Introduction

Follicular thyroid carcinoma (FTC), accounting for approximately 10% of clinically manifested thyroid carcinomas, is prone to hematogenous spread and confers an unfavourable prognosis [1, 2]. Ultrasonography is an established method for the initial evaluation of thyroid nodules, but it lacks reliable distinction of FTC due to the rare malignant features on ultrasonography [3]. The histopathological diagnostic criterion of FTC, which is based on the presence of the capsule or vascular invasion [4], cannot be determined by preoperative fine-needle aspiration (FNA), core-needle aspiration and intraoperative frozen section examination (IFSE) [3, 5, 6]. This limitation frequently leads to unnecessary diagnostic surgical procedures (hemithyroidectomy) to evaluate whether there is capsule or vascular invasion [7, 8]. For postoperative pathologically proven benign nodules (without capsule and vascular invasion), they could be treated conservatively or with radiofrequency ablation without hemithyroidectomy [9, 10]. Furthermore, a second surgical procedure with remnant thyroidectomy can be required in the majority of FTC cases to facilitate radioiodine therapy due to the high propensity for distant metastasis [3]. This formidable challenge highlights the urgent need to develop a new, potentially objective, accurate preoperative auxiliary approach to detect FTC following ultrasonography.

To address this diagnostic dilemma, molecular markers adjunctive to FNA, such as the gene expression classifier [11, 12] and protein detection [13, 14], have been proposed to differentiate FTC prior to surgery in decade-long practice. However, its clinical application remains limited due to the unsatisfactory diagnostic performance [11,12,13, 15,16,17] and the requirement for sample quality, i.e. the need for fresh tissue samples in nucleic acid-based testing and enough FNA tissue in protein detection [14, 18]. Recently, liquid biopsy has yielded novel insight into the detection of malignancies, which can be performed serially with ease by using circulating biomarkers [19,20,21]. Among these molecular elements, small extracellular vesicles (sEVs) and cell-free RNAs are the fastest-growing biomarkers that have been extensively validated in the clinical setting [22, 23]. However, a diagnostic test regarding circulating RNAs to help distinguish FTC from benign thyroid lesions is not yet well reported.

In this multicentre study, we aimed to discover and validate a novel circulating RNA signature that could serve as a noninvasive, convenient, specific and stable auxiliary approach to facilitate discrimination between FTC and follicular thyroid adenoma (FTA).

Methods

Study design and participants

This study was approved by the institutional ethics review board of West China Hospital of Sichuan University (number: 2019 [507]). Patients who signed their written informed consent for their clinical samples and information to be used and met the following inclusion criteria were prospectively enrolled in this study: (1) patients with thyroid nodules who planned to undergo thyroidectomy, (2) patients without a previous cancer history, (3) patients who were not pregnant or lactating, and (4) patients with no endocrine or metabolic comorbidities. The patients with FTC and FTA were enrolled based on postoperative pathological diagnosis, which was confirmed by two pathologists based on the standardised World Health Organisation classification [4]. Patients in whom the postoperative histological diagnosis was not unambiguous or other types of thyroid neoplasms were excluded from the study. Age and sex were then matched 1:1 (training cohort) and 1:2 (validation cohort) based on the results of propensity score matching. This study involved observational cohorts from four tertiary hospitals: a training cohort (n = 41) was from West China Hospital, and a validation cohort (n = 150) was obtained from West China Hospital, West China Fourth Hospital, West China Tianfu Hospital and Sichuan Province Cancer Hospital. The study design and inclusion/exclusion process are depicted in Fig. 1 and Supplementary File 1: Fig. S1.

Fig. 1: Flow diagrams showing the overall study design and patients in the training and validation cohorts.
figure 1

FTC follicular thyroid carcinoma, FTA follicular thyroid adenoma, sEV small extracellular vesicle, ROC receiver operating characteristic, RAI radioiodine.

Full size image

Statistical analysis

Data analysis was performed using the statistical package for the social sciences (version 25.0), statistical programming language python (version 3.8) and GraphPad Prism (version 8.0). Quantitative variables are presented as the mean ± standard deviation (SD)/standard error (SE) or median and interquartile range (IOR), which were compared by using Student’s t test, one-way ANOVA, nonparametric Mann‒Whitney U-test and Wilcox rank sum test. Comparisons of categorical variables were performed using the chi-square test or Fisher’s exact test. The diagnostic model was constructed using multivariate logistic regression analysis, and the predictive efficacy was determined by receiver operating characteristic (ROC) curves with sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). The optimal cutoff thresholds for the ROC curves were determined using Youden’s index. The tumour burden (TB) was calculated as previously described [24]: (maximum tumour diameter)2 + (number of tumours)2. A P value <0.05 (*P < 0.05, **P < 0.01, ***P < 0.001) was considered indicative of statistical significance, and all tests were two-sided.

