A total of 300 participants (150 non-PCOS and 150 PCOS) who underwent the conventional in vitro fertilization (IVF) in Shanghai First Maternity and Infant Hospital between October 2017 and December 2019 were recruited to obtain informed consent. Eligibility inclusion criteria were: 1) 25–35 years of age; 2) diagnosis of PCOS, according to the revised Rotterdam European Society of Human Reproduction and Embryology/American Society for Reproductive Medicine Criteria (Rotterdam, 2004), where at least two of the followings were fulfilled: chronic oligo-ovulation or anovulation, androgen excess, and polycystic ovaries. The control group had regular menstrual cycles, normal ovary sonographs and normal ovulating; 3) no history of drugs affecting glucose and lipid metabolism, and without any other diseases that affect endocrine hormones, including congenital adrenal hyperplasia, Cushing’s syndrome, endometriosis and androgen-secreting tumors; 4) undergoing the first IVF treatment with GnRH antagonist protocol; 5) males with normal sperm according to the WHO criteria to diminish the chances of failure on embryo development due to defects in the spermatozoa; 6) one blastocyst transferred in fresh cycle or freezing-thawing cycle afterwards, in order to track the pregnancy outcome. Namely, if the oocyte was abnormally fertilized or did not develop into a transferrable blastocyst, its corresponding FF was discarded and the patients was not included in this study. The clinical characteristics of the PCOS and non-PCOS controls are summarized in Table 1.

All patients who participated in the current study underwent the same long GnRH antagonist stimulation protocol. The details of the stimulation cycle procedure were previously described (Onofriescu et al., 2013). When two or more follicles were at least 18 mm in diameter and the serum E₂ levels were at least 300 pg/ml per dominant follicle, the patients received 250 μg hCG (Profasi; Serono). Follicular fluid (2–3 ml) was collected from one of the dominant follicles (>18 mm) by vaginal puncture under ultrasound echo-guidance 36 h after hCG administration. Then, the corresponding oocyte isolated from the collected FF was rinsed and inseminated individually to culture and evaluate the development potential. The collected FF samples were centrifuged at 1,000 g for 10 min at 4°C to remove the cells and debris. The supernatant was stored at −80°C for Raman spectra test.

Approximately 2–3 h after the oocyte retrieval, insemination was performed with Percoll-prepared spermatozoa. One hundred thousand sperms per milliliter were added to each oocyte. The fertilized oocytes were cultured in sequential media of SAGE (CooperSurgical, Leisegang Medical, Berlin) in a single droplet culture (25 μL) covered by mineral oil and cultured at 37°C with 6% CO2 to the blastocyst stage on Day 5–6. The morphological characteristics of the oocytes and embryos were individually recorded. On Day 1 (at 16–18 h after insemination), zygote exhibiting two pronuclei (2 PN) was regarded as normal fertilization and continued to be cultured to the blastocyst stage (Day 5/6), while the zygote exhibiting abnormal fertilization (0 PN or 3 PN) was discarded. And, the blastocysts were scored according to the Gardner scoring system (Gardner et al., 2000). Briefly, blastocysts were evaluated mainly through three indicators: the expansion of the blastocoel cavity, the cohesiveness of the inner cell mass, and trophectodermal cells. First, according to the expansion of the blastocoel cavity, the blastocyst was given a numerical score from 1 to 6. From 1 to 6 points, the expansion of the blastocyst cavity was indicated as follows: 1, the blastocoel was less than half of the volume of the embryo; 2, the blastocoel was more than half of the volume of the embryo; 3, the blastocoel completely filled the embryo; 4, the blastocyst was expanded larger than the early embryo, with a thin zona; 5, a hatching blastocyst with the trophectoderm started to herniate though the zona; 6, a hatched blastocyst which had completely escaped from the zona. Second, the inner cell mass and trophectoderm were assessed for blastocysts with scores at least 3. The scoring criteria for inner cell groups were as follows: A (the highest score), many cells were tightly packed; B, several cells were loosely grouped; C, very few cells existed. Lastly, the trophectoderm was also scored as A (the highest score), B, or C as follows: A, a cohesive epithelium with many cells; B, a loose epithelium with few cells; C, very few large cells. In this study, blastocysts with scores ≥ 3BC (blastocoel cavity score ≥3, and inner cell score ≥ B, and trophectoderm score ≥ C) are regarded as transferrable blastocysts. And, the blastocysts with scores below 3BC were discarded. In order to analyze the relationship between the metabolic components of follicle fluid and the development of the blastocyst, the transferrable blastocysts (score ≥3BC) were divided into two subgroups: high-quality (score ≥4BB) blastocysts and low-quality (score ≥3BC and ≤4BC). The transferrable blastocysts were transferred in fresh cycles or freezing-thawing cycles. Vaginal progesterone was given for 2 weeks after the oocyte retrieval. The presence of an intrauterine gestational sac is defined as clinical pregnancy.

