0), CER(18:0), DCER(16:0) and TAG56:2-FA20:1 were also enriched when comparing between tumour size (>2 cm and ≤2 cm); (Additional File 2: Fig.  S6 D). Collectively, this analysis revealed numerous lipid species which relate to tumour severity, highlighting their potential use as markers associated with prognostic features important for early detection and management of EC. Fig. 7 EC severity can be determined using lipid species identified by univariable receiver operating characteristic (ROC). Scatter plots display 25 lipid species with highest discriminatory potential for ( A ) other EC subtypes and grades 1 with 2 EEC vs healthy controls, ( C ) MMR-deficient EC and MMR-proficient EC vs healthy controls. For comparison, the potential of those same 25 lipid species to distinguish EC based on histological type and MMR status from benign conditions or combined benign conditions and healthy controls was also included. Violin plots showing cervicovaginal concentrations of key lipid species enriched in EC based on histological grade ( B ) and MMR status ( D ), in individual samples among disease groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001) EC severity can be determined using lipid species identified by univariable receiver operating characteristic (ROC). Scatter plots display 25 lipid species with highest discriminatory potential for ( A ) other EC subtypes and grades 1 with 2 EEC vs healthy controls, ( C ) MMR-deficient EC and MMR-proficient EC vs healthy controls. For comparison, the potential of those same 25 lipid species to distinguish EC based on histological type and MMR status from benign conditions or combined benign conditions and healthy controls was also included. Violin plots showing cervicovaginal concentrations of key lipid species enriched in EC based on histological grade ( B ) and MMR status ( D ), in individual samples among disease groups. Thick, horizontal line represents the mean, thinner line represents quartiles. P values were determined using one-way ANOVA with Tukey’s post hoc test (* p  < 0.05, ** p  < 0.01, *** p  < 0.001, **** p  < 0.0001)

Endometrial cancer (EC) is the fourth most diagnosed cancer among women in high-income countries, exerting substantial influence on women globally. EC is increasingly important due to its rising incidence rates. It is estimated that by 2030, EC will surpass colorectal cancer as the third most common and fourth leading cause of cancer-related deaths in women worldwide [ 1 ]. EC impacts women across their lifespan. Among all women diagnosed with EC, approximately 15% of EC cases occur in premenopausal women and 70% among postmenopausal women [ 2 ]. Increasing incidence rates of EC in women may partially be attributed to Westernised diets which are rich in fats. Additionally, socioeconomic disparities may limit access to nutritious food options and reliable transportation, contributing to risk and delayed diagnosis [ 3 ]. Coincidently, diets high in fat are closely associated with key risk factors of EC: high body mass index (BMI), obesity, and metabolic syndrome. Westernised diets also influence endocrine-related risk factors like increased oestrogen exposure, early menstruation, or late menopause. Unfortunately, there are no accurate and minimally invasive screening tools available for early EC detection and management [ 4 ]. The current diagnostic flow for EC still relies on endometrial tissue samples collected using methods dating back to 1984 [ 5 ]. Symptomatic women will be diagnosed by traditional methods subjective of post-menopausal bleeding. Endometrial biopsies coupled with hysteroscopy or dilation and curettage remain the gold standard for EC diagnosis. Despite their high accuracy, traditional methods pose challenges. Their invasive and painful approach coupled with the requirement for trained personnel, medical equipment, and travel to clinic may prevent women from receiving a timely diagnosis and subsequent treatment [ 6 , 7 ]. This highlights the need to establish more accessible and acceptable methods for early diagnosis of EC. Novel less invasive tests could improve screening adherence [ 8 ]. Early diagnosis of EC is critical for effective management of the condition as diagnosis is closely followed by assessing the likely outcome of disease. Prognosis of women diagnosed with EC depends on tumour