Massively Parallel Sequencing of Esophageal Brushings Enables an Aneuploidy-Based Classification of Patients with Barrett’s Esophagus

Abstract

BACKGROUND & AIMS:

Aneuploidy has been proposed as a tool to assess progression in patients with Barrett’s Esophagus (BE), but has heretofore required multiple biopsies. We assessed whether a single esophageal brushing that widely sampled the esophagus could be combined with massively parallel sequencing to characterize aneuploidy and identify patients with disease progression to dysplasia or cancer.

METHODS:

Esophageal brushings were obtained from patients without BE, with non-dysplastic BE (NDBE), low-grade dysplasia (LGD), high grade dysplasia (HGD), or adenocarcinoma (EAC). To assess aneuploidy, we employed RealSeqS, a technique that uses a single primer pair to interrogate ∼350,000 genome-spanning regions and identify specific chromosome arm alterations. A classifier to distinguish NDBE from EAC was trained on results from 79 patients. An independent validation cohort of 268 subjects was used to test the classifier at distinguishing patients at successive phases of BE progression.

RESULTS:

Aneuploidy progression was associated with gains of 1q, 12p, and 20q and losses on 9p and 17p. The entire chromosome 8q was often gained in NDBE, whereas focal gain of 8q24 was identified only when there was dysplasia. Among validation subjects, a classifier incorporating these features with a global measure of aneuploidy scored positive in 96% of EAC, 68% of HGD, but only 7% of NDBE.

CONCLUSIONS:

RealSeqS analysis of esophageal brushings provides a practical and sensitive method to determine aneuploidy in BE patients. It identifies specific chromosome changes that occur early in NDBE and others that occur late and mark progression to dysplasia. The clinical implications of this approach can now be tested in prospective trials.

Graphical Abstract

Esophageal adenocarcinoma (EAC) is a rapidly rising ¹, refractory cancer, with an overall five- year survival that remains below 20%.² Barrett’s esophagus (BE), a condition of intestinal metaplasia of the distal esophagus associated with chronic gastroesophageal reflux disease (GERD) and other risk factors, is the only known precursor to EAC. BE progresses stepwise from metaplasia (non-dysplastic BE, NDBE), to low grade dysplasia (LGD), then to high grade dysplasia (HGD), and finally to carcinoma.^3–5 Although most BE cases do not progress, those that do progress to cancer likely do so through acquired genetic and epigenetic alterations.^6–12 Current strategies for prevention and early detection of EAC are based on endoscopic detection of BE, with subsequent surveillance to identify dysplasia on random biopsies obtained during endoscopy, followed by endoscopic ablation of the dysplastic tissue.^{3–5, 13, 14} While surveillance and ablation of dysplasia is effective in reducing the risk of developing invasive cancers, sampling with random biopsies is inherently imprecise. In addition, there remains a well-recognized risk of development of interval cancers between surveillance examinations.^{15, 16} Perhaps most importantly, sole reliance on histopathologic criteria for recognition of dysplastic BE is imperfect, as reflected by poor inter-observer agreement even among expert pathologists.^17–19 Moreover, the related inability to identify the majority of patients with BE who are at low risk for progression to cancer leads to over-surveillance of the total population of BE patients.²⁰

Aneuploidy and tetraploidy have long been recognized to accompany progression from non-dysplastic BE to dysplasia and early EAC and have been proposed to be predictive biomarkers for identifying BE at high risk of such progression.^{11, 21–27} Methods for detecting aneuploidy have traditionally employed biopsies that capture only focal regions of the affected esophageal segment, and that therefore must be repeated across multiple locations to obtain, at best, a somewhat representative sampling. In part for this reason, aneuploidy has not been incorporated into routine clinical practice. Additionally, the relationship between aneuploidy, the risk of progression, and actual progression has not yet been defined.^{11, 21–27} Other approaches to assess the risk of progression in patients with BE include in vitro imaging and flow cytometry of biopsied material. These approaches have also not been widely used, either because they are technically cumbersome, are low throughput, or have special requirements.^28–30

Esophageal brushing is a method to conveniently sample an extensive area of esophageal epithelium. However, the cells collected from brushings represent a mixture of normal epithelium, non-dysplastic Barrett’s epithelium, and dysplastic epithelium, thereby substantially diluting the genomic signal originating from the dysplastic cells. To evaluate aneuploidy in such brushings, a technique that can sensitively detect aneuploidy in a mixed cellular population is required. The Repetitive Element Sequencing System (RealSeqS) is a recently described massively parallel sequencing (MPS) approach that was designed for the detection of aneuploidy in plasma samples containing low levels of DNA derived from neoplastic cells (Fig. 1).³¹ We reasoned that this method might be applicable to the detection and characterization of aneuploidy in esophageal brushings because of the same underlying challenge, i.e., detection of aneuploidy in DNA derived from a mixture of a low number of aneuploid cells with a much larger number of euploid cells.

