Introduction

Polycystic ovarian syndrome (PCOS) is a prevalent endocrine-metabolic disorder that impacts a large population globally, the “Rotterdam criteria”, defined in 20031, determined the diagnosis of PCOS based on oligomenorrhea or amenorrhea, hyperandrogenism, ovarian polycystic changes, and infertility. Although numerous studies have focused on PCOS, the precise origin and etiology remain incompletely comprehended.

The prime organ affected in PCOS is the ovary, in which macroautophagy/autophagy performs a pivotal role in directing the chain of events starting from oocytes origin until its fertilization. Defective autophagy in the follicular cells during different stages of follicles is observed in the PCOS ovary2. Lately, there has been considerable attention on the notion of selective autophagy prompted by the autophagy substrate acting as a catalyst. Mitophagy, being one of the extensive-studied forms of selective autophagy, plays a critical role in maintaining the function and genetic stability of mitochondria3. It is believed to be an important component in the onset of PCOS.

Moreover, the recent wide acceptance of functional mitochondrial disorders as a correlated factor of numerous diseases has led to the presupposition that abnormal mitochondrial metabolic markers are associated with PCOS4. Mitophagy serves as a cytoprotective mechanism to eliminate excess or malfunctioning mitochondria, ensuring a proper balance of mitochondrial numbers for intracellular stability. The PINK1-PRKN/Parkin pathway, acknowledged as the primary regulator of mitophagy5, involves the labeling of impaired mitochondria with ubiquitin chains, initiating their selective autophagy. Accumulation of PINK1 in damaged mitochondria leads to the recruitment of parkin, resulting in ubiquitination of mitochondrial proteins. These can then be bound by the autophagic proteins p62/SQSTM1 and LC3, leading to the degradation of mitochondria through mitophagy. MAP1LC3A, our subject, is precisely one of the members of the human autophagy-related LC3/GABARAP-protein family. In addition, there has been direct evidence that indicate the relationship between mitophagy and PCOS progression. Yi et al. observed that mitophagy is significantly enhanced in dihydrotestosterone (DHT) -induced PCOS-like mice, and melatonin treatment can significantly decrease the levels of PINK1/Parkin, thus improving mitochondrial dysfunction and PCOS phenotype both in vitro and in vivo6.

There are multiple perspectives on the screening of PCOS biomarkers, including multi-omics, DNA methylome, senescence-related genes and also autophagy-associated mRNA-miRNA-LncRNA network7,8,9,10, but the specific genes associated with mitophagy (mitophagy-related genes, MRGs) that are linked with PCOS have yet to be explored and clearly identified. In our study, to discern the candidate mitophagy-related biomarkers for PCOS, functional enrichment analysis, application of machine-learning algorithms (least absolute shrinkage and selection operator [LASSO], random forest [RF], and support vector machine-recursive feature elimination [SVM-RFE]), evaluation of receiver operating characteristic (ROC) curve, and analysis of immune cell infiltration were successively performed11. We came to the conclusion that MAP1LC3A may serve as a promising mitophagy-associated biomarker in PCOS.

Materials and methods

Data sources and processing

The study flow chart is presented in Fig. 1. We obtained three trustworthy transcriptomic datasets related to PCOS from the Gene Expression Omnibus (GEO) database12, namely GSE9572813, GSE16840414, and GSE15548915. Detailed information about the three datasets were listed in Table 1. Totally, 32 granulosa cell (GC) samples (16 PCOS, 16 Controls) and 12 oocyte samples (6 PCOS, 6 Controls) were taken into consideration in our research to assess the MRGs' expression levels. Gene symbols were matched to the array probes based on the corresponding annotation data. The normalized gene expression matrices for GSE95728 and GSE168404 were directly downloaded, whereas the total read count per sample in GSE155489 were required to be normalized to a common library size using the DESeqDataSetFromMatrix function from the R package “DESeq2”. Then we merged the training datasets together (GSE95728-GC and GSE155489-GC). The ComBat function from the R package “sva”16 was utilized to eliminate sequencing batch variation, and the impact of inter-sample correction was visualized by a two- dimensional principal component analysis (PCA) cluster plot.

Figure 1
figure 1

Study flow chart.

Full size image
Table 1 Detailed information about the collected datasets.
Full size table

Screening for differential expressed mitophagy-related genes

29 MRGs were taken out from the reactome pathway database (Table 2). The “limma” package17 was utilized to identify the differentially expressed mitophagy-related genes (DE-MRGs) between PCOS patients and controls. We also utilized the Benjamini-Hochberg-based False Discovery Rate method to modify P values and identified the noteworthy DE-MRGs with an adjusted P-value < 0.1. Subsequently, we employed the “ggplot2” package to depict a volcano plot and the “pheatmap” package to generate a heatmap.

