Pre-selection of candidates for cancer association studies

Association with tumour development

Our list of cancer-associated variants contains a number of genes possibly involved in the cellular capabilities that might be acquired by cancer cells [28], e.g., translationally controlled tumour protein TCTP, IL-4-R, HLA class II antigens, TIMP-3, CD147, CD44, and the jun-B, c-fos, AF4, Ki-Ras and RAF proto-oncogenes. Also included in our list are 38 novel sequences, i.e., entries for which no annotation was available at the time of the study (these transcripts are referred to as 'NULL' in Additional file 1). In particular, we identified a ~800 bp- long nucleotide sequence located in the 5'-UTR of ENST00000285718, which contained as many as ten cancer-associated UTR-SNPs. The corresponding gene (encoding a putative proline-rich protein) has been mapped to 2q13, a region defined as a tumor amplicon [29]. Furthermore, out of the 269 RNAs with UTR-SNPs, the screen returned 22 hits previously identified as bearing cancer-associated non-synonymous coding SNPs (nsSNPs) on the basis of a similar computer-based screen [7]. Among these transcripts exhibiting both cancer-associated nsSNPs and cancer-associated UTR-SNPs (highlighted in light grey in Additional file 1) are those encoding Heat shock cognate 71 kDa protein, polyadenylate binding protein (PABP)-3, translationally controlled tumour protein (TCTP), immunoglobulin gamma FcRIIIA, and dynein light chain 1 (DNCL1, see Table 2).

"Hot spots" for base substitutions were found for some transcripts, either as consecutive SNPs (e.g., 1286 c→a and 1287 t→c for ENST00000234617) or as 'nests' of SNPs (e.g., 991 g→t, 999t→c and 1005 c→t for ENST00000285718). However, most transcripts (~75%) displayed a unique cancer-associated SNP. We found a variant causing a g→c change at nucleotide 175 in the 5'-UTR of RhoH, a gene prone to aberrant hypermutation activity in lymphomas [30]. Interestingly, determination of the origin of the EST libraries revealed that this UTR-SNP was specific to lymphoid tissues. In addition to the previously reported 4 c→a and 956 t→c alterations in the 5'- and 3'-UTRs of Kruppel-like factor 6 (KLF6), an important DNA-binding transcriptional regulator [31], our analysis also revealed a 1206 c→t polymorphism in the 5'-UTR of this gene. Owing to the high mutation frequency of KLF6 in a number of pituitary tumors [32], knowledge of these KLF6 polymorphisms may be important for prostate cancer diagnosis.

Last, we found among the hits a series of UTR-SNPs concerning the SPARC gene, which encodes a multifunctional glycoprotein playing roles in tissue development, remodelling and fibrosis [24–27]. As a regulator of cell-extracellular matrix (ECM) interactions, SPARC is thought to represent a major factor in the ECM remodelling occurring during tumour invasion. Our in silico analysis revealed 4 UTR-SNPs located in the 3'-UTR of the SPARC gene, corresponding to 1474 g→a, 1551 g→c, 1922 t→g and 2072 c→t changes, which were significantly associated with the tumoral state. Noteworthy, of all the 'digital' hits, the 2072 SPARC polymorphism had the clearest association with cancer (see Table 2 and Additional file 1). This SNP is localized in a 44 bp- long conserved sequence between rodents and primates, suggesting that it might belong to a functionally constrained region.

Detection of SPARC variants in tumour samples

Patterns of substitution

In addition to a gene-centric view, SNPs can be characterized by type of nucleotide change and putative functional effect. The objective of this section was to examine the substitution patterns among the cancer-associated UTR-SNPs identified by our computer-based procedure.

