Cancers of unknown primary origin (CUP) are characterized by chromosomal instability (CIN) compared to metastasis of know origin

Sixty eight consecutive CUP patients were enrolled in the study, but since eleven samples did not meet the quality criteria the number of CUP samples ended at 57. The histological features of the 57 CUP that underwent expression profiling are summarized in Table 1. During the diagnostic work-up, a possible primary tumor site was eventually identified in 28 of the 57 patients (Additional file 2: Figure S1 and Table 1). Among these 18 samples were in accordance with diagnostic work-up or the Standard of Reference.

To examine if CUP exhibit particular genetic and phenotypic characteristics compared to metastases of known origin, that could warrant particular diagnostic procedures and treatment, we first generated a transcriptome-based signature that could classify 16 common tumor classes and predict the origins of CUP and metastases of known origin with high accuracy (Detailed in Additional file 1: Table S1). To allow detection of samples without sufficient tumor tissue, a group of normal tissues was also included. Since all CUP data were generated at our facility, we moreover examined a series of primary cancers and metastases from Rigshospitalet to exclude possible site- and batch-specific effects. The cross-validation accuracy during training of a 428 probe sets classifier was 92.2% (Additional file 3: Table S2) and the overall accuracy in the validation set was 90% and 83% for primary tumors and known metastases, respectively (Additional file 3: Table S2). The distribution of variables among the 16 tumor categories is depicted in the heat map (Figure 1). Since we suspected that the low accuracy in some of the classes, e.g. cholangiocarcinoma, was associated with the small number of samples in the training set, and because CUPs were supposed to be compared to metastases of known origin, we subsequently combined the training and validation sets and generated a second classifier, consisting of 641 probe sets (641 classifier). Furthermore a gender correction by renormalizing the prior class probabilities in the test situation was implemented because we noted that tumors from males incorrectly were classified as ovary, cervical and endometrial cancer. The accuracy in primary tumors, known metastases and normal samples of the 641 classifier was 92%, 87% and 89%, respectively (Additional file 3: Table S2) and this classifier was subsequently used for the prediction of CUP. The principal component analysis is shown in Figure 1B and the ten most selective transcripts and their gene ontology for each tumor class are listed in Additional file 4: Figure S2.

To provide a systematic overview of the expression of conventional tumor markers in the CUP samples, we also compiled a list of 45 common histopathological biomarkers and depicted their expression in a two-way hierarchal cluster (Figure 2). Whereas, about 85% of the primary cancers exhibited a characteristic expression of their individual histomarkers, only 10 of the 28 (35%) CUP - where a putative primary site was identified and 3 of the 29 (10%) CUP - where the primary site remained unknown - expressed one or more biomarkers at significant levels. The strongest overlap between histopathological markers and the LDA based CUP classifications was observed for CUP predicited as ovary and colorectal cancers, where 4 and 3 samples expressed WT1 or CEA/CEACAM5, respectively. Moreover, 6 samples were positive for TP63 and 2 samples were positive for surfactant proteins. Finally, one sample was positive for PAX2 in agreement with the LDA prediction as renal carcinoma. Compared to the primary cancers there was a limited concordance between markers within the same tumor category. Only two of the WT1 positive cancers were positive for CA125/MUC16, and only 3 of the TP63 positive samples expressed CK17 and CK5, characteristic of squamous carcinoma. If the histological markers were combined and used in an LDA based fashion, the concordance with the 641 signature LDA predictions or Standard of Reference was about 66% indicating that systematic application of the patomarkers may at least to some extent compensate for the modest predictive power of individual markers.

To determine the similarity between primary cancers, metastases of known origin and CUPs, we employed Quadratic Discriminant Analysis (QDA) to determine the likelihood that a particular sample belonged to one of the predefined tumor classes. Outlier scores were calculated in LOOCV fashion for one sample at a time using all remaining samples i.e. primary tumors and metastases to represent the classes. The outlier scores of the samples from normal tissues are not comparable to the primary tumors and metastases because of the heterogeneity among the many different tissues in the class.