Figure 1: Overview of the RealSeqS approach.

A) A single primer pair concomitantly amplifies ∼350,000 unique loci spread throughout the genome. B) The patient sample is matched to the 7 closest control samples C) The statistical significance of gains and losses for each of the 39 non-acrocentric chromosome arms is calculated. D) The 39 chromosome arms are integrated into a Genome Aneuploid Score (GAS) using a supervised machine learning model. E) Chromosome arm levels can be quantified and focal changes of interest queried.

The implementation of this concept required evaluation of DNA from a relatively large number of esophageal brushings as well as re-normalization of the basic algorithms used to evaluate the RealSeqS data from previous experiments on plasma. In this paper, we describe this implementation and apply it to the evaluation of samples from patients with BE at various stages of disease. The primary aim of the clinical component of this study was to develop a classifier based on RealSeqS data to distinguish the progressive phases of BE. The secondary aim was to identify specific chromosomal alterations that tracked with disease progression.

Methods

Patient Samples

The study was cross sectional in design. Patients were recruited prior to esophagogastroduodenoscopy (EGD) as part of an Institutional Review Board-approved, multi-institutional study to develop biomarkers for Barrett’s esophagus (BE) and esophageal adenocarcinoma (EAC). The six participating institutions are tertiary centers that care for patients referred for management of dysplastic BE and EAC. Brushings were obtained before any biopsies were taken, but because the study cohort was in general individuals referred for tertiary care, many patients had already obtained a diagnosis. Patients in general underwent sampling and endoscopy on the day they were enrolled in the study. Brushings of patients with BE or LGD were of the entire BE segment, while brushings of patients with known HGD and EAC sampled a 3 cm patch targeted to include any nodularity, depression, or irregularity, as these areas would most likely contain the highest grade lesion. Ninety-nine esophageal brushings from concurrently enrolled subjects without BE were also performed with an approximately 3-cm long brushing obtained that covered the gastroesophageal junction and distal esophagus. Following retrieval, brushes were cut with scissors into cryovials, snap frozen at bedside, then stored at −80°C.

Seventy-nine patients were sampled in a Training Set and 258 patients in a Validation Set. Duplicate esophageal brushings were obtained at the same clinical session in 22 participants and were used to assess assay concordance.

Histopathologic Diagnosis

The diagnosis of NDBE, LGD, or HGD was primarily determined by histopathology of the biopsy obtained after brushing at the time of study entry. Slides were retrievable for central pathology review for 76% of the Validation Set cases. For all other study cases, diagnoses were as determined by expert GI pathologists at the respective enrollment centers. The presence of surface epithelium was confirmed on all reviewed biopsies. When biopsies were not performed on the day of the brush sampling, the results of biopsy from EGDs performed within the previous three months were used. For study purposes, cases with pathology described as focal HGD were classified as HGD. Intramucosal cancer was classified as EAC.

Detection of Chromosome Arm Alterations

A single primer pair was used to amplify _∼350,000 loci spread throughout the genome (Fig. 1).³¹ One of the primers included a unique identifier sequence (UID) as a molecular barcode of 16 degenerate bases to reduce error rates associated with sequencing, performed on an Illumina HiSeq 4000. The average number of uniquely aligned reads was 11.2 Million (M) (interquartile range 9.7 M-12.8 M). Sequencing data were processed to identify single chromosomal arm gains or losses using the Within-Sample AneupLoidy Detection (WALDO) algorithm incorporated into the RealSeqS work flow.³¹ Fifteen esophageal brushings from individuals without BE were used as reference samples; these were excluded from all other analyses. Each experimental sample was then matched to the reference samples that were most similar to it with respect to the amplicon distributions generated by RealSeqS.³¹ The reference samples included four individuals with evidence of esophagitis, one as determined by histopathology and three as noted on endoscopy.