characteristics such as the size of tumour, presence of myometrial invasion, histological subtype [ 9 ], and molecular classification [ 10 ]. These prognostic features are evaluated through histopathological assessment of endometrial tissue resected by biopsy or dilation and curettage, followed by imaging tests such as a transvaginal ultrasound (TVUS) to determine the severity of the malignancy [ 8 ]. At present, clinical prognostic algorithms stratify women with EC into distinct groups based on National Comprehensive Cancer Network (NCCN) or European Society for Medical Oncology (ESMO), European SocieTy for Radiotherapy & Oncology (ESTRO), and European Society of Gynaecological Oncology (ESGO) guidelines. Cancer Genome Atlas (TCGA) molecular profiling stratifies women according to their mutation burden and copy number alterations and is being adopted into clinical practice in the US and European countries [ 11 ]. Despite the high discriminatory power for stratifying risk, these molecular classification systems continue to rely on samples collected in an invasive manner. This highlights the need for improved biomarkers measured preoperatively in non-invasive samples, to reduce need for invasive interventions, patient anxiety and healthcare costs, and simultaneously improve clinical outcomes [ 1 , 10 , 12 ]. Altered lipid metabolism is an established hallmark of many cancers [ 13 – 15 ], with rising evidence linking it with EC. To date, only a handful of early stage, pre-clinical studies have explored the etiologic link between lipidomic changes which coincide with the development and progression of EC [ 16 – 19 ]. Previous diagnostic models built to detect EC utilised blood-based lipid biomarkers which belonged to a broader class of phospholipids, and ceramides [ 18 , 20 , 21 ]. Concentrations of other lipid classes measured in blood plasma and serum have already been implemented in clinical practice to assist in management of lipid-associated conditions. Triacylglycerides (TAGs) and cholesterol-associated metabolites are key fat components found in blood plasma. Elevated levels of these lipids often reflect hyperlipidaemia, obesity or metabolic syndrome which are risk factors for EC [ 4 ]. As such, additional research is required to validate the idea of lipid biomarkers as an approach for the diagnosis and or prognosis of EC. In our previous global metabolomics study, we utilised minimally invasive cervicovaginal lavage (CVL) sampling which identified sterol lipids, and other simple lipid signatures to be significantly dysregulated in EC [ 22 ]. Herein, we investigated the association between complex lipids detected within the cervicovaginal environment in the context of EC. Based on our previous findings [ 22 ], we hypothesised that minimally invasive lavage sampling could be applied to detect complex lipids within the lower female reproductive tract. We aimed to distinguish women diagnosed with EC from women with benign conditions and healthy women based on variations in complex lipid concentrations. We also explored whether changes in complex lipids were reflective of disease severity with the aim of stratifying women with EC based on key tumour characteristics. To achieve this, we coupled CVL sampling with targeted lipidomic analysis and measured the concentrations of lipid classes, individual lipid species and fatty acid side chains bound to complex lipids within the CVL samples collected from women newly diagnosed with EC, benign conditions, and healthy controls.

Using MS-based lipidomics to measure complex lipids in minimally invasive CVL samples, we provided additional insight into changes across lipid classes, fatty acids and lipid species associated with EC. Our observations indicate EC favours enrichment of long-chained PUFAs, mainly within neutral lipid pools, potentially indicating a shift towards lipid uptake and storage. Using machine learning tools, we discovered CE(22:4) held promising univariable diagnostic and capability for distinguishing EC according to prognostic features such as MMR status. Incorporating CE(22:4) in multivariable models led to excellent discrimination of women with EC, relative to a combined group of women with benign conditions and healthy women. Overall, we demonstrated that complex lipids in CVLs can be applied to detect EC and predict its severity.

