Sequencing‐based microsatellite instability testing using as few as six markers for high‐throughput clinical diagnostics

Abstract Microsatellite instability (MSI) testing of colorectal cancers (CRCs) is used to screen for Lynch syndrome (LS), a hereditary cancer‐predisposition, and can be used to predict response to immunotherapy. Here, we present a single‐molecule molecular inversion probe and sequencing‐based MSI assay and demonstrate its clinical validity according to existing guidelines. We amplified 24 microsatellites in multiplex and trained a classifier using 98 CRCs, which accommodates marker specific sensitivities to MSI. Sample classification achieved 100% concordance with the MSI Analysis System v1.2 (Promega) in three independent cohorts, totaling 220 CRCs. Backward–forward stepwise selection was used to identify a 6‐marker subset of equal accuracy to the 24‐marker panel. Assessment of assay detection limits showed that the 24‐marker panel is marginally more robust to sample variables than the 6‐marker subset, detecting as little as 3% high levels of MSI DNA in sample mixtures, and requiring a minimum of 10 template molecules to be sequenced per marker for >95% accuracy. BRAF c.1799 mutation analysis was also included to streamline LS testing, with all c.1799T>A variants being correctly identified. The assay, therefore, provides a cheap, robust, automatable, and scalable MSI test with internal quality controls, suitable for clinical cancer diagnostics.


