Ethic statement

Written informed consent for CTC isolation and genomic characterization was obtained for all patients included. All experiments conformed to the principles set out in the WMA Declaration of Helsinki and were approved by the Ethical Committee Boards responsible for the corresponding studies (Azienda Ospedaliero Universitaria di Bologna, Policlinico S. Orsola Malpighi, Caratterizzazione Molecolare per la Medicina Personalizzata nel Paziente Oncologico, CAMMPPO, 82/2015/O/Tess; Royal Marsden Research Ethics Committee approved protocol CCR2472).

Cell culture

Cell lines (NCI-H1650, ATCC^® CRL-5883^™; NCI-H1563, ATCC^® CRL-5875^™; NCI-H2228, ATCC^® CRL-5935^™; NCI-H23, ATCC^® CRL-5800^™; NCI-H441, ATCC^® HTB-174^™; NCI-H661, ATCC^® HTB-183^™) were cultured in RPMI 1640 (ATCC modification) supplemented with 10% fetal calf serum (FCS, both Gibco^® /Life Technologies^™) and Penicilline/Streptomicine (Pen/Strep 100x, Euroclone). Cells were maintained at 37°C and 5% CO₂.

Cell lines single-cell isolation by micromanipulation

For single cell isolation of mononuclear cells, the bottom of a petri dish was coated with FCS, cell suspensions (cell lines NCI-H1650, NCI-H1563, NCI-H2228, NCI-H23, NCI-H441, NCI-H661) were diluted with 1X PBS to achieve a density of one cell per visual field under an inverse microscope with 10X magnification. Single cells were picked under visual control using a 1 μl pipette and transferred (with 1μl PBS) into a 0.2 ml PCR tube for subsequent whole genome amplification.

Isolation of patients’ pure CTCs and WBCs by DEPArray

Blood was collected from 1 prostate cancer and 2 lung adenocarcinoma patients. CTCs were enriched (CellSearch CTC Kit, CellTrack^® Autoprep, Menarini Silicon Biosystems Inc) and counted (CellTrack^® Analyzer II, Menarini Silicon Biosystems Inc) prior to being extracted from CellSearch cassettes and loaded on DEPArray^™ cartridge (Menarini Silicon Biosystems, SpA, Italy) [34]. With the DEPArray^™ system, CTCs and White Blood Cells (WBCs) were identified and isolated as pure cells according to the manufacturer’s protocol.

Ampli1^™ whole genome amplification, DNA library construction and whole genome sequencing

DNA of isolated cells was amplified using the Ampli1^™ WGA kit (Menarini Silicon Biosystems) according to manufacturer instructions. Quality of Ampli1^™ WGA products was checked using Ampli1^™ QC kit (Menarini Silicon Biosystems) and only products with at least 2 amplified bands were retained. 5 μl of Ampli1^™ WGA product were transferred into a new tube and cleaned up with 1.8X SPRIselect Beads (Beckman Coulter) according to manufacturer instructions and eluted in 12.5 μl TE. We designed a streamlined method (which we implemented as a kit called Ampli1^™ LowPass, commercially available from Menarini Silicon Biosystems) for preparing libraries for low-pass WGS by specifically exploiting the deterministic nature of Ampli1^™ WGA. In brief, starting from a 10–50 ng of purified primary Ampli1^™ WGA product, we perform a re-amplification using hybrid PCR primers, including barcoded adaptors compatible with the Ion Torrent^™ Systems on the 5’ end, and primary WGA universal adaptor on the 3’ end. Barcoded libraries were quantified using Qubit dsDNA HS Assay kit and Qubit 2.0 Fluorometer (Thermo Fisher Scientific) and pooled in equimolar concentrations to obtain 1500 ng in 44 μl of total volume. Pooled libraries were size selected (300–450 bp) using E-Gel SizeSelect^™ Agarose Gels, 2% on a E-Gel Agarose Gel Electrophoresis System (Thermo Fisher Scientific) according to manufacturer instructions. Size selected library pool was cleaned up with 1.2X SPRIselect Beads (Beckman Coulter) according to manufacturer instructions and quantified using Agilent High Sensitivity DNA Kit using the Agilent Bioanalyzer 2100 instrument (Agilent). Then, library pool was used for emulsion PCR amplification (400bp) and template-positive Ion Sphere Particles (ISPs) were enriched using the Ion Chef^™ System (Thermo Fisher Scientific). Sequencing was performed using 318 BC chips on the Ion PGM^™ and Ion 530^™ chip on Ion S5^™ System (525 flows).

