The Potential of Digital Image Analysis to Determine Tumor Cell Content in Biobanked Formalin-Fixed,Paraffin-Embedded Tissue Samples

Abstract

Introduction:

Best practices dictate that biobanks ensure accurate determination of tumor content before supplying formalin-fixed, paraffin-embedded (FFPE) tissue samples to researchers for nucleic acid extraction and downstream molecular testing. It is advisable that trained and competent individuals, who understand the requirements of the downstream molecular tests, perform the microscopic morphological examination. However, the special skills, time, and costs associated with these assessments can be prohibitive, especially in large case cohorts requiring extensive pathological review. Determination of tumor content reliably by digital image analysis (DIA) could represent a significant advantage if validated, utilized, and deployed by biobanks.

Materials and Methods:

Whole slide digital scanned images of colorectal, lung, and breast cancer specimens were created. The scanned images were imported into the DIA software QuPath and digital annotations were completed by biobank technicians, under the direction of trained histopathology senior scientists. Automated cell detection was conducted and tumor epithelial cells were classified and quantified.

Results:

DIA scores were highly concordant with the manual assessment for 376 of 435 samples (86%). A detailed review of discordant cases indicated digital scores had a higher accuracy than the manual estimation.

Conclusion:

Automated digital quantification has the potential to replace visual estimations with reduced subjectivity and increased reliability compared with manual tumor estimations. We recommend the use of DIA by biobanks involved in provision of FFPE tissue samples, especially in large research studies requiring high volumes of cases to be analyzed.

Introduction

Biobanks that supply formalin-fixed, paraffin-embedded (FFPE) tissue samples to researchers for nucleic acid extractions and downstream molecular testing should ensure accurate determination of tumor content in each sample. Traditionally, trained histopathologists, who are competent to interpret tissue morphology and who understand the requirements of the molecular tests, perform each microscopic morphological examination. Microscopic assessment of Hematoxylin and Eosin (H&E) slides is required to confirm sample suitability and ensure that annotated regions of tissue represent the true histopathological diagnosis.¹ However, macrophages and lymphocytes can also present intratumorally, although to varying extents, which may make assessment problematic. The percentage of tumor cells in an annotated region of interest (ROI) must be high enough to meet the minimum sensitivity of the downstream molecular testing requested.² Manual annotation and estimation of percentage tumor content in H&E slides by trained histopathologists remains a best practice,³ however, undertaking this in a large biobank-supported research project could be time consuming and therefore costly to manage.

Proposed downstream assays must be considered when assessing the percentage of tumor cells in a sample, as each molecular test has a limit of analytic sensitivity that is, a minimum percentage of tumor cells required to identify a mutation or copy number alteration.^4,5 Overestimation of tumor content may be more detrimental than underestimation, as it can result in false negative results due to the number of variant alleles in tumor cells being diluted out by wild-type alleles.^1,5 To allow molecular findings to be fully evaluated and confidently reported, it is essential that an accurate figure for tumor content of processed sections can be determined.⁶ This is not only important for FFPE samples that undergo macrodissection, but also for samples, such as needle core biopsies that are too small for macrodissection.

In a previous related study, 194 laboratories examined 10 colon cancer slides and the results revealed that pathologists were more accurate in estimating tumor cell percentage in cases of low tumor cells (<40%). Factors affecting accurate manual estimation included section thickness and H&E staining intensity. The study also concluded that histopathologists reasonably estimate tumor percentage, but are more likely to overestimate the result.¹

Digital scanning, on-screen annotation, and subsequent digital image analysis (DIA) promise objective and reproducible results. Viray et al.⁴ attempted to develop and test an algorithm to calculate tumor cell percentage, for which they used 25 colon adenocarcinoma cases. Validation images and results were compared with manual pathologist scores. Errors were due to misclassification of tumor and stroma by the software algorithm, suggesting in this study that DIA may be successful for small areas of tissue sections or tissue microarray cores, but less accurate in whole slide image (WSI) sections.⁴

Manual glass-slide annotation, even by trained individuals, can be challenging in cases with low tumor content and may lead to imprecise results. Hamilton et al.⁷ highlighted the importance of tumor annotation on H&E WSIs to enrich for tumor cell content for macrodissection and DNA extraction. Digitizing H&E glass slides could enable trained operators, who may not be clinically qualified, to annotate tumor regions of interest and subsequently classify viable tumor from stromal cells to calculate overall viable tumor content. Quality control by more experienced qualified histopathology staff could make the process less time consuming and more efficient.

