The lipidome of endometrial fluid differs between implantative and non-implantative IVF cycles

Abstract

Objective: To characterize the most relevant changes in the lipidome of endometrial fluid aspirate (EFA) in non-implantative cycles.

Design: Lipidomics in a prospective cohort study.

Settings: Reproductive unit of a university hospital.

Patients: Twenty-nine women undergoing an IVF cycle. Fifteen achieved pregnancy and 14 did not.

Intervention: Endometrial fluid aspiration immediately before performing embryo transfer.

Main outcome measures: Clinical pregnancy rate and lipidomic profiles obtained on an ultra-high performance liquid chromatography coupled to time-of-flight mass spectrometry (UHPLC-ToF-MS)-based analytical platform.

Results: The comparative analysis of the lipidomic patterns of endometrial fluid in implantative and non-implantative IVF cycles revealed eight altered metabolites: seven glycerophospholipids and an omega-6 polyunsaturated fatty acid. Then, women with a non-implantative cycle were accurately classified with a support vector machine algorithm including these eight lipid metabolites. The diagnostic performances of the algorithm showed an area under the receiver operating characteristic curve, sensitivity, specificity, and accuracy of 0.893 ± 0.07, 85.7%, 80.0%, and 82.8%, respectively.

Conclusion: A predictive lipidomic signature linked to the implantative status of the endometrial fluid has been found.

Keywords: Assisted reproduction; Endometrial fluid; Implantation; In vitro fertilization; Lipidomics; Machine learning algorithms; Pregnancy.

Fig. 1

Data analysis workflow. Detailed flowchart of data analysis procedure followed. Data matrix includes 29 samples and 294 metabolites (lipid species). Peak intensity for each metabolite was normalized to the sum of the peak intensities within each sample. Once normalized, outlier analysis was performed and imputation techniques were used for the missing values. Univariate analysis (Student’s t test) was then carried out. Eight metabolites were found to be significantly altered (p < 0.05) in the non-implantative endometrium group, compared to levels in the implantative endometrium. Different modeling approaches were tested taking into account all metabolites and only the metabolites selected after univariate analysis: decision trees, random forests, k-nearest neighbors, linear classifiers, a neural network, and support vector machines. A three-fold cross-validation was performed with a 10 resampling for each modeling. After ROC analysis, the support vector machine model with a linear kernel was selected

Fig. 2

Differential metabolite levels in implantative compared to non-implantative endometrial fluid aspirate (EFA). Volcano plot representation indicating the -log₁₀ (p value) vs. log₂ (fold-change) of each individual metabolic ion features for the comparison implantative vs. non-implantative endometrium. BA, bile acid; Cer, ceramide; CMH, monohexosylceramide; Cho, cholesterol; ChoE, cholesteryl ester; DG, diglyceride; LPC, lysophosphatidylcholine; LPE, lysophosphatidylethanolamine; LPI, lysophosphatidylinositol; MUFA, monounsaturated fatty acid; NAE, N-acyl ethanolamine; oxFA, oxidized fatty acid; PC, phosphatidylcholine; PE, phosphatidylethanolamine; PI, phosphatidylinositol; PUFA, polyunsaturated fatty acid; SFA, saturated fatty acid; SM, sphingomyelin; ST, steroid; TG, triglyceride

Fig. 3

Summary of models run without data pre-processing for the prediction of the endometrial implantative status. Boxplots show the distribution of a AUC, b sensitivity, and c specificity in the cross-validation analysis for the following methodologies: decision trees (treebag, rpart, PART, and J48), linear classifiers (Naïve-Bayes, naïve_bayes; generalized linear models, glm; and flexible discriminant analysis, mda), k-nearest neighbors, kknn; a neural network, nnet; random forests, rf; and a support vector machine with several kernels: linear, svmLinear; radial, svmradial; and polynomial, svmPoly. All these methodologies were tested including all variables detected (a total of 294 lipids) and only the eight significant metabolites selected under the t test criterion (p < 0.05). Models are ordered from the best to worst performance based on AUC. Missing boxplots indicate a failure in the model applied

Fig. 4

ROC analysis of the support vector machine models. a ROC curve by kernel. Analysis performed for the optimal tuning parameters: linear kernel: C = 0.28; radial kernel: C = 2.74 and sigma = 0.017; polynomial kernel: C = 1.09, scale = 0.024, and degree = 1. b Accuracy by kernel. A cutoff at 0.5 is considered. Models were calculated with data without pre-processing