The Renishaw inVia Raman spectrometer (Witec alpha300 R, German) connected to a Zeiss microscope (Axioscope, German) and equipped with a 532 nm laser was used for the spectra collection. The 50x dry lens objective (Zeiss EC Epiplan, German) focused the laser beam with a power of 5 mW. The size of the laser spot was 100 μm. The spectral range was from 50 to 2000 cm⁻¹ with a spectral resolution at 1 cm⁻¹. The accumulative times of a single point are 8 times (8 s per time). Before each scanning session, silicon wafer was used for calibration (adjusted to 520.5 ± 0.3 cm⁻¹ for silicon peak). Briefly, the FF samples from −80°C storage were thawed at room temperature (25°C) for 30 min before analysis. A total of 7 μL of the FF sample was dropped onto a clean place on the surface of the disposable aluminum slide without touching the substrate. For each drop of FF, 5 individual spectra were collected at random spots. The background control spectrum was subtracted from each sample spectrum. The process of Raman analysis took about 10 min per sample.

In this study, spectral analysis relied on machine-learning data sets and bioinformatics methods. The full-spectrum pattern recognition was drawn and the characteristic peaks of Raman spectrum which contains abundant material structure information were identified. The raw Raman spectra were preprocessed by subtracting the dark signal background from each spectrum. Spectral normalization was finished by vector normalization of the fingerprint region from 600 to 1,800 cm⁻¹ using the Labspec 6 software (Horiba). Mean-centered Raman spectra were obtained by subtracting the means of the spectra from each sample spectrum. Data analysis, statistics and visualization were done in the R environment with the use of in-house scripts. The Raman data was analyzed by the unsupervised PCA method, which could extract key features, reduce the high dimensionality of the data and determine principal components (PCs). The specific wavenumber of Raman shift derived by PCA analysis were corresponding to the molecular bonds or metabolites. Quantification of metabolite concentration was done by integrating individual Raman bands and presented as box plots.

Besides feature extraction to find biological information for PCOS, screening oocyte with high developmental potential and more promising clinical pregnancy rates are often desirable for IVF purposes. In this study, referring to the methods described in the previous report (Wang et al., 2020), the fully connected artificial neural network (ANN) was established to classify oocyte development potential and IVF outcome based on original Raman spectra of follicular fluid.

The schematic architecture of the fully connected ANN is illustrated in Figure 1. Because the Raman spectral data is not particularly complex, taking into account the speed and accuracy of the analysis, two full-connection layers were selected for constructing the ANN models in this study. The first layer was the input layer and the input data was one-dimensional intensity data after spectral standardization. The second layer was a fully connected layer of 64 neurons and the third layer was similar as the second layer. An activation function “relu” was associated with each neuron that summed up the outputs from that neuron and transferred them to the next layer in the network, while function “dropout” was used to prevent overfitting. The last layer was the output layer, and the activation function was “softmax.” During the model training process, the loss function was the “crossentropy” and the optimizer was “adam.” The code for building the ANN model was detailed in the Supplementary file. In addition, a very important pre-processing of the raw spectral data was done by using haar wavelets and db4 wavelets to reduce data noise in the process of building the ANN models.

The relationship between input (i.e., Raman spectral pattern of follicular fluid) and output (i.e., the high-quality versus low-quality blastocyst, or pregnancy success versus pregnancy failure) could be learned from the recorded data (training data set) and used to predict the unknown sample (prediction data set). The error threshold was set as 1%. The prediction ability of fully connected ANN analysis was evaluated by “leave-one-out” cross-validation. Briefly, all the data points except one were employed to train the model and a prediction was made for that one at each time. The information of sample ID, replicate number, morphological score and the pregnancy outcome were added as features to train the fully connected ANN models. All models were constructed in a Python 3.3 environment.

For each classification model, the precision, sensitivity, F1 score as well as an overall accuracy rate were calculated. Specifically, precision is the ability of a classifier to not label a negative sample as positive. Sensitivity is the ability of a classifier to find all of the positive samples in one class. F1 score is a weighted mean of the precision and the sensitivity, where 1 is the best value and 0 the worst. And, the performance of the established model was also evaluated by the area under the receiver operating characteristic (ROC) curve (AUC) based on a five-fold cross-validation strategy, which divided the data into five parts, training four parts in turn, and the remaining was used to estimate the performance of the models.

