Jiao Wu, Chunmei Liang, Xin Wang, Yida Huang, Wanshan Liu, Ruimin Wang, Jing Cao, Xun Su, Tao Yin, Xiaolei Wang, Zhikang Zhang, Lingchao Shen, Danyang Li, Weiwei Zou, Ji Wu, Lihua Qiu, Wen Di, Yunxia Cao,  Dongmei Ji, Kun Qian

Adv. Sci. 2023, 10, 2302023

Abstract

Ovarian reserve (OR) and fertility are critical in women’s healthcare. Clinical methods for encoding OR and fertility rely on the combination of tests, which cannot serve as a multi-functional platform with limited information from specific biofluids. Herein, metabolic fingerprinting of follicular fluid (MFFF) from follicles is performed, using particle-assisted laser desorption/ionization mass spectrometry (PALDI-MS) to encode OR and fertility. PALDI-MS allows efficient MFFF, showing fast speed (≈30 s), high sensitivity (≈60 fmol), and desirable reproducibility (coefficients of variation <15%). Further, machine learning of MFFF is applied to diagnose diminished OR (area under the curve of 0.929) and identify high-quality oocytes/embryos (p < 0.05) by a single PALDI-MS test. Meanwhile, metabolic biomarkers from MFFF are identified, which also determine oocyte/embryo quality (p < 0.05) from the sampling follicles toward fertility prediction in clinics. This approach offers a powerful platform in women’s healthcare, not limited to OR and fertility.

Introduction and Methods

Ovarian reserve (OR) and fertility reflect reproductive potential, which is critical in women’s healthcare.[1] Precise encoding of OR and fertility is vital in determining the appropriate treatments, affecting 10–32% of women at reproductive age, based on the estimates from the National Society for Assisted Reproductive Technology (SART) system in the United States. Current analytical methods in clinics rely on a combination of tests including biochemical analysis and ultrasound imaging. These methods are based on the selected protein biomarkers (such as follicle-stimulating hormone [FSH] and antimüllerian hormone [AMH]) or physical measurements (antral follicular count [AFC]), which cannot serve as a multi-functional platform with limited information from specific biofluids

like follicular fluid (FF). Accordingly, a single test that offers comprehensive metabolic information would be desirable in encoding OR and fertility, to engage various clinical applications toward women’s healthcare.

Key Results and Conclusions

For the future perspective of this work, the performance of metabolic biomarkers would be further enhanced by incorporating additional biomarkers into a multi-modal database. Moreover, biological validations could be conducted to understand the underlying mechanism and could lead to the development of new therapeutic strategies to improve fertility outcomes.

In summary, we performed efficient MFFF using PALDI-MS and constructed a biomarker panel for encoding OR and fertility. We achieved efficient MFFF with fast speed (≈30 s), high sensitivity (≈60 fmol), and desirable reproducibility (CVs < 15%) by PALDI-MS. Further, we applied machine learning of MFFF to diagnose dOR with an AUC of 0.929 and identify high-quality oocytes/embryos (p < 0.05) by a single PALDI-MS test. Subsequently, we constructed a biomarker panel showing an AUC of 0.849 for dOR diagnosis and effective determination of highquality oocytes/embryos (p < 0.05). Our work would provide a powerful tool for women’s healthcare including but not limited to OR and fertility.

Fig.1：Characterization of the PALDI-MS.

a) Digital images of the microarray chip with 384 sample spots and particle solution dispersed by water in a 600 μL tube for the insert image. The distance between the centers of two spots on the chip was 4.5 mm. b) Scanning electron microscopy (SEM) image and transmission electron microscopy (TEM) image of particles (inset) showing the surface roughness of the particles. The scale bars for SEM and TEM images were 100 nm. c) The comparison for sample pretreatment of dilution and deproteinization, displaying more metabolites detected by dilution treatment. *** represented p < 0.001. d) The intensities of metabolites detected using particles, CHCA and DHB as the matrix, and typical mass spectra (insert) of 1.12 nmol Ala. The standard deviation (s.d.) of three tests was obtained as an error bar. e) The TIC of mass spectra for detecting standard mixture (including 10 ng mL−1 Glc, Suc, Ala, and Arg) using particles, CHCA, and DHB as the matrix. The error bars represented the s.d. of three replicates. *** represented p < 0.001. f) Molecular trapping ratios for Glc and BSA using particles as the matrix through elemental mapping results showing the selective trapping of particles for metabolites. The error bars represented ±s.d. *** represented p < 0.001. g) Cosine similarity scores of mass spectra from raw FF and its dilutions with dilution folds of 2–100 using 10–500 nL of raw FF samples. The error bars showed the s.d. of three samples. h) CVs of intensities for a standard mixture containing Lys, Glc, and Suc in 15 tests, showing the high reproducibility of PALDI-MS using 9particle as the matrix. i) The CVs of intensities for FF samples from nOR and dOR subjects with five independent tests. The error bars represented ±s.d.