The other materials and methods used in this study are described in Supplementary File 2.

Results

Clinical characteristics of the study population

To identify circulating sEV-associated and cell-free RNAs as potential diagnostic biomarkers for FTC patients following ultrasonography which lacks reliable distinction of FTC (Supplementary File 1: Fig. S2), plasma was initially collected from 1076 patients during the study phase. After the elimination of patients with unambiguous histological diagnosis (n = 9) or other types of thyroid neoplasms (n = 790) postoperatively and matching for age and sex, a total of 21 FTC patients and 20 FTA patients from a single centre were enrolled in the training cohort (n = 41). In addition, a matched validation cohort (n = 150) from four tertiary hospitals, including 50 patients with FTC and 100 patients with FTA, was used for classifier validation (Fig. 1 and Supplementary File 1: Fig. S1). The basic clinical and pathological features of the enrolled patients are summarised in Table 1.

Table 1 The baseline demographics and clinical characteristics of enrolled subjects.
Full size table

Profiling of circulating sEV-associated and cell-free RNAs in the training cohort

To identify potential high-quality circulating RNA biomarkers, each plasma sample was divided into two aliquots: one was used for the sequencing of sEV RNAs, while the other was used directly for the sequencing of cell-free RNAs from whole plasma according to the workflow shown in Fig. 2a. Circulating sEVs were isolated and validated by nanoparticle tracking analysis (NTA), western blot (WB) and transmission electron microscopy (TEM) (Fig. 2b–d). Unsupervised hierarchical clustering based on the expression levels of circulating sEV-associated and cell-free RNAs categorised the samples in a similar pattern as the clinicopathologic classifications did (Fig. 2e, f). Machine learning suggested that the diagnostic potential of cell-free miRNAs is inferior to that of the other three kinds of RNAs (Fig. 2g). Given the diagnostic power of sEV long RNAs, cell-free long RNAs and sEV miRNAs is comparable, sEV miRNAs are stable due to the protection of the membranous structure [19, 25] and are easier to detect than long RNAs in the circulatory system [26, 27], which may be convenient for subsequent clinical application. The top 10 sEV miRNAs (Supplementary File 1: Table S1) with the largest diagnostic potential in the training cohort were selected for subsequent investigation.

Fig. 2: Identification of circulating candidates for FTC in the training set.
figure 2

a Workflow of RNA sequencing. b–d Circulating sEVs were isolated and validated by nanoparticle tracking analysis (NTA), western blotting (WB) and transmission electron microscopy (TEM), scale bar, 100 nm. e, f Heatmap of sEV long RNAs, sEV miRNAs, cell-free long RNAs and cell-free miRNAs in the training cohort by unsupervised hierarchical clustering (n = 41). g The comparison indicated that the diagnostic properties of sEV long RNAs, cell-free long RNAs and sEV miRNAs were comparable for the detection of FTC and were superior to those of cell-free miRNAs in RNA sequencing. The results are presented as the median and interquartile range (IOR). Note: Several cases failed sequencing library construction, so the number of enrolled cases was <41 in certain heatmaps.

Full size image

Construction and validation of a circulating sEV-based miRNA classifier

The candidate sEV miRNAs were further detected in the training cohort by qRT‒PCR, which confirmed five upregulated sEV miRNAs in FTC (Fig. 3a). The expression of the other five miRNAs either failed to be detected (almost all CT valueså 35) or showed the opposite results to the sequencing data between the two groups of patients (data not shown). Then, to identify a suitable sEV miRNA classifier, ROC analysis was performed, and the area under the curve (AUC), sensitivity, specificity, PPV and NPV were calculated (Fig. 3b and Table 2). miR-127-3p, the miRNA with the best performance, achieved an AUC of 0.889. Next, two or more miRNAs were integrated to further improve the performance. Integrating miR-127-3p and miR-223-5p resulted in an improved AUC of 0.905; integration of three miRNAs (miR-127-3p, miR-223-5p, and miR-432-5p) exhibited an AUC of 0.919; integration of four miRNAs (miR-127-3p, miR-223-5p, miR-432-5p, and miR-146a-5p) exhibited an AUC of 0.921; and integration of five miRNAs (miR-127-3p, miR-223-5p, miR-432-5p, miR-146a-5p and miR-151a-3p) achieved the highest AUC of 0.924. The sensitivity, specificity, PPV and NPV of these five miRNAs, called the CirsEV-miR classifier in our study, were 0.810, 0.900, 0.895, and 0.818, respectively.

