CRABP1, C1QL1 and LCN2 are biomarkers of differentiated thyroid carcinoma, and predict extrathyroidal extension

Two wFTC (cases 1 and 2) and two mFTC (cases 3 and 4) were collected from tumour tissue biobank (Porto). These four tumours from patients with 55–82 years of age were subjected to high-throughput paired-end RNA-seq analyses. Briefly, the patient/case 1 had a tumour measuring 2 cm in diameter and presenting a predominant follicular growth pattern and oncocytic features; patient/case 2 had a tumour measuring 5 cm in diameter and presenting a predominant solid/trabecular growth pattern; patient/case 3 had a tumour measuring 3.7 cm in diameter and presenting a follicular growth pattern; and patient/case 4 had a tumour measuring 4 cm in diameter and presenting a follicular growth pattern.

For comparative and validation purposes, 137 frozen thyroid samples were collected from tumour tissue biobank (Porto): 98 DTC [15 FTC (including the four used in RNAseq), 23 follicular variant of PTC (FVPTC) and 60 PTC], 20 FTA, and 19 normal samples of adjacent/contralateral thyroid tissues from patients with DTC from this series. The histology of all DTC was revised by two pathologists (CE and MSS) and the final classification was made according to the WHO criteria [3]. The clinicopathological features and genetic alterations of the DTC are presented in Table 1, and a briefly description of the series is available in the Additional file 1. Furthermore, ten thyroid cancer cell lines were also used in this study including one cell line derived from one FTC with oncocytic pattern (XTC1), three derived from PTC (BCPAP, K1 and TPC1), and six derived from undifferentiated thyroid carcinoma (C643, HTH74, KAT4, T238, T241, 8505C).

Library construction was performed using the TruSeq RNA Sample Prep Kit v2 according to protocol (Illumina Inc., San Diego, CA, USA), including poly-A mRNA isolation, fragmentation, and gel-based size selection. Shearing to about 250 bp fragments was achieved using the Covaris S2 focused-ultrasonicator (Covaris Inc., Woburn, MA, USA). Sequencing was performed according to the paired-end RNA-seq protocols from Illumina for Solexa sequencing on a Genome Analyzer IIx with paired-end module (Illumina Inc). Seventy-six bps were sequenced, from each side of a fragment of about 250 bp long.

Gene expression levels in FTCs were computed by using Cufflinks v1.1.0 [10] with the Illumina iGenomes Ensembl GRCh37 data set (2011–06-20) as reference, on reads aligned with TopHat v1.3.3 [11]. Gene expression levels were compared between the two wFTC (case 1 and 2) and two mFTC (case 3 and 4). Results were represented as fold change of gene expression in wFTC comparing with mFTC. Values of fold change in gene expression were at logarithmic scale (log2). Genes and transcripts were considered differentially expressed between mFTC and wFTC whenever Q value <0.05.

Frozen tissue samples available from thyroid tumours (n = 118) and adjacent/collateral normal thyroid tissues (n = 19) were crushed and homogenized. Total RNA from tissues and thyroid cancer derived cell lines was extracted using TRIzol® Reagent (Ambion®, Life Technologies™, CA, USA), according to the manufacturer’s protocol. Genomic DNA was extracted using the Genomic DNA Purification Kit (Citomed, Lisbon, Portugal) according to the manufacturer’s protocol.

cDNA was synthesized from 1 μg of RNA at 42 °C for 60 min, using oligo (dT) primers and M-MuLV reverse transcriptase (Fermentas, Thermo Scientific, St. Leon-Rot, Germany). The reverse-transcription PCR (RT-PCR) products were analysed by cDNA Sanger sequencing using the Big Dye terminator version 3.1 cycle (Applied Biosystems, Foster City, CA, USA). The samples were analysed in an automated sequencing machine (ABI Prism 3100 Genetic Analyser, Applied Biosystems, Foster City, CA, USA).

The 98 DTC were further characterized for selected molecular alterations known to be common in thyroid tumours. PCR followed by Sanger sequencing was used on genomic DNA for detecting mutations in the most frequent hotspot regions of BRAF (exon 15), NRAS (exon 2), and promoter region of TERT, according to previously described procedures [12, 13]. In our previous works we verified that HRAS and KRAS mutations were very rare (or absent) events. For that reason, we only screened mutations for NRAS gene.

The presence of PAX8-PPARG and RET/PTC rearrangements were determined by RT-PCR followed by Sanger sequencing of PCR products. Some positive cases were confirmed by fluorescence in situ hybridisation (FISH), following the procedures described previously [12].