Table 2 29 mitophagy-related genes extracted from reactome pathway database.
Full size table

Annotating the functional aspects of DE-MRGs

The gene ontology (GO) and kyoto encyclopedia of genes and genomes (KEGG) enrichment analyses were conducted for functional annotation of DE-MRGs using the “clusterProfiler” R package18. The GO analysis identified three categories, namely biological process (BP), cellular component (CC), and molecular function (MF)19. The investigation of potential biological pathways was carried out utilizing KEGG20. Significant enrichment was defined as P-value < 0.05.

Identification of optimal hub genes for PCOS

To filter feature genes, three machine learning algorithms were performed: LASSO, RF, and SVM-RFE. LASSO regression, serving as a technique for reducing dimensionality, outperforms regression analysis when dealing with high-dimensional data and employs regularization to enhance prediction accuracy. Using the “glmnet” R package, a tenfold cross-verification method with a turning or penalty parameter was conducted21. The RF algorithm employed the “Random Forest” package to calculate the error rate and accuracy rate of the combination in each iteration. Additionally, it used the RFE approach to determine the importance and important ranking of each gene. The related genes in the ideal combination with the lowest error rate were the characteristic genes. Meanwhile, the SVM-RFE model was evaluated by calculating the average misclassification rates from their tenfold cross-validations using the R package “e107”22. SVM-RFE, a novel technique in machine learning, can avoid overfitting by recursively ranking features23. The final feature importance was determined by using the average importance of each feature in every iteration. The genes that fell between the intersections of the three subsets were then chosen as hub genes for further examinations.

Next, we established the ROC curve to assess the discriminatory capacity of the hub genes in the merged training set (GSE95728-GC + GSE155489-GC); the diagnostic performance of each model was measured by the area under the ROC curve (AUROC). The accuracy of these gene predictions would be separately confirmed in the independent validation datasets (GSE168404 [GC, Control vs. PCOS 5: 5] and GSE155489 [Oocyte, Control vs. PCOS 6: 6]). The ROC analyses were completed by the R package “pROC”24.

Assessment of immune cell infiltration

Based on the principle linear support vector regression, we used CIBERSORT algorithm to estimate the proportion of 22 different types of immune cell infiltration of each oocyte samples in GSE15548925. The LM22 matrix file contains 547 genes, which serve as the standard reference leukocyte gene signature. These genes are capable of accurately distinguishing 22 mature human hematological populations that have been isolated from peripheral blood or in vitro cultures26. All estimations of immune cell type fractions for each sample added up to 1. To visualize the disparities in immune cell infiltration between the PCOS and control samples, PCA clustering and boxplots were utilized. Correlation analysis and visualization of infiltrating immune cells were performed using the “corrplot” package27.

Clinical correlation between hub genes and testosterone levels in PCOS patients

According to the expression form of testosterone levels in GSE168404 samples as mean ± standard deviation, we determined that this variable type was in line with normal distribution and homogeneity of variance. Based on the provided values, SPSS software (Rv. Normal function) was applied to generate random numbers that match the mean and standard deviation of testosterone data of individual samples in GSE168404. We then assessed the correlation between hub genes and testosterone levels in patients with PCOS and drew a scatter plot using the “ggplot2” package.

Statistical analysis

R (version 4.3.1) was utilized for all statistical analyses. Group comparisons were undertaken for continuous variables using Student's t-test for normally distributed variables or the Mann–Whitney U-test for variables with an abnormal distribution. Pearson correlation analyses were applied for the necessary tasks. All statistical analyses were two-sided with P-value < 0.05 were regarded statistically significant.

Results

Panoramic view of MRGs in PCOS

Currently, the role of 29 MRGs has been extensively studied and their interaction was shown in (Fig. 2A). The density distribution of the groups shown in (Fig. 2B) was basically consistent, indicating that the normalized gene expression matrix of each dataset could be applied for subsequent analysis. Next, we merged the GSE95728-GC and GSE155489-GC datasets as a training set (GSE95728-GC + GSE155489-GC) and removed the sequencing batch effects. The PCA cluster plots in (Fig. 2C) showed that the clustering of the two datasets was more obvious after batch removal, indicating that the source of the samples was reliable. Figure 2D showed that 26 MRGs were figured out in the GC samples, of which 12 genes exhibited a marked expression difference between PCOS and the matched controls (P < 0.05), including TOMM5 and MTERF3 (P < 0.0001).