Fig. 3: Classifier development and validation process for FTC.
figure 3

a Relative expression levels of five identified sEV miRNAs in the training cohort measured by qRT‒PCR. b ROCs of classifiers constructed with different combinations of the five sEV miRNAs in the training cohort. Only ROCs with the best diagnostic powers of different miRNA combinations are shown. c ROCs of the CirsEV-miR classifier for total FTC, minimally invasive FTC, encapsulated angioinvasive FTC, and widely invasive FTC. d The CirsEV-miR classifier scores of FTA and FTC in the validation cohort. e The CirsEV-miR classifier scores of FTC with low TB and FTC with high TB in the validation cohort. f The relationship of clinical features with the CirsEV-miR score. TB tumour burden, VI vascular invasion, DM distant metastasis, RAI radioiodine. The results are presented as mean ± SE.

Full size image
Table 2 Performance of classifiers in the training and validation cohorts.
Full size table

Then the differences in these five sEV miRNAs were further confirmed in the validation cohort (Supplementary File 1: Fig S3), and ROC analysis showed that the CirsEV-miR classifier achieved AUCs of 0.844, 0.824, 0.821, and 0.929 for total FTC, minimally invasive FTC, encapsulated angioinvasive FTC, and widely invasive FTC in the validation set, respectively (Fig. 3b, c and Table 2), suggesting that this classifier may not only have diagnostic value for total FTC but also apply to different FTC subtypes. The CirsEV-miR score was then generated through logistic regression analysis (Supplementary File 1: Table S2). The CirsEV-miR score of FTC was significantly higher than that of FTA in the validation cohorts and increased with TB in FTC (Fig. 3d, e), suggesting that the CirsEV-miR score increases with malignancy. Furthermore, our data analysis indicated that FTC patients with higher CirsEV-miR scores were more likely to have advanced stage disease (Fig. 3f). Taken together, these results revealed that the CirsEV-miR classifier may be an effective and promising approach to help the differential diagnosis of FTC.

Biological function enrichment analysis of five sEV miRNAs in the CirsEV-miR classifier

To evaluate the functional relevance of the five miRNAs used in the CirsEV-miR classifier, we performed Gene Ontology (GO) and Kyoto Encyclopaedia of Genes and Genomes (KEGG) analyses of their target genes identified through the miRNA target database. We found that the target genes were significantly enriched in biological processes, such as cellular polysaccharide metabolic processes and organ- or tissue-specific immune responses, and in molecular functions, such as cytokine activity and tumour necrosis factor receptor binding (Supplementary File 1: Fig S4A). In addition, the target genes were significantly enriched in pathways related to tumourigenesis, such as the Wnt and Hippo signalling pathways (Supplementary File 1: Fig S4B). The results suggested that these five miRNAs may be involved in the tumourigenesis and progression of FTC.

Parallel expression and stability of circulating sEV miRNAs

To determine whether the levels of these circulating sEV miRNAs were consistent with sEV miRNAs from other samples, sEVs from cell lines, organoids and tissues were isolated and validated (Supplementary File 1: Fig S5). The results showed higher levels of sEV miRNAs in the culture media of the FTC cell line than in the normal thyroid follicular epithelial cell line. Additionally, FTC organoid- and tissue-derived sEVs expressed higher levels of these sEV miRNAs (Fig. 4a). In addition, the instability of miRNAs in plasma remains a major limitation for clinical application [22]. Thus, we randomly collected five FTC plasma samples to investigate the stability of sEV miRNAs. We found that the levels of sEV miRNAs in plasma were constant after treatment with RNase A. Furthermore, prolonged exposure to room temperature and repeated freezing and thawing had no influence on the plasma levels of sEV miRNAs (Fig. 4b). We next analysed 10 paired plasma samples acquired from FTC patients with distant metastasis before and one month after thyroidectomy and radioiodine therapy. Most patients showed a significant decrease in the levels of circulating sEV miRNAs in postoperative plasma (after thyroidectomy) compared with those in preoperative plasma, which was also observed in postradioiodine therapy plasma compared with postoperative plasma (Fig. 4c). Taken together, these results indicated that the levels of circulating sEV miRNAs were consistent with those of FTC cell lines, organoids, and tissues and could serve as FTC-specific miRNAs that confer sufficient stability for use in the clinic.