Real-time quantitative PCR (qPCR) for C1QL1, LCN2, IL22RA1, MAMDC2, CILP, ASXL3, CRABP1, and SCUBE3 genes was performed in the two wFTC (cases 1 and 2) and two mFTC (cases 3 and 4) used in paired-end RNA-seq, and their corresponding normal thyroid tissue. qPCR for C1QL1, LCN2, CILP, and CRABP1 were also performed in the remaining series of DTC, in FTA and in normal thyroid tissues, as well as in thyroid cancer cell lines.

cDNA was amplified for the C1QL1, LCN2, IL22RA1, MAMDC2, CILP, ASXL3, CRABP1, and SCUBE3 by qPCR using SYBR® Green PCR Master Mix (Applied Biosystems, Foster City, CA, USA). A qPCR assay targeting Hypoxanthine phosphoribosyltransferase 1 (HPRT) was used as the housekeeping control. Primer sequences are available on request. Expression levels were obtained as the average cycle threshold (CT) of at least two replicas. A CT = 35 was assigned for target genes that were not expressed in samples with positive expression of the housekeeping control gene (HPRT).

The relative quantification of target genes was determined using the comparative CT method (2^-ΔΔCT) which was previously validated by Livak’s Linear Regression Method (Sequence Detector User Bulletin 2; Applied Biosystems) [14]. This provided the fold changes (2^-ΔΔCT) in gene expression normalized to an internal control gene (HPRT), and relative to one pool of normal thyroid tissues (calibrator). Expression values (2^-ΔΔCT) were normalized at logarithmic scale (log2). For C1QL1 and LCN2 genes, fold change [log2 (2^-ΔΔCT)] > 1 was considered as gain of expression, and fold change [log2 (2^-ΔΔCT)] ≤ 1 was established as normal gene expression. For CRABP1 and CILP genes, fold change [log2 (2^-ΔΔCT)] < −1 was considered as loss of expression, and fold change [log2 (2^-ΔΔCT)] ≥ −1 was established as normal gene expression. The diagnostic performance (sensitivity and specificity) of those cut-off values are presented in Table 2.

RNASeq (version v2) information were extracted from TCGA [15] for 393 DTC and 59 thyroid normal samples, for C1QL1, LCN2, CRABP1 and CILP genes. Gene expression values for each sample are normalized read counts, estimated by using RSEM software [16].

Statistical analysis was conducted with SPSS version 22.0 (SPSS Inc., Chicago, IL, USA). The results are expressed as percentage or mean ± SE. Statistical analysis was performed both on the whole DTC series and on the different subgroups: FTC, FVPTC and PTC. Receiver Operating Characteristics (ROC) curves for individual biomarkers were generated using log2 (2^-ΔΔCT) gene expression values and thyroid tissue type (DTC against normal or FTA) as input. For evaluation of the combined biomarker panel, the sum of log2 (2^-ΔΔCT) expression values from genes with gain (C1QL1 and LCN2) and loss (CRABP1 and CILP) in DTC were used. Spearman’s rho test (non-parametric) was used for evaluating the correlation of the expressions between different genes. Fisher’s exact test, t-test (unpaired, two-tailed), Mann–Whitney U test, and ANOVA were used when appropriate. The predictive value of C1QL1, LCN2, CRABP1 and CILP expression as a prognostic factor in thyroid cancer and their correlation with clinicopathological factors – age, tumour size, and extrathyroidal extension were assessed using univariate and multivariate logistic regression models. Test results with P-values <0.05 were considered statistically significant.

Differential expression analysis of the paired-end RNA-seq data identified 17 genes with significantly differential expression between mFTC and wFTC (Fig. 1, Additional file 2: Table S1 and Additional file 3: Table S2). Increased gene expression of KLK1, NEFL, KIAA1239, SLC6A17, NXPH4, LCN2, C1QL1 and TRIB3, and decreased gene expression of ST6GAL1, SCUBE3, IL22RA1, MAMDC2, CILP, CYSLTR2, ASXL3, CRABP1 and LINC00887 were observed in wFTC in comparison with mFTC. C1QL1, LCN2, IL22RA1, MAMDC2, CILP, ASXL3, CRABP1, and SCUBE3 genes were selected for validation through a customized filtering based on gene expression in thyroid tumours and normal thyroid tissues datasets [17‐19], assessed from Oncomine™ and in the literature available. Validation of expression levels was done in the four FTCs from paired-end RNA-seq and in their matched normal tissue by qPCR. Gain of C1QL1 and LCN2 expression, and loss of IL22RA1, MAMDC2, CILP, ASXL3, CRABP1, and SCUBE3 expression were confirmed in wFTC compared with mFTC and between the four FTC and their respective corresponding normal thyroid tissue (Additional file 4: figure S1). Based on the accuracy of gene expression levels obtained from paired-end RNA-seq and qPCR in the FTC and in the gene expression levels in their normal thyroid tissues by qPCR, C1QL1, LCN2, CRABP1 and CILP genes were selected to be tested as biomarkers in a well-characterized series of DTC (Table 1).