Figure 2
figure 2

Landscape of mitophagy-related genes in PCOS. (A) Protein-protein interaction (PPI) network made up of 29 mitophagy-related genes. (B) The density distribution plots of the GSE95728, GSE168404 and GSE155489 datasets in granulosa cell or oocyte samples. (C) The PCA plots of the GSE95728 and GSE155489 datasets before and after sample correction in granulosa cell samples. (D) Box plots showing the expression levels of 26 mitophagy-related genes in granulosa cell samples of PCOS and the matched control. Differences between groups are represented by “*”. *P < 0.05; **P < 0.01. GC, granulosa cell. Data were analyzed by wilcoxon tests.

Full size image

Volcano in Fig. 3A depicted these 12 DE-MRGs, including 6 up-regulated genes and 6 down-regulated. TOMM5 exhibited the greatest fold-change among these downregulated genes, while MAP1LC3A had the greatest fold-change among those upregulated. The heatmap in Fig. 3B showed the expression of DE-MRGs among GC samples. In the correlation analysis (Fig. 3C), we found that these genes were closely related, indicating that they may work together. The scatterplots of Fig. 3C displayed the genes with the highest positive and negative correlation, specifically, CSNK2A2 and VDAC1 turned out to be the most negative correlation, whereas MTERF3 were most positively correlated with TOMM5.

Figure 3
figure 3

Variance analysis of mitophagy-related genes in PCOS. (A) Volcano plot showing a summary of the expression differences of 12 mitophagy-related genes between control and PCOS patients’ granulosa cell samples. (B) The clustering heatmap exhibiting the expression pattern of 12 PCOS‐related DE-MRGs among granulosa cell samples. (C) Correlations between DE-MRGs in PCOS granulosa cell samples and the respective scatterplots showing the two pairs of MRGs with the highest correlation. Correlation analyses were assessed using Pearson correlation.

Full size image

GO and KEGG analysis of the DE-MRGs

Based on the GO and KEGG databases, we analyzed the functional enrichment of DE-MRGs. Figure 4A showed the 15 highest-ranking GO terms, including organelle disassembly, autophagy of mitochondrion, mitochondrion disassembly, protein targeting to mitochondrion, establishment of protein localization to mitochondrion and macroautophagy (Fig. 4B). The KEGG analysis revealed that the DE-MRGs were involved in the process of neurodegeneration-multiple disorders and the process of mitophagy-animal (Fig. 4C,D).

Figure 4
figure 4

GO and KEGG analysis of 12 PCOS‐related DE-MRGs. (A) Bar plot of enriched GO terms. (B) Chord diagram of enriched GO terms. (C) Bubble plot of enriched KEGG terms. (D) Chord diagram of enriched KEGG terms. BP biological process, CC cellular component, MF molecular function.

Full size image

Identification of hub genes

For a better understanding of the diagnostic potential of DE-MRGs, we then constructed a prediction model for the diagnosis of PCOS applying three different algorithms to distinguish the PCOS patients from healthy controls. The 12 candidate genes were successively submitted into LASSO, RF and SVM-RFE. 9 out of 12 PCOS-related features of non-zero coefficients were filtered by the means of LASSO algorithm (Fig. 5A,B). Next, we identified feature importance using RF and the top 8 genes were selected as diagnostic genes, as shown in (Fig. 5C,D). And then, features were selected and 3 genes were identified as the best candidates for PCOS based on SVM-RFE (Fig. 5E,F). Finally, we crossed the candidate genes obtained from LASSO, RF, and SVM-RFE models and identified 2 hub genes (TOMM5 and MAP1LC3A) for follow-up steps (Fig. 5G).

Figure 5
figure 5

2 DE-MRGs were identified as potential marker genes for PCOS. (A,B) Regression coefficient path plot and cross-validation curves in LASSO regression algorithm. (C,D) The identification of feature importance based on RF algorithm. (E,F) The curve of change in the true and error value of each gene prediction in SVM-RFE algorithm. (G) Venn diagram showing the intersection of selected markers obtained from the three algorithms.

Full size image

Performance of hub genes to diagnose PCOS in the training and validation sets

In the training set (GSE95728-GC + GSE155489-GC), MAP1LC3A was significantly overexpressed in PCOS compared with the control (P < 0.01, Fig. 6A). The AUROC of MAP1LC3A was 0.860 (95% CI 0.692–1.000), with a sensitivity of 90.9% and a specificity of 81.8% (Fig. 6B). Notably, the AUROC of TOMM5 was 1.000 (95% CI 1.000–1.000), with a specificity of 100.0% and a sensitivity of 100.0% (Fig. S1). The small sample size included in our study may account for this distortion, so we decided that TOMM5 was not suitable for further validation and generalization.