| INTRODUCTION
Increased microsatellite instability (MSI) is a hallmark of mismatch repair (MMR) deficiency, which affects approximately one in six colorectal cancers (CRCs; Boland et al., 1998). Lynch syndrome (LS) is an inherited predisposition to cancer caused by germline pathogenic variants affecting one allele of an MMR gene and accounts for approximately one in five MMR deficient CRCs (Hampel et al., 2008).
MMR deficiency is also associated with tumor response to immune checkpoint blockade therapy, irrespective of tissue of origin (Le et al., 2017). Therefore, the assessment of MSI, or MMR, status can inform patient management and is recommended in all CRCs by national and international guidelines to screen for LS (Balmana, Balaguer, Cervantes, & Arnold, 2013;Newland et al., 2017;Stoffel et al., 2015). Once identified, patients with LS benefit from surveillance colonoscopy, prophylactic surgery, and chemoprevention (Burn et al., 2011;Vasen et al., 2013).
MMR status of tumors is commonly assessed by immunohistochemistry (IHC) of MMR proteins, or polymerase chain reaction (PCR) fragment length analysis (FLA) of microsatellites to detect increased MSI. MMR deficiency is inferred from the absence of at least one MMR protein, or high levels of MSI (MSI-H). MSI-H is defined by mutation of ≥30-40% of microsatellites analyzed (Boland et al., 1998). These methods are highly sensitive and specific, with reported sensitivities and specificities of 93% and 95% for IHC of all four MMR proteins (Shia, 2008), and 97% and 100% for FLA of mononucleotide repeats (MNRs; Bacher et al., 2004). IHC and FLA also perform well with respect to other demands of diagnostic tests. FLA is considered highly reproducible, with 98% concordance of results observed between independent laboratories (Zhang, 2008), although IHC shows some heterogeneity due to discordant interpretation of variable staining, and use of different antibodies (Shia, 2008). FLA has been shown to be reliable when sample tumor cell content is ≥10% (Berg et al., 2000), and IHC can detect focal MMR deficiency (Chapusot et al., 2002). Both are also considered to be relatively cheap and cost-effective for LS screening (Snowsill et al., 2014).
However, the uptake of MMR deficiency testing has been poor; only 28% of 152,993 CRC cases were analyzed during [2010][2011][2012] in the USA (Shaikh, Handorf, Meyer, Hall, & Esnaola, 2018), with a similar proportion being analyzed in the UK This is despite guidelines recommending testing and estimates that only 1.2% of LS gene carriers were known to clinical services in the US in 2011 (Hampel & de la Chapelle, 2011). We estimate that only 5% of carriers are currently known in the UK.
Automated sequence analysis is better suited to high-throughput diagnostics than FLA, or IHC, leading to the development of nextgeneration sequencing (NGS)-based MSI assays that analyze microsatellites captured by gene panel sequencing. These determine the mutation status of each microsatellite from the frequency of length variants detected, and then use the proportion of microsatellites that are mutated to classify a sample. Several such classifiers have reported sensitivities and specificities >95% (Kautto et al., 2016;Zhu et al., 2018), and have identified samples misclassified by conventional MMR deficiency tests, highlighting that there is no gold standard reference method (Hause, Pritchard, Shendure, & Salipante, 2016). Gene panel sequencing also allows additional clinically actionable markers to be simultaneously assessed. For example, separate testing for the BRAF c.1799T>A variant (p.V600E), following FLA or IHC, is recommended to increase the specificity of LS screening (Newland et al., 2017), but both MSI and BRAF can be analyzed by a single tumor sequencing assay (Hampel et al., 2018).
However, the high cost of gene panel sequencing (Marino et al., 2018) may be a barrier to its widespread deployment for MSI testing, or for the detection of LS by MMR gene sequencing. Targeted NGSbased MSI assays that use multiplex amplification of specific panels of microsatellites have been developed that, similar to gene panelbased methods, classify samples by the proportion of microsatellites that are mutated (Gan et al., 2015;Hempelmann et al., 2018;Hempelmann, Scroggins, Pritchard, & Salipante, 2015;Waalkes et al., 2018). However, even when using the same method, different marker proportions can be used as a classification threshold with different marker sets (Hempelmann et al., 2015;Hempelmann et al., 2018;Kautto et al., 2016;Waalkes et al., 2018;Zhu et al., 2018), and thresholds can be uncertain when relatively few microsatellites (<20) are analyzed (Hempelmann et al., 2015).
We have previously used amplicon sequencing of short (7-12 base pairs [bp]), monomorphic MNRs to classify the MSI status of CRCs, without needing matched normal tissue (Redford et al., 2018). Short MNRs were selected as longer (>15 bp) microsatellites are associated with increased PCR and sequencing error (Fazekas, Steeves, & Newmaster, 2010), and it has been reported that 9-15 bp microsatellites give the greatest differences in mutation frequencies between MSI-H and microsatellite stable (MSS) samples using NGS (Maruvka et al., 2017). Our method for MSI detection accounts for the individual sensitivity and specificity of each marker, and achieved >97% accuracy in 209 CRCs with only 17 markers, using FLA as the reference method (Redford et al., 2018).
However, the protocol required singleplex PCR amplification, followed by the second round of PCR to prepare the amplicons for sequencing. Here, we modify this method to develop an MSI assay suitable for clinical cancer diagnostics. We use singlemolecule molecular inversion probes (smMIPs; Hiatt, Pritchard, Salipante, O'Roak, & Shendure, 2013) to amplify in multiplex and sequence 24 short MNRs, and show that the assay achieves 100% accuracy with as few as six markers. We also include BRAF c.1799 sequencing for streamlined LS screening (Newland et al., 2017). To establish the assay is suitable for clinical practice, we follow joint guidelines from the Association for Molecular Pathology and the College of American Pathologists (Jennings et al., 2017). This includes validation of diagnostic accuracy using independent sample cohorts, assessment of reproducibility and detection limits, the definition of quality control criteria, and deployment in an independent diagnostic laboratory. were considered equivalent to MSS samples (Halford et al., 2002).