Libraries from gDNAs (100ng) were prepared using Ion Xpress^™ Plus gDNA Fragment Library preparation kit (Thermo Fisher Scientific). Briefly, samples were fragmented for 200-base-read libraries, end repaired, ligated with adaptors, nick repaired and bead purified prior to amplification of size selected (E-Gel SizeSelect^™, Thermo Fisher Scientific) fragments around 250 bp long. Fragment sizes were assessed using the Bioanalyzer system and quantified using the Ion Library TaqMan^® Quantitation Kit (Thermo Fisher Scientific). Pooled libraries were used for emulsion PCR amplification (200bp) using the Ion Chef^™ System (Thermo Fisher Scientific). Sequencing reactions were run on the Ion Proton^™ System using Ion PI^™ version 3 chips (Thermo Fisher Scientific).

Sequence alignment, read counting and normalization

Signal processing, base calling and alignment to Homo sapiens hg19 reference sequence was performed with the Torrent Suite^™ v4.6 with—g 0 parameter for the alignment step with tmap. Genome binning was performed using WindowMaker tool from BEDTOOLS suite [35]. Read counting and assignment to genomic bins were performed using the HTSeq library [36]. Reads spanning more than one bin were assigned to the one with the longest overlap. Read counting and assignment to MseI fragments were performed by BEDTOOLS IntersectBed tool, filtering out reads with more than one fragment match. GC-based normalization was performed by LOWESS fitting of per-bin GC content versus read count on each bin. Calculation of bin mappability value was performed using bigWigAverageOverBed (http://hgdownload.cse.ucsc.edu/admin/exe/) using mappability track for 100mers produced by Encode/CRG (wgEncodeCrgMapabilityAlign100mer; downloaded from https://genome.ucsc.edu/).

Identification of problematic genome regions

For determination of problematic genome regions, read counts from 21 control WBCs over 500 Kbp bins were GC-normalized and mappability-normalized and divided by median normalized read count. For each bin, the median of normalized read counts across the 21 control WBCs was calculated and bins with median values > 1.4 or < 0.6 were flagged as problematic regions, potentially leading to false positive calls.

CNA calling

Control-FREEC (Control-Free Copy number caller) software was used to obtain copy-number calls, using the mode without control sample [37]. Read counts were corrected by GC content and mappability (uniqMatch option). Bin size was manually set in order to match the desired resolution. To determine significant CNA calls, Wilcoxon test and Kolmogorov-Smirnov test (p value < 0.01) were performed using the script assess_significance.R provided with Control-FREEC software.

ROC curves

To assess the sensitivity and specificity of single cell low-pass experiments, the altered copy number status on each single cell was compared, in windows of 500Kbp, to the CNA calls of their corresponding reference WGS of non-amplified gDNA of the respective cell line by means of a receiver operating characteristic (ROC) curve. The comparison refers only to the presence of a CNA in the single cell data versus the reference. Type (gain or loss) and actual copy number were not considered in the comparison. Computation of true and false positive rates for various Wilcoxon non-parametric p-value thresholds and the area under the curve (AUC) were performed using scikit-learn python library. Analogous analyses were performed also to assess sensitivity and specificity at variable read depths, using a 3.5 million reads dataset as reference, and to assess sensitivity and specificity of Ampli1^™ LowPass protocol respect to aCGH.

Ploidy determination

To determine the ploidy of single cells, raw BAM data were processed as follows:

1. GC-normalized, mappability-normalized and median centered read counts over 500 Kbp bins were multiplied by the ploidy to be tested and were smoothed using the method implemented in smoothseg R package [38] to reduce random noise of NGS data obtained at shallow coverage, which may affect the later determination of ploidy levels [39].
2. A probability density function was estimated from the smoothed data using kernel density estimation (KDE); KDE bandwidth is estimated by Silverman’s ‘rule of thumb’ [40] and, if necessary, manually tweaked after visual inspection of the density plot to best reflect underlying data distribution.
3. Estimates of the copy numbers are obtained by finding peaks on the KDE fitted data as described by Du et al. [41]; peaks with a relative probability contribution lower than 2% are excluded as potential false positives.
4. Copy number estimates were rounded to the nearest integer and resulting values are assumed to be the putative underlying copy numbers. Given the discreet nature of read counts, which are expected to be directly proportional to DNA content, the copy number estimates should increase linearly with the underlying copy number. The estimates were thus fitted to a linear regression y = aP* where a is the slope for P, which is a vector of the putative copy numbers.
5. Process was repeated for each ploidy to be tested (from 2 to 8)

