Samples

DNA was extracted from 54 CRC cell lines (C10, C32, C70, C75, C80, C99, C106, C125PM, C170, CACO2, CL11, COLO201, COLO205, COLO320, COLO678, COLO741, CX1, GP5D, H508, H716, HCA7, HCA46, HCT8, HCT116, HCT15/DLD1, HRA19, HT29, HT55, HT115, HUTU80, LOVO, LIM1863, LS1034, LS123, LS174T, LS411, MBDA8, NCI-H747, PC/JW, RKO, SKCO1, SNUC2B, SW1116, SW1222, SW1417, SW403, SW48, SW620, SW837, SW948, T84, VACO4S, VACO5, VACO10MS).

Research was undertaken in compliance with the Helsinki Declaration and with full ethical approval (Oxfordshire REC B, 05/Q1605/66). 268 anonymised, formalin-fixed, paraffin-embedded (FFPE) samples of colorectal adenomas from thirteen FAP patients with FAP were identified. In order to increase the frequency of tumours showing LOH, we selected cases with germline mutations around codon 1300. We also used larger lesions (> 0.5 cm diameter) to minimise problems of polyclonality in smaller colorectal adenomas [8]. For 10 of the larger polyps (diameter ≤ 0.9 cm), more than one sample was obtained from different parts of the same polyp in order to check for consistency of the molecular data, although in each case, complete concordance was observed. After enrichment for dysplastic epithelium using a fine gauge needle and the dissecting microscope, DNA was extracted from each adenoma using a standard proteinase K digestion and the Qiagen DNeasy kit, except that elution with water was undertaken twice to increase the yield. Normal tissue from the same block was extracted using the same method.

For analysis of associations between mitotic recombination breakpoints and colorectal cancer risk, two case-control series were analysed: (i) cases and controls from the UK CORGI study of familial colorectal tumour patients [9]; and (ii) cases from the VICTOR http://www.octo-oxford.org.uk/alltrials/infollowup/vic.html and QUASAR 2 http://www.octo-oxford.org.uk/alltrials/trials/q2.html clinical trials and controls from the UK 1958 Birth Cohort http://www.b58cgene.sgul.ac.uk/. All cases and controls, comprising about 4,000 samples in total, were of white UK ethnic origin. Further details of ascertainment, inclusion criteria and exclusion criteria can be found in [9]. Genotyping data were derived from the Illumina Hap300, Hap370 or Hap550 arrays (see below). Each series was analysed separately for significant allele or genotype frequency differences between cases and controls, followed by a weighted meta-analysis using the Mantel-Haenszel method in STATA9.0.

SNP microarray analyses

Patient and colorectal cancer cell line DNAs were prepared and hybridised to arrays (Affymetrix 10 K HuSNP and Illumina Hap300, Hap370 and Hap550) using the manufacturer's standard protocols. Genotyping calls were made using the manufacturer's software, resulting in call rates of over 98% and sample failure rates of < 2% using good quality DNA. LOH and copy number changes were scored using the manufacturer's software in each case, supplemented by visual inspection of allele frequency plots along each chromosome.

For the custom Goldengate arrays, designed to assess LOH on 5q in FFPE tumour samples, the Illumina custom probe design software was used to test the suitability of all SNPs from dbSNP126 that mapped between the chromosome 5 centromere and the APC locus (~112 Mb). After elimination of SNPs with low design scores, a panel of 360 remained. An additional 11 SNPs distal to the APC gene were chosen to provide evidence that LOH extended as far as the telomere; 7 were located on chromosome 5p in order to give evidence of whole-chromosome LOH; and 10 SNPs mapped elsewhere in the genome to give evidence of copy number changes. 200–300 ng of each DNA was hybridised to the arrays using the manufacturer's standard protocols. After excluding SNPs with Gentrain scores of less than 0.3, the Illumina Beadstudio software was used to indicate LOH and copy number change on 5q, either in unpaired sample mode (for CRC cell lines) or, after training on FFPE samples, in paired mode for the colorectal adenomas. We found, however, that visual inspection of allelic binning and manual corrections of clear errors by two or more independent observers (SR, KH, IT) improved data quality and allowed breakpoints to be mapped more closely.

