Enhanced neoplasia detection in chronic ulcerative colitis: the ENDCaP-C diagnostic accuracy study

Patient recruitment

Patients were identified from archived histology biopsy samples in six hospitals across the West Midlands area. Patients were identified through tracing endoscopy records and correlation with histology reports. Searches were restricted to endoscopies after January 1996 because of changes to formalin fixation at that time, and to before January 2014 to minimise missed neoplasia through identification during the follow-up period. Mucosal biopsies were classified as one of:

• neoplasia – defined as any of adenocarcinoma, high-grade or low-grade dysplasia

• matched non-neoplastic – defined as non-neoplastic chronically inflamed colonic mucosal biopsies taken distant from areas of neoplasia (Figure 1)

• control – defined as colonic mucosa, sampled from patients with chronic UC of duration of > 8 years and extending at least to the splenic flexure or patients with a diagnosis of both UC and PSC who had been screened for neoplasia without it being found.

FIGURE 1.

Diagram of colonoscopic sampling from patients for the ENDCaP-C study.

Ethics approval was from South Birmingham Research Ethics Committee (reference 08/H1207/104).

Sample processing

Biopsy samples from identified patients were retrieved from the histopathology archives at the six collaborating hospitals. For those patients with neoplasia in the large bowel, separate biopsies from different colonic segments were selected alongside the neoplastic biopsy. All blocks were reviewed centrally by the lead ENDCaP-C histopathologist, with representative sections undergoing DNA extraction. Histological diagnosis of dysplasia (the gold standard) was defined as an altered nuclear/cytoplasmic ratio, increased cell size and/or an increase in mitotic figures. DNA extraction of neoplasia was performed by needle macrodissection to enrich for tumour material; macrodissected material was extracted using the FFPE protocol of the Qiagen DNEasy FFPE kit (Qiagen, Hilden, Germany). Extracted DNA was quantified by both NanoDrop spectrophotometry (Nanodrop ND 8000 spectrophotometer; Thermo Fisher Scientific, Waltham, MA, USA) and Qubit fluorimetry (Qubit 2.0 Fluorometer; Invitrogen, Life technologies, Carlsbad, CA, USA).

Neoplasia, matched non-neoplastic and control samples were included in mixed batches to ensure that test performance could be analysed at sequential analyses across the study duration. Each sample was labelled with only a study sample identification number and assays undertaken blinded to neoplasia status.

Deoxyribonucleic acid methylation analysis

A total of 500 ng of extracted DNA was bisulphite converted using a Zymo EZ-DNA methylation bisulphite conversion kit (Zymo Research, Irvine, CA, USA) according to the manufacturer’s protocol. Two microliters of eluted bisulphite converted DNA was utilised in a pyrosequencing polymerase chain reaction (PCR) using the Qiagen PyroMark PCR kit (Qiagen, Hilden, Germany) according to the manufacturer’s protocol in a 25-µl reaction volume. PCR products were run on a 2% agarose gel with a DNA size ladder and successful amplification was defined as the presence of a band at the appropriate size for the marker run. PCR products were cleaned using streptavidin beads, washed and mixed with the requisite pyrosequencing primers. These were then sequenced on a Qiagen PyroMark Q96 instrument (Qiagen, Hilden, Germany). All reactions were run with 100% methylated and unmethylated DNA positive and negative controls, as well as a water reaction. Methylated DNA was generated by incubating 1 µg of blood-derived control DNA with CpG methyltransferase (M.SssI) (New England Biolabs, Ipswich, MA, USA). Unmethylated DNA was generated by whole-genome amplification of 10 ng of blood-derived control DNA using the Qiagen Repli-G Mini kit (Qiagen, Hilden, Germany).

The marker panel chosen for this experiment was based on previous reported findings and consisted of the following markers: SFRP1, SFRP2, SFRP4, SFRP5, WIF1, TUBB6, SOX7, APC1A, APC2, MINT1 and RUNX3. Primer sequences, chromosomal positions and reaction conditions are shown in Appendix 1, Table 17. After each run, sample data were examined using Qiagen PyroMark Q96 software (Qiagen, Hilden, Germany). Samples that had failed Qiagen quality metrics were marked as failed on the sample sheet.

Sample size

The original planned sample size of 160 neoplastic and 320 control samples was determined to provide adequate events to develop a robust model (with > 10 events per marker) and provide estimates of sensitivity and specificity with adequate precision [with 95% confidence interval (CI) width < 16% for sensitivity and < 12% for specificity]. The 2 : 1 sampling ratio is determined based on access to sample banks.

Statistical analysis

Where a biomarker was examined at multiple CpG sites, the mean CpG value across sites was computed and used in all analyses. This was undertaken for consistency and was considered appropriate, as methylation within small regions tends to be distributed homogenously. 32 Statistical analysis was then undertaken in three stages.

In stage 1, biomarkers that were futile with respect to their ability to discriminate between neoplastic samples and control samples were identified and testing was terminated early. The sample size required 320 control samples and 160 neoplastic samples (480 in total). To evaluate the performance of each biomarker as the samples were being processed in batches of 30 (20 control samples and 10 neoplastic samples), we planned to undertake 16 interim analyses for each biomarker (420 total samples divided by 30 batches of samples). To account for multiple testing, the results for each batch of samples for each biomarker were evaluated in a group sequential analysis following the O’Brien and Fleming method36 to assess whether further testing of each biomarker was justified or considered futile. Sequential boundaries were constructed according to the O’Brien and Fleming method,36 the t-statistic from the t-test computed for each biomarker at each analysis step, and the comparison made to the predefined boundary values to test for statistical significance or futility (see Appendix 1, Figure 18). The boundary values were computed using the SAS^® statistics software (SAS Institute Inc., Cary, NC, USA) and the SAS code is given under Appendix 1, Figure 18. (SAS and all other SAS Institute Inc. product or service names are registered trademarks or trademarks of SAS Institute Inc. in the USA and other countries. ^® indicates USA registration.)

In this first stage, the distributions of the biomarker values were unknown and data were analysed without transformation. Once all data were accumulated, visual inspection of histograms demonstrated a positive skew (see Appendix 1, Figure 19) and the mean biomarker values for each sample were log-transformed prior to further analysis.

In stage 2, markers with responses suitable for inclusion in the predictive models were selected. Biomarkers that showed significant (p < 0.05) discrimination and attained amplification rates of > 85% were selected for inclusion. The ability of each marker to discriminate was described by computing the ratio of geometric means (with 95% CI) and statistical significance was assessed by two-sample t-tests undertaken on the log-transformed scale. Comparisons were made between (1) neoplasia samples and control samples and (2) matched non-neoplastic samples and control samples.

In stage 3, predictive models were fitted using logistic regression with outcome (1 = sample or 0 = control) and the mean log-CpG value for each patient for the biomarkers selected from stage 2. Only samples that had complete data for the selected biomarkers were initially included in these analyses. Three separate models were constructed: (1) differentiating neoplasia samples from control samples, (2) differentiating dysplasia samples from control samples and (3) differentiating matched non-neoplastic samples from control samples. Model 2 used the same patients and biomarker selection as model 1, but excluded any samples that were classed as adenocarcinomas.