Only main ploidies for which R² > 0.98 were considered further and best fitting main ploidy was selected based on sum of squared residuals (SSR). Since ploidies multiple of the real main ploidy would produce similar fittings and SSR values, results are manually reviewed and the lowest possible plausible ploidy with similar SSR and R² values was selected.

Comparative genomic hybridization with oligonucleotide microarrays (aCGH)

aCGH analyses on oligonucleotide arrays were performed according to the manufacturer’s instructions (Agilent Oligonucleotide Array-Based CGH for Genomic DNA Analysis, Version 6.4, August 2011, G4410-90010) with slight modifications as described in [42]. All CGH arrays were processed using the Microarray Scanner G2565CA by Agilent Technologies with 3 μm resolution and 16 bit color depth. The output image files were imported, normalized and fluorescent ratios for each probe were determined using Feature Extraction software (Agilent Technologies, Version 10.7.3.1, Protocol CGH_107_Sep09). Feature Extraction output files were imported into the Genomic Workbench 5.0.14 software. aCGH data were examined using the aberration detection method 2 (ADM-2) algorithm with a threshold of 6.0. No centralization was applied. An aberration filter was defined for identifying copy number alterations, where changes only were considered as true positive events with a minimum log2ratio of 0.3 and a minimum of 50 consecutive probes with the same polarity per region.

Development of an Ampli1^™-based protocol for low coverage whole genome sequencing (Ampli1^™ LowPass)

Ampli1^™ WGA, based on LM-PCR, was employed for DNA amplification because it showed accurate and more even representation of the original single-cell genomic DNA compared to available methods, as shown by previous reports [27,43,44]. Ampli1^™ WGA has already been used for low-pass WGS by Hodgkinson C. L. et al. [31]. However, according to the workflow used in that paper, the creation of Illumina^®-compatible libraries requires several steps including i) digestion of WGA adaptors, ii) DNA fragmentation, iii) EndRepair iv) A-Tailing v) barcoded adaptor ligation, vi) sample pooling of barcoded NGS libraries and vii) sequencing. To avoid complex processing steps and streamline the protocol we devised a method, named Ampli1^™ LowPass, which exploits the universal sequences at the end of Ampli1^™ WGA DNA to incorporate Ion Torrent^™ compatible adapters (Fig 1). A single PCR amplification step is employed to produce barcoded libraries which are ready to be pooled for sequencing, thus skipping laborious and costly processing steps. The same amplification also introduces barcodes incorporated into one of the primer sequences as shown in Fig 1. Finally only a size selection step is needed to make the libraries compatible with the sequencing platform (Ion Torrent^™ PGM or IonS5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic overview of Ampli1^™ LowPass approach.

DNA is amplified through primers complementary to Ampli1^™ WGA universal adapters through a single PCR reaction. Primers incorporate Ion Torrent^™-compatible adapter sequences and barcodes. Libraries are then pooled and subjected to standard processing for sequencing on PGM or Ion S5 platforms.