While earlier studies have reported an association between complex lipids and EC, they have not explored their diagnostic utility in minimally invasive CVL samples. Our previous global metabolomics study of CVLs demonstrated EC progression and severity was associated with upregulated simple lipids. Glycerophosphoethanolamine, glycerophosphocholine, glycerophosphoserine and glycerophosphoglycerol previously discriminated between EC and benign conditions with good ability (AUC range 0.837–0.807) [ 22 ]. Herein, we coupled a targeted lipidomic approach with CVL analysis to explore lipidome changes in women with newly diagnosed EC, women with benign conditions and healthy women. This study demonstrates complex lipids in CVLs can be applied to detect EC and predict its severity. Herein, top discriminatory biomarkers were CEs and glycerides like TAGs. The AUC values of top three most discriminatory species ranged from 0.865 to 0.826 when discriminating EC from benign conditions or a combined group of benign and healthy controls; reaching 0.907 for EC vs healthy controls only. As women often undergo several procedures to receive a definite diagnosis, biofluid biomarkers collected in a minimally invasive manner would be beneficial as a pre-surgical approach to reduce multiple surgical procedures [ 25 ]. Peripheral blood, urine [ 26 ], and cervicovaginal fluid are potential sources of EC biomarkers. While the procedures used to obtain them are non- to minimally invasive, these approaches remain under development. Tumour-derived components within blood are limited by their insufficient concentrations during early stages of EC resulting in low sensitivity [ 8 , 27 ]. In contrast, urinary biomarkers are currently not yet specific enough for EC and require validation [ 28 ]. In a review article summarising current non-invasive strategies for EC, Shen and colleagues [ 8 ] endorsed cervicovaginal fluid as an optimal biofluid sample used to collect biomarker specimens for EC. CVLs provide a rich source of differential compounds which could be exploited as biomarkers for various malignancies, including EC. CVL offers several advantages over current standard of care procedures for EC diagnosis: including potential for self-administration, better patient tolerability, culturally appropriate and offer a sufficient volume for long term storage and future use [ 8 , 29 ]. CVLs already proved to be a promising collection technique for EC biomarkers as shown by our previous work [ 22 , 23 , 30 , 31 ] and that of others [ 29 , 32 ]; metabolites incorporated into multivariate diagnostic models accurately distinguished EC (AUC range between 0.780–1.00). Consequently, the novel biomarker test should utilise a sample which is collected in a culturally appropriate manner, be as minimally invasive as possible, easily accessible for all women and possess an appropriate sensitivity to specificity ratio [ 8 ]. Herein, lipid classes PE, LPE, LPC, CER, SM, HCER and CE were enriched between EC, benign conditions, and healthy controls. Phospholipids possess roles in autophagy, cell division, apoptosis, influence membrane fluidity, lipid raft formation and receptor localisation. Alterations in these lipids correspond with cell proliferation, migration, and cancer survival in tumour cells [ 33 ]. Changes in PE concentration vary across studies, demonstrating enrichment in prostate cancer cells [ 34 ], follicular fluid of endometriosis patients [ 35 ], while others describe depletion in EC tissue [ 36 ]. How PE contributes to early development of cancer remains unclear; however, within the tumour microenvironment, enriched levels were connected to autophagy-dependent therapy resistance and metastasis [ 37 ]. Elevated levels of LPE were previously observed in ovarian cancer tissues [ 35 ]. LPE functions in a similar manner to growth factors, signalling via G protein-coupled receptors to induce migration in ovarian cancer cells [ 38 ]. LPC in colorectal cancer have been associated with inflammation and their down-cycling to PCs in lipid droplets [ 39 ]. Sphingolipid classes CER, SM and HCER were previously enriched in tissue [ 40 ], plasma [ 41 ], and serum [ 42 ] of women with EC. While sphingolipid implication in EC was attributed to inflammation, growth and spread [ 41 ], the exact role may vary according to EC subtype, reflecting differences in disease biology across subtypes. Blahnio-Zabielska et al. [ 40 ] observed variations in enriched