Estimation of the adjustments for shrinkage and optimism in the model fit were computed using standard bootstrap internal validation techniques as recommended by Steyerberg. 37 The final models were produced including all samples, using multiple imputation using chained equations to impute missing biomarker data. Multiple imputation models used 50 iterations with pathology categorisation, sample type and measurements of all other biomarkers as predictors. The model coefficients were corrected for optimism by application of the shrinkage factor. To facilitate application of the model when individual or pairs of biomarkers are unavailable, reduced models were computed omitting each biomarker and possible pair of biomarkers from the multiple imputation model. Discriminatory performance for each model was measured by the area under the receiver operating characteristic curve (AUC).

One biomarker (TUBB6) was not selected for inclusion in models 1 and 2 in the stage 1 O’Brien and Fleming method36 analysis on untransformed data, but did show significant differences in stage 2 once log-transformations had been applied. Models 1 and 2 were fitted with and without this biomarker.

The clinical team considered that a positive test result should have a positive predictive value of at least 20% to be of clinical value. Given an assumed background incidence of 4% this corresponds to the point on the receiver operating characteristic (ROC) curve with a positive likelihood ratio [sensitivity/(1 – specificity)] of 6. The threshold at this point was identified from the ROC tabulation of each predictive equation, and estimates of sensitivity and specificity were obtained. We also identified thresholds for each model that corresponds with 90% of cases being detected.

Selection of biomarkers

Eight out of the 11 methylation markers had an amplification success rate of > 85% (Table 3 and see Appendix 1, Table 18). The three remaining primer sets were within the promoter regions of SFRP1, MINT1 and RUNX3. Because of the reduced reliability, these three primer sets were not taken forward for further analysis.

In the stage 2 analysis, five markers accurately discriminated between neoplasia and control samples with a p-value of < 0.0001: SFRP2, SFRP4, WIF1, APC1A and APC2 (see Table 3) (with between 40% and 76% increases in geometric mean values). TUBB6 showed a smaller (31%) increase, but this was also strongly significant (p = 0.003).

Comparing methylation in the samples of background mucosa (from patients with colitis-associated neoplasia) with control patients (with chronic UC only) showed some discrimination in four out of the eight promoter regions: SFRP4, APC1A, SFRP5 and SOX7 (see Table 3). Two of these markers (i.e. SFRP4 and APC1A) showed increases of in geometric mean values of 7% and 28%, respectively; the other two (i.e. SFRP5 and SOX7) showed decreases of 23% and 27%, respectively.

Performance of predictive models

Predictive models to discriminate neoplasia samples from controls had good discrimination (Table 4). The optimism-adjusted AUC for model 1 detecting all neoplasia was 0.86 (95% CI 0.81 to 0.91) for the complete-case analysis (with a shrinkage factor of 0.93) but lower at 0.83 (95% CI 0.79 to 0.88) for the model using multiple imputation (see Table 4 and Figure 2). The addition of TUBB6 increased the AUC by only 0.001. When considered together in the panel, all markers other than WIF1 showed significant independent predictive value.

FIGURE 2.

The ROC curves for final fitted predictive models (after multiple imputation) for model 1 neoplasia vs. control, with the biomarkers used within this model.

The discrimination of model 2, predicting only dysplasia (excluding adenocarcinoma cases), was higher with an optimism corrected AUC of 0.92 (95% CI 0.88 to 0.96) for the analysis of complete cases (with a shrinkage factor of 0.91) and 0.88 (95% CI 0.84 to 0.92) for the model using multiple imputation (see Table 4 and Figure 3). Again, adding TUBB6 made little difference, decreasing the AUC by 0.001. The coefficients for model 1 based on the multiple imputation data set are reported in Appendix 1, Table 19, and for model 2 in Appendix 1, Table 20. When considered together in the panel, all markers other than WIF1 showed significant independent predictive value.

FIGURE 3.

The ROC curves for final fitted predictive models (after multiple imputation) for model 2 dysplasia vs. control, with the biomarkers used within this model.

The predictive model to discriminate samples where there is methylation in the background mucosa (the matched non-neoplastic samples, model 3) from controls had poorer discrimination, with an optimism-adjusted AUC of 0.66 (95% CI 0.59 to 0.73) for the complete-case model (shrinkage factor 0.91) and 0.68 (95% CI 0.62 to 0.73) for the model using multiple imputation (Figure 4). We report coefficients for model 3 based on the multiple imputation data set in Appendix 1, Table 21. For SFRP5 and SOX7, lower levels of methylation were associated with neoplastic change in background mucosa. When considered together, only AP1CA and SOX7 showed significant independent predictive value.

FIGURE 4.

The ROC curves for final fitted predictive models (after multiple imputation) for model 3 matched non-neoplastic vs. control, with the biomarkers used within this model.

The calibration plot for model 1 after multiple imputation suggested that our final model for neoplasia detection after multiple imputation was reasonably well calibrated, with slight overestimation of probability at lower risk and overestimation at higher risk (see Appendix 1, Figures 19 and 32).

Identification of a diagnostic threshold

We identified the value of the predictive model corresponding to a likelihood ratio of at least 6, to identify thresholds that would have a positive predictive value of at least 20% when incidence of neoplastic disease was 4%. This corresponded to a threshold of 0.40 for model 1, 0.28 for model 2 and 0.50 for model 3. Sensitivity and specificity (with 95% CIs) at this threshold were 58.4% (48.8% to 67.6%) and 90.38% (86.8% to 93.3%) for model 1, 79.5% (71.0% to 86.6%) and 86.9% (82.8% to 90.3%) for model 2, and 7.1% (3.1% to 13.5%) and 98.8% (97.0% to 99.7%) for model 3.

To achieve a sensitivity of at least 90%, model 1 would use a threshold of 0.11 with a specificity of 46.4% (95% CI 41.1% to 51.8%), positive predictive value of 6.6% (95% CI 5.9% to 7.3%) and diagnostic odds ratio of 8.0 (95% CI 4.1 to 17.1); model 2 would use a threshold of 0.11 with a specificity of 68.2% (95% CI 63.0% to 73.1%), positive predictive value of 10.7% (95% CI 9.11% to 12.3%) and diagnostic odds ratio of 18.8 (95% CI 8.6 to 46.5); and model 3 would use a threshold of 0.19 with a specificity of 27.1% (95% CIs 22.5% to 32.1%), positive predictive value of 4.9% (95% CI 4.5% to 5.3%) and diagnostic odds ratio of 3.4 (95% CI 1.7 to 7.4).

Models for missing data

For model 1, to provide models for scenarios where data on fewer than five of the chosen biomarkers amplify, separate models accounting for all possible scenarios, where at least three out of the five biomarkers amplified, were created through re-analysis of the multiple imputation model with reduced sets of predictor variables. This created an additional 15 models (see Appendix 1, Table 22).

Bisulphite DNA conversion

Bisulphite DNA conversions were performed using the EZ-96 DNA Methylation Gold MagPrep Kit (Zymo Research; Irvine, CA, USA) according to the manufacturer’s instructions. The amount of input DNA was variable, dependent on the sample source, ranging between 1 ng and 30 ng.

Target enrichment using the Fluidigm Access Array system

Multiplex PCR amplification was carried out on a 48.48 Fluidigm Access Array (Fluidigm Corporation, South San Francisco, CA, USA) following the methods of Lange et al. 44 using a Roche High Fidelity Fast Start Kit (Roche Holding AG, Basel, Switzerland). The target specific primer sequences are given in Table 5.