https://doi.org/10.1371/journal.pone.0193689.g001

Ampli1^™ LowPass produces a comprehensive representation of the human genome

To assess the presence of biases and determine the uniformity of genome representation by Ampli1^™ WGA size-selected fragments, genomic DNA from 21 WBCs from 8 individuals was amplified and sequenced on Ion Torrent^™ PGM platform. On average, 528,836 reads were generated per sample (S1 Table). Genome was divided into 500kb fixed-size, non-overlapping bins. Read counts per bin showed a strong dependency on GC content (Fig 2a). This is however expected as it is well known that read counts are affected by polymerase biases in presence of high or low GC content [45–47]. Another potential source of read count bias is the non-homogeneous distribution of MseI sites (TTAA) along the genome, also dependent on GC content (Fig 2b), which leads to different numbers of fragments generated from different genomic regions (Fig 2c). Moreover, fragments generated by MseI restriction employed in Ampli1^™ WGA kit are size selected prior to PCR amplification, potentially leading to further bias in read counts. To evaluate this source of bias and effect of GC-based normalization, we calculated the number of fragments per bin, weighted on the fragment length probability (S1 Fig) and evaluated the correlation with read counts before and after GC normalization. As expected, plot of raw read counts against the weighted number of fragments per bin showed a strong bias (Fig 2d). However, biases in read counts were effectively corrected by GC-based normalization (Fig 2e), currently implemented in available software for CNA detection from WGS data [37,39,48], improving the distribution of read counts and producing an even and tight normalized read count distribution along the genome (Fig 2f). Corrected data showed also a high consistency of read count distribution across different control WBCs (S2 Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Effect of normalization on read counts distribution.

a) Scatter plot of read counts, normalized on 1 million of reads, versus GC content in 500 Kbp bins obtained by sequencing of a single WBC; number of MseI fragments per bin is plotted b) respect to GC content and c) along the 22 autosomes; scatter plots of read counts in a single WBC versus number of MseI fragments per bin, weighted on per-fragment probabilities, before d) and after e) GC normalization, three standard deviations are used to discriminate outliers (red dots); f) GC-normalized read counts plotted along the 22 autosomes.

https://doi.org/10.1371/journal.pone.0193689.g002

After normalization, a minor number of bins corresponding to about 1.5–2% of the genome still showed a high deviation (> 3*standard deviation) from the median of normalized counts, likely due to poor mapping in correspondence of repeated and low complexity regions (Fig 2e, red dots). Indeed, the analysis of the distribution of mappability values for highly deviating bins confirmed that they mainly correspond to regions with low mappability values (S3 Fig). Mappability-based normalization is also currently implemented in available software for CNA detection from WGS data [37,48].

To evaluate the homogeneity of genome representation, we calculated a uniformity value defined as the fraction of bins whose normalized read count is at least 20% of the normalized read count averaged across all the bins. On average, uniformity greater than 99% was obtained for each sample (S4 Fig) and, furthermore, more than 90% of bin bases were covered at higher than 60% of the mean coverage in all the control samples, implying a very tight distribution of normalized read counts around the mean.

Finally to identify problematic regions we analyzed the median of the normalized fold change across 21 WBC normal controls. Bins, with normalized read depths highly deviating (FC ≥ 1.4 or FC ≤ 0.6) from the genome median value, were mostly located near pericentromeric and telomeric regions, which are typically rich of repeated and low-complexity regions (S5 Fig). A list of 50 problematic regions with size up to 2 Mbp, and globally accounting for 28 Mbp, was built and was used in following analyses to filter false positive CNA calls.

Ampli1^™ LowPass produces distortion-free accurate copy number profiles

To verify absence of distortions and residual biases in copy number profiles due to Ampli1^™ WGA of single-cell DNA, we compared the profiles generated by Ampli1^™ LowPass of 2 single cells from each of 4 aberrant cell lines (NCI-H1650, NCI-H1563, NCI-H23, NCI-H441) with those generated by WGS sequencing of the corresponding bulk genomic DNA. WGS of genomic DNA from the 4 cell lines generated between 20,4 to 31.9 million reads, while from 633,049 to 1,284,763 reads were generated from sequencing of the amplified DNA from the single cells (S1 Table). Copy number profiles were generated from GC-normalized and mappability-normalized read counts in 500 Kbp bins. Visual inspection of the profiles showed a strong agreement between copy number profiles generated from single cells and bulk genomic DNA (S6–S9 Figs). To analyze more in depth the agreement between CNA calls in single cells and corresponding bulk DNA we performed a ROC analysis, using Wilcoxon non-parametric test as classifier, to call for copy gains and losses. For all the 4 cell lines analyzed, AUCs ≥ 0.91 were obtained indicating a strong agreement between CNA calls from Ampli1^™ LowPass of single cells and corresponding bulk DNA (Fig 3). Above data confirms WGA does not introduce any significant bias in whole genome copy number profiles and produces accurate CNA calls.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Performance of CNA calling in amplified vs. non-amplified DNA in 4 aberrant cell lines.