The WALDO algorithm compares the normalized read counts of 500 kb intervals to intervals on other chromosome arms in the same sample. Its normalization is therefore internal, “within-sample.” ^{31, 32} The intervals are aggregated across the entire length of the chromosome arm to produce an arm level statistical significance score (Z_w). The 39 non-acrocentric Z_w values serve as features that are integrated and modeled with a support vector machine (SVM) to provide a summary Global Aneuploidy Score that discriminates between aneuploid and euploid samples. The SVM classifier was trained on 1,334 normal euploid plasma samples and 2,016 in silico aneuploid samples generated from the normal plasma samples.³¹ The in silico samples were generated to mimic recurrently altered chromosome arms observed in cancers, including esophageal cancers.³¹ In addition to generating arm level statistical significances, the circular binary segmentation algorithm was applied to identify sub-chromosomal focal alterations.³³ Note that the SVM classifier and segmentation algorithm were identical to those used to evaluate previous data on plasma.³¹ The only difference in the algorithmic component used in the current study was that the 500 kb clusters used to define aneuploidy in the test sample were generated from seven matched esophageal samples from normal individuals rather than seven matched plasma samples from normal individuals.^{31, 32} RealSeqS, WALDO, and the SVM classifier were all performed by investigators blinded to the clinical classification of the associated samples. The software used to generate the scores reported in this manuscript are available at (https://doi.org/10.5281/zenodo.3656943)

Post-hoc Review of Clinical Follow-up

Although the study was cross sectional in design, limited follow-up data were available and were retrieved following completion of all aneuploidy analysis and classification. Follow-up results were retrieved by an investigator (AC) via review of the electronic health and pathology results on biopsies performed during surveillance endoscopies subsequent to the research brush sampling.

Results

Patient Characteristics

The cohort consisted of 79 patients in the training set (15 samples from normal gastroesophageal junctions, 19 with NDBE, 15 with HGD, and 30 with EAC) and 268 patients in the validation set (84 samples from normal gastroesophageal junction, 41 with NDBE, 32 with LGD, 28 with HGD, and 83 with EAC). LGD samples were not included in the training set because of known challenges in the reproducibility of expert pathologists in classification of LGD.^17–19 There were no statistically significant differences between the demographic compositions of the training and validation sets except for the racial makeup among the unaffected controls. The general demographics of the training and validation sets including race, gender, smoking history, and age are presented in Supplementary Tables S1, S2 and S3).

Training Set

We first evaluated aneuploidy in the Training Set. Z_w scores for each of the 39 non-acrocentric chromosome arms in each sample were calculated. These chromosome arm level Z_w scores were then integrated into a single Global Aneuploidy Score (GAS) reflecting the number and similarity of the alterations to those commonly observed in cancers (as per Methods). Our major goal was to discriminate patients with advanced lesions that require clinical intervention (HGD and EAC) from those with earlier lesions that do not (i.e. samples with NDBE). Sensitivity was determined by the fraction of patients with advanced lesions who had GAS greater than a given threshold. Analogously, specificity was determined as the corresponding fraction of patients with early lesions having GAS values less than this threshold. The receiver operating characteristic (ROC) curve of sensitivity versus 1-specificity is shown for thresholds ranging from highest to lowest and demonstrates an area under curve (AUC) of 0.864, consistent with GAS providing good discrimination between NDBE versus EAC and HGD (Fig. 2A).

Figure 2: Performance of the Genome Aneuploid Score (GAS) to discriminate samples from patients with HGD or EAC from samples from individuals with NDBE.

(A) The Receiver Operating Characteristic (ROC) curve and area under the curve (AUC) for the GAS metric as applied to the Training Set. (B) Violin plot of the GAS distribution among the clinical subsets of the Training Set. Individuals with LGD were excluded from the Training Set. (C) ROC curve and AUC for the GAS metric as applied to the Validation Set. (D) Violin plot of the GAS distribution among the clinical subsets of the Validation Set.

Violin plots of the individual distributions of GAS scores are shown in Fig. 2B. The distributions observed in samples from patients with a normal esophagus, with BE without dysplasia (NDBE), with BE with high-grade dysplasia (HGD), and with carcinoma (EAC) were each significantly different from one another (P <0.01 by the Kolmogorov-Smirnov test). Interestingly, a clear bimodal distribution of aneuploidy scores was observed in the BE patients without dysplasia (NDBE). One mode was centered at a GAS of 0.2 and the other at a GAS of 0.85. This suggested that a biologically appropriate threshold for a positive GAS should distinguish between these two groups of patients with NDBE. From inspection of the violin plots, a GAS value of 0.6 was chosen for this threshold. At the threshold of 0.6, GAS identified the patient group with advanced lesions of HGD or EAC at a sensitivity of 86.7% and showed specificity for not detecting patients with NDBE of 73.7% (Supplementary Table S4). We considered that for clinical applications, the medical consequence of missing an advanced lesion (false negatives) was higher than the cost of false positives, and accordingly decided to employ this threshold of 0.6 for all subsequent studies. Hereinafter, samples with a GAS>0.6 are referred to as “aneuploid” and those with GAS≤0.6 as “non-aneuploid”. Note that the non-aneuploid designation is relative; some of the samples with GAS≤0.6 had a small number of chromosome arm alterations, while all of the samples with GAS >0.6 had a larger number of chromosome alterations (Dataset 1).