Sequencing adaptor and barcode primer incorporation

A bidirectional sequencing protocol was adopted as provided by Fluidigm (Access Array™ Barcode Bidirectional Library for Illumina Sequencers – 100–3771; Fluidigm Corporation, San Francisco, CA, USA) to facilitate simultaneous sequencing of the forward and reverse directions, which is important to provide sufficient base diversity content during subsequent MiSeq Illumina sequencing (Figure 5).

FIGURE 5.

Schematic representation of the bidirectional sequencing strategy adopted to overcome the low diversity issues.

For each sample, 1 µl of the 50-fold diluted PCR products was added to each of two PCR plates containing 15 µl of pre-sample mastermix containing 0.05 U/µl of FastStart™ High Fidelity Enzyme blend (Roche Holding AG, Basel, Switzerland), 1 × FastStart High Fidelity Enzyme Buffer (Roche Holding AG, Basel, Switzerland), 200 µM each of dNTPs, 4.5 mM MgCl₂ and 5% dimethyl sulphoxide. In the first plate, 4 µl of one pair of primers containing an individual 10-base barcode (BC) sequence, and sequence tags for reading in one direction (PE2-BC-CS1+PE1-CS) were added to each well. In the second plate, 4 μl of primers containing PE2-BC-CS2/PEI-CS2 were added to each well. Reaction products in plates were amplified for 15 cycles: 95 °C for 10 minutes, 15 cycles of 95 °C for 15 seconds, 60 °C for 30 seconds and 72 °C for 4 minutes, 1 cycle of 95 °C for 3 minutes.

Quantification and clean-up of deoxyribonucleic acid sequencing library

The barcoded products were pooled together and purified using AMPure XP beads (Beckman Coulter Life Sciences, Brea, CA, USA) using a bead to amplicon ratio of 1 : 1. The library was analysed and quantified using microfluidic electrophoresis on Agilent 2200 TapeStation (D1000 DNA screen tape) (Agilent Technologies Incorporated, Santa Clara, CA, USA) to ensure that the expected size was obtained (a main peak of 330 bp). A 2-nM library dilution was prepared, checked on the Agilent 2200 TapeStation (high-sensitivity tape) and sequenced on MiSeq sequencing platform according to the Illumina standard protocol.

Bioinformatic pipeline

An in-house bioinformatics pipeline was used to automate the cytosine methylation state detection ‘Bismark’ software (Babraham Bioinformatics; Cambridge, UK). 47 Briefly, CS (consensus sequences) were removed using cutadapt (v1.4.1) and target-specific sequences were analysed using Bismark package (version v0.10.0). Bismark utilised bowtie (v1.1.1) to align the trimmed reads allowing only one mismatch per 50 (high quality, Q > 28) bases. Methylation data were extracted using the bismark_methylation_extracter tool to generate a table of methylation counts. Methylation count data were mapped and converted to genomic locations using a custom Python script.

Raw data obtained from the MiSeq was quality assessed and filtered against a minimum Phred score of 30, using a sliding four-base window across each read. Reads were trimmed to remove the primer sequences using the cutadapt v1.4.1 tool.

Reads were processed using Bismark v0.10.1, which simultaneously aligned against in silico converted DNA and unmodified DNA reference sequences (the 11 amplicon sequences as defined by the target specific primers, hg19 genome assembly).

After alignment the ‘bismark_methylation_extractor’ script computed the number of counts corresponding to methylated and unmethylated bases to give the percentage of methylated reads. The ‘local’ alignment co-ordinates were then converted into genomic co-ordinates using a custom Python script to create a bed file for viewing methylation data on the Integrative Genomics Viewer 2.3 (The Eli and Edythe L Broad Institute of MIT and Harvard; Cambridge, MA, USA).

Primer design and validation

Primers used in module 1 for the 11 chosen targets (i.e. SFRP1, SFRP2, SFRP4, SFRP5, WIF1, RUNX3, MINT1, SOX7, APC1A, APC2 and TUBB6) were single nucleotide polymorphisms checked to exclude common population variants with a frequency of > 1%. Primers included CS tags, which were required for the Fluidigm protocol.

Amplification performance of the primers was tested on anonymised normal control genomic DNA samples (using the Fluidigm 48.48 Access Array system) as described in the above methods. Products of the expected size were seen for all targets.

Alternative primers were designed for three poorly performing targets (RUNX3, MINT1 and SOX7). Sequence complexity prohibited the redesign of alternative primers for the APC2 locus. Primers were designed using the MethPrimer online tool (UroGene S.A., Genopole, France). 48 The new primers were tested in triplicate using normal FFPE tissue and resulted in a marked improvement in test performance.

Amplification metrics

A total of 729 samples were processed through the ENDCaP-C NGS workflow in duplicate. Samples for which there was insufficient DNA remaining for duplicate analysis were excluded. Targets for which < 100 reads were obtained have been excluded from the analysis (a marker of poor quality).

The average methylation percentage at each CpG site between replicates was calculated and this output was forwarded to the Birmingham Clinical Trials Unit (BCTU) for statistical analysis.

The DNA amplification performance (a necessary first step for successful NGS) for each of the 11 targets is illustrated in Figure 6. On average, the success rate fell below 35% for each target. This low amplification rate is anticipated as small formalin-fixed biopsy samples will yield only small amounts of DNA strands of sufficient length for the amplification process. The variation between the different targets (from 34% to < 5%) reflects the variation in length of the amplicons and also the reduced sequence complexity (i.e. enrichment for adenine-thymine base pairs) during DNA bisulphite conversion, which affects the binding of the primers.

FIGURE 6.

Amplification success at each of the 11 targets for the 729 samples processed (in duplicate).

The reasons for poor amplification performance are multifactorial, including low DNA integrity, low concentration (below the minimum threshold required for the Fluidigm Access Array workflow) and the amplicon length. The relationship between DNA quality and performance of the NGS test is shown by comparing Figures 7 and 8. The relationship between amplicon length performance of the NGS test is shown by comparing Figures 9–11.

FIGURE 7.

Q-ratio of the ENDCaP-C input samples.

FIGURE 8.

The ENDCaP-C input samples performance of the NGS test. Green shading means pass, blue shading means fail and pale green shading means a result from only one duplicate.

Sample 3	Sample 2	Sample 1	Sample 6	Sample 7	Sample 5	Sample 4	Sample 8	Sample 9	Sample 10
SFRP2	WIF1	WIF1	RUNX3	RUNX3	SFRP1	WIF1	RUNX3	WIF1	SFRP4
WIF1	RUNX3	SOX7	SFRP5	WIF1	SFRP4	SFRP2	WIF1	RUNX3	SFRP2
SFRP5	SFRP2	SFRP1	WIF1	APC1A	RUNX3	SFRP4	APC1A	SFRP1	WIF1
APC1A	SFRP4	RUNX3	APC1A	MINT1	SFRP2	RUNX3	MINT1	SOX7	RUNX3
RUNX3	APC1A	SFRP4	MINT1	SFRP4	WIF1	SFRP1	SFRP2	APC1A	APC1A
SFRP4	SFRP1	SFRP2	SFRP2	SFRP1	APC1A	SOX7	SFRP4	APC2	MINT1
SOX7	SOX7	APC1A	SFRP4	SOX7	MINT1	APC1A	SFRP1	SFRP2	SFRP1
SFRP1	TUBB6	MINT1	SFRP1	TUBB6	SOX7	APC2	SOX7	SFRP4	SOX7
TUBB6	SFRP5	SFRP5	TUBB6	SFRP2	SFRP5	SFRP5	SFRP5	SFRP5	SFRP5
MINT1	MINT1	APC2	APC2	SFRP5	TUBB6	TUBB6	TUBB6	TUBB6	TUBB6
APC2	APC2	TUBB6	SOX7	APC2	APC2	MINT1	APC2	MINT1	APC2

FIGURE 9.

