Exosomal let-7d-3p and miR-30d-5p as diagnostic biomarkers for non-invasive screening of cervical cancer and its precursors

We first performed retrospective analyses of medical records of cervical cancer patients to evaluate the accuracy of current cytology tests (Additional file 2: Supplementary Methods). A total of 456 of 608 patients had at least one TCT or Pap smear record, and 498 of the total patients had tissue biopsy results; 468 of 608 patients had an HPV test, of which 445 were HPV positive and 23 were HPV negative (Additional file 3: Figure S2A). The pathological stages of tissue specimens obtained from the operation were used as the diagnostic criteria for each patient. TCT or Pap smears that were negative for Intraepithelial Lesion or Malignancy (NILM), or Low-grade Squamous Intraepithelial Lesion (LSIL) were classified as true positive results for CIN I patients. High-grade Squamous Intraepithelial Lesion (HSIL) or Atypical Squamous Cell of Undetermined Significance (ASC-US) /Atypical Squamous Cell—cannot exclude high-grade squamous intraepithelial lesion (ASC-H) / Atypical Glandular Cells—not otherwise specified (AGC-NOS) were classified as true positive results for CIN II-III, adenocarcinoma (ACC), or squamous cell carcinoma (SCC) patients. Based on the generally accepted gold standard described above, the overall detection rate of the cytology tests in all the 465 patients with cytology results was approximately 68.86% (CIN I 67.65%, CIN II-III 65.57%, SCC 73.71%, and ACC 65.71%) (Additional file 3: Figure S2B). The overall detection rate of the biopsy tests in all the 498 patients with biopsy results was approximately 93.17% (CIN I 76.92%, CIN II-III 92.76%, SCC 96.52%, and ACC 94.59%) (Additional file 3: Figure S2C). Retrospective analyses demonstrated that the accuracy of current cytology tests is relatively low when compared with cervical biopsy, but there is still much room for improvement in CC screening.

To develop a more accurate screening method based on circulating exosomal miRNAs, miRNA sequencing was performed in 121 plasma samples from healthy volunteers, cervical carcinoma patients, and precancerous patients. The miRNA expression levels were quantified by Reads Per Million (RPM) mapped reads and then normalized with log2(RPM + 1), which is the commonly used method for miRNAs quantification and normalization [10]. Detailed methods regarding plasma exosomal miRNA sequencing analysis are provided in Additional file 2: Supplementary Methods.

Proper classification of the studied subjects was not only important for identifying DEmiRs, but also critical for developing powerful diagnostic biomarkers for CC screening. According to clinical guidance, CIN I patients have a reversible disease response and may return to normal, and thus do not have to be treated with surgery and medication. Therefore, CIN I patients and healthy volunteers were combined into one group named CIN I-. A high-grade CIN (i.e., CIN II-III) patients and CC patients (i.e., ACC and SCC) need treatment and were thus combined into another group named CIN II+. Our aim was to identify circulating exosomal DEmiRs as diagnostic biomarkers for screening CC and high-grade CIN. This grouping strategy increased sample sizes in each group and maximized the possibility of discovering the diagnostic miRNAs. The average age of all 608 studied subjects was 50 ± 24 years, and the average age of the CIN I- and the CIN II+ groups was 50 ± 27 and 50 ± 24 years, respectively; thus, there was no significant age difference between the groups of patients.

A total of 312 miRNAs with mean log2(RPM + 1) values > 1 were detected from miRNA sequencing of exosomes derived from 121 plasma samples. Among these miRNAs, CIN I- samples were used as a reference data to compare with the other sample groups (CIN II-III, CC, SCC, and ACC). As a result, a total of 69 DEmiRs were identified in these four comparisons (false discovery rate, FDR < 0.01), of which 29 were identified in at least two comparisons (Table 1 and Fig. 1a). Specifically, 61 DEmiRs were identified between CIN I- and CIN II-III. Thirteen and eight DEmiRs were identified between CIN I- and SCC, and between CIN I- and ACC, respectively, of which four were common. Thirty-six DEmiRs were identified between CIN I- and CC and 28 were also identified between CIN I- and CIN II-III (Fig. 1a).