Based on the results from primary tumors and metastases we plotted the predictive error rates versus the outlier scores and demonstrated a clear relationship between errors and outlier scores (Figure 3). Samples with outlier scores below 800 exhibited less than 10% risk of being erroneous, whereas, outlier scores above 1000 had more than 25% risk of being incorrect. However, even in the high end of outlier scores with only 75% accuracy, prediction is far from random, since we are working with 16 different classes. As shown in the box plot (Figure 3), CUP samples had significantly higher outlier values than primary tumors and metastases. To ensure that the difference was not related to our platform, we compared our own samples of known metastases and primary tumors and observed the same difference. CUPs, moreover, consisted of biopsies that may contain more normal tissue than samples obtained during surgery. We therefore plotted the percentage of normal tissue as estimated from the relative expression of markers of lymphoid, liver, and muscle tissue versus the outlier scores, but observed no correlation between the amount of normal tissue in the biopsies and the outlier scores (Additional file 5: Figure S3). A number of samples that expressed conventional histopathological biomarkers exhibited low scores, but if we compared CUPs where a primary cancer was identified during the clinical processing with CUPs where no primary site could be identified, there was no difference between the outlier scores (mean 991 vs mean 1031, P = 0.24). Taken together, the results demonstrate that CUPs are more distantly related to the predefined tumor classes, than known metastases.

To identify differentially expressed transcripts, we performed a class comparison between CUP and metastases of known origin. The analysis was performed as a paired analysis with respect to the LDA predictions to eliminate differences between tumor classes. Metastases from uterine, testis, prostate, melanoma and thyroid cancers were excluded from the analysis because no CUPs had been allocated to these groups. CUPs predicted as normal tissue were also excluded. Moreover, cholangiocarcinomas were omitted from the calculations because they were not represented in the LDA predicted metastases group. In total 41 CUP and 186 metastases comprising 10 different cancer groups were included in the analysis. To define the most up- and down-regulated CUP transcripts, a cut-off of p < 10⁻⁸ corresponding to a false discovery rate of q < 1.9*10⁻⁷ was employed. This resulted in 1550 down- and 1390 up-regulated probe sets corresponding to 1117 and 934 unique annotated genes, respectively. These two lists comprised our CUP core set of differentially expressed transcripts. The 40 most significantly down- or up-regulated mRNAs are shown in Additional file 6: Table S4. The lists of genes was subsequently subjected to a Gene Set Enrichment Analysis (GSEA) using the Broad Institute’s GSEA database (http://​www.​broadinstitute.​org/​gsea/​msigdb). Initially, we searched for enriched gene ontology terms, and this revealed that up-regulated transcripts were associated with GO-terms (q < 0.01): DNA_INTEGRITY_CHECKPOINT,DNA_DAMAGE_CHECKPOINT,DNA_REPLICATION_INITIATION,DNA_PACKAGING,NEGATIVE_REGULATION_OF_DNA_METABOLIC_PROCESS,CELL_CYCLE_CHECKPOINT;NEGATIVE_REGULATION_OF_DNA_REPLICATION,CHROMATIN_REMODELING,DNA_DAMAGE_RESPONSESIGNAL_TRANSDUCTION. There were no particular enrichments among the down-regulated mRNAs.

To depict CUP enriched molecular pathways, we further examined if the CUP core set exhibited overlaps with the Molecular Signature Database (MSigDB) curated gene sets. Overlaps between the CUP core set (p < 10⁻⁸) were computed by submission of up- and down-regulated probe sets separately (Table 2). Gene sets consisting of transcripts that were positively correlated to BRCA1, ATM and CHECK2 expression were highly enriched in the up-regulated CUP core set. The down-regulated CUP mRNAs showed fewer significant overlaps but SHEN_SMARCA2_TARGETS_DN gene set, which depict transcripts that are negatively correlated with SMARCA2 expression in prostate cancer was clearly overlapping with the CUP set.