CER profiles between EC type I-Ishikawa and EC type II-HEC-1A EC cell lines, with type I EC exhibiting higher CER enrichment. However, ceramide metabolism in EC remains poorly understood and warrants further investigation [ 42 ]. CERs are substrates for HCERs and SMs, therefore EC-associated enrichment of CERs may reflect metabolic shifts in sphingolipids within the same pathway. HCERs were previously enriched murine models with stage V ovarian cancer, reflective of disease severity [ 43 ]. Sphingomyelin species previously been reported as potential biomarkers for ovarian cancer, and associated with endometriosis [ 43 , 44 ]. Elevated levels of CEs were reported in EC plasma and tissue samples [ 17 ]. In EC, excessive CE biosynthesis and storage may sustain active signalling pathways, maintain metabolic demand enabling continued disease development and protect cancer cells from toxicity otherwise caused by excess free cholesterol [ 17 ]. Excessive CE biosynthesis may stem from mutation or loss of p53 expression, leading to hepatocellular carcinoma development in murine models [ 45 ]. Significant (FC > 2.0; q  < 0.05) enrichment of total FA concentrations was observed in lipidomes of women with EC, with preference for long-chained PUFAs (20:3 and 20:4). Surveyor (Metabolon Inc.) identified these FAs as dihomo-γ-linolenic acid (20:3n-6) and arachidonic acid (20:4n-6). Dihomo-γ-linolenic acid was previously identified and correlated with tumour size and myometrial invasion in our global metabolomics EC study [ 22 ]. Dihomo-γ-linolenic acid and arachidonic acid belong to the same metabolic pathway and are essential omega (ω)-6 type PUFAs obtained directly from diet or by endogenous synthesis from ω-6 PUFAs precursors. In healthy tissues [ 46 ], and blood [ 47 ], dihomo-γ-linolenic acid concentrations are low, often determined by cellular metabolism rather than dietary intake [ 48 ]. Dihomo-γ-linolenic acid is frequently incorporated into phospholipid reservoirs, and released by phospholipase A 2 , to be metabolised into anti-inflammatory eicosanoids [ 49 ].The association between enriched dihomo-γ-linolenic acid incorporated in complex lipids and EC remains poorly understood. Enriched levels of dihomo-γ-linolenic acid in EC tissue samples were previously hypothesised as compensatory mechanisms to reduce pro-inflammatory effects associated with arachidonic acid-derived series-2 prostaglandins, thromboxanes and series-4 leukotrienes [ 50 ]. Eicosanoids may employ autocrine or paracrine signalling between transformed epithelial- and adjacent cells [ 13 , 19 , 51 ]. Prostaglandin PGF2 α , an arachidonic acid derivative has previously been shown to enhance proliferation, angiogenesis, and migration in EC cells. Its concentrations were also elevated in tissues of other oestrogen sensitive cancers (breast and colorectal cancer) [ 52 ]. This study also examined fatty acid concentrations within different lipid classes in EC, relative to benign and healthy controls. In EC, FA 22:4 concentrations were most enriched within CE. Given the above findings, it may be reasonable that FA 22:4 represents adrenic acid (22:4n-6). Adrenic acid is synthesised from arachidonic acid and metabolised down to eicosanoids by the same enzymatic reactions as dihomo-γ-linolenic and arachidonic acid. Prolonged high dietary fat intake may cause enrichment of PUFAs in membrane lipids and other complex lipids [ 53 ], enhancing membrane fluidity and chronic inflammation [ 54 , 55 ]. Western dietary patterns high in PUFAs influence metabolic conditions which are risk factors for EC: obesity, metabolic syndrome, and hormonal dysregulation. After adjusting for age and BMI between women with EC and healthy women, statistical significance for dihomo-γ-linolenic acid persisted but not for arachidonic acid. Whether enrichment of PUFAs induces onset of EC or merely coincides with disease development is not yet know and warrants further investigation. Furthermore, while BMI was a confounder of several PUFAs enriched in EC, it may not fully account for the observed changes in lipid metabolism. The enrichment of PUFAs and complex lipids in EC may result from the interplay between obesity, dyslipidaemia, and insulin resistance–key components that constitute metabolic syndrome. Insulin resistance, associated with dyslipidaemia, impairs lipid catabolism and promotes lipid storage. Furthermore, the expansion of adipose tissue contributes to dyslipidaemia