Using all the DEmiRs detected above, principal component analysis (PCA) and clustering analysis were performed to assign these plasma samples into groups with similar miRNA expression patterns (Fig. 1b and c). Interestingly, CIN II-III and CC subjects shared common miRNA expression profiles. Furthermore, we also compared the expression of miRNAs between healthy and CIN I subjects in the discovery set, but none of the DEmiRs were found (FDR > 0.05). These results also justified our grouping strategy by which CIN II-III, ACC, and SCC patients were combined into one group, while healthy and CIN I subjects were combined into another group. Finally, the comparison of CIN I- with CIN II+ group identified 37 DEmiRs (FDR < 0.01), including 9 up-regulated and 28 down-regulated DEmiRs (Additional file 4: Table S1).

Next, a set of miRNAs were selected from these 37 DEmiRs in the 121 plasma exosomal sequencing samples using the Random Forest algorithm. This led to the identification of the best panel with eight miRNAs (let-7a-3p, let-7d-3p, miR-30d-5p, miR-144-5p, miR-182-5p, miR-183-5p, miR-215-5p, and miR-4443) that are the strongest predictors in clinical diagnosis (i.e., CIN I- versus CIN II+). PCA was performed using these eight miRNAs; the first two principal components explained the 60% of total variance in the discovery set. They were visualized to show the groupings of these exosomal samples, indicating that samples in CIN I- and CIN II+ groups were nicely separated (Fig. 1d). Hierarchical clustering of these eight miRNAs indicated that only two CIN II+ patients were incorrectly classified into CIN I- group (Fig. 1e). ROC analysis was further performed to evaluate the sensitivity and specificity of the eight-miRNA signature to discriminate CIN II+ from CIN I- subjects. This yielded a very high AUC value of 0.992 (Fig. 1f). The AUC value of individual DEmiRs ranged from 0.797 to 0.890 in the discovery set (Fig. 1g and h). Furthermore, there were no significant differences in the expression of miRNAs in the best panel between different HPV types (P > 0.05). In summary, the newly identified eight-miRNA signature is highly predictive of CIN I- and CIN II+ irrespective of HPV types.

To gain further insight into the molecular function of these diagnostic miRNAs in CC, we performed enrichment analysis of Gene Ontology categories and Kyoto Encyclopedia of Genes and Genomes pathways on these miRNA targets (Additional file 2: Supplementary Methods). There were 25 significant pathways (FDR < 0.01) involved in at least five of the eight DEmiRs in the best panel (Fig. 1i) and most of them were cancer-related pathways, such as adherens junction, hippo signaling pathway, cell cycle, p53 signaling pathway, AMPK signaling pathway, and so on. Interestingly, the top targeted pathway was viral carcinogenesis, which was consistent with CC caused by HPV. Oocyte meiosis and estrogen signaling pathways were also significant. The connection network showed genes targeted by at least three miRNAs. Notably, miR-30d-5p, miR-182-5p, and miR-183-5p simultaneously regulate genes RAC1, IGF1R, NRAS, TP53, and CCND1, which were involved in more than 10 of the 25 significant pathways, and also regulate CDC27 and YWHAG, which were involved in 5 to 10 of the 25 significant pathways. Furthermore, these eight miRNAs also regulated several other important cancer genes, including CDK6, which was involved in more than 10 pathways, and MAP3K1, which was involved in 5 to 10 pathways (Fig. 1j). These results demonstrated that the exosomal miRNAs detected in our sequencing study can not only serve as potential diagnostic biomarkers, but can also be identified as potential anti-cancer drug targets because they are functionally involved in the development and progression of CC.

We next used qRT-PCR to evaluate eight DEmiRs in the best panel for discriminating CIN I- from CIN II+ (Fig. 1g and h) in paired cancerous and para-carcinoma tissues from 46 new CC patients (Additional file 5: Figure S3A and B). Five miRNAs (let-7a-3p, let-7d-3p, miR-30d-5p, miR-183-5p, and miR-182-5p) showed consistent variation trends in plasma exosomes, among which three (let-7a-3p, let-7d-3p and miR-30d-5p) showed significant differences in expression between cancerous and para-carcinoma tissues. However, other three of the eight miRNAs (miR-215-5p, miR-144-5p, and miR-4443) showed either no changes or reversed trends in the tissues compared with plasma exosomes.