Identifying Chromosome Changes That Distinguish Histologic Progression of BE

We next characterized which specific chromosome arm changes accompanied progression of BE to EAC, reasoning that emphasizing these specific chromosome alterations might further improve accuracy of the GAS for discriminating histologic progression of disease. To minimize overfitting, we restricted our investigation to the samples that were aneuploid (as defined by a GAS > 0.6) in the Training Set.

There were five aneuploid BE samples and 28 aneuploid EAC samples in the Training Set (Supplementary Tables S1 and S4). As there was a low number of samples and large number of possible chromosome arms involved in progression (39 possible gains and 39 possible losses), we sought to determine a minimal set of alterations enabling EAC discrimination. From careful manual inspection of the Training Set data (Supplementary Table S5), informed by previous studies of mutations and copy number alterations in cancer, we selected a panel of 5 candidate chromosome arms for this purpose: 1q gain, 9p loss, 12p gain, 17p loss, and 20q gain. All five of these arms had been previously implicated in cancer: 1q and 20q gains are very commonly found in cancers; 9p contains tumor suppressor p16; 12p contains oncogene KRAS; and 17p contains tumor suppressor TP53.^{32, 34} Further inspection of the Training Set data highlighted focal amplifications surrounding 8q24 as present in 50% of aneuploid EAC and as absent in the 5 aneuploid NBDE (Representative Plots in Fig. S1). The 8q24 region contains the driver gene CMYC and has been reported as one of the most common focal amplifications in EAC; we therefore also incorporated focal amplifications of 8q24 into our panel of alterations used for classification.^{7, 8, 35} The presence of any of these six specific chromosomal alterations distinguished all but one the five aneuploid NDBE samples from the 28 aneuploid EAC samples, with most EACs having at least two of these specific chromosomal changes (Fig. 3B). Consideration of other chromosome alterations did not further improve the accuracy of distinguishing EAC from NDBE in the training set, satisfying our aim of developing a minimal set of alterations for discriminating EAC from NDBE.

Figure 3. The BAD Molecular Classification of progression to dysplasia in patients with Barrett’s Esophagus.

A) The BAD Decision Tree Algorithm. B) Heatmap of predictive features used in the BAD classifier depicted for Training Set samples that do or do not meet criteria as Very-BAD. C) Heatmap of predictive features used in the BAD classifier depicted for Validation Set samples that do or do not meet criteria as Very-BAD.

Based on the GAS score and this panel of six specific chromosomal alterations, we developed a simple decision tree classifier, termed BAD (Barrett’s Aneuploidy Decision), for distinguishing stages of BE progression (Fig. 3A). BAD sorted samples into three categories. Not-BAD cases had GAS≤0.6, indicating relative non-aneuploidy. Maybe-BAD cases had GAS>0.6 but none of the six specific chromosome alterations, possibly indicating a greater potential risk of progression. Very-BAD cases had GAS>0.6 and losses of 9p or 20q, gains of 1q, 12p, or 20q, or a focal gain of 8q24 (Fig 3A, 3B, Supplementary Table S5, and DataSets 1 and 2). The BAD classification system, which used both specific chromosome changes plus GAS scores, outperformed GAS scores alone. In particular, compared to the aneuploidy classification of GAS, the Very-BAD classification markedly improved both the specificity for rejecting NDBE and the positive predictive value (PPV) for identifying HGD plus EAC cases. Moreover, compared to aneuploidy alone, the Very-BAD classification produced only minimal decreases in the sensitivity for detecting HGD or EAC or in the negative predictive value (Fig. 3B, Supplementary Table S4).

Validation Set

The Validation Set provided an opportunity to independently assess the sensitivity and specificity of the BAD classifier. Note that the patients recruited for the Validation Set were entirely distinct from those in the Training Set, and that the GAS and BAD assignments of validation cases were performed by investigators blinded to the clinical status of the samples. The first important observation in the Validation Set was that the ROC curve for the GAS component of the BAD classifier was strikingly similar to that in the Training Set (Fig. 2A versus 2C). For example, the AUCs were 0.86 and 0.87 in the Training and Validation Sets, respectively. Violin plots of the GAS for each of group of patients in the Validation Set are shown in Fig. 2D, and closely resemble those obtained in the Training Set (Fig. 2B). Importantly, samples from the patients with NDBE in the Validation Set exhibited the same bimodal distribution as observed in the Training Set, with the GAS threshold of 0.60 again cleanly separating Validation Set NDBE into two populations, 36.6% as aneuploid and 63.4% as non-aneuploid (Figure 2D and Supplementary Table S4).