We explored the distribution of the various types of simple substitution SNPs in the different sets of candidate UTR-SNPs, i.e. the complete dataset of UTR-SNPs (n = 20,304), the total pool of cancer-associated UTR-SNPs (n = 358), and the subset of UTR-SNPs which were positive after the resampling procedure (n = 92) and that are less likely to correspond to false positives. The transition rates were around 70 % and the transversion rates were ~30% in the different categories, in accordance with previous genome-wide estimates [33, 34]. In all cases, the most common substitution was C→T (see Additional file 3 and Additional file 4 for a graphical representation); however, this type of change was 1.5 times less frequent in the pool of UTR-SNPs positive after the resampling procedure as in the total dataset (18.5 % versus 27.8 %, respectively). At the same time, the T→C transition accounted for 16.3 % of all single nucleotide substitutions within this pool versus 9.6 % within the total dataset. The couple of complementary substitutions A↔T followed a similar distribution in the total and cancer-associated datasets. Similarly, G→T and A→C frequencies were of similar magnitude in the three datasets; however, one can see that the frequencies for the complementary substitutions T→G and C→A behave in opposite manner: T→G substitutions were over-represented in the pool of UTR-SNPs positive after the resampling procedure (8.7 % versus 4.2 % in the total sampling) whereas only ~2 % of UTR-SNPs were of type C→A in the cancer-associated datasets (versus 4.9 % in the total pool of UTR-SNPs). Last, while the global frequencies of C↔G did not differ significantly between the different datasets (see Additional file 4, panel A), when the SNPs are reported respective of the direction of change, the frequencies of the pairs C→G and G→C showed a pattern reversal in the pool of UTR-SNPs positive after the resampling procedure compared to the total dataset (1.1 % versus 6.3 % for C→G, and 6.5 % versus 2.6 % for G→C, respectively). Together, these results show that the ratios of several types of substitutions differ between the entire dataset of UTR-SNPs and cancer-associated alleles.

Possible impact of cancer-associated UTR-SNPs on mRNA secondary structure and UTR regulatory elements

Although many of the UTR-SNPs identified in our experiment are not expected to be functional, but rather to act as markers for functional variants yet to be discovered elsewhere in the gene or even possibly in a nearby gene, it is possible that at least a fraction represent functional SNPs. Therefore, we decided to assess the putative structural and functional consequences of the tumor-associated UTR-SNPs on mRNA metabolism (mRNA secondary structure and putative regulatory sites).

Altogether, these results provide evidence that at least a subset of cancer-associated SNPs might have functional consequences on mRNA stability and/or expression.

Data preparation

We have used an EST-based pipeline to scan for UTR-located polymorphisms associated with libraries of cancerous origin. Human ESTs from dbEST [70] (October 2004 release) were first extracted using the ACNUC sequence retrieval system [71]. ESTs were classified according to their UNIGENE library features [72] as previously described [6]. The eVOC ontology [73] (October 2004) for anatomical sites and pathology types was then used to classify the libraries through a number of criteria such as tissue origin and pathological context including tumor state. This well-accepted hierarchical vocabulary provided us with a mean to determine when a specific tissue was part of an organ and when a specific label was part of the 'tumoral' state. A total of 5135 'tumor' and 2503 'normal' (i.e. non-pathological) libraries were catalogued. Our approach to EST clustering used the human genome as a reliable guide. ENSEMBL RNAs [74] annotated on human genome assembly (release 16.3) were used as a backbone for the clustering of dbEST sequences using MEGABLAST (alignment length ≥ 100 bp and similarity ≥ 95%) [75]. In order to avoid paralogous false positive assignation, only best EST hit matches were subsequently selected. RNA clustering of ESTs in both normal and tumoral tissues was the starting point for in silico mining of UTR-SNPs associated with tumoral phenotype.