To validate these diagnostic miRNAs by ddPCR, four stably expressed miRNAs (i.e., miR-128-3p, miR-129-5p, miR-320a, and Let-7i-5p) were chosen from the exosomal miRNA sequencing data in the discovery set (Additional file 2: Supplementary Methods). These four miRNAs had relatively high expression levels (log2(PRM + 1) > 10) and small variability among samples (coefficient of variation < 5%) (Additional file 6: Figure S4). They were used as endogenous references for normalizing exosomal miRNA expression levels in ddPCR analysis. The application of endogenous reference with stable expression is beneficial in ddPCR data normalization, especially when the sample quality is highly variable. Furthermore, the expression stability of the four endogenous references were evaluated by ddPCR in 203 independent plasma samples, including 50 healthy volunteers, 34 CIN I patients, 56 CIN II-III patients, and 63 CC patients. The four inner control miRNAs showed invariable expression levels across these independent samples in both CIN I- and CIN II+ groups. The difference in expression of each of these inner control miRNAs was not significant between CIN I- and CIN II+ groups (P > 0.05) (Additional file 7: Figure S5), demonstrating the robustness and suitability of these inner controls in ddPCR analysis.

Then, the expression levels of the above three miRNAs (i.e., let-7a-3p, let-7d-3p, and miR-30d-5p) that were validated in tissues were quantified and normalized by ddPCR in 203 independent plasma samples. In our preliminary ddPCR analysis, let-7a-3p only produced about 2–10 copies /μl positive droplets using 4 ng input of exosomal miRNA sample, whereas the four inner controls and the other two miRNAs produced similar numbers of positive droplets with 0.02–0.1 ng input. Therefore, the less significantly expressed let-7a-3p in plasma exosomes was not pursued further in subsequent validations.

The expression levels of let-7d-3p and miR-30d-5p was significantly decreased in the CIN II+ group when compared with the CIN I- group (Fig. 2a and b), which is consistent with both the plasma exosomal sequencing results and the qRT-PCR results of patients’ tissues as described above. The combination of the expression of let-7d-3p and miR-30d-5p from plasma exosomal sequencing in 121 training samples gave a distinguishing performance, with an AUC value of 0.922, whereas the AUC value based on the combined expression of these two DEmiRs from ddPCR in 203 validation plasma samples was 0.828 (Fig. 2c). A total of 166 of 203 validated samples had cytology test results. The AUC value based on cytology tests was 0.766, AUC value based on the miRNAs increased to 0.834, and integration of the two miRNAs in a cytological test-based model further achieved a higher AUC value of 0.887 (Fig. 2d). The positive predictive value and negative predictive value of the two-miRNA test were 0.95 and 0.75, respectively, in the discovery set, and the corresponding values were 0.80 and 0.81, respectively, in the validation set.

To determine the spatial distribution and expression levels of miRNAs, we also measured seven miRNAs in the best panel (miR-320a, miR-128-3p, miR-129-5p, let-7d-3p, let-7i-5p, miR-30d-5p, and miR-30a-5p) in both plasma exosomes and in cancerous and adjacent normal tissues of CC patients using ddPCR analysis. There were different expression profiles between tissues and exosomes in the same validation system. The proportions of these miRNAs were, respectively, 57.5, 0.5, 2.9, 5.7, 9.1, 10.3, and 13.9% in tissues, and, respectively, 23.8, 22.8, 5.3, 7.5, 4.5, 18.2, and 17.9% in exosomes (Additional file 8: Figure S6). The Spearman correlation coefficient of miRNAs expression abundances between circulation and tissues was only 0.321 (P = 0.498), suggesting that exosomal miRNAs are likely selectively secreted from tumor cells. Although tumor cell content of sections from cervical tissue samples was confirmed by H&E staining to be > 70% in our study (Additional file 2: Supplementary Methods), these bulky tissue samples contained various amounts of stromal cells. It might be possible, although less likely, that exosomal miRNAs can be secreted by other cell types such as cancer-associated stromal cells in the tumor microenvironment.