Concentration (ng/µl) of the different fragment sizes (40, 129 and 350 bp) within the ENDCaP-C input samples as measured by qPCR using the KAPA hgDNA Quantification and QC Kit.

FIGURE 10.

The ENDCaP-C input samples performance of the NGS test. Green shading means pass, blue shading means fail and pale green shading means result from only one duplicate.

Sample 1	Sample 2	Sample 3	Sample 4	Sample 5	Sample 6	Sample 7	Sample 8	Sample 9	Sample 10
WIF1	WIF1	SFRP2	WIF1	SFRP1	RUNX3	RUNX3	RUNX3	WIF1	SFRP4
SOX7	RUNX3	WIF1	SFRP2	SFRP4	SFRP1	WIF1	WIF1	RUNX3	SFRP2
SFRP1	SFRP2	SFRP5	SFRP4	RUNX3	WIF1	APC1A	APC1A	SFRP1	WIF1
RUNX3	SFRP4	APC1A	RUNX3	SFRP2	APC1A	MINT1	MINT1	SOX7	RUNX3
SFRP4	APC1A	RUNX3	SFRP1	WIF1	MINT1	SFRP4	SFRP2	APC1A	APC1A
SFRP2	SFRP1	SFRP4	SOX7	APC1A	SFRP2	SFRP1	SFRP4	APC2	MINT1
APC1A	SOX7	SOX7	APC1A	MINT1	SFRP4	SOX7	SFRP1	SFRP2	SFRP1
MINT1	TUBB6	SFRP1	APC2	SOX7	SFRP1	TUBB6	SOX7	SFRP4	SOX7
SFRP5	SFRP5	TUBB6	SFRP5	SFRP5	TUBB6	SFRP2	SFRP5	SFRP5	SFRP5
APC2	MINT1	MINT1	TUBB6	TUBB6	APC2	SFRP5	TUBB6	TUBB6	TUBB6
TUBB6	APC2	APC2	MINT1	APC2	SOX7	APC2	APC2	MINT1	APC2

FIGURE 11.

Amplicon length of the targets.

Assessment of deoxyribonucleic acid quantity and quality

To better understand the reasons behind the high failure rate in the NGS platform, an analysis of the input DNA was carried out on a small number of representative samples, with a wide range of different DNA concentrations.

The concentration and quality of 10 DNA samples was assessed by:

• quantitative polymerase chain reaction (qPCR) analysis to quantify the percentage of amplifiable DNA in the samples, performed using the KAPA hgDNA Quantification and QC Kit (Roche Holding AG)

• NanoDrop (spectrophotometric analysis to assess purity; Thermo Fisher Scientific, Waltham, MA, USA)

• fluorometric quantitation (Qubit platform; Thermo Fisher Scientific).

The qPCR assay determines the relative copies of three amplicons of different lengths (i.e. 350 bp, 129 bp and 40 bp) to assess DNA integrity and assign a Q-ratio. High-quality DNA has a 350 : 40 ratio and a 129 : 40 ratio close to 1 (Q-ratio), whereas for damaged DNA the ratio would be much lower.

Table 6 gives a summary of the three methods used to assess quantity and quality of DNA.

Qubit quantification showed that for the majority of samples the amount of input DNA fell below that recommended for both the bisulphite conversion stage (> 200 ng) and the Fluidigm amplification (> 50 ng).

The Q-ratio scores for both the 129-bp and the 350-bp fragments are below 1 in all 10 samples assessed, indicating significant sample degradation. The Q-ratio scores have been compared directly with the NGS analysis performance (see Figure 5). This comparison shows a clear correlation between the Q-score and NGS success. For example, sample 3 has the highest 129-bp Q-score and also the best NGS performance (6 out of 11 fragments passed), whereas sample 10 has the lowest 129-bp Q-score and a very poor NGS performance (no fragments passed). It is important to note that the Q-score for the 350-bp fragment is at least 10-fold less than that of the 129-bp Q-score for each test sample, indicating low amplification potential for fragments of this size. This is significant because the NGS assay includes fragments up to 350 bp in size.

The concentration of input DNA was also shown to correlate closely with NGS performance, as illustrated in Figure 6 comparing the concentration of each fragment (40 bp, 129 bp and 350 bp) within the 10 representative input samples. For example, sample 1 gave the highest concentration of the 40-bp and 129-bp fragments (see Figure 9) as well as the best NGS performance (see Figure 10). As described above, the length of each NGS target fragment correlated well with overall performance (see Figure 11).

Assessment of predictive value of next-generation sequencing

Data produced from the module 2 NGS assay were analysed using the methods established in module 1 (see Chapter 2, Statistical analysis) to compare amplification performance with predictive ability.

Data analysis included an assessment of amplification performance for each target, comparing directly the NGS and pyrosequencing technologies. The statistical model derived from module 1 was then applied to the model 2 data set to assess whether or not the data obtained where amplification was successful maintained the same predictive ability as data from pyrosequencing used in module 1.

Overall, the NGS assay showed a significantly lower amplification success than the pyrosequencing method used in module 1 (Figure 12 and Table 7).

FIGURE 12.

Percentage amplification of target biomarkers (total biopsies) using pyrosequencing (module 1) and NGS (module 2).

We fitted the model on data from the 60 participants for whom biomarker data were available for SFRP2, SFRP4, WIF1 and APC1A. Owing to very low levels of amplification for biomarker APC2 in module 2, this marker was not included in the model. The results are shown in Table 8 and Figure 13.

FIGURE 13.

The ROC curve of the fitted final logistic regression model for module 2.

This fitted model suggests that, despite the poor amplification success obtained in module 2, the data from the NGS platform have predictive ability as in module 1 (pyrosequencing), although the values of estimated coefficients differ (see Appendix 3, Tables 23–25).

Rationale for study

The clinical effectiveness of colonoscopy surveillance for colitis-associated neoplasia has been under scrutiny for some time. 6 Unlike colonoscopic examination for sporadic neoplasia, the inflamed or distorted background mucosa associated with chronic UC can compromise the detection of neoplastic change. The recognised ‘miss rate’ for neoplasia has reduced patient and clinician confidence in the procedure. Although some recent advances in optical technology hold promise of improvement, there is a need for an adjunctive test that can primarily increase the sensitivity of colonoscopy to detect neoplasia and secondarily to provide some stratification of future risk to enable colonoscopy to be better targeted at the highest risk population. In module 3, we prospectively assess whether or not the methylation marker panels generated in module 1 can detect neoplasia in this setting.

Selection of patients

The purpose of the study was to assess if there might be added value from methylation testing above that of colonoscopy. Study subjects were therefore selected from patients within surveillance programmes. One concern was that there would be insufficient cases of neoplasia (not recognised and missed at the index examination) to address the question. No similar study had been carried out on which to base the assessment. We therefore set out to select a high-risk population (see Chapter 4, Inclusion criteria). We also recognised that the surveillance population is heterogeneous and, therefore, to overcome centre bias we would seek to recruit from multiple hospitals across the country.