In this quality assurance project, trained Northern Ireland Biobank (NIB) technical staff utilized the open-source DIA software, QuPath, to annotate scanned pathology images on a digital screen. Automated cell detection was applied and tumor cells were classified and quantified using published methodology.^8–11 A combination of digital on-screen annotations, automated tumor percentage calculations, and specific quantitation of malignant nuclei cell numbers under annotated areas is a rapid and robust solution for provision of accurate data from biobanks to researchers, maximizing the potential to capture all molecular aberrations with accurate determination of assay sensitivity levels. In addition, this approach will be faster and therefore good for large case cohorts.

Materials and Methods

Tissue samples

Four hundred thirty-five representative tumor H&Es, all from surgical resections, were available for analysis, 149 from colorectal cancers (CRC), 138 from lung cancers, and 148 from breast cancers (BC). Cases were prospectively collected following patient consent by NIB (ethics approval number 21-NI-0019).¹² Each slide was microscopically reviewed by experienced histopathology staff to confirm the presence of tumor tissue before entry to NIB. The CRC cases consisted of 144 primary colorectal adenocarcinomas and 5 metastatic adenocarcinomas. The lung tumor cases were composed of 113 primary adenocarcinomas, 15 squamous cell carcinomas, 2 adenosquamous carcinomas, 4 large cell neuroendocrine tumors, 3 poorly differentiated carcinomas, and 1 carcinoid tumor. The breast tumors consisted of 118 infiltrating ductal carcinomas (IDC), 16 lobular carcinomas, 6 with mixed IDC and lobular features, 3 micropapillary carcinomas, 2 mucinous carcinomas, and 1 each of phyllodes tumor, tubular carcinoma, and noninfiltrating intracystic carcinoma.

As the slide collection spanned a period of 3 years, an initial quality control review of H&E staining was conducted. In the vast majority of cases, the quality of the H&E staining, while variable, did not impact the ability of the DIA software to classify tumor versus stroma. Only four cases were excluded from analysis as the H&Es were considered too thick or too lightly stained to be suitable for image analysis (IA). Therefore, this led to a high level of confidence in the calculated tumor cell content within the annotated area.

From 30 of the cases (10 from each tumor type), a second slide was annotated for an area of tumor and the DIA classifiers, which had been set for the first tumor block of the case was applied.

Digital image analysis

All slides were digitized using an Aperio AT2 slide scanner at × 40 magnification. Images were imported into QuPath (version 0.1.2). Annotation tools were utilized by trained biobank technicians to digitally annotate representative tumor areas onto WSI H&Es, ensuring removal of unsuitable areas for analysis, for example, tissue folds, necrosis, and normal epithelium, as described. Annotations were quality controlled by experienced histopathology staff, following a robust workflow.^11,13–15 Briefly, within the annotations, cell detection was undertaken using default parameters within QuPath. Training objects for tumor and stroma were created over detected cells to train a classifier using the random forest method. The percentage of viable tumor nuclei, stromal nuclei, and total number of cells detected within the annotated area of tissue were exported from the software, once the classifier passed extensive quality control steps.

For the purpose of this study, regions of in situ disease in all cancer types were excluded from analysis. For lung samples, where it was possible, regions of extensive macrophage infiltration were also excluded.

Manual assessment of tumor percentage within digital annotations

Each representative area of digitally annotated tumor was assessed visually by trained histopathology senior scientists and comparative estimation of tumor content was recorded. Assessors were blinded to the tumor percentage score determined by IA. This ensured that the same area was used to estimate tumor by both methods.

An upper limit of 20% difference was set as the threshold for acceptable concordance based on previous related research. Cases in the present study were classified as discordant if there was >20% difference between the DIA-generated score and manual estimation.

Results

This study compared two methods to determine the percentage of tumor cell content in resection blocks sampled from three different tumor types (colorectal, lung, and breast). DIA software was used to create an annotated ROI on digitized images and the percentage of tumor cells was calculated within the same ROI either by (1) using QuPath software following cell detection and classification of tumor or (2) manually following visual estimation. Those undertaking the manual assessments were blinded to the DIA results.