Moreover, when the BAD classifier was applied to the Validation Set, it again was more accurate than GAS alone. For example, it more accurately discriminated the majority of samples with histologic evidence of disease progression from those with NDBE. Specifically, 96.4% of EAC and 67.9% of HGD were classified as Very-BAD (Fig. 3C, Table 1, Supplementary Tables S2, S4, and DataSet 3). In contrast, only 7.3% of NDBE were classified as Very-BAD (Fig. 3C, Table 1, Supplementary Table S4). Moreover, intramucosal cancer (IMCA), the earliest stage of EAC, largely curable with endoscopic techniques, was detected as Very-BAD in three of three cases (Supplementary Table S2). Furthermore, the sensitivity for detecting HGD as Very-BAD increased to 72.7% when considering only cases with more than 1 mm of HGD present in the diagnostic biopsy (i.e. cases with more than “focal” HGD extent, N=22) (Supplementary Table S2).

Table 1:

Performance of the RealSeqS BAD Classifier

	Training Set N=79			Validation Set N=268			Training Plus Validation Sets. N=347
	VERY-BAD	MAYBE-BAD	# Patients	VERY-BAD	MAYBE-BAD	# Patients	VERY-BAD	MAYBE-BAD	# Patients
GEJ from Unaffected controls	0.0%	0.0%	15	0.0%	3.6%	84	0.0%	3.0%	99
NDBE	5.3%	21.0%	19	7.3%	29.3%	41	6.7%	26.7%	60
LGD			0	50.0%	21.9%	32	50.0%	21.9%	32
HGD	60.0%	13.3%	15	67.9%	3.6%	28	65.1%	7.0%	43
EAC	90.0%	3.3%	30	96.4%	2.4%	83	94.7%	2.7%	113
EAC+HGD combined**	80.0%	6.7%	45	89.2%	2.7%	111	86.5%	3.8%	156

^**Not an independent category, but the sum of the above HGD and EAC groups.

Among Validation Set NDBE cases, 7.3% were classified as Very-BAD, 29.3% as Maybe-BAD, and 63.4% as Not-BAD (Table 1). Thus, the Validation Set resembled the Training Set in that approximately one third of the patients with NDBE harbored aneuploid cells, but the BAD classifier could distinguish nearly all these aneuploid NDBE from the aneuploid HGD and EAC cases (Table 1). Among NDBE patients studied, acquisition of either Very-BAD or Maybe-BAD status showed no significant relation to BE length, with mean BE segment lengths of 3.9 cm for Not-BAD NDBE (95% confidence interval 3.0 to 4.9 cm), versus 4.0 cm for Very-BAD NDBE (95% confidence interval 0 to 9.7 cm), and 4.6 cm for Maybe-BAD NDBE (95% confidence interval 2.5 to 6.7) (P = 0.787 by one-way ANOVA) (Fig. S2). Additionally, there was no significant difference between performance of the BAD classifier in the total Validation Set versus among the 76% subset of cases for which central pathology review was available (Supplementary Table S6). None (0%) of 84 Validation Set normal controls were scored as Very-Bad, three (3.6%) were scored as Maybe-BAD, and the remaining 96.4% were scored as Not-Bad (Table 1 and Supplementary Table S2).

The Validation Set included samples from BE patients with LGD. These samples also demonstrated a bimodal distribution, with one group having high GAS scores and the other low GAS scores (Figure 2). Among these LGD cases, 50.0% were classified as Very-BAD, 21.9% as Maybe-BAD, and 28.1% as Not-BAD (Table 1). This difficulty in classifying the status of LGD cases with respect to aneuploidy mirrors the well-known difficulties in histopathological classification of LGD. As noted above, some experienced pathologists classify LGD as NDBE and others classify the same sample as HGD.^17–19

In the Validation Set, as in the Training Set, the BAD classification system outperformed GAS scores alone. Compared to the aneuploidy classification of GAS, the Very-BAD classification again markedly improved both the specificity for rejecting NDBE and the positive predictive value (PPV) for identifying HGD plus EAC validation cases. And again, compared to aneuploidy alone, the Very-BAD classification produced only minimal decreases in the sensitivity for detecting Validation Set HGD or EAC or in the negative predictive value (Fig. 3C, Supplementary Table S4).