Introduction into routine practice

To demonstrate clinical value, we set up the DNA testing within a recognised NHS laboratory, rather than a research institute. This would help demonstrate the clinical applicability of the test as well as its generalisability.

Objectives

The primary objective in module 3 was to prospectively evaluate the ability of the methylation assay to detect pre-cancerous lesions (dysplasia) missed by histology within a surveillance programme for colitis-associated neoplasia.

A secondary objective was to estimate the incremental accuracy of methylation testing in addition to colonoscopy and histology assessment within the existing colitis-associated neoplasia surveillance programme, and thereby gain experience of its applicability in the clinical setting.

Trial design

Module 3 of the ENDCaP-C study was a multicentre cohort test accuracy study for enhanced neoplasia detection and cancer prevention in patients with chronic colitis. In the first stage of the study (Figure 14) all participants underwent baseline colonoscopy, which assessed baseline histology and baseline methylation status. Baseline histology results were made available to patients and clinicians and were used to inform their immediate health care. Methylation test results were not released. This baseline methylation test is the index test under evaluation in the study. Those found to be histology negative at baseline were considered to be at risk of having dysplasia missed by colonoscopy and formed the key group of interest in the study to assess whether or not the methylation test can identify cases missed by colonoscopy. To assess this, this subgroup of participants progressed to the second stage (Figure 15).

FIGURE 14.

Study schema: baseline colonoscopies. PIS, patient information sheet. −ve, negative; +ve, positive.

FIGURE 15.

Study schema: repeat colonoscopies. −ve, negative; +ve, positive.

In the second stage, a reference colonoscopy was undertaken 4–12 months after the baseline colonoscopy to identify dysplasia missed at baseline. The histological assessments made from the biopsy samples taken at the second colonoscopy form the reference standard in the study, with which the baseline methylation results were compared. The methylation tests were also repeated in these patients. Participants were selected for invitation to the second stage according to their baseline methylation results – all patients with positive baseline methylation status were invited as well as a random sample of methylation-negative patients (see Sample size for the justification for this design). The second stage reference colonoscopies, histology and methylation tests were undertaken blinded to information from the baseline methylation tests (it was known that the baseline histology was negative in all these cases).

Three additional comparisons were made within the study and are reported:

1. Accuracy of the methylation index test at baseline compared with baseline histology. This comparison assesses the accuracy of methylation testing compared with a reference standard obtained from the histological assessments made at the contemporaneous baseline colonoscopy, and provides a ‘proof of principle’ assessment as to whether or not methylation tests identify clinically evident neoplasia.

2. Agreement of methylation testing at the baseline and reference assessments to assess the stability of methylation assessments.

3. Accuracy of a combined baseline plus follow-up methylation assessment compared with a reference colonoscopy. This comparison redefines the index test according the combination of repeated assessments across two time points.

Participants

Module 3 of the ENDCaP-C study aimed to recruit 1000 patients with chronic colitis. The inclusion and exclusion criteria were as follows.

Consent and registration

Tests and procedures

Serious adverse events

Any adverse events (AEs) meeting the definition of a serious adverse event (SAE) were recorded on a standardised SAE form and faxed to the BCTU within 24 hours of the local principal investigator or member of their research team becoming aware of the event. The principal investigator was responsible for assigning causality to the SAE.

For the purposes of ENDCaP-C module 3, SAEs included (but were not limited to):

• bowel injury/perforation

• post-colonoscopy bleed requiring admission to hospital

• inpatient admission for exacerbation of chronic colitis.

The following are SAEs that could be reasonably expected for this group of patients during the course of the study and did not require reporting on a SAE form:

• hospitalisations for routine treatment or monitoring of the studied indication, not associated with any deterioration in condition

• hospitalisations for treatment, which was elective or pre-planned, for a pre-existing condition that is unrelated to the indication under study, and did not worsen

• admission to a hospital or other institution for general care, not associated with any deterioration in condition

• treatment on an emergency, outpatient basis for an event not fulfilling any of the definitions of serious given above and not resulting in hospital admission

• hospitalisations for planned surgery following non-response to UC treatment.

Disease-related morbidity and routine treatment or monitoring of a pre-existing condition that has not worsened were not considered as SAEs and were not reported.

The SAEs occurring within 1 week from the date of the colonoscopy (baseline or repeat colonoscopy) and not listed as ‘expected’ as defined above were always reportable. SAEs outside this time frame could also be reported if it was felt by the principal investigator that there was a possible causal relationship to an aspect of the study.

Assessments of relatedness and expectedness were undertaken by the module 3 chief investigator (or delegated clinical co-ordinator).

Data collection

Data were collected via paper CRFs (see www.journalslibrary.nihr.ac.uk/programmes/eme/1110029/#/documentation; accessed September 2019) completed, signed/dated and returned to the ENDCaP-C study office by the investigator (or an authorised member of the site research team). Data were input centrally into a bespoke database with built-in validation so that range, date and logic checks could be performed at the point of data entry. Data reported on each CRF were to be consistent with the source data or the discrepancies had to be accounted for. If information was not known, this was to be clearly indicated on the CRF. All missing and ambiguous data were queried. It was requisite to complete all sections on the CRFs.

Statistical considerations

Patient and public involvement

Independent patient representatives were members of the TMG (Jane French) and Trial Steering Committee (TSC) (Heather Beard).

Representatives contributed to the design of the patient pathway and the PIS. Patient representatives attended management meetings before and during the trial. They will also be involved in the development of a lay report for participants.

Ethics approval, regulation and trial registration

Ethics approval for the work that constituted module 1 was granted by South Birmingham Research Ethic Committee (reference 08/H1207/104). Ethics approval for module 3 was granted by South East Coast – Surrey Research Ethics Committees (reference 14/LO/1842).

The trial was conducted in accordance with the recommendations guiding physicians in biomedical research involving human subjects, adopted by the 18th World Medical Association General Assembly, Helsinki, Finland, June 1964 (and its subsequent amendments),52 the Research Governance Framework for Health and Social Care,53 and the applicable UK Statutory Instruments including the Data Protection Act 199854 and the International Conference on Harmonisation Guidelines for Good Clinical Practice. 55

Screening

A total of 4827 patients were assessed for eligibility. Of these patients, 3433 did not meet the eligibility criteria, 522 were eligible but were not recruited and 53 had been identified as potential participants but their eligibility was not established (Table 9).

Reasons for ineligibility were available for 2147 out of the 3433 (62.5%) ineligible patients (see Figure 17). The largest subgroup of these 2147 excluded patients were patients undergoing endoscopic procedures for reasons other than as part of surveillance for cancer prevention (n = 532, 24.8%).

Recruitment

Recruitment commenced in November 2014 and formally ended on 31 March 2017, with a small number of pre-identified patients being entered in April 2017, finishing with 818 recruited patients. There were 32 centres opened to recruitment, of which 31 recruited at least one patient (see Appendix 5, Figure 33 and Table 26). June 2016 was the best recruiting month with 59 patients registered.

Participant flow

The flow of all participants from screening to completion of the baseline colonoscopy is shown in Figure 16. The flow of the participants identified to undergo a repeat colonoscopy is shown in Figure 17.

FIGURE 16.

The Standards for Reporting of Diagnostic Accuracy (STARD) flow diagram; baseline colonoscopy. a, i.e. not enough biomarkers amplified to give meaningful result.