Figure 1 shows the concordance between the DIA and manual assessments of tumor percentage in (1) CRC, (2) lung, and (3) BC cohorts. Three hundred seventy-six of 435 cases analyzed by manual and DIA were concordant (86%), with 59 cases discordant (14%) at a 20% threshold.

FIG. 1.

Scatterplots for the distribution of tumor percentage scores from manual and DIA assessment of (a) 149 CRC cases (b) 138 lung cancer cases, and (c) 148 BC cases. Green data points represent cases whose percentage tumor content were found to be concordant. Red data points represent cases whose scores were discordant defined as >20% difference between assessments. BC, breast cancer; CRC, colorectal cancers; DIA, digital image analysis.

In CRC 142 of 149 cases were concordant (95%) with a p-value = 0.004 and R² = 0.702. For lung, 108 of 138 cases were concordant (78%) with a p-value = <0.001 and R² = 0.44 and in the BC cases 126 of 148 were concordant (85%) with a p-value = <0.001 and R² = 0.498. All correlations were assessed by Spearman's rank.

Across all three tumor types, in the discordant cases, 43 of 59 (73%) were assessed as containing higher tumor cell content by manual estimation than by DIA. However, of these cases, 40 (68%) were from the lung and BC cohorts. Only seven discordant CRC cases were discordant with three assessed as having higher tumor cell content by manual estimation than by DIA.

Within these 59 discordant cases, 41 cases were discordant between 20% and 29% (20 lung, 15 breast, and 6 CRC), 15 cases were discordant between 30% and 39% (8 lung, 6 breast, and 1 CRC), and 3 cases were discordant between 40% and 50% (2 lung and 1 breast). The majority of the discordant cases were lung (30 cases). This is likely a reflection of the heterogeneity of lung tumor pathology, which tends to be less uniform than either breast or CRC tumor.

Reproducibility of DIA across slides from the same case

In a subset of 30 cases, a second tumor slide was annotated and DIA classifiers for tumor and stroma, which were used for the first slide from the case were applied. In all cases, the tumor and stroma classifiers applied to the second slide were assessed as accurate in an expert manual review. Examples from each tumor type, breast, lung, and CRC, are shown in Supplementary Figures S1–S3, respectively.

Directionality and quality control of discordant cases

Looking at directionality of the differences between manual and QuPath estimates across all cases, manual estimates were higher than DIA in 250 of 435 (57%), lower in 166 of 435 (38%), and the same in 19 cases (4%). Splitting the data into primary cancer subtypes showed that for breast and lung, manual estimates were in general higher than those of QuPath (70% and 66% of cases showed higher manual estimates, respectively), whereas for CRC, the opposite was true, with only 37% of cases having higher manual estimates (Fig. 2). CRC also had less discrepant cases overall than the other two cancer types.

FIG. 2.

Graph showing the % difference between manual and QuPath estimates across all cases, split into the three primary tumor types: breast, CRC, and lung, respectively. Differences are calculated by subtracting QuPath scores from manual—anything above the zero line shows manual estimates were higher than those from QuPath, while bars below the line show lower manual estimates. Green bars represent concordant and red discordant cases. Gray shaded areas show the concordance thresholds (±20%).

Extensive quality control review of all discordant cases highlighted that overestimation of tumor cell content by manual estimation was invariably due to an under-recognition of high numbers of weakly stained stromal lymphocytes or fibroblasts when viewed at low magnifications on-screen. In a specific CRC case, manual estimation was 50% tumor cells but only 18% as scored by DIA (Fig. 3a); in a lung case, manual estimation was 65% tumor cells but only 42% by DIA (Fig. 3b); and in a breast case example, manual estimation was 85% and only 61% by DIA (Fig. 3c). In all these examples DIA with robust classifiers is able to more readily detect all cells and therefore to provide more accurate data. Furthermore, in cases where a tumor was more generally composed of small noncohesive tumor islands dispersed widely across the sample, tumor/stroma classifiers in DIA were able to provide specific cell numbers unlike subjective manual estimations. An example whereby manual estimation (40% tumor cells) was much lower than DIA (61% tumor cells), is highlighted in Figure 3d. In this example, the acellular nature of the stromal compartment is highlighted by the IA classifier leading to a higher tumor content calculation.

FIG. 3.