Fig. 4: Parallel expression and stability of circulating sEV miRNAs.
figure 4

a Differential expression of miRNAs between sEVs from the FTC cell line, organoids and tissue and sEVs from the normal thyroid follicular epithelial cell line, organoid and tissue. b Relative expression of circulating sEV miRNA levels in patients with FTC when plasma was treated with RNase A, subjected to prolonged incubation at room temperature and subjected to repeated freezing and thawing. c Differential expression of circulating sEV miRNAs between patients with FTC before surgery and 1 month after surgery and radioiodine (RAI) therapy. The results are presented as mean ± SE.

Full size image

Integrating the CirsEV-miR classifier improves the current surgical strategy

FNA and ultrasonography are established methods for the evaluation of thyroid nodules. Thyroidectomy may be considered based on the results of FNA or large nodules (>4 cm) or in patients with familial thyroid carcinoma or a history of radiation exposure [28]. Given the importance of clinical decision-making, we asked whether combined analyses with the CirsEV-miR classifier would improve the surgical strategy. We found that if currently established tools were used to guide surgical intervention, including FNA, imaging features or high-risk stigmata, then 14/50 cases (28%) with FTC would not have proceeded to surgery. Of those 14 patients (12 minimally invasive FTC and 2 encapsulated angioinvasive FTC), the preoperative cytological result was category III, and thyroidectomy was performed mainly due to the patient’s preference, continuous growth of the nodule or compression symptoms in the follow-up. However, when the CirsEV-miR classifier was added, all patients would have been correctly identified at diagnosis (Supplementary File 1: Fig. S6). In addition, we evaluated the decisional regret score (DRS) of patients in the validation cohort after surgical intervention and found that 16/100 FTA cases (16%) and 7/50 FTC cases (14%) met our definition of heightened decisional regret (Supplementary File 1: Tables S3 and S4, Fig. S7A). The main reasons for regret for FTA and FTC patients were daily thyroxine replacement therapy and the second surgery with remnant thyroidectomy to facilitate radioiodine therapy due to the findings of circulating tumour cells (CTCs) after surgery (Supplementary File 1: Fig. S7B), respectively. In those cases, 14/16 FTA cases (87.5%) were diagnosed as “low-risk”, and all FTC cases were diagnosed as “high-risk” by the CirsEV-miR classifier. Hence, those FTA patients would not have considered surgery, and those FTC patients would perform total thyroidectomy in the first surgical intervention through CirsEV-miR classifier analysis. Taken together, the addition of this classifier as a biomarker improves the current surgical strategy.

Discussion

The detection rate of thyroid nodules has dramatically increased in recent decades due to advances in high-resolution ultrasound and liberal screening practices [1]. Adequate evaluation of the malignancy risk of thyroid nodules is essential for management decision-making. Unfortunately, FTC, characterised by capsular or vascular invasion, lacks adequate approaches for its preoperative detection [3, 6], which results in a large number of unnecessary diagnostic operations performed in patients with benign thyroid nodules [7, 8] and the second surgical procedure in patients with FTC to facilitate radioiodine therapy due to the high risk of distant metastasis [3].

Previous studies have explored the diagnostic performance of gene and protein signatures for the detection of FTC by FNA samples and found that a six-gene panel (ECM1, RXRG, SDPR, SLC26A4, TIFF3, TIMP1) [11], two miRNA classifier (miR-7-5p/miR-7-2-3p) [12] and protein markers (HBME-1, Ki67, FAM172A, APLP2, RRM2 and PRC1) [13, 15,16,17] could distinguish patients with FTC. However, several potential challenges related to those previous studies still exist, including its inherent limitations, i.e. the need for tissue samples to be obtained under ultrasound guidance, the need for fresh tissue samples in nucleic acid-based testing, and especially the unsatisfactory sensitivity or specificity for detection of patients with FTC, i.e. the sensitivity of 100%, but specificity of 45% only [15]. While gene- and protein-based approaches continue to be refined, for example, with successive iterations of ThyroSeq panels [29], there is an evident need for alternative approaches to address this diagnostic dilemma. Recently, accumulating evidence has suggested that sEV and cell-free RNA-based liquid biopsy assays offer a promising strategy for the detection of multiple human cancers [19, 22, 23]. A previous study reported TPO-positive extracellular vesicle miR-let-7f as a candidate circulating biomarker to identify FTC [25]. However, this study was primarily performed in a relatively small sample size (n = 60) and lacked validation in an independent clinical cohort, and the profiling assay was limited by only 85 cancer-associated vesicular miRNAs. Cell-free RNAs, which are often released in systemic circulation from multiple cellular sources, were not investigated in their study, and whether they represent adequate diagnostic potential for FTC remains unclear.