Clinical Review for Disease Progression

Following analysis of the validation set, an expert central reviewer (AC), who was blinded to the RealSeqS results, reviewed the medical records of all NDBE cases in both the Training and Validation Sets for any evidence of disease progression subsequent to study entry. Four of the 60 NDBE cases in the total cohort were Very-BAD, and 2 of these 4 were found to have progressed within 36 months of study entry, one to HGD and one to EAC. In contrast, no instances of histopathologic progression were identified during equivalent follow-up of the other 56 NDBE (nominal P<0.004), including 40 Not-BAD and 16 Maybe-BAD cases. Progression to HGD within three years was also identified in two individuals with LGD, despite both having undergone disease ablation. At study entry one of these individuals was classified as Very-BAD and the other as Maybe-BAD. No progression was identified in the 10 Not-BAD LGD cases, including 5 individuals who did not undergo disease ablation. In total, of 4 individuals who developed disease progression within 36 months of study entry, 3 had antecedent brushings testing as Very-BAD.

Expanded Analysis of Chromosome Arm Changes During Progression

We designed only one classifier on the basis of the Training Set data so that we could rigorously evaluate the Validation Set in a statistically sound fashion. Had we designed several classifiers, then we would have had to correct performance in the Validation Set for multiple hypothesis testing. However, after evaluating the Validation Set with our single BAD classifier, we thought it of interest to evaluate how alternative classifier panels might be constructed for future studies.

To explore this question, we conducted a secondary analysis that used all aneuploid samples from the combined Training and Validation Sets to characterize alterations during progression of each of the individual chromosome arms. We observed that the total number of chromosome arms lost or gained steadily increased during disease progression. The average number of altered chromosome arms was 6 for aneuploid NDBE, 10 for aneuploid LGD, 17 for aneuploid HGD, and 22 for aneuploid EAC (Fig S3A). The chromosome arm abnormality counts in the four stages of disease were each significantly different from one another (P value < 0.05 by the Student’s t-test). Inspection of the individual chromosomes showed that 16 specific chromosome arm gains and 14 specific chromosome arm losses had a difference of greater than 23% in EAC compared to aneuploid NDBE, and each of these differences was statistically significant (P value < 0.05 by the Binomial Proportions test)(Supplementary Table S5). All five chromosome arms selected from Training Set data for inclusion in the BAD decision tree were among those. The most prominent chromosome arm changes observed with progression are depicted in Fig. 4, and, as expected, included the five chromosome arms selected from the Training Set data for inclusion in the BAD decision tree (Figs. 3B, 3C, S3B, S3C, and Supplementary Table S7).

Figure 4:

Schematic of the progression of aneuploidy, chromosomal arm alterations, and the BAD classification going successively from Non-dysplastic Barrett’s Esophagus to Adenocarcinoma. Chromosome alterations shown in red contribute to the BAD classifier algorithm.

Chromosome 8q Gains are found in Aneuploid NDBE and Evolve During Progression to EAC.

We next investigated what chromosome alterations accounted for the aneuploidy detected in 37% of NDBE. Examination of these aneuploid NDBE cases revealed that chromosome alterations were not entirely random and were typified by gains of 8q. Of the 20 aneuploid NDBE samples in the complete study cohort, 15 (75%) had an 8q gain. This was by far the most common alteration in the aneuploid NDBE samples and is one of the most common chromosome alterations in cancer in general.^{32, 36, 37} Furthermore, the fraction of aneuploid BE samples with an 8q gain was not significantly different in patients with NDBE, LGD, HGD, or EAC (Fig. S3B and S3C). Chromosome 8q in NDBE was different from all other chromosome arms, for which the frequency of alteration increased concomitantly with disease progression (Fig. S3B, S3C, and Supplementary Table S7).

Overall, of 60 patients with NDBE, 18 cases (15 with GAS>0.6 and three cases with GAS≤0.6) harbored a gain of 8q (DataSet 1). In 100% of these 18 NDBE instances, the entire 8q chromosome arm was gained (Fig. S1). Moreover, a gain of the entire chromosome 8q was detected as the sole chromosomal change in five (8.3%) of the 60 NDBE cases evaluated. In EACs, this pattern (“only 8q gain”) was seen in only two (1.8%) of 113 cases (nominal P value =0.05). This suggested that gain of the entire chromosome 8q arm as a sole chromosomal alteration is a biomarker of early aneuploidy development in BE. As implied above, none of the brushings from the four patients who progressed to HGD or EAC, starting from BE or LGD, showed this “only 8q gain” pattern.