SNP detection

We have developed an algorithm to identify exonic SNPs in multiple alignments of various ESTs associated to a particular annotated transcript. This algorithm takes advantage of EST library redundancy and performs four filters to reduce the effect of sequencing inaccuracies at each position. The first filter required that each position within a multiple alignment of ESTs should have an exact match with the reference RNA (windows length = 10 bp around each variant position). The second filter considered a position as informative if the number of libraries in the multiple alignment was superior to a fixed minimum threshold (library number ≥ 5). The third filter of the algorithm required the variant to be found at least two times independently i.e., in two different libraries. A last independent filter that required a minimum of two variant ESTs in one of the libraries was subsequently added in order to increase further the stringency of the cancer association mining strategy. We then combined detection method information (library and EST depth coverage) and nucleotide substitution features (e.g., transition/transversion, position in 5'- or 3'-UTR) for the UTR-SNPs that have been filtered out. Statistical analysis was performed using R [76]. Genomic data were stored in a local PostgreSQL database (GeMCore) [77] using PERL and Java script.

Cancer association

Finally for each informative SNP that has both normal and tumoral EST coverage, an exact Fisher's test was performed in order to statistically evaluate the association of a particular variant with the tumoral state. We privileged the counting of ESTs rather than a count per library because of the frequent lack of precision concerning the origin of the source tissues and the use of pooled samples. To adjust p-values produced by the Fisher's exact test, three approaches were used: (a) Bonferroni, and (b) Benjamini and Hochberg corrections, which are very conservative methods for controlling the false discovery rate, and (c) a resampling procedure. The standard Bonferroni correction multiplies the uncorrected p-value by the number of statistical tests. The Benjamini and Hochberg correction consists of ranking all p-values and adjusting each by multiplying by the total number of tests and dividing by the rank of that p-value. The resampling procedure simulates the distribution of the minimum p-value that we would expect if there was no association with cancer. To do this, reference and variant margins were fixed at each SNP; Fisher's exact test was then performed for 1,000 resampled datasets, and the smallest p-value was recorded. This resampling procedure was repeated for n = 10,514 SNPs, from which an empirical distribution of the minimum p-value was obtained. From this distribution, we estimated the p-value that corresponded to the conventional 5% threshold. The intensity of the bias of tumoral versus normal allele frequency was calculated according to the following formula:

Ib = (a/V) - [(T - a)/R], where 'a' is the number of tumoral variants, 'V' the total number of variants, 'T' the sum of tumoral counts (variant plus reference) and R the total number of reference alleles (Ib being close to 1 in case of strong association).

In silico characterization of UTR-located SNPs

UTRScan [78] was used to identify putative cis acting elements patterns in the regions containing cancer-associated SNPs. UTRScan looks for UTR functional elements by searching through user submitted query sequences for the patterns defined in the UTRsite collection [44]. To test the potential effects of the detected polymorphisms on mRNA folding, we took advantage of the RNAMute application [41].

In vitro detection of SPARC variants

Genomic DNA was extracted from bone marrow in patients with acute myeloblastic leukemia (AML) and peripheral blood in healthy donors. The phenol/chloroform method was used for DNA extraction according to standard procedures. Primers to explore SPARC polymorphisms detected by computational procedures were designed based on the DNA sequence from GenBank (entry: BC072457). The sequences of the primers are listed in Additional file 2. Amplification was carried out in an iCycler Thermal Cycler (Bio-Rad, Hercules, CA, USA): 1 μg of DNA was amplified in a 25 μl of final reaction volume containing 10 × buffer II, 2 mM MgCl2, 0.2 μl of 25 mM dNTP's mixture, 0.4 μl of 20 pmol/μl forward and reverse primers and 0.2 μl of Fast-Taq polymerase (5 u/μl) (Roche Diagnostics, Indianapolis, IN, USA). PCR procedure consisted of 35 cycles of denaturation at 94°C for 15 sec, annealing at 60°C for 60 sec, with an initial denaturation of the DNA at 94°C for 5 min before PCR, and a final extension at 72°C for 5 minutes. The PCR products were sequenced in an ABI PRISM 3100 Genetic Analyzer (Applied Biosystems, Foster City, CA, USA).