Data obtained by low-pass WGS (0.5-1M reads) of DNA from single cells amplified with Ampli1^™ WGA kit were processed for CNA calling. CNAs detected in non-amplified bulk gDNA (20-30M reads) were used as reference. For all the 4 cell lines considered ROC analysis showed an excellent agreement (0.91≤AUC≤0.97) between CNA calls from single cells and bulk gDNA.

https://doi.org/10.1371/journal.pone.0193689.g003

Optimization of coverage and resolution

To determine the number of reads necessary to reach high specificity and low number of false positives we merged the alignments from the 21 WBCs and randomly subsampled the dataset in subsets ranging from 100,000 mapped reads to 8 million mapped reads. Data analysis was performed at different bin sizes, corresponding to different resolutions, ranging from 100 Kbp to 2 Mbp. Resulting copy number profiles are expected to be free from CNAs. Any call was treated as a false positive call and specificity was calculated. Interestingly, at 200 Kbp resolution, 400,000 reads are sufficient to get specificity = 0.994 (S10 Fig), which shows a dependency on both read number and bin size.

To estimate sensitivity, specificity and accuracy of CNA calling depending on read number and resolution, DNA from two single cells from the aberrant cell line NCI-H1650, deriving from metastatic site in state 3B adenocarcinoma, was processed with Ampli1^™ LowPass protocol (S1 Table). Following mapping to hg19 reference genome, subsets ranging from 100,000 to 3.5 million mapped reads were extracted by random sampling of alignments and CNAs were called at different resolutions ranging from 100 Kb to 2 Mbp (S11–S18 Figs). It is to note that, while longer bins may miss smaller CNAs, on average 93.88% and 92.74% of the total length of CNAs detected at a resolution of 100 Kb were also detected at resolutions of 500 Kbp and 1 Mbp respectively, using a dataset of 3.500.000 mapped reads. This is expected as most CNAs in tumors are in the order of megabases [18]. A ROC curve analysis was performed using the dataset at 3.5 million mapped reads as a reference. For both cells, 200,000 mapped reads were sufficient to get an excellent accuracy at all the resolutions tested with AUCs ranging from 0.94 to 0.99 (Fig 4) as confirmed also by visual analysis of copy number profiles (S11–S18 Figs). For subsamples of 100.000 reads, we observed a decrease of the AUC by 25% at 100 Kbp bin size, thus indicating that lower bound for accurate CNAs detection at 100.000 reads is approximately 200 Kbp. From a quantitative point of view, copy number changes at a resolution of 500 Kbp were also maintained consistently across the different subsets, showing a mean R² of 0.94 and 0.89 between copy number profiles obtained from 3.5 million mapped reads with those obtained from 1 million and 0.5 million mapped reads respectively (S19 Fig). Correlation, however, decreased rapidly at lower resolutions and read counts likely due to stochastic noise in read counting.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. CNA detection by low-pass experiments at different read depths and resolution.

Two cells (a-d & e-h) from cell line NCI1650 were analyzed at different window size/resolutions (a,e = 100Kb; b,f = 200Kb; c,g = 500Kb; d,h = 2,000Kb). A dataset at 3,500,000 reads served as reference for ROC analysis.

https://doi.org/10.1371/journal.pone.0193689.g004

To conclude, at a resolution of 500 Kb, sufficient to resolve the majority of CNAs in tumors, 500,000 mapped reads are enough to get an accurate CNA calling both from a qualitative and quantitative point of view. At lower bin sizes it is still possible to get an accurate detection of aberrations even if profiles start to get noisier from a quantitative point of view for low read coverages.