HCT116 and K562 cells were grown in the Roswell Park Memorial
Institute growth medium containing 2 mM L-glutamine (Gibco), 10% fetal bovine serum (Gibco), 60 µg/ml penicillin, and 100 µg/ml streptomycin (Gibco) at 37°C and 5% CO 2 . HCT116 cells were passaged or harvested at 80-90% confluence by decanting expired growth medium, washing in 5 ml phosphate-buffered saline (Gibco), and detaching the cells using 0.05% trypsin-ethylenediaminetetraacetic acid (Gibco). K562 cells were passaged or harvested at a density of 1 × 10 6 cells/ml. DNA extracted from HCT116 CRC cell line (MLH1 deficient) was used as an MSI-H control. DNA extracted from K562 chronic myeloid leukemia cell line was used as an MSS control.

| DNA extraction and quantification
DNA was extracted from FFPE CRC tissue using the GeneRead DNA FFPE Kit (Qiagen). DNA was extracted from cell lines using the Wizard Genomic DNA Purification Kit (Promega). DNAs were quantified using QuBit 2.0 Fluorometer (Invitrogen) and QuBit dsDNA BR/HS Kits (Invitrogen).

| Markers and smMIP design
The marker panel includes 24 MNRs, previously published by Redford et al. (2018), for MSI classification, as well as BRAF c.1799 to screen for sporadic MSI-H CRCs (Newland et al., 2017). MIPgen (Boyle, O'Roak, Martin, Kumar, & Shendure, 2014) was used to generate smMIP sequences for each marker. MIPgen parameters were: Tag size 6, 0, minimum capture size 120, and maximum capture size 150. smMIP designs were selected by the following criteria: No common single nucleotide polymorphisms (SNPs) in the smMIP extension or ligation arms, a logistic score >0.8, and successful amplification of loci.
Marker loci and smMIP sequences are detailed in Table S2.

| Oligonucleotide synthesis
smMIPs and primers for amplification and sequencing (Table S3) 92.5-100.0%; Figure 1a). Data filtering using smMIP molecular barcodes to reduce sequencing error (Hiatt et al., 2013) did not improve sample separation by the classifier (Supporting Information S1), and therefore was not employed for MSI classification. The 15 markers remaining from the 17-marker panel of Redford et al. (2018) also achieved 100% sensitivity and specificity (data not shown), indicating redundancy in the marker panel. Backward-forward stepwise selection was used to define a subset of six short MNRs (Table S2) with accuracy equal to the 24 marker panel ( Figure 1A).
An independent validation cohort of 50 MSI-H and 49 MSS was sequenced and analyzed, and 100% sensitivity (95% CIs: 92.9-100.0%) and 100% specificity (95% CIs: 92.8-100.0%) was again achieved using all 24 markers, and the 6-marker subset ( Figure   1b). To assess assay reproducibility, 16 MSI-H and 16 MSS CRCs from the validation cohort were amplified, sequenced, and classified a second time. The classification was 100% concordant, and scores were strongly correlated between sample repeats, using both 24 markers (β = .97, R 2 = .97), and the 6-marker subset (β = 1.01, R 2 = .97).  (Jennings et al., 2017). Of the 32 CRCs that tested negative, 30 had VAFs ≤0.60%, in line with observed MiSeq base-calling error rates of 0.62% (May et al., 2015). The remaining two samples had VAFs of 1.67% and 1.72%, suggesting they may contain low-frequency variants below the detection limit of HRM (Nikiforov et al., 2009). Consistent with this, when molecular barcodes were used to reduce sequencing error (Hiatt et al., 2013), these variants were found at frequencies of 1.82% and 0.46%, respectively ( 3.4 | MSI classification is accurate when 10 or more molecules are sequenced per marker Whilst we found no improvement to classifier performance using molecular barcodes to correct sequencing error (Supporting

| 337
and validation cohorts had a mean molecular barcode detected per marker <75 (Table S1), and all were correctly classified with either the 24-or 6-marker panels.
To explore the minimum number of template molecules that need to be sequenced for accurate classification, we also performed an in silico resampling of sequencing data. Analysis of the nine-sample dilution series gave a strong correlation between classifier scores from empirical observations and from resampling (β = .92, R 2 = .96; Supporting Information S3). Resampling of the CRCs included in the validation cohort was used to increase the number of observations, and it was found that using the 24-and 6-marker panels, sequencing of ≥10 and ≥15 molecular barcodes per marker, respectively, gave a correct classification of >95% of samples (Figure 3b). It should be noted that these estimates were obtained from resampling highquality sequencing data. Therefore, the mean of 75 molecular barcodes per marker obtained empirically provides a more conservative threshold for diagnostic use.