Microsatellite and analysis

Chromosome 5q microsatellite markers (D5S623, D5S664, D5S407, D5S398, D5S2107, D5S624, D5S1990, D5S2089, D5S647, D5S2019, D5S2003, D5S2041, D5S2029, D5S107, D5S644, D5S669, D5S346) and in-del polymorphisms (rs2067135, rs1305058, rs3087334, rs2307799, rs1610940) were chosen according to their location from the human genome March 2006 build http://genome.ucsc.edu. PCR genotyping used a single dye-labelled primer, the ABI 3730 sequencer and the Genescan/Genotyper software. For analysis of CRC cell lines, absence of one allele was required to score possible LOH. The fact that paired constitutional DNA was not available in most cases meant that this approach was more useful for indicating definite retention of homozygosity than mapping LOH.

Pyrosequencing

According to the manufacturer's standard protocols, a Pyrosequencing assay was designed to discriminate between the SMN1 and SMN2 genes using a C>T substitution [10] corresponding to SNP rs4916 (location chr5:69,408,109 in the March 2006 Human Genome Build). Briefly, primers were designed using the proprietary Pyrosequencing assay design software, giving: (i) biotinylated forward PCR primer, TCCTTTATTTTCCTTACAGGGTTT; (ii) reverse PCR primer, ATGCTGGCAGACTTACTCCTTAAT; and sequencing primer, TCCTTCTTTTTGATTTTGT. Allelic intensities were derived from the standard Pyrosequencing software and used to derive genotypes (see below).

Mitotic recombination breakpoint mapping in colorectal cancer cell lines

We initially screened 49 colorectal cancer cell lines for copy-neutral LOH at APC using the Affymetrix 10 K HuSNP array, alongside array-CGH analysis using a 1 Mb density BAC-based array. Of the 14 cases that had undergone copy-neutral LOH involving APC, 10 showed break-induced replication/mitotic recombination events, extending from a location between the centromere and the APC locus on chromosome 5q to the long-arm telomere. In 4 of these 10 lines (C10, VACO4S, CACO2 and C70), the LOH breakpoint was probably within or close to a small region, 67.5–72.0 Mb, the site of a complex low copy number repeat (LCR). Further genotyping using highly polymorphic microsatellite markers and in/del polymorphisms confirmed the locations of these breakpoints and showed that they were likely to lie within or just distal to the LCR (see Additional File 1). However, poor coverage of polymorphic markers within the LCR prevented finer scale mapping.

Mitotic breakpoint mapping using a custom SNP array in colorectal tumours

The preliminary observation from the colorectal cancer cell lines that there might be a non-random distribution of mitotic recombination breakpoints between the chromosome 5 centromere (~50 Mb) and APC prompted us to examine this phenomenon in a larger set of tumours. In order to enrich for tumours with LOH [2] and to reduce the frequency of confounding background events – such as copy number changes related to chromosomal instability in late lesions – we analysed early colorectal adenomas from FAP patients who harboured germline mutations close to codon 1300 of APC. Since these lesions were generally available as FFPE specimens, we set up a custom SNP array (see Additional File 2) using the Illumina Goldengate platform that has previously proved suitable for LOH analysis [11].

After verification using the colorectal cancer cell lines and having excluded a small number of samples with 5q copy number changes, 75/268 adenomas were both successfully analysed and showed LOH involving APC (Figure 1a). Apart from one loss of the whole chromosome, all LOH events began at a breakpoint between the centromere and APC and extended as far as juxta-telomeric SNPs, consistent with a mechanism of mitotic recombination through break-induced replication (Figure 1b). No copy-neutral LOH involving an interstitial region of the chromosome arm was detected. Moving along the chromosome arm from the centromere, we mapped each breakpoint to a region between the last constitutionally heterozygous SNP showing no LOH to the first heterozygous SNP with unequivocal LOH. In some cases, localisation to a few kb was possible, but in other cases, such as regions where SNP density was low owing to repeated sequences, where SNPs were poorly informative, or where some SNPs had ambiguous or discordant allelic calls, breakpoints could only be mapped to a few megabases. Breakpoint positions were therefore placed into 'bins' of 5 Mb regions.