Similar to NDBE, 82 of 110 (75%) EAC cases also demonstrated an 8q gain. However, in contrast to NDBE, in only 36 of these 82 EAC cases (44%) was all of 8q gained (Fig. S1). In the remaining 46 EAC cases (56%), 8q gains were all sub-chromosomal, in every case encompassing 8q24 (nominal P value=0.0004) (examples depicted in Fig. S1)(Fig. 3B and 3C, DataSets 2 and 3). Intriguingly, the pattern of 8q gains in LGD and HGD more closely resembled that of NDBE than EAC (Fig. 3B and 3C. DataSets 2 and 3). In LGD, 13 cases demonstrated 8q gain, and 11 of these 13 cases showed gains of the entire 8q. Similarly, in HGD, 19 cases demonstrated 8q gains, and 17 of these 19 cases showed gains of all of 8q. Thus, alterations of 8q may be initiated in NDBE, beginning as whole arm gains, and may evolve in EAC to focus on 8q24, accompanied by stepwise increases in the copy number of CMYC.

Reproducibility

Duplicate esophageal brushings obtained at the same clinical session were available from 22 participants in this study. These included eight unaffected controls, four cases of LGD, five cases of HGD, and five cases of EAC. In 21 of the 22 instances (95%), the determinations from the independent duplicate brushings were fully concordant, both with respect to their GAS determination of aneuploidy as well as their further classification as Not-BAD, Maybe-BAD, or Very-BAD (Table 2 and DataSet 4). In the other case, an LGD, the duplicate brushings were discordant, with one classified as Not-BAD and the other as Very-BAD (Table 2, Supplementary Table S8 and DataSet 4).

Table 2:

Concordance between duplicate brushes.

Diagnosis	Concordant positive	Concordant negative	Discordant
Unaffected Control		8
LGD	2	1	1
HGD	4	1
EAC	5
Totals	11	10	1

Concordance status for both GAS status and BAD classifications was identical in all samples.

Discussion

Multiple related observations and insights were garnered from this study. First, the combination of esophageal brushings with RealSeqS analysis provides a practical and sensitive method for detecting chromosome arm alterations in BE patients. Second, aneuploidy typifies EAC and HGD, but also can be identified in a small proportion of NDBE and a larger proportion of LGD patients. Third, alterations of specific chromosome arms containing well-known driver genes are found commonly in late stage but rarely in early stage disease. Fourth, there is a special role for chromosome 8q in BE progression, with gains of the whole of 8q appearing to be an early event, present in 75% of aneuploid NDBE, and with selective focused gains on 8q24 often found in EAC. And fifth and most importantly, these chromosomal alterations can be used to design a molecular classifier called BAD, and the BAD classification of DNA from esophageal brushings is highly correlated with histopathologic classification of the same patients (visually depicted in Fig 4).

The power of the approach described here comes from the juxtaposition of the two major components described above. Brushings can sample a much more extensive region of the esophagus than conventional biopsies, even when multiple biopsies are performed. But this extensive and convenient sampling comes with a cost: the aneuploidy present in the dysplastic cells within any individual lesion is diluted with non-dysplastic cells from the remaining esophagus. Fortunately, RealSeqS technology is well-suited for the detection of a relatively small fraction of aneuploid cells admixed with a much larger number of non-aneuploid cells. The chromosomal alterations identified by RealSeqS analysis of esophageal brushings are consistent with previous studies of EAC tissues that employed more traditional genetic methods and/or fluorescence in situ hybridization (FISH) analysis.^{6–8, 10, 11, 22, 24, 35, 38–40}

Shallow whole genome sequencing (WGS), in principle, should also be able to detect aneuploidy in esophageal brushings. Shallow whole genome sequencing has been productively used for many purposes, particularly for non-invasive prenatal testing.⁴¹ RealSeqS has some advantages over WGS. First, it requires only a tiny amount of input DNA and is exceedingly simple to perform, as it employs PCR with a single pair of primers to prepare samples for massively parallel sequencing. WGS requires several steps, including shearing and library preparation, prior to sequencing. Additionally, at the same sequencing depth, whole genome sequencing is not as sensitive as RealSeqS for detecting relatively small fractions of aneuploid cells, particularly when aneuploidy is of small chromosomes.³¹ On the other hand, an advantage of WGS is that it can reveal information about chromosome regions that are not queried in RealSeqS because the latter evaluates only _∼350,000 repetitive elements rather than the entire genome. In this regard, Fitzgerald and colleagues recently reported a landmark longitudinal study in which they used shallow whole-genome sequencing to assess the risk of BE progression in multiple, conventional esophageal biopsies from patients.³⁸ It would be informative in the future to compare the two technologies - RealSeqS and shallow whole genome sequencing - on DNA from esophageal brushings.