Figure 6
figure 6

The performance of MAP1LC3A to discriminant PCOS in the training and validation set. (A,C,E) Expression difference of MAP1LC3A in PCOS and control groups. (B,D,F) The ROC curve of MAP1LC3A in PCOS and control groups. (G) Diagnostic values of MAP1LC3A for differentiating PCOS from control groups. GC granulosa cell, PPV positive predictive value, NPV negative predictive value, AUROC area under the receiver operating characteristics curve, CI confidence interval. Data were analyzed by Wilcoxon tests or Student’s t-tests.

Full size image

In the validation set (GSE168404-GC), (Fig. 6C,D) showed the value of MAP1LC3A in the diagnosis of PCOS. The expression of MAP1LC3A was also significantly higher in PCOS groups than in controls (P < 0.05, Fig. 6C). The ROC curve demonstrated that MAP1LC3A performed exceptionally well in diagnosing PCOS, with the AUROC of 0.960 (Fig. 6D). Similarly, in another validation set (GSE155489-Oocyte), MAP1LC3A exhibited the excellent diagnostic value with the AUROC of 0.944, as shown in (Fig. 6E–G).

Analysis of immune infiltration

To investigate whether the expression levels of MRGs were related to immunity, the CIBERSORT algorithm was used to evaluate the immune infiltration of PCOS. The analysis of PCA clusters showed that there was a huge distinction between the PCOS and control samples for immune cell infiltration (Fig. 7A). Using the par function, the immune cell percentage was calculated and the stacked histogram was presented (Fig. 7B). Correlation heatmap drawn to assess the correlation among 22 immune cell infiltrations showed that M1 macrophages, CD4 memory resting T cells, and naive B cells had a significant positive relation. Moreover, M1 macrophages and CD4 memory resting T cells also had a positive relation. Activated NK cells, gamma delta T cells, and follicular helper T cells had a significant positive relation. Moreover, activated NK cells and gamma delta T cells also had a positive relation. A positive correlation was also observed between M0 macrophages and memory B cells, and also between CD8 T cells and plasma cells, neutrophils and resting dendritic cells, respectively. CD4 naive T cells had a significant negative correlation with naive B cells (Fig. 7C). Figure 7D showed the difference among 22 immune cell infiltrations, plasma cells in PCOS had a high infiltration compared with control sample. Additionally, it was discovered that the expression of MAP1LC3A was positively related to monocytes (r = 0.615, P = 0.033) (Fig. 7E).

Figure 7
figure 7

Evaluation and visualization of immune cell infiltration. (A) The PCA plot showing immune cell infiltration between PCOS and control samples. (B) Stacked histogram comparing PCOS and control samples for the immune cell proportion. (C) Correlation heatmap of 22 types of immune cells. (D) Boxplots showing 22 types of immune cells in proportion. (E) Lollipop diagram showing the correlation between MAP1LC3A and infiltrating immune cells. Scatter diagram indicating the correlation between MAP1LC3A expression and Monocytes. Data were analyzed by Wilcoxon tests; Correlation analyses were assessed using Pearson correlation.

Full size image

Clinical correlation of MAP1LC3A with testosterone levels

To further illustrate the status of MRGs in PCOS, correlation analysis between MAP1LC3A and testosterone levels was conducted. MAP1LC3A was positively related to testosterone levels (r = 0.795, P = 0.006) (Fig. 8), revealing that MAP1LC3A may exert an effect on ovulation disorders in PCOS.

Figure 8
figure 8

Scatter diagram indicating the relationship between MAP1LC3A expression and testosterone levels. Correlation analysis was assessed using Pearson correlation.

Full size image

Discussion

Polycystic ovarian syndrome is a common endocrine and metabolic syndrome that accounting for 75% of cases of anovulatory infertility28. The development of hyperandrogenemia is the characteristic biochemical feature of the disease and the primary reason behind the majority of PCOS clinical symptoms29. Recently, evidence has shown that androgens could impact the cellular metabolic pathways, potentially leading to risks within the mitochondria30. This indicates that the mitochondrial dysfunction of follicular cells (granulosa cell and oocyte) caused by hyperandrogenemia may partly account for the PCOS ovulation disorders. Therefore, we proposed a novel proposal for finding potential biomarkers possessing high specificity and sensitivity, capable of delineating the extent of mitochondrial quality control in follicular cells to better understand PCOS pathogenesis.