FIGURE 17.

The STARD flow diagram; repeat colonoscopy. a, i.e. not enough biomarkers amplified to give meaningful result.

Figure 17 shows that out of the 191 methylation positives identified, 108 were recalled. In addition, out of the 496 methylation negatives identified, 238 were recalled. The reason for not recalling all of the methylation positives was because, as per the sample size calculation for module 3, we needed 66 methylation positives to be recalled and twice as many methylation negatives to be recalled. However, once a patient was requested to return for a repeat colonoscopy it was not possible to know with certainty how many of the recalled patients will actually attend their repeat colonoscopy, as quite a large proportion of recalled patients declined. Hence, patient recalls were sent until the target of at least 66 methylation-positive patients having had their repeat colonoscopy was reached. We also tried to get as close as possible for the number of repeat colonoscopies for methylation negatives to be twice as many as the methylation positives by the time the study had closed.

Data completeness

The return rates for the CRF, FFPE blocks and pathology reports separated by time point are shown in Appendix 6, Table 27.

Baseline data

The baseline data for module 3 participants are shown in Table 10.

The study population in module 3 represents a group of patients with UC at a higher risk of colorectal neoplasia at the time of baseline colonoscopy. Coexisting risk factors included 24.5% of the population additionally suffering from PSC; the median age was > 50 years (owing to longer duration of disease); and the disease extended beyond the splenic flexure in 62.7% (n = 505) of patients. The proportion of patients on steroids was 10.1%, perhaps reflecting the selected patients representing chronic rather than active disease.

There was an attempt to select for a high-risk population for the purposes of powering the study. The proportion of patients with PSC, a family history of colorectal cancer and their median age were all higher than in the general population of patients with UC. The incidence of dysplasia (11.2%) was almost three times that predicted from historical published data.

Compliance

Of the 193 patients who underwent the reference standard, 12 (6.2%) did not use dye spray and 9 (4.7%) did not follow the standard endoscopic technique (Table 11). Thus, 21 patients were excluded from the sensitivity analysis, which was based on 172 (89.1%) of the patients scoped.

Seventy-seven (39.9%) of the reference colonoscopies were undertaken after 12 months, with a maximum delay of 26.9 months. Twelve of these were also undertaken without dye spray or not using the standard technique. Hence, the per-protocol analysis was based on 104 (53.9%) of the patients scoped.

A final intention-to-treat analysis was undertaken, which included all 193 patients regardless of compliance with the reference standard protocol.

Serious adverse events

There was a single SAE reported in one participant. The event was related to the trial procedure, the colonoscopy, and consisted of a bowel perforation during the reference procedure. The rate is comparable to the 0.1% reported in the PIS: 1 out of 1007 colonoscopies.

There were complications reported for 18 colonoscopies with six pertaining to AEs: bleeding (n = 4), hypotension (n = 1) and stricture (n = 1).

Outcomes

Successful delivery

Module 3 was embedded within the existing surveillance programme for UC within the NHS. We demonstrated that the additional test was acceptable to patients and clinicians and could be incorporated within existing care pathways. One challenging aspect was the collection and secure transfer of samples for centralised processing. This was successfully carried out within/across NHS histopathology and DNA laboratories.

A variety of colonoscopic techniques are currently used across the UK for surveillance of patients with UC. To standardise the late reference examination (carried out in 193 patients), the protocol stipulated the use of dye spray as well as obtaining untargeted biopsies. This protocol was followed in > 90% of reference examinations, providing additional quality assurance to the study.

The protocol stipulated the reference examination was performed between 4 and 12 months after the baseline examination. This was completed in 60% of all reference colonoscopies (n = 116). The delay largely reflected the pressure on NHS endoscopic services. An emerging risk of bias from potential differences in the length of follow-up was identified when difficulties were noted in obtaining reference colonoscopies within the 12-month time-frame stated in the protocol. To overcome this, a matching process was introduced to select and invite methylation-negative reference colonoscopies that matched methylation-positive reference colonoscopies in terms of time from index colonoscopy. The median time to reference standard colonoscopy was slightly longer in the methylation-negative group of the study (median 11.5 vs. 10.1 months). These factors would tend to increase the number of ‘new’ dysplastic lesions occurring in the methylation-negative group, but the impact would be anticipated to be negligible.

Prevalence higher than expected at baseline: neoplasia/dysplasia risk

The study population in module 3 was selected to be at high risk of neoplasia as defined by the associated risk factors seen in the study population. The proportion of patients with identified dysplasia (11.2%) at the baseline colonoscopy was considerably higher than predicted. The high prevalence was not explained by inclusion of PSC patients as their observed risk was lower (9.6%), although their age-matched colorectal cancer risk is substantially higher. This lower dysplasia rate may reflect the younger median age of PSC patients in the study.

Some changes to the eligibility criteria were made during the study. These relaxed the criteria, specifically around the duration of disease and the extent of disease as this was often not recorded in patient records prior to colonoscopy. The high proportion of patients with dysplasia would indicate that this had little impact on the selection of a high-risk population.

Incidence higher than expected at reference colonoscopy

The reference colonoscopy population excluded patients who had neoplasia diagnosed by histological examination of biopsies (at baseline colonoscopy). The rate of dysplasia in the reference colonoscopy population was substantially higher than we anticipated. This may be reasonably predicted within those with a methylation-positive result, as they had been shown to have a significant association with neoplasia at baseline (DOR = 2.4; p < 0.0001). The high incidence of dysplasia in the methylation-negative reference group is more difficult to explain, as it approaches that found at baseline (9% vs. 11.2%). The reason for this is not obvious. It is not explained by de novo lesions as the time frame is too short.

If the reference colonoscopy neoplasia rate correctly reflects the background neoplasia rate in the study population, it suggests that colonoscopy was missing as many patients with dysplasia as were being identified. Given the high rate of identified dysplasia at baseline, this seems unlikely, and the use of central review of histology provided quality assurance to this diagnosis. Alternatively, the rate of progression to neoplasia may be higher than previously recognised. However, as cancer risk in this population does not rise significantly until after 10 years, such an increase would be expected to take place over a much longer time frame.

Acceptance rates for the repeat colonoscopy averaged 56% but were higher in those methylation positive (64%) than those methylation negative (52%), which is a slightly larger difference than could be explained by chance (p = 0.04). Only the study statistician was aware of the methylation result, thus there should not have been any selection bias by methylation status. However, patients and clinicians were aware of the clinical and histological findings of the baseline colonoscopy, and one explanation for the difference in acceptance rates is that the clinician/patient interaction around the baseline colonoscopy may have ‘subliminally’ caused reference colonoscopies to be declined where the index examination was deemed ‘normal’ or accepted otherwise. This may have resulted in a subgroup of higher-risk patients being selected into the methylation-negative group. As it is impossible to blind the clinician and the patient to the findings at the initial examination, it is difficult to see how to avoid this potential source of bias.

Methylation marker relationships

The association between background mucosal methylation changes and associated distant neoplasia in patients with UC was established for the first time in a prospective multicentre study (DOR = 2.4; p < 0.0001). This is a proof of principle for methylation testing as an adjunct to colonoscopy for these patients in increasing the sensitivity of colonoscopic surveillance. In fact, methylation changes were seen in only 50% of patients harbouring neoplastic changes (at baseline). It may be that testing a wider range of methylation sites would increase the sensitivity of the test. It is also possible that there is a subset of patients without field change. If this is the case, these patients would be more suitable for local treatment of dysplasia and so avoid radical surgery (total colectomy). We would plan to investigate the remaining biopsy samples by genome-wide testing to investigate this further. Longitudinal follow-up of this patient population would also go some way to addressing the residual neoplasia risk in the study population.