Additionally, we identified and experimentally verified a set of 21 fusion transcripts expressed by DTC (Additional file 5: figure S2; Additional file 6: Table S3; Additional file 7: Table S4). However, the fusion transcripts are not likely to be biomarkers because they were also detected in normal thyroid tissues. Further details can be found in the Additional file 8.

A series of 98 DTC (15 FTC, 23 FVPTC and 60 PTC) and a pool of 19 normal thyroid tissues and 20 FTA were used to evaluate the performance of C1QL1, LCN2, CRABP1, and CILP as biomarkers of DTC. Measurements of the expression levels of those genes were done by qPCR. C1QL1, LCN2, CRABP1 and CILP were differentially expressed in DTC in comparison with normal thyroid tissues (Additional file 9: figure S3), with statistical significance (P = 0.002, P = 0.005, P < 0.001 and P = 0.018, respectively). Notably, such significant differences were also observed for LCN2 and CRABP1 expression when comparing DTC against FTA (P < 0.001 and P = 0.002, respectively), and FTA against normal thyroid tissues (P = 0.013 and P = 0.022, respectively).

In order to test the aforementioned biomarkers as putative diagnostic tools for thyroid cancer, ROC curves were performed (Fig. 2, Additional file 10: Table S5). In the ROC analysis, comparing DTC versus normal thyroid (Fig. 2a), the areas under of the ROC curve (AUC) for C1QL1, LCN2, CRABP1 and CILP were 0.799 (P = 2.1E-3), 0.783 (P = 3.6E-3), 0.902 (P = 3.4E-5), and 0.687 (P = 5.5E-2), respectively. When these potential biomarkers were combined in a panel of gene expression, the AUC value was 0.927 (P = 1.1E-5) in DTC versus normal thyroid (Fig. 2c), providing a slight improvement over the AUC results obtained in each individual gene. Comparing DTC with FTA, the AUC values for C1QL1, LCN2, CRABP1 and CILP were 0.566 (P = 4.8E-1), 0.898 (P = 1.9E-5), 0.746 (P = 8.3E-3), and 0.675 (P = 6.1E-2), respectively (Fig. 2b). When these potential biomarkers were combined in a panel of gene expression, the AUC value was 0.839 (P = 2.7E-4) in DTC versus FTA (Fig. 2d), not improving the result in comparison with those of the AUC of LCN2 gene. CRABP1 was the best biomarker for the detection of DTC when tested against normal thyroid. On the other hand, LCN2 was the best biomarker for the detection of DTC when tested against FTA.

We also tested gene expression levels from an additional series of 393 DTC and 59 normal thyroid tissues available in TCGA [15]. C1QL1, LCN2 and CRABP1 were differentially expressed in DTC in comparison with normal thyroid tissues (Additional file 11: figure S4), with statistical significance (P < 0.001).

C1QL1 expression levels were successful measured by qPCR in 89 DTC (15 FTC, 20 FVPTC and 54 PTC), 10 thyroid cancer cell lines, 17 FTA, and 18 matched thyroid normal tissues (Fig. 3a). Gain of C1QL1 expression in relation to thyroid normal tissues was found in FTC, FVPTC and PTC (Fig. 3b), but the differences only achieved the threshold of statistical significance in FVPTC (P = 0.015) and PTC (P = 0.001). Additionally, eight out of ten (80%) thyroid cancer cell lines showed gain of C1QL1 expression (Fig. 3c). Gain of C1QL1 expression was significantly associated with extrathyroidal extension (P = 0.003) and lymphocytic thyroiditis (P = 0.003) in the DTC samples (Table 3). All the cases with BRAF mutations present gain of C1QL1 expression (P < 0.001). In FTC, gain of C1QL1 expression was present in 71% of wFTC and absent in mFTC (P = 0.007; Additional file 12: Table S6), and lymphocytic thyroiditis was only present in FTC with gain of C1QL1 expression (P = 0.026). In FVPTC, absence of NRAS mutations was significantly associated to the gain of C1QL1 expression (P = 0.014; Additional file 13: Table S7). PTC with C1QL1 gain were significantly larger (2.96 ± 0.33 cm) than PTC without gain of expression (1.78 ± 0.25 cm; P = 0.021; Additional file 14: Table S8). Gain of C1QL1 expression in PTC was also significantly associated with the presence of extrathyroidal extension (P = 0.006), and BRAF mutations (P = 0.001).

