Shallow whole genome sequencing for robust copy number profiling of formalin-fixed paraffin-embedded breast cancers

Pathology archives with linked clinical data are an invaluable resource for translational research, with the limitation that most cancer samples are formalin-fixed paraffin-embedded (FFPE) tissues. Therefore, FFPE tissues are an important resource for genomic profiling studies but are under-utilised due to the low amount and quality of extracted nucleic acids. We profiled the copy number landscape of 356 breast cancer patients using DNA extracted FFPE tissues by shallow whole genome sequencing. We generated a total of 491 sequencing libraries from 2 kits and obtained data from 98.4% of libraries with 86.4% being of good quality. We generated libraries from as low as 3.8 ng of input DNA and found that the success was independent of input DNA amount and quality, processing site and age of the fixed tissues. Since copy number alterations (CNA) play a major role in breast cancer, it is imperative that we are able to use FFPE archives and we have shown in this study that sWGS is a robust method to do such profiling.

Comparative Genomic Hybridisation (CGH) 1 has had a significant impact in the study of cancer genomes. Chromosomal regions gained or lost in the tumour could be easily visualised by hybridization onto normal human metaphase spreads, allowing characterisation of genome-wide copy number alterations (CNA) in tumours 1 .
Microarrays with DNA probes (cloned DNA or oligonucleotides) spotted onto glass slides representing the entire genome soon replaced normal chromosomes 2 making it faster and easier to profile. The importance of characterizing somatic CNAs in cancer is now well established, with a recent TCGA pan-cancer analysis showing that human tumours can be classified into mutation driven (M-class) or copy-number driven (Cclass) subtypes. Breast cancer is a C-class cancer type 3 and we have previously shown that CNAs are the main determinants of the expression architecture of breast cancers.
Using gene expression driven in cis by CNAs, we have generated a new molecular taxonomy of breast cancer with 10 genomic driver-based subtypes termed Integrative Clusters. The samples used in this analysis were derived from the METABRIC cohort, which encompassed a large biobank of fresh frozen tumour samples collected across five major teaching hospitals in the UK and Canada 4 .
Formalin-fixed paraffin-embedded (FFPE) tissue samples are more routinely collected and hence more representative of cancer in the general population. These FFPE archives are a valuable resource for molecular profiling in cancer research.
Whilst the fixation process is essential to protect cellular morphology and protein expression, it is detrimental to nucleic acids and results in their chemical modification and degradation. As a result, extraction of DNA from FFPE tissues results in lower yields when compared to extraction from fresh frozen tissues. DNA extracted from FFPE works well for downstream applications using polymerase chain reaction (PCR), particularly for small size amplicons (less than 300 base pairs), but for other applications, including microarray based CGH, where efficient labelling of the DNA is dependent on its integrity, its use is more challenging. There have been several studies describing different methods for DNA extraction 5 , quality control 6,7 , labelling 8 and other optimisation protocols 9 to improve the performance of FFPE DNA on microarrays. In the past, we have tried to profile CNAs using FFPE DNA on microarrays with limited success. Only Illumina Infinium and Molecular Inversion Probe (MIP, Affymetrix) arrays yielded good results but these required good quality and at least 200ng of DNA 10 .
Next generation sequencing has revolutionised cancer genomics. It is now relatively easy and inexpensive to sequence an entire genome. However, as with microarrays, the robustness of the results obtained are dependent on the quality of the input DNA. Two recent studies have demonstrated the feasibility of doing shallow whole-genome sequencing (sWGS) for CNA profiling using DNA extracted from FFPE tissue material 11,12 . The first report used 250ng of DNA from FFPE tissues and a breast cancer cell line to produce libraries and developed an analytical method for sWGS. The second study compared several sequencing library production kits and reported generating successful sequencing libraries with low input DNA in a small number of FFPE samples.
Here we present extensive sWGS data generated from DNA extracted from FFPE breast cancer samples to describe steps to ensure successful libraries.

Sequencing library generation
Sequencing libraries were generated using either the beta testing version of the Illumina

Bioinformatics
Alignment against the GRCh 37 assembly of the human genome was performed using BWA ver. 0.7.9 13 or NovoAlign ver. 3.2.13 (NovoCraft, Malaysia). PCR and optical duplicates were identified using Picard tools ( https://broadinstitute.github.io/picard) or Novosort (NovoCraft, Malaysia). Circular binary segmentation on the aligned files was performed in 100kb windows using the QDNAseq R package available on Bioconductor, which corrects for mappability and GC content 11 . All statistical analyses were performed in R using the functions lm() for fitting linear models and t.test() for Welch two-sample t-test.