In this study, Raman spectra of 150 PCOS and 150 non-PCOS patients were obtained by analyzing follicular fluid samples. The fingerprint region from 600 cm⁻¹ to 1800 cm⁻¹ was calculated for analysis, which typically contained the most essential biological information (Li et al., 2012). In accordance with the previous study (Zhang et al., 2021), the raw Raman spectra data of the FF samples from PCOS patients showed no significantly different standard deviation/variation by compared to those of non-PCOS patients (Figure 2A). Then, Raman spectra from all samples were averaged and subtracted from each sample at each wavenumber, and the mean values were assembled for each group to obtain mean-centered spectra. The results show that the mean-centered spectra are significantly different between the two groups (Figure 2B). The corresponding Raman bands and their tentative assignments are summarized in Table 2. Some signatured bands can be assigned to phenylalanine (1,003 cm⁻¹), C-C, C-N stretching (protein) (1,156 cm⁻¹), β -carotene accumulation (C-C stretch mode) (1,516 cm⁻¹), and C=O (1,668 cm⁻¹) (Tarcea et al., 2007).

All of the spectra were analyzed by unsupervised PCA to reduce the high-dimensional Raman dataset and transform data into appropriate variables that conferred biologic information. A spectral range of 600–1,800 cm⁻¹ was used to minimize the effects of uneven baseline. As shown in Figure 2C, PCOS and non-PCOS have two clusters with some overlap. The overlap may be due to the close similarity of follicular fluid between PCOS and non-PCOS patients. In order to find biomarkers which can best separate the two groups and provide diagnostic advice for PCOS, a loading plot, which is along dimension 1, was made to assign each Raman wavenumber to indicate significant Raman bands. Then, by scanning all available zones based on the bands and calculating the significance of each zone on discriminating PCOS vs. non-PCOS, two Raman zones (993–1,165 cm⁻¹ and 1,439–1,678 cm⁻¹) were identified as the ones with the strongest significance (Figure 2D). The quantification of metabolite concentration between PCOS and non-PCOS groups were also presented at box plots by integrating individual Raman bands at the two Raman zones (993–1,165 cm⁻¹ and 1,439–1,678 cm⁻¹). The results showed that the quantification of integration at 993–1,165 cm⁻¹ was found to be higher in PCOS FF samples (0.0286 ± 0.001) compared with non-PCOS FF samples (0.0274 ± 0.001; p < 0.001). And, the quantification of the integration at 1,439–1,678 cm⁻¹, on the other hand, showed a higher content in non-PCOS FF samples (0.0337 ± 0.001) than in PCOS FF samples (0.0319 ± 0.001; p < 0.001) (Figure 3). All the results suggest that there are real different metabolic patterns in follicular fluid of PCOS patients.

As the follicular fluid is the direct environment of oocyte, the changes of metabolism in FF will affect oocyte development. In order to investigate whether the Raman spectra were related to the oocyte development, the special Raman spectra were further compared between two subgroups of the transferrable PCOS blastocysts (n = 150), according to their morphological scores: 1) high-quality blastocysts group (HQ, ≥ 4BB) (n = 75), which included 58 blastocysts formed on Day 5 and 17 ones formed on Day 6, and 2) low-quality blastocysts group (LQ, ≥3BC and ≤4BC) (n = 75), which were all formed on Day 6. The raw Raman spectra data and the mean-centered spectra of the FF samples from the two subgroups (HQ group or LQ group) of PCOS were shown in Figure 4 A and B. It indicated that the mean-centered spectra are significantly different between the two subgroups, while the raw Raman spectra data had no significant difference. Unsupervised PCA was also applied to find the significant Raman bands between the two subgroups (Figure 4C) and a loading plot was made along dimension 1 to assign each Raman wavenumber to indicate significant Raman bands. Two Raman zones (993–1,165 cm⁻¹ and 1,439–1,678 cm⁻¹), same as in the above section (Figure 2D), were also identified as the most significant differences between HQ and LQ subgroups (Figure 4D). As shown in Figure 5, the quantification of the integration at 993–1,165 cm⁻¹ was lower in HQ group (0.0316 ± 0.001) compared with LQ group (0.0328 ± 0.001; p = 1.07 × 10^–5). On the other hand, the quantification of the integration at 1,439–1,678 cm⁻¹ showed a higher content in HQ group (0.0292 ± 0.001) than LQ group (0.0279 ± 0.001; p = 1.92 × 10^–6).