LCN2 expression levels were successful measured by qPCR in 86 DTC (14 FTC, 19 FVPTC and 53 PTC), 10 thyroid cancer cell lines, 18 FTA, and 16 matched thyroid normal tissues (Fig. 4a). Gain of LCN2 expression was found in FTC, FVPTC and PTC (Fig. 4b), but the differences only achieved the threshold of statistical significance in PTC (P < 0.001). Three out of ten (30%) of the thyroid cancer cell lines showed gain of LCN2 (Fig. 4c). At variance with DTC, significant loss of LCN2 expression was found in FTA (P = 0.013; Fig. 4b). Gain of LCN2 expression in DTC samples was significantly associated with the presence of extrathyroidal extension (P = 0.020), oncocytic features (P = 0.018), and the presence of BRAF mutations (P = 0.031; Table 3). Oncocytic pattern of the FTC was significantly associated with the gain of LCN2 expression (P = 0.016; Additional file 12: Table S6). PTC with gain of LCN2 expression were significantly larger (2.72 ± 0.30 cm) than PTC without gain of expression (1.76 ± 0.34 cm; P = 0.023; Additional file 14: Table S8). Although not statistically significant (P = 0.130), extrathyroidal extension was more frequent in PTC with gain (54%) than without gain (31%) of LCN2 expression. No significant associations that could be related with gain of LCN2 expression were observed in FVPTC.

CRABP1 expression levels were successful measured by qPCR in 89 DTC (15 FTC, 19 FVPTC and 55 PTC), 10 thyroid cancer cell lines, 14 FTA, and 19 matched thyroid normal tissues (Fig. 5a). Significant loss of CRABP1 expression was detected in FTC, FVPTC and PTC (P < 0.001), and in FTA (P = 0.023; Fig. 5b). Notably, all the 10 thyroid cancer cell lines tested showed loss of CRABP1 expression (Fig. 5c). Loss of CRABP1 expression was associated with encapsulated DTC (P = 0.037). BRAF mutations (P = 0.025; Table 3) were only present in DTC with loss of CRABP1 expression. No significant associations were found in FTC, FVPTC and PTC with loss of CRABP1 expression.

CILP expression levels were successfully measured by qPCR in 88 DTC (14 FTC, 20 FVPTC and 54 PTC), 10 thyroid cancer cell lines, 17 FTA, and twelve matched thyroid normal tissues (Fig. 6a). Loss of CILP was detected in FTC, FVPTC and PTC (Fig. 6b), but differences only attained the threshold of statistical significance in PTC (P = 0.013). All the ten thyroid cancer cell lines tested showed loss of CILP expression (Fig. 6c). NRAS mutations (P = 0.001; Table 3) were only present in DTC with loss of CILP expression. The same association was also found in FVPTC with loss of CILP expression (P = 0.043). No significant associations were found in FTC and PTC with loss of CILP expression.

A regression model was performed with C1QL1, LCN2, CRABP1 and CILP expression values for thyroid cancer prognostic factors: age of patients (> 45 years), tumour size (> 4 cm), and presence of extrathyroidal extension of the tumour (Table 4). Distant metastasis as prognostic factor was not considered for the computation in the regression models due to the reduced number of thyroid cancers with distant metastasis in the present series.

In the univariate analysis, gain of the C1QL1 [odds ratio (OR) = 5.34; P = 0.006] and LCN2 (OR = 3.48; P = 0.029) expression were significantly associated with the extrathyroidal extension of the tumour. In the multivariate model, gain of C1QL1 (OR = 4.86; P = 0.011) and LCN2 (OR = 3.39; P = 0.039) expression were independent predictive factors for extrathyroidal extension in DTC. Additionally, gain of C1QL1 expression was observed in the larger (> 4 cm) DTC (OR = 4.60), but the difference was only borderline in terms of statistical significance (P = 0.056). No associations were found between CRABP1 and CILP expression and prognostic factors of thyroid cancer.