Representative image sets in annotated regions of tumor that were discordant between manual estimation and DIA score. (a) Colorectal cancer, (b) lung cancer, and (c) BC. In examples, (a–c) the manual method underestimated the number of stromal cells, which are more readily recognized by the DIA stromal cell identifiers shown in (iii). (d) Example from BC in which areas of the stroma are acellular and therefore the percentage of tumor estimated by DIA was much higher than that assessed manually.

Of all cases correlated between manual and DIA 376 were concordant and 59 were discordant. Using only the DIA cell counts, we assessed whether the total cell number within the sample contributed to the concordance of the results. Frequency plots were evaluated for all DIA cases. The distribution of total cell number does not differ between concordant and discordant cases regardless of tumor type (Fig. 4a). Within each tumor type, the concordance and discordance are not markedly different (Fig. 4b). These data indicate that total cell number, high or low, does not negatively affect the ability of DIA to determine tumor percentage reliably.

FIG. 4.

Frequency graphs displaying the distribution of total cell number by DIA. (a) Total cell number analyzed regardless of tumor type. Top panel displays concordant data while the bottom panel displays discordant data. (b) Total cell number analyzed by specific tumor types. Top three panels display concordance data for BC cases, CRC cases, and lung cases while the bottom three panels display discordance data for each tumor type.

Discussion

The percentage tumor evaluation and through enrichment of tumor cell populations through macrodissection, where possible, in FFPE samples underpins molecular testing both in diagnostics and research.^5,7 For biobanks, the ability to supply precise information on tumor content within annotated areas of tumors has a bearing on test accuracy in downstream technologies such as Next Generation Sequencing (NGS) or transcriptomics and therefore can assist in correct interpretation of complex data. The potential for a DIA approach for tumor determination would be a major advantage for biobanks. It would mean that tumor content in cases from large cohorts of FFPE tissue samples could be quickly and accurately determined while minimizing hands-on time.

In the present study, a blinded visual estimation was performed on a large series of colorectal, lung, and breast tumor resection samples using percentage tumor DIA data generated by experienced and trained biobank technicians, which was then compared with manually assessed percentage tumor within the same annotated area in digitized H&E slides by trained histopathology senior scientists. Statistically significant positive correlations were found between the manual estimation and DIA calculation for all three tumor types under investigation. Manual assessment of tumor content has been shown in many studies to be subjective and lacking in reproducibility.^4,5 We chose less than 20% discrepancy for concordance based on a previous report, which assessed automated tumor analysis for molecular profiling that indicated, in 136 H&E slides, the discrepancy between two pathologists was over 20%.⁷ The present study shows that DIA software programs, such as QuPath, have the potential to be a more reproducible method than manual tumor estimations.

The main advantage of DIA is to allow the user to define an algorithm to identify tumor cells versus stromal cells within a digitally annotated area for all cancer types. However, in this study it was not possible to set one classifier identifying tumor cells against stromal cells in annotated regions of interest and then to apply it to all digitized H&E slides for a given cancer type. This resulted in the classifier having to be constantly reoptimized. Automated recognition of tumor with potential for the use of artificial intelligence (AI) is an area of much ongoing research. For example, in a recent article assessing breast tumor cellularity deep neural networks were used by the authors in an end-to-end approach that predicted the cellularity score without any intermediate segmentation steps.¹⁶ Although attractive and apparently robust, the authors concluded that unsupervised segmentation lacked interpretability, which is particularly important for the application of AI approaches ahead of downstream testing. We believe our method of combining manual annotations on digitized images and subsequent application of automated tumor cell quantification presents an ideal balance.

Few research studies to date have used DIA methods for calculating the percentage tumor in H&E scanned slides of cancer. Hamilton et al.⁷ used DIA to generate a percentage tumor score for 10 lung cancer resection cases in a small validation study. This study recorded manual counts of tumor and stromal cells in 1 mm² areas as “baselines” to which automated scores were compared. They demonstrated significant concordance with manual cell counts (p < 0.001).⁷ However, this study has not been verified in a larger cohort by more recent literature.