We investigated the impact of adherence to the protocol for reference colonoscopy in per-protocol analyses restricted to those who complied with the required colonoscopic technique and with compliance to timing, and in an analysis of all patients. It was notable that the discriminatory performance of the methylation panel was highest in analyses that were restricted to the most compliant patient group and lowest when all patients were included, suggesting that the suitability of the reference standard did depend on how and when the follow-up investigation was completed.

Repeat methylation results

At reference colonoscopy, we were able to assess the impact of repeat methylation testing. The numbers were necessarily small, but did show evidence of additional risk stratification, with repeated positive methylation tests being associated with a rising risk of neoplastic change distant to the biopsy site. As surveillance for UC is undertaken over decades for most patients, the ability to better risk stratify patients would be of real clinical value. It would not only select patients more accurately for intensive surveillance, but also select a high-risk population for evaluation of chemopreventative therapy. It was also seen in a subset of patients who methylation changes could be ‘lost’. Whether this reflects a limitation of sampling or perhaps a subgroup of lesser risk patients without true field change is not possible to discern from this study.

This prospective multicentre study has shown for the first time, to our knowledge, a clear association between methylation changes in the background mucosa and distant neoplasia. Additional reference colonoscopy demonstrated that this positive association was sustained, although failed to reach clinical significance. These background mucosa changes were seen in only 50% of patients with neoplasia, but the clinical significance of this remains unclear. Secondary analyses also showed that repeat testing could help refine stratification for surveillance and the true added value of this will be investigated through longitudinal follow-up of this cohort. The assay, as tested, would not yet be recommended for routine clinical use, but this finding may be modified as follow-up data become available.

Outcomes

We have summarised discrimination using the DOR, which combines both sensitivity and specificity into a single number. Values of the DOR for accurate diagnostic tests can be high numbers, for example, when both sensitivity and specificity are 90% the DOR is 81, when they are both 95% it is 361, and when they are both 99% it is 9801. Although such values of diagnostic accuracy were not anticipated, nor are necessary for a test to be a useful stratification tool, the methylation panels were found to have poorer discriminatory performance with observed DOR values of between 1.5 and 4 than hypothesised or observed in module 1 (where observed values were DOR = 8 for model 1, DOR = 19 for model 2 and DOR = 3 for model 3). In most analyses, observed DORs were not shown to be significantly different from an uninformative test (DOR = 1). The sample size for the second stage of the study was based on preliminary data, which predicted that the DOR would be around 19, and thus the study was underpowered to detect less discriminatory tests.

However, the comparison of baseline methylation with the findings of baseline colonoscopy was based on the larger cohort of 773 participants and, thus, was powered to detect less discriminatory tests. This analysis conclusively demonstrated that the relationships between methylation and neoplasia is not explicable by the play of chance, even if it is weaker than hypothesised. The diagnostic odds ratio of 2.37 indicates that a positive methylation test (of a biopsy taken distant from the neoplastic lesion and showing no histological evidence of neoplasia, in a blinded central analysis by Philippe Taniere) was able to select for a population at higher risk of associated concurrent dysplasia. Although the sensitivity of the test, 47.7% (41/86), appears low, it must be emphasised that the methylation test was seeking to compliment, not substitute for, a colonoscopic examination (we were seeking to investigate the ‘added value’ of methylation testing above that observed by colonoscopy alone).

When planning the study, we considered that we expected the methylation test to detect around 50% of missed cases (i.e. with a sensitivity of 50%) and that it would only give false-positive results in the 5% of non-neoplastic cases. Although we observed sensitivities in the range stated, false-positive rates were much higher, with up to 35% of patients without evidence neoplasia being classified as methylation positive. Although this could be reduced by increasing the test positive threshold and increasing the positive predictive value, it would also reduce the proportion of cases detected as positive (as is seen in the baseline plus follow-up combined methylation test).

In technical comparisons in the laboratory, the methylation test was found to be robust and reproducible under stringent testing of > 800 patients. However, methylation status varied between baseline and reference colonoscopies, which may explain why the observed accuracy was lower than expected. The source of this variation is unclear – whether it is a result of within-patient variation in methylation status between sample sites or over time, or undetected laboratory issues. Samples were collected from 31 different hospitals across the country and processed centrally. Despite challenges around transferred samples, and the small size of the biopsies taken, interpretable methylation results were available for over 95% of patients in the study.

The positive methylation rate (191/773, 24.7%) was above our projected positive rate of 20%, based on data from module 1. This may be because of the analysis of multiple biopsies in module 3. The module 1 analysis was performed on single blocks.

Ultimately, the study was underpowered. However, the unprecedented high rate of dysplasia in the methylation-negative group had a considerable impact on this. The high reference dysplasia rate demonstrates the real clinical need for an adjunctive test to colonoscopy for these patients. It also implies that there was some occult clinical bias in the selection of the methylation-negative population.

Further studies

Further analysis using markers that were specific to the background mucosa identified in model 3 is planned. In models 1 and 2, the markers were selected by their occurrence within neoplastic lesions. This is rational if you are seeking to detect neoplasia at its earliest stage of development. However, it is possible to hypothesise that the optimal markers in background mucosa may not be present in the neoplastic tissue. In practice, these markers that seemed to select for neoplasia but were not represented in neoplastic tissue overlapped with those identified in module 1, but included two additional methylation markers.

In addition to the model 3 analysis, we are also planning an exploratory analysis of stool and blood samples, to see if the methylation markers can be identified through non-invasive testing (i.e. without the need for a colonoscopy). Such a test may prove of value to patients deemed at low risk of neoplasia and for whom a colonoscopy may be avoided. The low sensitivity of this current marker panel would suggest that a different marker panel may need to be developed, particularly if we seek to define a low-risk subpopulation.

The most exciting question raised by this study is whether or not methylation changes may precede neoplastic change. The ENDCaP-C study established a unique cohort of patients with a characterised methylation profile. By follow-up of this cohort, it would be possible to determine whether or not methylation changes may be identifying a high-risk population. A longitudinal study is now proposed to follow these patients for the next 5 years. If such a predictive value can be assigned to methylation markers, this will open the opportunity to instigate chemopreventative strategies that might prevent the onset of neoplasia and perhaps reverse these early methylation changes.

Project overview

A unique multidisciplinary team of research scientists, NHS laboratory scientists and clinicians were brought together across over 30 hospitals to undertake this study. This has enabled the first ever multicentre prospective evaluation of predictive molecular biomarkers for patients at risk of UC-associated dysplasia, to our knowledge. The collaboration enabled a large set of historical samples to be collected to evaluate all existing methylation markers in the published literature. It enabled the development of a high-throughput, clinically applicable testing platform to be developed within the NHS. It then enabled this to be tested prospectively across a population of > 800 high-risk patients for the first time.

This work will enable a longitudinal follow-up study to be undertaken and determine the predictive value of these markers. The sample set will also allow some limited testing of new improved markers that might emerge to be rapidly evaluated and incorporated into clinical practice if appropriate, potentially without a further prospective study.