Comparison with aCGH

aCGH is a widely used and accepted method for screening CNVs and CNAs in clinical diagnostics [49]. Moreover a previous report has shown that, in conjunction with Ampli1^™ single-cell WGA technology, aCGH provides precise and high resolution assessment of copy number changes in single cells [42]. To compare CNA calling by Ampli1^™ LowPass with aCGH, DNA from 6 aberrant cell lines (NCI-H1650, NCI-H23, NCI-H2228, NCI-H1563, NCI-H441, NCI-H661) was amplified with Ampli1^™ WGA kit and processed with both aCGH (G3 4x180k Agilent aCGH microarrays) and low-pass protocols. Based on the above results we aimed at producing about 500.000 reads per cell and we analyzed copy number profiles with a bin size of 500 Kb. Visual analysis of Ampli1^™ LowPass profiles confirmed the high concordance to aCGH profiles (Fig 5a and 5b; S20–S25 Figs). Moreover, Ampli1^™ LowPass showed high accuracy in calling of aCGH-detected CNAs with AUCs ranging between 0.81 and 0.91 for cell lines NCI-H1650, NCI-H23, NCI-H2228 and NCI-H1563 (Fig 5c and 5l). A lower agreement between Ampli1^™ LowPass and aCGH was observed for lines NCI-H441 and hyperhexaploid NCI-H661. However, visual analysis of the former line shows noisy and flatter profiles for aCGH, which may contribute for problems in CNA callings (S24 Fig); the latter mainly shows differences in segmentation and CNA calling despite the similar profiles and a pretty good correlation of copy number profiles (0.84≤R²≤0.87; S25 and S26 Figs). This is likely due to a compression of copy number alterations due the multiploid nature of cell line NCI-H661, which might hinders an accurate calling of CNAs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Comparison of LowPass copy number profiles and CNA calling with aCGH.

Example profiles from one single cell of aberrant cell line NCI-H23 generated by Ampli1^™ LowPass (a) and aCGH of Ampli1^™ amplified DNA (b). In c-p): ROC curves comparing Ampli1^™ LowPass CNA calls with aCGH calls from single cell of 6 cell lines of the NCI-H series.

https://doi.org/10.1371/journal.pone.0193689.g005

Finally, low-pass shows a high concordance with aCGH also from a quantitative point of view as indicated by the high correlation between copy number values of low-pass CNA calls compared with corresponding aCGH fold changes (R² = 0.89; S27 Fig).

Determination of single cell ploidy

Different bioinformatic methods are already available to determine cancer ploidy and purity such as ABSOLUTE [50], ASCAT [51], THetA [52] and PyLOH [53]. The first two are however designed for SNP array data and do not formally model DNA sequencing data, THetA is designed to identify subclonal CNAs in mixed samples data obtained by high throughput sequencing (40X coverage) and is thus not suitable to low-pass sequencing data, PyLOH uses allelic information which is not available for low-pass sequencing data.

The method we present exploits the linear relationship between read counts and underlying copy numbers [39]. Indeed, it is expected that, for a given cell ploidy, normalized ratios of gains and losses will scale linearly with read counts. We illustrate this in Fig 6 where copy number profiles were generated from the analysis of a single cell from aberrant cell line NCI-H23 by using 2 different main ploidies (Fig 6a). Frequency distribution of smoothed copy number data, obtained by multiplying normalized ratio by the main ploidy, show a multimodal distribution where different modes ideally correspond to different copy number levels in the genome of the cell analyzed and highest peak correspond to the main ploidy (Fig 6b). Kernel density estimation and determination of modes by a peak detection method based on wavelet transform [41] clearly shows that a main ploidy of 2 produces a better fit to hypothetic underlying copy numbers, compared to 3. Indeed peaks for a main ploidy of 2 are = (1.0, 1.97, 2.94, 3.82, 4.67) and are reasonably centered around the putative underlying ploidies (1, 2, 3, 4, 5). On the contrary peaks detected with a main ploidy = 3 are = (1.58, 2.98, 4.49, 5.79, 6.52, 7.12). As expected, regression analysis of peak positions vs putative underlying copy numbers shows a ploidy of 2 produces a better fit with higher R² (1.0) and SSR (0.008) compared to a main ploidy of 3 (R² = 0.97; SSR = 0.668). We tested the method on Ampli1^™ LowPass data from 2 single cell of the hyperesaploid cell line NCI-H661 (S28 Fig) using 7 different main ploidies (2, 3, 4, 5, 6, 7, 8). Best fits were obtained with main ploidies set to 6, in agreement to what suggested from cytogenetic data available for the cell line. Absolute copy number plots obtained from one single cell from cell line NCI-H661 by setting alternatively a main ploidy of 2 and 6 clearly show that segments better represent the underlying copy number profiles with a main ploidy of 6 (Fig 7, S29 Fig). Moreover, setting a ploidy of 6 improves CNA calling by doubling (108% increase) regions called as gains or losses which were missed because of compression effects as can be clearly seen from visual analysis of profiles.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Determination of single cell ploidy.