Results
The majority of the FFPE samples available were core biopsies collected as part of a neoadjuvant clinical trial (GEICAM/2006-03, n=107) yielding low amounts of DNA (range=4 -61ng, median 30ng). Therefore, to successfully generate libraries for CNA profiling using limited input DNA, we needed to understand how different variables could influence the quality of libraries and steps that can be taken to ensure good sequencing results ( Figure 1).

Assessment of the copy number plots
We examined the copy number plots by manual inspection and categorised them based on the variance in the CN data for each case into categories: "Very Good", "Good", "Intermediate" and "Poor" (Figure 2a). We also used QDNAseq 11 which calculates the expected (estimated from read depth) and measured (using read depth and influenced by DNA quality) standard deviation of the summarised reads, as a measure of variance. Both measures increased as the quality of library decreased and validated our categorisation of library quality (measured standard deviation shown in Figure 2b).

Assessment of different sequencing kits
We tested two kits (Illumina FFPE TruSEQ kit and Rubicon Genomics Thruplex DNASeq) using four FFPE samples to generate sequencing libraries and found comparable results (Supplementary Figure 1a-b). The CNA profiles obtained using DNA processed with the ILMN kit had less variance (noise) than those processed using the RGT kit however the ILMN libraries were generated using more input DNA (200-500ng (ILMN) versus 50ng (RGT)) and were sequenced deeper (average coverage 0.9X (ILMN) versus 0.08X (RGT)). For a more comparable evaluation, we downsampled ILMN sequencing data to a similar read depth as RGT; this showed comparable copy number profile qualities between the two library preparation technologies.
In theory, increasing the sequencing depth should improve the copy number results by reducing the variance. We examined this by increasing the sequencing depth of 23 RGT kit libraries which had less reads (from 0.08X up to 0.15X) and found improvement in the data quality in 20 out of 23 libraries (examples shown in Supplementary Figure 2a). To examine the association between sequencing depth and variance, we down-sampled the number of reads (in steps of 1x10 6 reads) for six libraries with high read counts (up to 24x10 6 reads). We found a significant improvement in the quality of copy number plots with increasing number of reads (p<2.2e-16; Supplementary Figure 2b). It is interesting to note that the noise reduction levels off at approximately 7x10 6 reads suggesting that increasing the read depth more than 7x10 6 reads provides little benefit to variance reduction.

Performance of sWGS for copy number profiling using the RGT kit
Due to the limited amount of DNA available for most samples, we chose the RGT kit as it required less input DNA due to fewer processing steps, in particular purifications. Sequencing libraries were generated from as little as 3.8ng of DNA, and out of 16 libraries prepared from less than 10ng of DNA, only one failed, 13 generated good quality CNA plots, and 2 generated intermediate quality CNA plots. Information for all the libraries generated are summarised in Supplementary Table 3.

Recovery of under-performing RGT libraries
Eight (1.8%) libraries failed and 12 (2.7%) generated poor quality libraries out of 446 libraries. To recover some of these failed/poor samples, we prepared fresh libraries from samples with sufficient DNA (n=6) or repeated the sequencing using three-fold more library material for samples with insufficient DNA to generate new libraries (n=8). Thirteen of these new/re-sequenced libraries generated good quality data. The one repeat sample that failed was from the re-sequencing group.
Consequently, only two out of 446 RGT libraries (taking into consideration the repeated libraries and re-sequencing) failed, resulting in a 99.5% success rate. Good sWGS data produced from 379/446 (84.9%) samples.

Association between FFPE storage time, site, and sequencing quality
The FFPE samples were collected from three different tissue banks, spanning 20 years ( Table 1). The effect of storage time on the DNA extracted was analysed ( Figure   3). DNA from older FFPE blocks (>5years) was generally of poorer quality: higher Δ Ct values, shorter fragment size, generating lower yield sequencing libraries. We compared the quality metrics for each banking site and found that overall FFPE samples from different sites were comparable (Table 1 and Supplementary Figure 3).

Association between input DNA characteristics and sequencing library yield
We used the Illumina FFPE QC kit, a quantitative-PCR assay to estimate the quality of  Using the RGT kit, we found no correlation between amount of input DNA and sequencing library yield (r 2 = -0.002, p=0.81). This is probably due to the fewer library-washing steps using the RGT kit (six washing steps in the ILMN protocol versus one in RGT).