Embryo quality is a key factor affecting pregnancy outcome. In this study, we tracked the pregnancy outcome of each transferred blastocyst. As shown in Figure 6, there are a total of 150 transferrable blastocysts from PCOS patients, including 75 high-quality blastocysts (≥4BB) and 75 low-quality blastocysts (≥3BC and ≤4BC). Of the 75 high-quality blastocysts, 62 successfully conceived, and the clinical pregnancy rate was about 83%. This indicated that even if a high-quality blastocyst was transplanted, 17% of patients failed to conceive. On the other hand, of the 75 low-quality blastocysts, 23 successfully conceived, and the clinical pregnancy rate was about 31%. In order to investigate the relationship of the special Raman spectra and IVF outcome in PCOS, the 150 Raman spectra of PCOS samples were divided into pregnancy success group (n = 85) and pregnancy failure group (n = 65). The raw Raman spectra data and the mean-centered spectra of the FF samples from the pregnancy success group and pregnancy failure group are roughly the same as those of high-quality blastocysts group and low-quality blastocysts group (data is not repeated). In order to further determine the Raman spectral difference between the two groups (pregnancy success group and pregnancy failure group), the quantified spectral intensity values were also compared. It indicated that the quantification of Raman spectra (993–1,165 cm⁻¹ and 1,439–1,678 cm⁻¹) showed the same significantly different between pregnancy success and pregnancy failure groups (p < 0.001) (Figure 7). In line with the trend of the same two Raman zones correlating with blastocyst development, the quantification of the integration at 993–1,165 cm⁻¹ was lower in pregnancy success group (0.0316 ± 0.001) compared with pregnancy failure group (0.0335 ± 0.002; p = 1.36 × 10^–5), and the quantification of the integration at 1,439–1,678 cm⁻¹ was higher in pregnancy success group (0.0289 ± 0.001) than that in pregnancy failure group (0.0278 ± 0.001; p = 1.98 × 10^–6).

In this study, based on the 150 Raman spectra of PCOS follicular fluid samples, the two fully connected ANN classification models were computed to predict the blastocyst development and clinical pregnancy, respectively. The spectra of different subgroups (HQ or LQ blastocyst; pregnancy success or pregnancy failure) were randomly split into training set and the testing set in a ratio of 4:1. The training set was used to train a classification model and the testing set was used to evaluate the model performance. Specifically, for the classification model to predict blastocyst development, Raman spectra of 150 PCOS samples were divided into high-quality blastocysts (HQ) (n = 75) and low-quality blastocysts (LQ) (n = 75) groups. Of them, Raman spectra of 100 samples (50 high-quality blastocysts and 50 low-quality blastocysts) were used to train a classification model and the remaining 50 spectra (25 high-quality blastocysts and 25 low-quality blastocysts) were input as the testing set to evaluate the model performance. At the same time, to construct the classification model to predict the IVF outcome, the same Raman spectra of 150 PCOS samples were divided into pregnancy success (n = 85) and pregnancy failure (n = 65) groups. Of them, 100 spectra (50 pregnancy success and 50 pregnancy failure) were used to train a classification model and the remaining 50 spectra (35 pregnancy success and 15 pregnancy failure) were input as the testing set to evaluate the model performance.

The results of ANN models were presented in Table 3. As shown in Table 3A, 23 out of 25 high-quality blastocysts (F1 score 0.9020) and 22 out of 25 low-quality blastocysts (F1 score 0.8980) were assigned correctly. For classifying clinical pregnancy outcome (Table 3B), the ANN model was able to correctly assign 25 out of 35 pregnancy success spectra and 12 out 15 pregnancy failure spectra, with F1 scores of 0.7937 and 0.6486, respectively. Noticeably, the accuracy of the ANN model for predicting the blastocysts development is 90.00%, while 74.00% for predicting the IVF outcome. Furthermore, ROC analysis resulted in AUCs of 0.89 ± 0.039, sensitivity of 0.692 ± 0.037 and specificity of 0.948 ± 0.031 for the blastocyst developmental ANN model (Figure 8A) and ROC analysis resulted in AUCs of 0.72 ± 0.023, sensitivity of 0.738 ± 0.049 and specificity of 0.63 ± 0.043 for the ANN model predicting the IVF outcome (Figure 8B).