Analysis of one single cell from the near-diploid cell line NCI-H23 analyzed using a main ploidy of 2 (red) and 3 (blue): a) copy number profiles along 22 chromosomes; b) copy number levels distribution; c) density estimated by KDE; peaks detected are indicated as dashed vertical lines; d) linear regression of peak values over putative underlying copy numbers: clearly peaks obtained with a main ploidy of 2 better approximate the regression line compared to those obtained at a main ploidy of 3.

https://doi.org/10.1371/journal.pone.0193689.g006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Absolute copy number CNA calling in a single cell of hyperhexaploid cell line NCI-H661.

Plots of copy number profiles along the 22 autosomes expressed as absolute copy numbers. In a) and b) profiles obtained from the same sequencing data with main ploidy parameter set to 2 and 6 respectively. Significant copy number gains and losses are highlighted in red and blue respectively. Clearly a main cell ploidy = 6 provides a better fit of profiles with segmented data (black lines) and improves CNA calling. CNA calls only detected with main ploidy = 6 are shaded in green.

https://doi.org/10.1371/journal.pone.0193689.g007

Copy-number profiling of CTCs

To test the assay on real clinical samples, we analyzed single CTCs and corresponding control white blood cells (WBC) from 1 patient with prostate cancer and 2 patients with lung cancer for a total of 15 CTCs and 7 WBCs. Cells were sorted with DEPArray^™ system and genomic DNA was processed with Ampli1^™ LowPass workflow. In average 854,484 reads were sequenced for each sample and data were analyzed according to the bioinformatics protocol outlined (S1 Table).

Prostate cancer CTCs and WBCs were collected from a patient with late stage metastatic castration- resistant disease with increasing CTC counts despite therapy prostate cancer. For all the CTCs a main ploidy of 2 was predicted and CNA analysis showed, on average, 660 Mbp of the genome impacted by copy gains and losses (S30 Fig). Conversely, WBCs showed a flat profile. Notably, total amount of copy number alterations ranged from 536 Mbp to over 1.2 Gbp. Cluster analysis of copy number profiles highlighted a main group of 6 CTCs with small or no differences among different cells (Fig 8a; cluster A), a cluster (B) corresponding to the 2 WBCs and a single CTC (CTC 6) showing a divergent profile compared to the main group of CTCs. Systematic analysis of CNAs revealed 15 losses common to all the cells of cluster. Most of these ‘core’ alterations were also present in CTC 6 (74% of length) indicating a common aberrant genetic background for all the cells analyzed. At the same time the analysis shows also alterations specific to single CTCs: single cell CTC 3, while belonging to the cluster of the 6 most similar CTCs and sharing all the losses with them, have a large copy gain accounting for more than 120 Mbp on chromosome 8 and CTC 6 has 534 Mbp of copy gains not present on the other cells (Fig 8a).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Cluster analysis of copy number profiles for CTCs and WBCs from 3 patients.

a) single cells (CTCs and WBCs) from a patient affected by prostate cancer; cluster A represents 6 CTCs with small or no differences in copy number profiles; cluster B is formed by WBCs clustering, as expected, on a distinct branch of the tree. b,c) single cells (CTCs and WBCs) from 2 patients affected by lung adenocarcinoma. Values are expressed as fold changes respect to the main ploidy.

https://doi.org/10.1371/journal.pone.0193689.g008

The second patient studied suffered from lung adenocarcinoma. All the 4 CTCs were assigned ploidies > 2 with alterations covering from 1.18 Gbp to 1.64 Gbp of genome (S31 Fig, Fig 8b). Profiles clearly showed a common genomic background with 42% of losses with respect to the main ploidy and 23% of gains with respect to the main ploidy shared among all the cells.

The third dataset was also obtained from a patient affected by lung adenocarcinoma; in this case all the 4 CTCs were assigned a ploidy > 2. Despite the limited number of cells analyzed the method highlights a huge cell-to-cell variation (S32 Fig; Fig 8c). Globally 2.06 Gbp of genome have a variation respect to the main ploidy in at least 1 of the 4 CTCs. However only 127 Mbp ‘core’ gains are shared among all the CTCs and CTC-specific CNAs range in size from 617 Mbp to 1668 Mbp. Again this demonstrates that Ampli1^™ LowPass is able to capture the huge heterogeneity and cell-to-cell variation between different CTCs of the same patient.