Along with this study, other studies have also advanced alternative technologies to improve the identification of BE patients at risk of progressing to EAC. These include: the combination of enhanced image analysis together with molecular markers to predict progression risk from individual BE biopsies⁴² and the combination of extensive and deep tissue brushing of BE with enhanced image analysis to increase the identification of dysplasia.⁴³ Future studies will be required to further evaluate the role of these different technologies in the early identification of high risk BE patients and to provide direct comparison of their performance to the technology we advance here.

There were several limitations of our study. Though the total number of patients studied was reasonable, the number within each phase of BE progression was relatively small. Additionally, the sensitivity for detecting HGD was lower than for detecting EAC, suggesting that perhaps even wider sampling of the esophageal epithelium may add further benefit. In this regard, while duplicate brushings at the same time point were in all but one case concordant with respect to the BAD classification (Table 2), the exception demonstrates the potential for increased detection by using two brushings from each patient. Furthermore, only one time point in each patient was evaluated. Our conclusions about progression are therefore inferential and conjectural, based on differences between patients assessed at a single time point rather than based on differences in the same patient at various times after the diagnosis of BE. And most importantly, confirmation of clinical value will require future prospective patient trials.

On the other hand, a particularly intriguing observation of the current study was that patients with relatively early phase BE are heterogeneous with respect to chromosomal changes, with a minority of NDBE (7%) cases and half of LGD cases classified as Very-BAD. This suggests the possibility that it is the Very-BAD subset of patients with NDBE or LGD that are at greatest risk for progression. Indeed, three of the four individuals who were later shown to progress to HGD or EAC were classified as Very-BAD in our study.

The observations described above, in combination with prior studies establishing the association of aneuploidy with BE progression, suggest several hypothetical clinical implications that will be worthy of exploration in future clinical trials. First, RealSeqS analysis of esophageal brushings may provide a technique that could be combined with the current Seattle protocol of multiple random biopsies to augment the effectiveness of BE surveillance for detecting early progression to dysplasia or cancer.^3–5 This combination could minimize the development of interval cancers between BE surveillance sessions, which is at present a widely-recognized challenge.^{15, 16} Second, the NDBE cases classified as Very-BAD may benefit from intensified surveillance, while the Not-BAD cases may require less surveillance. Similarly, the LGD cases classified as Very-BAD may benefit from ablation therapies, whereas Not-BAD cases could potentially be followed with continued endoscopic surveillance. If confirmed, this would potentially enable molecular classification to add to current morphologic criteria for risk stratification.

In overview, the current study demonstrates that the combination of esophageal brushing with RealSeqS can molecularly discriminate different stages during BE progression and can detect the majority of prevalent histologically advanced lesions. Additional cross sectional studies with larger cohorts will be of value to further refine the estimates of specificity and sensitivity of the BAD classifier for detecting LGD, HGD, and EAC. Most importantly, longitudinal studies will be required to test the inference that NDBE and LGD lesions detected as Very-BAD are at the highest risk for progression.The present findings provide the technical means and conceptual basis for exploring each of these hypotheses and for enabling their future testing in larger prospective clinical trials.

What You Need to Know.

BACKGROUND AND CONTEXT:

Aneuploidy has been suggested as a means to identify disease progression, and risk of progression, in patients with Barrett’s Esophagus, but the need for multiple biopsies and complex technology has impeded its clinical application.

NEW FINDINGS:

Esophageal brushing combined with massively parallel sequencing sensitively identifies both global aneuploidy and individual chromosome alterations. A classifier based on selected alterations accurately discriminates patients with non-dysplastic Barrett’s esophagus versus those with progression to dysplasia or cancer, and identifies a 7% subset of Barrett’s patients with the molecular signature of progression.

LIMITATIONS:

Prospective studies are needed to fully evaluate the clinical implications of these findings.

IMPACT:

This technology and classifier may enable molecular based improvement in identifying individuals who have developed actual disease progression and those who are at high progression risk.

Short Summary:

Detecting specific chromosomal alterations in esophageal brushing samples provides a practical and sensitive aneuploidy-based method to determine disease progression in patients with Barrett’s Esophagus.

Abbreviations:

BAD

Barrett’s Aneuploidy Decision

BE

Barrett’s Esophagus

EAC

Esophageal Adenocarcinoma

EGD

Esophagogastroduodenoscopy

FISH

Fluorescence In Situ Hybridization

GAS

Global Aneuploidy Score

HGD

High Grade Dysplasia

LGD

Low Grade Dysplasia

MPS

Massively Parallel Sequencing

NDBE

Non-Dysplastic Barrett’s Esophagus

RealSeqS

Repetitive Element Aneuploidy Sequencing System

SVM

Support Vector Machine

WALDO

Within-Sample AneupLoidy Detection