| Validation in an independent clinical laboratory
Assessment of an assay's performance in an independent clinical laboratory supports that it is a reproducible method, suitable for wider adoption (Jennings et al., 2017). To test this, our smMIPbased MSI assay was set up by the Northern Genetics Service (Newcastle Hospitals NHS Foundation Trust, Newcastle upon Tyne, UK) using our protocols, and our smMIP and primer oligonucleotides. All other reagents and equipment were distinct from those used during assay development, and the personnel running the assay were independent of our research team. Once established, a further 23 independent CRCs were analyzed using the assay, and it again achieved 100% sensitivity (95% CIs: 79.4-100.0%) and 100% specificity (95% CIs: 59.0-100.0%) relative to the MSI Analysis System v1.2, when classifying samples with both the 24-and 6-marker panels ( Figure 4). Although four samples had <75 molecular barcodes per marker detected (Table S1), they were accurately classified in agreement with reading sampling predictions, and so were not resequenced at a higher depth.

| DISCUSSION
The MSI assay presented here achieved 100% accuracy of MSI classification in 220 CRCs, relative to the MSI Analysis System v1.2 (Promega), using only tumor DNA and as few as six microsatellite markers. We found no improvement to classifier performance using molecular barcodes for sequencing error correction (Hiatt et al., 2013; Supporting Information S1). This is likely due to our use of short MNRs with flanking SNPs, selected from genome-wide data, and classification method. Shorter microsatellites have lower PCR and sequencing error rates compared with longer microsatellites (Fazekas et al., 2010), while the SNPs flanking the microsatellites provide additional discrimination between error and true microsatellite mutations. Classification by a naïve Bayesian approach accounts for individual marker sensitivity, specificity, and sequencing error rate (Redford et al., 2018). However, molecular barcodes are used in our assay to provide a quality control metric by estimating the number of independent molecules sequenced (Jennings et al., 2017). We have also shown, previously, that molecular barcodes are useful for the detection of much lower frequency microsatellite variants, found in the PBLs of patients with constitutional mismatch repair deficiency (CMMRD; Gallon et al., 2019).
To show that the assay is suitable for clinical practice, we tested its clinical validity according to published guidelines (Jennings et al., 2017 avoids expenditure on additional tests as a single tumor assay is required before germline testing. It also demonstrates the modularity of the assay, which can be expanded to cover additional clinically relevant markers, or adapted to different tumor types, with ease since thousands of smMIPs can be multiplexed (Hiatt et al., 2013;Oud et al., 2017). MLH1 promoter methylation is an alternative marker to the BRAF c.1799T>A variant to exclude sporadic MMR deficient patients with CRC from germline testing and has superior specificity for LS detection (Pérez-Carbonell et al., 2010). However, this also excludes MLH1 mutation carriers who have methylation as the second hit in their tumor (Moreira et al., 2015), or have germline epimutations (Suter, Martin, & Ward, 2004).
In summary, the MSI assay outlined here is accurate, reproducible, robust to sample heterogeneity, and includes both internal quality controls and sample identification. The automatable laboratory workflow and analysis, and the need for as few as six microsatellite markers at moderate read depths provide a cheap and scalable option for high-throughput MMR deficiency testing.

ACKNOWLEDGMENTS
The authors would like to thank Illumina Cambridge Ltd. for their inkind support of MiSeq™ Reagent Kits, which were used in the performance of this study. The authors would also like to thank Dr. Research UK (A15934) is gratefully acknowledged. The funders had no role in the study design, sample collection, and data analysis, decision to publish, or preparation of the manuscript.