Mitophagy, which is a form of selective autophagy within mitochondria, serves as a crucial mechanism for maintaining cellular mitochondrial quality and therefore is crucial for sustaining energy production and responding to energy stress. When mitophagy is overstimulated under certain stressful conditions, the essential components for cell survival can be digested and lead to cell dysfunction31. So far, a series of studies have shown that excessive mitophagy contributes to the advancement of PCOS. One study observed the autophagy activation in the ovarian tissues of both PCOS individuals and PCOS-liked rats32. Furthermore, there was an observed elevation in mitophagy and a higher presence of injured mitochondria in the cumulus cells of individuals with PCOS. Yi et al. also proposed that the granulosa cells of PCOS patients experienced mitochondrial injure due to the excessive activation of PINK1/Parkin-mediated mitophagy6. Therefore, it is necessary to speak out here that our finding is consistent with those previous study conclusions. We found an obvious up-regulation of MAP1LC3A (a kind of MRGs) expression in PCOS individuals. Moreover, the positive correlation between MAP1LC3A and testosterone levels in PCOS patients supports the notion that mitophagy contributes significantly to the manifestations of PCOS4.

The application of machine learning algorithms in creating decision models that support disease diagnosis and treatment is growing33. A total of 2 differential MRGs were identified in our study, namely TOMM5 and MAP1LC3A. In the merged training set (GSE95728-GC + GSE155489-GC), the AUROC value of TOMM5 was 1.0, with a specificity of 100.0% and a sensitivity of 100.0%. We speculated that the model distortion may be due to the limitations of sample size, making it unsuitable for further validation and generalization. MAP1LC3A, on the other hand, of which the AUROC values were all greater than 0.8 both in the merged training set and validation sets (GSE168404-GC and GSE155489-Oocyte), indicating that MAP1LC3A have the certain accuracy and specificity for distinguishing PCOS from the matched controls. Additionally, it is worth emphasizing that we selected 2 types of follicular cells (granulosa cell and oocyte) for modeling and validation, aiming to compensate for the limitations of the sample size. The quality of oocyte can directly characterize the follicular microenvironment, thereby predicting the ovulation ability of PCOS patients. And the granulosa cells specifically located around the oocyte play an important role in oocyte maturation and ovulation6. Abnormal granulosa cell function may indirectly affect follicular development and alter many symptoms of PCOS34.

Growing proof suggests that the self-clearance of malfunctioning mitochondria is an effective strategy to keep the immune system in check. Mitophagy restricts the secretion of inflammatory cytokines and directly regulates mitochondrial antigen presentation, thereby maintaining the immune cell homeostasis35. Moreover, by regulating the adaptive immune response of memory NK cells, CD8 T cells, and dendritic cell-T cell synapses, mitophagy can shield cells against chronic inflammation36. Currently, several studies have jointly demonstrated the link between PCOS and low-grade chronic inflammation37. Furthermore, the persistent presence of inflammation in PCOS can exacerbate the obstruction of energy supply to oocytes, resulting in ovum quality impairment and subsequently impacting ovulation38. In our study, the immune infiltration analysis displayed a positive correlation between the presence of monocytes and MAP1LC3A levels. Based on the consistent trend of changes in MAP1LC3A and testosterone levels mentioned above, this suggested that immune cells and cytokines interact with androgens may result in the disruption of ovarian immune balance in PCOS. As González’s findings suggested, the infiltration of monocytes into the ovary could potentially initiate a localized inflammatory response, leading to the stimulation of ovarian androgen synthesis in PCOS women39. In addition, the CIBERSORT analysis revealed an increased infiltration of plasma cells in PCOS. In Ewa Rudnicka’s review, it was also noted that PCOS women display higher serum concentration of TNF and C-reactive protein (CRP) as well as monocyte and lymphocyte circulating levels37. To summarize, infiltrating immune cells contributes to the initiation and advancement of PCOS, and targeting MAP1LC3A may help correct this aberrant immunological status in the coming times.

In our study, the MRGs were acquired from the Reactome database, an up-and-coming resource that has been extensively utilized by numerous studies40. However, several specific mitophagy receptors, including BNIP3, p62, OPTN, etc., were absent from it. Thus, it may be preferable to combine the Reactome database with other databases such as KEGG to obtain more thorough MRGs for future studies. Furthermore, the absence of experimental verification for the samples poses a constraint on our study. The limited sample size of the three datasets in our study necessitates selecting more datasets and confirming our findings in a larger PCOS cohort.

Conclusions

Our study identified the mitophagy-related gene MAP1LC3A as a promising biomarker in PCOS. Additionally, we discussed the possible correlation between MAP1LC3A and infiltrating immune cells, shedding new perspective on its significant contribution to the progress of PCOS. This offers a new insight into the prevention and treatment of PCOS.