Fig. 2：Characterization of OR and fertility using MFFF. 

a) Flowchart for the enrollment of subjects, highlighting the strict selection of 344 subjects out of 520 for inclusion in the study. Notably, the MFFF was performed on all 344 FF samples collected. b) Representative mass spectra of FF from dOR and nOR samples, showing TIC of ≈8.34–8.53 × 107 at m/z of 100–400 Da. c) Frequency distribution of intragroup similarity scores of typical spectra from FF of the dOR group (50 samples) and nOR group (50 samples), indicating high levels of intra-group similarity scores with 95% of FF samples having scores >0.85. d) Heatmap visualization of MFFF including 304 features from 141 dOR and 203 nOR samples. The FF samples were diluted at tenfolds using water. The color bars represented intensities corrected by logarithms. e) Unsupervised clustering by PCA revealing rough separation between dOR (cyan dots) and nOR (purple dots) groups.

Fig.3：Encoding of OR and fertility by machine learning of MFFF.

a) A study design based on machine learning to diagnose dOR. There were 275 FF samples (111/164, dOR/nOR) included in the discovery cohort for cross-validation. An independent validation cohort was used to assess the optimized model (30/39, dOR/nOR). b) The model performance of RR, NN, SVM, and RF for dOR diagnosis in terms of AUC, Sen, Pre, Acc, and F1. A min-max normalization was used to normalize these parameters. c) The ROC curves of dOR diagnosis in the discovery cohort (111/164, dOR/nOR, blue line) and the independent validation cohort (30/39, dOR/nOR, orange line). d) The sample-level plot depicting the probability of RR in the discovery cohort for differentiating dOR (111, purple dots) from nOR (164, cyan dots). e) Fertility information (the number of subjects with HQO and HQE) of the enrolled ubjects (141/203, dOR/nOR). f) The probability of the HQO group and no oocyte group, showing a significant difference in predicting oocyte quality (p < 0.05). The p-value was calculated by a two-tailed t-test. * represented p < 0.05. g) The probability of HQE group and no embryo group, displaying a significant difference in predicting embryo quality (p < 0.05). The p-value was calculated by a two-tailed t-test. * represented p < 0.05.

Fig.4：Biomarker panel construction and related pathway analysis.

a) Feature number (orange line) and AUC of RR (red line) calculated at various thresholds of RR rank score in the discovery cohort. Blue dotted lines indicated the optimized AUC of 0.919 at the threshold of 0.25. b) Intensity heatmap of 7 m/z features showing distinct expression levels between dOR and nOR. The color bars represented intensities corrected by logarithms. c) The violin plot displaying the intensity levels of identified biomarkers for dOR and nOR. A two-tailed t-test was used to determine the p values (* represented p < 0.05, ** represented p < 0.005, *** represented p < 0.001). The biomarker intensities were normalized. d) The ROC curves for the biomarker panel with an AUC of 0.849, superior to the single biomarker with an AUC of 0.571–0.803 (p < 0.05 by Delong test). e) The comparison of probability for the HQO group and no oocyte group in predicting oocyte quality using constructed biomarker panel (* represented p < 0.05). A two-tailed t-test was conducted to calculate the p values. f) The comparison of probability for the HQE group and no embryo group in predicting embryo quality using constructed biomarker panel (* represented p < 0.05). A two-tailed t-test was conducted to calculate the p values. g) The potential pathways associated with dOR and nOR. The size and color of the circles represented the p-value and pathway impact, respectively.

https://doi.org/10.1002/advs.202302023