Examination of the locations of the mitotic recombination breakpoints (Figure 2a) showed them to be non-random compared with the expected flat distribution (χ² = 33.8, 11 df, p < 0.001). There was a maximum frequency of breakpoints between 65 Mb and 75 Mb, with an elevated frequency extending as far as 80 Mb. There was actually a deficiency of breakpoints close to the APC locus. These findings were consistent with our colorectal cancer cell line data. Owing to poor SNP density within the 68–71 Mb LCR itself – because of the presence of multiple repeated sequences and the inability to distinguish true SNPs from variants occurring in paralogous sequences in cis – we were unable to fine-map our mitotic recombination breakpoints to determine whether they lay within the LCR or just outside it.

Comparison between the mitotic and meiotic recombination maps

We compared the locations of the mitotic recombination breakpoints with the meiotic recombination rate reported by deCODE http://genome.ucsc.edu/cgi-bin/hgTrackUi?hgsid=93924978&c=chr15&g=recombRate. We calculated the mean meiotic recombination rate within 5 Mb blocks between 50 Mb and 110 Mb on chromosome 5 (Figure 2b) and compared it with the frequencies of mitotic recombination breakpoints that we had found. Although the meiotic recombination map was more uniform than the mitotic map along the chromosome arm, linear regression analysis (Figure 2c) showed that there was a strong association between the locations of the breakpoints of the two forms of recombination (r = 0.729368, t = 3.449, d.f. = 10, p = 0.0062, 2-tailed).

Comparison between mitotic recombination and gain/deletion of chromosome 5q in colorectal tumours

Although mitotic recombination is the foremost mechanism of loss of heterozygosity in colorectal tumours, some colorectal carcinomas acquire deletions or gains of 5q around APC as an additional, later event during tumorigenesis [12]. Were the mitotic recombination map to be influenced primarily by chromosome breakage, we would expect the copy number and mitotic recombination breakpoint maps to be similar. Using array-CGH data from the panel of 54 CRC cell lines and from 99 fresh-frozen cancers (most samples taken from [12]), we determined the positions of breakpoints associated with regions of copy number change between the chromosome 5 centromere and APC (Figure 2d). There was no correlation between the locations of mitotic recombination and copy number breakpoints (r = 0.73, t = 3.459, d.f. = 10, p = 0.0062, 2-tailed), primarily because many copy number breakpoints were close to the 5 centromere (~50–56 Mb). Apart from the peri-centromeric region, the site with the highest frequency of copy number breakpoints was 70–75 Mb, as it was for the mitotic and meiotic recombinbation maps, but only a small proportion of all copy number breakpoints was at this site.

Local sequence context and mitotic recombination breakpoints

We tested whether local G+C content in 5 Mb regions between the chromosome 5 centromere and APC was associated with the frequency of mitotic recombination breakpoints. A mean of 38.0% G+C content was found, with a maximum of 40.7% at 75–80 Mb and a minimum of 35.5% at 100–105 Mb. There was a strong correlation between the location of mitotic recombination breakpoints and G+C content (r = 2.85, t = 3.67, d.f. = 10, p = 0.004, 2-tailed).