Project limitations

The project was limited to the available technology at the time. If carried out now, more advanced methylation assessment would be possible and assessment of the whole methylome would have been undertaken in module 1. There have also been advances in NGS that might enable more efficient platform testing to be developed.

Performing a repeat endoscopy within the already stretched resources of the NHS has proven to be challenging for the service, the clinicians and the patients. We are extremely grateful for all of their efforts.

The main limitation of the study design would appear to be the potential bias in the selection of the methylation-negative reference population. Unfortunately, it is difficult to see how this might be overcome (even in hindsight). The endoscopist and the patient would expect to discuss the findings at their endoscopy. This will inevitably inform judgement of the undertaking of a further reference examination. Patients and clinicians will inevitably be less enthusiastic to undertake a further examination in the knowledge that the initial examination appeared low risk. This limitation may, however, be mitigated to some extent if we can successfully follow-up the patients to determine whether or not future risk of neoplasia is increased by the initial presence of methylation abnormalities. Funding for follow-up is being sought as an extension to this project.

Bart’s Health NHS Trust

Konstantinos Giaslakiotis, Ceri Guarnieri, Syed Samiul Hoque,* Bianca Petri and Attia Rehman.

Basildon and Thurrock University Hospitals NHS Foundation Trust

Nazar Alsanjari, Wilson Alvares, Georgina Butt, Natasha Christmas, Mark Jarvis,* Denis Lindo, Zia Mazhar, Pushpakaran Munuswamy, Javaid Subhani and Gavin Wright.

Blackpool Teaching Hospitals NHS Foundation Trust

Senthil Murugesan,* Greta Van Duyvenvoorde, Suboda Weerasinghe and Rachael Wheeldon.

Cambridge University Hospitals NHS Foundation Trust

James Yiu Hon Chan, Miles Parkes,* Konstantina Strongili and Merry Jay Jimenez-Smith.

County Durham and Darlington NHS Foundation Trust

Sarah Clark, Anjan Dhar,* Gillian Horner and Sree Mussunoor.

Gateshead Health NHS Foundation Trust

Jamie Barbour,* Sophie Gelder, James A Henry, Paul O’Loughlin, Jitendra Singh and Ann Wilson.

Hampshire Hospitals NHS Foundation Trust

Matthew Brown,* Susan Farmer, Asmat Mustajab, Caroline Palmer and John Ramage.

Heart of England NHS Foundation Trust

Safia Begum, Gerald Langman and Naveen Sharma.*

Hull and East Yorkshire Hospitals NHS Trust

Sarah Ford, Laszlo Karsai, Sally Myers and Shaji Sebastian.*

Imperial College Healthcare NHS Trust

Robert Goldin, Bridget Hradsky, Christwishes Makahamadze, Mike Osborn, Simon Peake and Julian Teare.*

Isle of Wight NHS Trust

Leonie Grellier,* Kamarul Jamil, Andres Kulla, Norman Mounter, Chris Sheen and Joy Wilkins.

Kettering General Hospital NHS Foundation Trust

Laura Aitken and Ajay Verma.*

Leeds Teaching Hospitals NHS Trust

Suzie Colquhoun, Ruth Fazakerley, Olorunda Rotimi, Venkat Subramanian* and Charlotte Wilson.

London North West Healthcare NHS Trust

Ibrahim Al-Bakir, Naila Arebi, Pooja Datt, Ailsa Hart,* Dennis Lim, Ravi Misra, Morgan Moorghen and Lawrence Penez.

Northumbria Healthcare NHS Foundation Trust

Shoba Abraham, Helen Bailey, Sonya Beaty, Jane Dickson, Anthoor Jayaprakash* and Linda Patterson.

Oxford University Hospitals NHS Trust

Nathan Atkinson, Amirah Bouraba, James East, Satish Keshav,* Conor Lahiff, Mary Lucas, Simon Travis, Lai Mun Wang, Kate Williamson and Jean Wilson.

Portsmouth Hospitals NHS Trust

Michelle Baker-Moffatt, Pradeep Bhandari,* Jennifer Hale, Maria Hayes, Kesavan Kandiah, Lisa Murray, David Poller, Sharmila Subramaniam, Sreedhari Thayalasekaran and Prokopios Vogiatzis.

Royal Liverpool and Broadgreen University Hospitals NHS Trust

Tim Andrews, Fiona Campbell, Martyn Dibb, Tracey Forshaw, Alvyda Gureviciute, Hayley Hamlett, Chris Probert,* Sreedhar Subramanian and Clare Shaw.

Royal United Hospital Bath NHS Foundation Trust

Leigh Biddlestone, Dawne Chandler, Ben Colleypriest* and Lucy Howie, Joyce Katebe.

Salford Royal NHS Foundation Trust

Gordon Armstrong, Abby Conlin* and Melanie Taylor.

Sandwell and West Birmingham Hospitals NHS Trust

Ala Ali, Julie Colley, Edward Fogden,* Ann Francis, Anne Hayes, Samia Hussain, Imtiyaz Mohammed, Suhail Muzaffar, Nithin Nair, Mirza Sharjil Baig and Nigel Trudgill.

Sherwood Forest Hospitals NHS Foundation Trust

Stephen Foley,* Cheryl Heeley, Samiya Ibrahim and Wayne Lovegrove.

South Tees Hospitals NHS Foundation Trust

Andrea Boyce, Wendy Jackson, David Oliver, Kolanu Prasad, Arvind Ramadas* and Julie Tregonning.

The Dudley Group NHS Foundation Trust

Clare Allcock, Shanika de Silva,* Sarah Fullwood, Shahkee Garai, Sauid Ishaq, Rizwan Mahmood, Ali Sherif, Veena Shinde and Claire Whitcombe.

The Princess Alexandra Hospital NHS Trust

Deb Ghosh,* Hazel Guth, Nikki Staines and Vasi Sundaresan.

The Royal Bournemouth and Christchurch Hospitals NHS Foundation Trust

Nina Barratt, Emma Gunter, Joan McCutcheon and Simon McLaughlin.*

The Royal Wolverhampton NHS Trust

Matthew Brookes,* Jill Brown, Dragana Cvijan, Marie Green, Shingankutti A M Mangalika, Jayne Rankin, Karen Sedgewick, Asit Shah and Monika Widlak.

University Hospitals Birmingham NHS Foundation Trust

Olufunso Adedeji, Claire Arthur, Lillie Bennett, Neeraj Bhala, Louise Bowlas, Ralph Boulton, Ruth Buckingham, Tamar Cameron, Rachel Cooney, Sheldon Cooper, Christopher Dobson, Jason Goh, Tariq Iqbal,* Kate Kane, Shrikanth Pathmakanthan, Zarqa Shoaib, Jessica Simmons, Nigel Suggett, Philippe Taniere, Tanya Van der Westhuizen and Robert Walt.

University Hospitals of Leicester NHS Trust

John de Caestecker,* Angus McGregor, Alison Moore and Sarah Nicholson.

Walsall Healthcare NHS Trust

Katherine Johanna Arndtz, Ashish Awasthi,* Kiran Desai, Victoria Foster, Rachael Howe, Rangarajan Kasturi and Jasbir Nahal.

West Middlesex University Hospital NHS Trust

Shaila Desai, Joel Mawdsley,* Kevin Monahan,* Metod Oblak, Krishna Sundaram and Anne Thorpe.