and elevated oestrogen levels [ 4 , 56 , 57 ]. Given the substantial differences in BMI between groups, some observed lipid alterations may reflect obesity-related metabolic dysregulation rather than EC-specific processes. In this study, highly altered pathways in EC were steroid and bile acid biosynthesis, both of which require cholesterol as a precursor [ 45 ]. Bile acids and steroids may utilise the myometrial oxytocin receptor pathway to influence EC cell growth [ 50 , 58 ]. Dysregulated glycerolipid and sphingolipid metabolism pathways were observed in our global metabolomics EC study [ 22 ], and associated with reprogramming of cellular metabolism. Notably, phospholipid biosynthesis and arachidonic acid metabolism pathways were also perturbated in EC, both of which are associated with cell growth. Using univariable and multivariable biomarker discovery tools we identified several CE species, of which CE(22:4) held most potential as an EC biomarker. The potential of CE(22:4) to accurately predict EC from benign conditions and healthy controls reflects a suitable biomarker candidate in clinical settings. To our knowledge, CE(22:4) was associated with cancer once by Sorvina et al. [ 34 ]. A two-fold increase but lack of statistical significance in CE(22:4) abundance was observed between prostate cancer cell lines and benign controls. Reported cancer-associated CE species include: CEs (13:0), (18:1), (18:2), (20:1), (20:3), (20:4), (22:6), (24:4), (24:5), (28:5) in prostate cancer [ 13 , 34 ], and (17:0), (22:0), (22:3) in colorectal cancer [ 39 , 59 ]. Differences in enriched CE species between cancer types may be attributed to cancer heterogeneity; CE(22:4) may be specific to EC. CEs detected in CVL samples may originate from lipid droplets which have migrated down from the upper female reproductive tract where EC manifests. Enriched CEs may reflect enhanced lipid storage in lipid droplets to sustain metabolic demand required for rapid proliferation of cancer cells. Lipid droplets may be degraded by cancer cells to free the constituent lipids and fatty acids needed for lipid raft, energy production, membrane and hormone biosynthesis [ 14 , 50 , 60 ]. CE(22:4) displayed high sensitivity for other EC subtypes and women with MMR-deficient EC, suggesting it may also serve as a marker of EC severity, reflecting the metabolic burden associated with more profound EC types. Enriched CEs and cholesterol acyltransferase (ACAT)1 were associated with prognostic marker Ki67 in EC tissues. This supports our findings in which CE enrichment may be tied with EC presence and poorer prognostic outcomes [ 17 ]. Avasimibe, a therapeutic agent targeting ACAT1, has shown to decrease CE-mediated proliferation in prostate cancer cells [ 61 ]. Avasimibe is yet to be investigated in the context of EC. The association between cholesterol metabolism and EC pathogenesis is further demonstrated by experiments with lipid-lowering drugs. Targeting serum lipids downstream of cholesterol esterification appeared to be effective in reducing risk of EC development [ 62 , 63 ]. Hyperactive signalling in EC may induce sterol regulatory element binding protein (SREBP)-driven upregulation of genes involved in synthesis and uptake of free cholesterol. Additionally, overactive SREBP may cause upregulated gene expression required for synthesis of FAs. This reveals a mechanism whereby EC cells may reprogramme lipid metabolism, enhancing ACAT1 expression to accommodate elevated levels of free cholesterol and ω-6 PUFAs [ 17 ]. Future research expanding on this study should consider utilisation of lipophilic dyes to visualise lipid droplets in specimens and quantifying activities of lipogenic and lipolytic enzymes to validate and comprehend mechanisms behind EC pathology and the cervicovaginal environment in relation to cholesterol metabolism. Another key diagnostic biomarker which accurately distinguished women with EC, relative to women with benign conditions and healthy women was TAG54:7-FA20:4. TAG54:7-FA:20:4, which was previously enriched in colorectal cancer [ 59 ], and hepatocarcinomas tissues [ 64 ]. In these studies, enrichment of TAG54:7-FA20:4 was attributed to storage in cancer cells, primarily in lipid droplets. Fatty acids released from TAGs can undergo beta oxidation to produce energy required for tumour growth or be metabolised into signalling molecules [ 65 ]. Hence, enrichment of TAGs in lipid droplets may contribute to