QuPath DIA system, developed in-house at Queen's University Belfast, has been previously demonstrated to generate tumor and stromal percentage scores in CRC resection cases, linked to patient prognosis.⁸ Although in this study, scores were not compared with any other method to estimate percentage tumor, it highlights the potential of DIA to generate robust percentage tumor analysis on H&E images. Furthermore, QuPath has been used successfully in several IHC studies to score a wide variety of biomarkers in different sample types. In one study by Bankhead et al., QuPath was applied to BC tissue microarrays to score 5 biomarkers, including ER, PR, and Her2. QuPath scores were compared with pathologist manual scores and demonstrated good concordance between both methods for all biomarkers.^9,11

In the present study, QuPath classifiers for tumor and stroma cells provided specific cell counts that were more likely to be accurate and reproducible than manual estimations and therefore provided more robust tumor identification and tumor estimation values. This was especially true in lung and BC samples, whereby lightly stained stromal cells were underestimated in manual assessments but clearly identified by DIA.

Beyond the scope of the present study, the ability to use DIA may be important in the determination of specific cell counts in cell blocks or indeed in evaluating tissues stabilized in methods other than FFPE, such as frozen sections, with the proviso that a H&E of sufficient quality can be produced from these alternative methods. Future work would also aim to investigate the extent of time required by histopathologists and technicians to undertake the proposed DIA workflow, to more specifically report on the utility of such a process. Numerous studies have shown that cell blocks, especially from NSCLC cases are suitable for molecular testing.^17,18 Currently, due to the variability and potential inaccuracy of tumor content estimates from H&E slides, many efforts have concentrated on attempting to determine tumor purity from molecular data. However, many of these require data from matched normal samples,^19,20 high-throughput sequencing,^21–23 or SNP data.²⁴ To date, concordance between these algorithms is poor, with no gold standard yet identified.²² An accurate value for the percentage of tumor cells within a sample before molecular analyses would eliminate the need for these complex algorithms.

Accurate percentage tumor content has consequences in terms of variant detection when samples are used for NGS. If a variant allele fraction (VAF) is below 3%, in our experience, it is commonly unreported. For example, a case containing 6% tumor cells (taking a 3% VAF detection threshold), would require 100% of tumor cells containing a heterozygous variant to be sequenced. The likelihood of this attainment is small, particularly for FFPE sample types. Furthermore, this example assumes tumor homogeneity. In reality, the clonal evolution of tumors results in intratumoral heterogeneity requiring higher tumor content for VAF detection. It is therefore vital that tumor content is accurately determined upstream of NGS. Overestimation of tumor content within a sample can lead to suboptimal samples being processed and has the potential to lead to false negative results—the lack of, or low proportion of, variants found in a sample with a reportedly high tumor content would likely be below the threshold for reporting. If that sample had a low tumor content that variant may be a true finding, and potentially be reportable and actionable. Conversely, underestimations of tumor content can lead to samples that could yield reportable results not being processed and repeat samples being requested. The results of this study show that manual estimates of tumor content were generally higher than those from DIA, although this appeared to be somewhat cancer type specific. This appears to support previous studies showing that pathologists may overestimate the percentage of tumor cells within a sample.⁵

In conclusion, the present study highlights that DIA tools used by trained biobank technicians to perform annotations and tumor content determinations, quality control by experienced histopathology staff, has the potential to provide a reliable quantitative workflow that is beneficial to a busy biobank. DIA is not only able to calculate the percentage of tumor cells, but also the total number of tumor cells within the area for macrodissection before molecular analyses or other downstream applications. Replacing visual estimations with automated methods could solve the issues of subjectivity and unreliability in manual tumor estimations.

Footnotes

Acknowledgments

The samples used in this quality assurance project are housed in the Northern Ireland Biobank, which has received funds from HSC Research and Development Division of the Public Health Agency in Northern Ireland and the Friends of the Cancer Center.

Author Disclosure Statement

No conflicting financial interests exist.

Funding Information

Northern Ireland Public Health Agency, Health & Social Care Research & Development Division Ref: SPI/5151/15.

Supplementary Material

References

Viray

, Li

, Long

, et al. A prospective, multi-institutional diagnostic trial to determine pathologist accuracy in estimation of percentage of malignant cells. Arch Pathol Lab Med, 2013; 137:1545–1549.

Salto-Tellez

, James

, Hamilton

. Molecular pathology—The value of an integrative approach. Mol Oncol, 2014; 8:1163–1168.

Aeffner

, Wilson

, Martin

, et al. The Gold Standard Paradox in digital image analysis: Manual versus automated scoring as ground truth. Arch Pathol Lab Med, 2017; 141:1267–1275.