A genotyping assay to distinguish between the two major haplotypes within the polymorphic chr5:68–71 Mb LCR region

The 68–71 Mb LCR appeared to be one site with a high frequency of mitotic recombination breakpoints in colorectal tumours. However, SNP density is very poor in this region, owing to the exceptionally large number of highly similar paralogous sequences. We therefore developed a new genotyping assay to distinguish between the two major haplotypes that have been described at this site, hap1 of length 2.2 Mb and hap2 of length 1.6 Mb. The SMN1 gene is present on both haplotypes as a single copy and SMN2 is present as a single copy on haplotype 1 only. Thus, hap1 is SMN1–SMN2 and hap2, SMN1 only. (Although there have been reports of SMN1 and SMN2 alleles that have copy numbers different from those found on haplotypes hap1 and hap2 [13], these occur at a very low frequency (< 1%) and their relationships to haplotypes hap1 and hap2 are uncertain; we therefore did not consider the rare SMN1 and SMN2 copy number alleles further.) We developed a high-throughput genotyping assay based on the C>T difference (rs4916, chr5:69,408,109) in exon 7 that differentiates the SMN1 gene from the SMN2 gene. In theory, allelotype hap1/hap1 corresponds to an equal dosage of C and T alleles, allelotype hap1/hap2 to a 2:1 C:T ratio and allelotype hap2/hap2 to the presence of the C allele only. Using Pyrosequencing, we initially genotyped a test series of 188 individuals for the two haplotypes. As controls, we used colorectal cancer cell lines, that had known 5q copy number and were homozygous for the 70 Mb repeat region owing to somatic loss of heterozygosity. Although the C:T allele ratios deviated slightly from those expected (with the C allele dosage typically being over-estimated by about 5%), the C:T ratios allowed us to distinguish three genotype groups corresponding to hap1/hap1, hap1/hap2 and hap2/hap2 (Figure 3).

We initially determined LD between SNPs present on chromosome 5q on the Illumina Hap550 arrays and rs4916 (Figure 4). We found that there was no significant LD between rs4916 and any other SNP in the region from the Hap550 panel. Even the two SNPs on the Hap550 arrays that flanked rs4916 – rs11955686 at chr5: 68,776,409 and rs13168712 at chr5:70,715,406 – were isolated from rs4916 and other SNPs around. Although the absence of LD around the LCR may reflect hyper-recombination within the region or simply normal rates of crossover, it was possible to conclude that the LCR does not grossly suppress meiotic recombination.

Testing regions with high frequencies of mitotic recombination breakpoints for associations with colorectal cancer risk

If mitotic (or meiotic) recombination breakpoints tend to cluster in certain regions, one possible explanation is that the locations of recombination crossovers are selected in normal cells to occur at sites that have no effects on gene function. However, an alternative explanation for recombination breakpoint clustering in mitotic recombination is that the chromosome tends to have an intrinsic fragility at that site. If this latter model were correct, inherited variants near the most common breakpoints might modulate LOH frequencies and hence be associated with differential CRC susceptibility. There is evidence that such variation can modulate meiotic recombination [14, 15].

We tested whether inherited variation within close to mitotic recombination breakpoints was associated with colorectal cancer risk. Specifically, we determined SNP genotype frequencies from the 65–80 Mb region of chromosome 5 in our primary series of about 1,000 colorectal tumour cases and 1,000 controls. We included all SNPs on the Hap550 arrays (see Additional File 3), plus the following 4 additional SNPs: rs4916; rs2561182, flanking the LCR on the centromeric side; rs28409706, in a small region of sequence with no paralogues within the LCR; and rs7447545, flanking the LCR on the telomeric side. We excluded all SNPs that failed Hardy-Weinberg equilibrium testing at P < 0.01. We then compared genotype frequencies in cases and controls at the remaining 3,843 SNPs using the allelic, recessive and Cochran-Armitage tests. The strongest evidence of association (P_{allele test} = 3.2 × 10^-4, Figure 5) was found at rs2056169 (chr5:71,095,468). This association was suggestive, but did not meet requirements for significance at P = 0.05 after correction for multiple testing. Subsequent analysis in the VICTOR/QUASAR2/1958 series failed to confirm any association (P_{allele test} = 0.86). None of the 4 additional SNPs showed any association with colorectal cancer risk under any of the tests of association, with P_{allele test} > 0.18 in all cases.