metabolic hallmarks of cancer: hypoxia, resisting cell death, sustaining proliferative signalling and tumour-promoting inflammation [ 39 , 66 ]. TAGs can also be stored in adipose tissue, where they’re involved in oestradiol production and inhibition of sex hormone-binding globulin (SHBG). Both of which contribute to increased levels of biologically active oestrogen; a risk factor for EC development. Furthermore, metabolic conditions involving dysregulated serum lipids, such as hypertriglyceridaemia, are accompanied by obesity, another EC risk factor [ 67 ]. Herein, numerous TAGs species were associated with other EC subtypes and MMR-deficient EC. A prominent TAGs signature was also observed between women with MMR-deficient EC and MMR-proficient EC. However, there are no studies that would directly explain how TAGs promote the development of specific molecular and histological EC subtypes or how their accumulation reflects metabolic reprogramming in different EC subtypes other than to sustain high metabolic demand of cellular processes. Further exploration of TAGs signatures identified herein would be of clinical significance for application of biomarkers for potential risk stratification of EC in the future. A key strength of using a targeted panel capable of measuring a broad range of lipids is its ability to provide the most accurate representation of the complete set of lipids (classes, fatty acid, and lipid species) within the lipidome. Secondly, the Complex Lipids Targeted Panel measures the concentrations of individual fatty acids (e.g. FA 20:6 and 16:0), rather than the sum composition of several fatty acids in a given lipid (e.g. PC(36:6)). This provided a detailed representation of changes in fatty acid and complex lipid metabolism. Women with benign conditions often present with similar symptoms to EC, such as abnormal bleeding, which contributes to both conditions are treated with hysterectomy. We observed a lower number of differentiating lipid species for the EC vs benign comparison relative to EC vs healthy controls, which may be associated with overlap in clinical symptoms. We were still able to distinguish between EC and benign conditions with high predictive accuracy. Notably, diagnostic accuracy for the EC vs benign comparison was maintained even in the presence of blood in samples. This strengthens the clinical relevance of our findings and highlights potential diagnostic methods which would reduce unnecessary invasive biopsies in the future. A limitation of this study is that complex lipids identified by the panel were pre-selected based on most abundant lipids within human plasma suggesting that the heavy neutral lipid signature (CE, TAGs) observed herein may partially represent the composition of complex lipids involved in lipid transport within the circulation. Regardless, these complex lipids were still detected at high enough concentrations in CVLs to accurately discriminate EC, thereby validating their utility in conditions localised to the upper female reproductive tract. Additionally, incorporating methodology to resolve between structural isomers of FA side chains would improve clarification into EC pathology. Our healthy control group consisted of a relatively small number of women ( n  = 49). Therefore, future studies should utilise a more extensive sample size to better define standard physiological ranges of these complex lipids. Moving forward, the most discriminatory biomarkers and multivariable models should be validated in an independent cohort to confirm their performance is not cohort-specific. This approach would mitigate the risk of overfitting and ensure the reproducibility of our results. Notably, several proposed biomarkers lost significance after adjusting for the confounding effects of age and BMI. This suggests some alterations in lipid metabolism may be influenced by age and BMI rather than EC alone. Future lipidomic studies should consider longitudinal studies, incorporating health questionnaires to gain better insight on the influence of dietary intake on cellular lipid metabolism in EC.

Supplementary Material 1: Additional File 1: Tables S1–S7. Supplementary Material 2: Additional File 2: Figures S1–S6. Supplementary Material 3: Additional File 3: Raw lipidomics data. Supplementary Material 1: Additional File 1: Tables S1–S7. Supplementary Material 2: Additional File 2: Figures S1–S6. Supplementary Material 3: Additional File 3: Raw lipidomics data.