In the present study, to address this formidable challenge, a relatively large plasma sample among the available studies was enrolled from four tertiary hospitals to identify potential circulating RNA biomarkers for FTC, including sEV-associated and cell-free RNAs. A systematic and comprehensive biomarker discovery approach by small RNA sequencing showed that sEV RNAs may demonstrate a superior diagnostic performance for its ability to identify FTC patients than cell-free RNAs, especially higher than cell-free miRNAs. It has been reported that bioactive molecules encapsulated in sEVs can reflect the cellular origin and physiological state, representing the “fingerprint” or “signature” of the donor cell [19]. Moreover, the membranous structure of sEVs can ensure the stable expression of RNAs in the bloodstream, avoiding degradation by extracellular enzymes. The remarkable activity and stability strengthen the potential of circulating sEV RNAs to be reservoirs for biomarkers. Consistent with our study, sEVs provide a more consistent source of RNA than whole plasma for RNA biomarkers, which was also observed in oropharyngeal squamous cell carcinoma [30], non-small cell lung cancer [31] and prostate cancer [32]. To develop a clinically feasible and convenient detection assay for FTC, sEV miRNAs were validated and applied to construct a CirsEV-miR classifier based on five circulating sEV miRNAs, which performed robustly with AUCs of 0.924 and 0.844 in the clinical training and validation cohorts, respectively. In summary, based upon our findings, we proposed and highlighted the potential intended use of a circulating sEV-based miRNA signature for screening FTC. In such a scenario, if a specific individual is diagnosed as “high risk” by this noninvasive and inexpensive sEV-based transcriptomic assay, operation may be considered for further clinical intervention.

The CirsEV-miR score in this study markedly increased with TB and advanced stage of FTC, which led us to hypothesise that the CirsEV-miR classifier might be an FTC-specific biomarker, as FTC patients with high TB and advanced stage could secrete more sEV miRNAs in the circulatory system. To verify this hypothesis, all five miRNAs were further validated in FTC and normal thyroid follicular epithelial cell lines, organoids, tissue and plasma at different stages. Among those samples, organoid- and tissue-derived sEVs have been reported to well reflect the original state of cell communication in the tumour microenvironment because they maintain the landscapes of the parental tumours from which they were derived in terms of histopathology, expression features, genomic profiles, mutational signatures, and intratumour heterogeneity [33,34,35]. As expected, the levels of these sEV miRNAs were higher in sEVs from the FTC cell line, organoid and tissue and were gradually decreased in plasma postoperation and postradioiodine therapy. Since these five miRNAs were enriched in the sEVs of FTC patients, the source of enrichment in the sEVs of FTC patients was more likely to be FTC. This potential link between FTC-derived sEVs containing these miRNAs and increased amounts of them in circulating sEVs from FTC patients confirmed that the CirsEV-miR classifier could have FTC-specific diagnostic value.

Several limitations of this study are worth mentioning. First, we found that the CirsEV-miR score is related to advanced tumour stage. However, the prognosis of FTC has not been investigated due to the limited follow-up time. Thus, the prognosis of the identified sEV miRNAs in the CirsEV-miR classifier may need to be further validated in FTC. Second, the biological function enrichment indicated that these miRNAs may be involved in the tumourigenesis of FTC, and the potential mechanisms should be clarified in future work. Finally, the sEVs and cell-free long RNAs identified in our study may need to be verified in the future with advanced detection technology, such as droplet digital PCR (ddPCR) [36] due to the subtle expression of long RNAs in the circulatory system, especially those capsulated in sEVs. Additionally, the diagnostic power of the CirsEV-miR classifier in our study still needs to be further verified in the future.

In conclusion, using systematic and comprehensive biomarker discovery followed by robust clinical validation, our study provided a CirsEV-miR classifier that may represent a noninvasive, convenient, specific and stable biomarker with the potential for assisting in the diagnosis of FTC following ultrasonography.