Association between input DNA characteristics and sequencing library quality
Next we sought to determine if sequencing quality was influenced by the nature the sequencing quality categories. Reassuringly, we found no biases in sampling that contributed to the sequencing quality. In other words, each copy number plot quality group had samples from all DNA quality (ΔCt) groups, fragment sizes and input quantity groups, suggesting that we could generate good quality libraries from most of our FFPE DNA regardless of these features.
Using our copy number output categorisation scoring, we examined if quality of the libraries (analysed as "all sequencing quality groups" versus "very good") can be attributed to the different features of the input DNA and library yields (Table 2, Figure   5). We found that the quantity of template was only significantly different in the good quality libraries. Meanwhile, the quality of input DNA was significantly different in the intermediate libraries only when compared to the "very good" libraries. Therefore, the lesser quality sequencing libraries (I, P and F) cannot be attributed simply to either quantity or quality of the template DNA. The DNA fragment sizes, which should reflect the length of the template as the DNA was sheared under similar conditions, were found to be significantly different in all groups (progressively becoming shorter) except the failures. We found that low quality DNA was associated with shorter DNA fragments, lower library yield and higher number of unmapped reads but no association with the total number of unique reads aligned (Figure 6a-d). The recovery of most of the poor/failed libraries described previously, was achieved by either repeating the library generation or re-sequencing to generate more reads. Consequently, we suspect the poor/failed libraries could be due to a loss of DNA during the purification steps or that the Q-PCR quantification of the libraries prior to normalisation, over-estimated the library concentration resulting in inadequate amount of library being used for sequencing. This would explain why by simply increasing the quantity of libraries for sequencing and reducing the number of samples in a single pool, ensured adequate read counts and successful sequencing.

Discussion
In this study, we have looked at the effect that quantity and quality of DNA from FFPE tissues has on successful sWGS library preparation for CN profiling of human breast cancers. Both the quantity and quality of DNA have always been an important consideration for sample selection and in deciding which genomic application to use. For example, microarrays require 100ng -2.5μg of DNA depending on the resolution of the arrays whereas PCR based methods require only 10ng of DNA. In our hands, we have not had much success in obtaining CN data with DNA extracted from FFPE DNA using microarrays, especially when the extracted DNAs are more fragmented and of lower quality (judging from absorbance ratios of 260nm to 280 nm and multiplex PCR for quality control).
Here we have robustly shown that we can generate CN data from virtually all archived FFPE samples using sWGS. We show good CN profiling data irrespective of the quality of input DNA, as inferred by whether it can be amplified with Q-PCR (ΔCt).
Previous work has extensively tested the utility of FFPE DNA for mutation analysis 14-18 but to date no comprehensive study has shown its use for CN profiling. Since many human cancer types, including breast and ovarian cancers, are driven mostly by CNA (C-class) rather than point mutations or indels (M-class), we believe more effort should be focussed on characterizing the copy number landscapes of these cancers 3 . We found sWGS to be very robust in generating these CN profiles, independently of the kits used, quantity and quality of DNA. sWGS is also significantly cheaper (~50%) than microarray-based methods (Supplementary Table 2).
Another advantage of generating sWGS libraries is the ability to use the same library for targeted sequence enrichment to identify mutations. There have been other methods reported for CN profiling using DNA extracted from FFPE samples but these methods do not generate sequencing libraries that can then be used for target enrichment and sequencing 19 or if they do, are expensive 20 . In addition, sWGS will also serve as a quality control for the libraries, given its relative low cost when compared to that of generating targeted sequencing libraries. Only libraries that generate good CN profiles should be used for target enrichment and mutation detection 21 . Whilst we haven't performed target enrichment on our FFPE libraries, we expect the performance of these FFPE libraries for mutation analysis to be similar to that of published data, including known artefacts caused by formalin-based fixation effects on the DNA template 15,17,22-24 .
In summary, we have shown that sWGS is a robust and cost-effective method for obtaining good quality CN data from FFPE cancer samples, irrespective of the DNA quality and quantity used. In the case of breast cancer, CN profiles can be used to stratify breast cancers into one of the 10 Integrative Clusters 25 , reiterating the importance of FFPE tumour archives. The methods described here are also of relevance to other cancers, e.g. ovarian cancers where CN profiling is essential to characterise their genomic landscapes.  Table 1 Features of input DNA and libraries generated from FFPE blocks collected at three different sites. Data provided in minimum-maximum range and median in brackets. Age = years since blocks were generated. Δ Ct = difference between the cycle threshold of test to the control template ACD1 provided in the kit. ng=nanogram, PCR= polymerase chain reaction, bp=base pairs, nM=nanomoles. * denotes the site where there is a significant difference to the index group (ie Cambridge).

Table 2
Features of input DNA and libraries for the different categories of copy number data. Data provided in minimum-maximum range and median values in brackets. Δ Ct = difference between the cycle threshold of test to the control template ACD1 provided in the kit, ng=nanogram, bp=base pairs, nM=nanomoles, SD=standard deviation. * denotes the site where there is a significant difference to the index group (ie Cambridge).  Dot plots represent the range (minimum-maximum) of observed values for each of the following categories and the red dot represents the median.
A. The quality of input DNA inferred by Δ Ct. B. Fragment sizes of the libraries in base pair C. The library yield in nanomoles D.