Viray

, Coulter

, Li

, et al. Automated objective determination of percentage of malignant nuclei for mutation testing. Appl Immunohistochem Mol Morphol, 2014; 22:363–371.

Smits

, Kummer

, de Bruin

, et al. The estimation of tumor cell percentage for molecular testing by pathologists is not accurate. Mod Pathol, 2014; 27:168–174.

Morris

, Subramanian

, Gel

, et al. Performance of next-generation sequencing on small tumor specimens and/or low tumor content samples using a commercially available platform. PLoS One, 2018; 13:e0196556.

Hamilton

, Wang

, Boyd

, et al. Automated tumor analysis for molecular profiling in lung cancer. Oncotarget, 2015; 6:27938–27952.

Bankhead

, Loughrey

, Fernández

, et al. QuPath: Open source software for digital pathology image analysis. Sci Rep, 2017; 7:16878.

Bankhead

, Fernández

, McArt

, et al. Integrated tumor identification and automated scoring minimizes pathologist involvement and provides new insights to key biomarkers in breast cancer. Lab Invest, 2018; 98:15–26.

10.

Humphries

, Hynes

, Bingham

, et al. Automated tumor recognition and digital pathology scoring unravels new role for PD-L1 in predicting good outcome in ER-/HER2+ breast cancer. J Oncol, 2018; 2018:2937012.

11.

Humphries

, McQuaid

, Craig

, et al. Critical appraisal of programmed death ligand 1 reflex diagnostic testing: Current standards and future opportunities. J Thorac Oncol, 2019; 14:45–53.

12.

Lewis

, McQuaid

, Clark

, et al. The Northern Ireland Biobank: A cancer focused repository of science. Open J Bioresour, 2018; 5:9.

13.

Loughrey

, Bankhead

, Coleman

, et al. Validation of the systematic scoring of immunohistochemically stained tumor tissue microarrays using QuPath digital image analysis. Histopathology, 2018; 73:327–338.

14.

Humphries

, Bingham

, Abdullahi Sidi

, et al. Improving the diagnostic accuracy of the PD-L1 test with image analysis and multiplex hybridization. Cancers (Basel), 2020; 12:1114.

15.

Humphries

, Craig

, Kacprzyk

, et al. The adaptive immune and immune checkpoint landscape of neoadjuvant treated esophageal adenocarcinoma using digital pathology quantitation. BMC Cancer, 2020; 20:500.

16.

Rakhlin

, Shvets

, Kalinin

, et al. Breast tumor cellularity assessment using deep neural networks. bioRxiv, 2019; 1:1–10. DOI: 10.1101/628693

17.

Malapelle

, Bellevicine

, Zeppa

, et al. Cytology-based gene mutation tests to predict response to anti-epidermal growth factor receptor therapy: A review. Diagn Cytopathol, 2011; 39:703–710.

18.

Lindeman

, Cagle

, Beasley

, et al. Molecular testing guideline for selection of lung cancer patients for EGFR and ALK tyrosine kinase inhibitors: Guideline from the College of American Pathologists, International Association for the Study of Lung Cancer, and Association for Molecular Pathology. Arch Pathol Lab Med, 2013; 137:828–860.

19.

, Zhang

, et al. PurityEst: Estimating purity of human tumor samples using next-generation sequencing data. Bioinformatics, 2012; 28:2265–2266.

20.

Carter

, Cibulskis

, Helman

, et al. Absolute quantification of somatic DNA alterations in human cancer. Nat Biotechnol, 2012; 30:413–421.

21.

Yuan

, Li

, Zhao

, et al. Accurate inference of tumor purity and absolute copy numbers from high-throughput sequencing data. Front Genet, 2020; 11:458.

22.

Patel

, Jo

, Eberhard

, et al. Improved tumor purity metrics in next-generation sequencing for clinical practice: The Integrated Interpretation of Neoplastic Cellularity and Sequencing Results (IINCaSe) Approach. Appl Immunohistochem Mol Morphol, 2019; 27:764–772.

23.

Oesper

, Satas

, Raphael

. Quantifying tumor heterogeneity in whole-genome and whole-exome sequencing data. Bioinformatics, 2014; 30:3532–3540.

24.

Van Loo

, Nordgard

, Lingjærde

, et al. Allele-specific copy number analysis of tumors. Proc Natl Acad Sci U S A, 2010; 107:16910–16915.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

6.63 MB