To examine the BRCA1 and SMARCA2 pathway networks defined by the SHEN_SMARCA2_TARGET_DN, SHEN_SMARCA2_TARGET_UP and PUJANA_BRCA1_PCC_NETWORK in greater detail, we generated two way clusters using the complete gene sets on our CUP core set (Figure 4). The clusters were based on a paired analysis with respect to their LDA predictions and with the same inclusion criteria, as described above. The SHEN_SMARCA2_TARGET_DN; SHEN_SMARCA2_TARGET_UP and PUJANA_BRCA1_PCC_NETWORK gene symbols were translated into probe sets and to exclude non-functional redundant probe sets, only the probe sets with the 50% highest variance were included. We moreover applied a p-value cut-off of 0.001 to filter probe sets that differed among the two groups (Figure 4). The PUJANA_BRCA1_PCC_NETWORK set of genes consists of 1671 gene symbols that translated into 3897 probe sets. Following filtering 705 probe sets corresponding to 519 up-regulated and 66 down-regulated genes were clustered (Figure 4). From the cluster it is apparent that the BRCA1 profile is strongly enriched in CUP compared to the corresponding metastases. A schematic representation of the BRCA1 and non-homologous repair networks showing the enriched factors is depicted in Additional file 7: Figure S4. Following the same procedure, we subsequently looked at the SMARCA2 networking (Figure 4). The SHEN sets consist of 360 SMARCA2 negatively- and 430 SMARCA2 positively- correlated genes that translated into 772 and 1211 probe sets respectively. In the SMARC2A negatively correlated group, we observed 20 genes that were up-regulated and 95 that were down-regulated in CUP compared to metastases, and amongst the SMARCA2 positive correlated genes we saw 161 up-regulated genes and 19 down regulated after filtering (top 50% variance probes and p < 0.001). Taken together, the GSEA shows that CUPs are characterized by enrichment of the double strand break DNA repair system and the SMARCA2/BRM chromatin dependent remodeling system.

Since the observed enrichment of genes involved in DNA double-strand break repair (Additional file 7: Figure S4) indicated that CUPs were more chromosome unstable than known metastasis and primary cancers, we examined the status of signatures involved in DNA repair and genome instability. Signatures of chromosomal instability (CIN), DNA double-strand break repair, nucleotide excision repair (NER), base excision repair (BER) and mismatch repair (MMR) were included to obtain a complete overview of DNA- repair and stability in CUP (Figures 5 and 6). The predefined gene sets were examined with the Broad Institute GSEA v2 software. The expression data matrix was preprocessed in Qlucore Omics Explorer™ and expression values were normalized within LDA predictions - so the expression values became expressed as a relative value compared to the mean expression of a gene within its group. The data set was analyzed against the 10 selected gene sets (Figure 5) employing 1000 permutations with standard GSEA settings. The most significant scores were observed for the signature of double strand break repair and for signatures of unstable sarcoma [9] and CIN [10]. Moreover, the KEGG signature of NER was enriched but not to a significant level (p = 0.123). The remaining nucleotide excision and mismatch repair signatures were not enriched in CUP and we infer that CUPs primarily distinguishes themselves from metastasis of known origin by signatures of chromosome instability. The signature of chromosome unstable sarcoma was finally employed to generate an instability score providing an index of the chromosomal instability for comparison of normal tissue, primary cancers and metastasis and CUP (Figure 6). The instability score was calculated as the mean of the expression values from the included probe sets of the signature following variance filtering (206 probe sets). As shown in Figure 6 panel B CUP exhibited a significantly higher score than paired metastasis of known origin. Metastases were significantly more chromosomal unstable than primary cancers and as expected normal tissues exhibited very low instability scores. We moreover examined the correlation between outlier scores and instability scores, because chromosome instability is envisioned to promote evolution and phenotypic variations. In a linear correlation analysis including primary cancers, metastasis and CUPs the outlier scores were positively correlated with the chromosomal instability score (p < 7.24E-33; q < 2.16E-30), so we infer that chromosomal instability is likely to be implicated in the phenotypic traits of CUPs.