This article is mentioned in:

Introduction

Numerous guidelines recommend tumor screening via microsatellite instability (MSI) analysis in all patients with colorectal cancer (CRC), as MSI status has substantial implications in CRC diagnosis (1), prognosis (2) and responses to fluorouracil (3). Fast progress in cancer immunotherapy also confirmed the good response of MSI tumors to checkpoint inhibitor blockade (4,5). Recently, the US Food and Drug Administration (FDA) approved pembrolizumab as the first drug for any solid tumor with MSI-high status (https://www.fda.gov/newsevents/newsroom/pressannouncements/ucm560167.htm); the clinical detection of MSI status is in great demand for patients with CRC or other cancer types.

MSI status is a molecular fingerprint for DNA mismatch repair (MMR) system deficiency. Dysfunctional MMR genes (MLH1, MSH2, MSH6 and PMS2) may introduce length-altering mutations in microsatellites (MS), which are short tandem repeat sequences dispersed throughout the human genome and could be used as deficient MMR (dMMR) markers (6,7). Two of the most common methods for MSI detection have been developed and are still considered the gold standards. MSI-PCR assesses length variability at several standardized MS loci, while immunohistochemical (IHC) staining inspects the protein expression of MMR genes. High frequency MSI (MSI-H) status is determined by the fraction of unstable MS markers (MSI-PCR) or the loss of one or more MMR proteins (IHC). However, considering the various outcomes of dMMR, these traditional methods may have limitations. The indirect IHC method cannot detect all abnormalities of MMR genes and depends heavily on specialized techniques (8). A recent study demonstrated that 5-marker MSI-PCR had inferior sensitivity in 91 prostate tumors among other methods with more markers (9).

With the drastic development of next-generation sequencing (NGS) technologies, the clinical management of cancer patients has been facilitated in a number of ways (10,11). A prominent advantage of NGS detection is that sufficient genetic information can be obtained with much lower costs than low-throughput methods by well-designed tumor sequencing. Therefore, more MS loci could simultaneously be assessed by NGS sequencing to provide new possibilities to evaluate MSI status. Recently, several computational methods for MSI detection, which can be integrated into existing NGS pipelines, have been designed and proposed to provide comprehensive information on tumor genomics. By comparing tumor DNA sequences to normal DNA sequences, such methods can infer MSI status according to either tumor mutation burden (12–14) or read count distribution. Two typical approaches based on read count distribution are mSINGS (15) and MANTIS (16). These methods evaluate each MS locus within a target marker set and report the MSI status by specifying certain cut-offs. However, these approaches differ in their determination of tumor instability. mSINGS compares tumor-only samples to a pre-constructed baseline-control and evaluates tumor instability by the fraction of unstable MS loci within a target set. By contrast, MANTIS compares each target locus in both tumor and matched healthy samples and calculates tumor instability as the averaged stepwise difference across all target loci. These two methods might have disadvantages due to their algorithms. mSINGS expends more time to analyze one sample for the pileup step and statistical test for every locus. However, since MANTIS does not employ a statistical test to calculate the MSI score, this method may be weak at exploiting informative MS loci and may lead to biased results.

Herein, a robust and rapid algorithm called NovoPM-MSI was developed by using a targeted tumor-sequencing panel to reliably assess the MSI status in CRC patients. The read count distribution from paired tumor-healthy samples was used to develop an improved strategy in order to determine tumor instability by examining target MS loci. We also validated this algorithm against conventional MSI-PCR in 113 CRC cases. Finally, we demonstrated the high performance of NovoPM-MSI over mSINGS and MANTIS in MSI status detection and runtime efficiency.

Materials and methods

Sample source and next-generation sequencing

To develop the NovoPM-MSI algorithm, retrospective CRC specimens from 113 patients were utilized. The study was conducted in accordance with the Declaration of Helsinki and the protocol was approved by the Ethics Committee of Shengjing Hospital of China Medical University. All the recruited patients had signed informed consent for their samples to be used in the study and all clinical data and specimens were received anonymously. These specimens were previously analyzed in our CAP-certified laboratory (Tianjin Novogene Med LAB). All cases were tumor tissue DNA extracted from formalin-fixed paraffin-embedded (FFPE) specimens with matched white blood cell (WBC) DNA. These paired cases were sequenced using a targeted gene panel, NovoPM, between January 2016 and October 2017. This panel covers all coding and 21 non-coding regions of the 548 genes associated with multiple tumor types. In the sequencing assay, DNA from tumor and matched WBCs was used to construct sequencing libraries by a hybrid-capture selection method. These libraries were sequenced with a highly uniform depth (targeting >1,000X coverage by non-PCR duplicated read pairs) on the Illumina Hi-Seq X Ten platform as paired-end reads. The protocols and reagents have been optimized to ensure uniform coverage and robust performance for a wide variety of specimens.

Selection of microsatellite marker loci

To identify MS repeats across the human genome with high confidence, a reference FASTA file (hg19/GRCh37) was first scanned using an in-house pipeline. A qualified MS locus was defined as follows: i) The base number within a microsatellite locus should range from 10 to 100; and ii) the minimum length of each motif should be 5. The total number of MS loci across the reference genome was 2,946,833. The mononucleotide repeats covered by the NovoPM panel were extracted from the MS loci in order to identify MS markers for MSI detection among those repeats. Only mononucleotide repeats were selected as MS markers, since these sequences are more sensitive in traditional MSI detection.

Workflow of NovoPM-MSI

A brief illustration of the NovoPM-MSI detection workflow is presented in Fig. 1. During the data pre-processing stage, the sequencing reads obtained after the in-house QC pipeline assessment were aligned using the Burrows-Wheeler Aligner (version 0.7.8) (17) against the reference genome (hg19/GRCh37) and indexed by SAMtools (version 0.1.19) (18). Duplicate reads were removed using Picard Tools (version 1.96) (http://broadinstitute.github.io/picard/) from BAM files. A reference genome and target loci were required as well as paired tumor-normal BAM files in NovoPM-MSI.

Figure 1. Overview of the NovoPM-MSI pipeline. Microsatellite (MS) loci are mononucleotide repeats loci screened from the NovoPM target region (Bed file). The per-locus read length distribution for the repeat types are collected from tumor and normal bam files. Supporting reads are filtered by minimum read length and minimum average base quality, whereas MS locus is filtered by minimum average base quality, abnormal locus length (mean ± 3*SD) and the minimum supporting reads for locus repeat types. The supporting read counts at each MS locus are then normalized for paired samples and a Mann-Whitney U Test between each distribution is performed. Finally, the pass P-value coverage (PPC) is defined as the instability score for paired samples.

Low-quality reads were removed step by step if they met the following conditions: i) reads from paired tumor-normal BAM files that were either too short (<35 bp) or had low average base quality (<25.0); ii) reads covering the MS locus that had delimited length and lower average base quality (<30.0); and iii) reads with no clipped parts that were shorter than a second length threshold (<35). At each MS locus, minimum coverage was required in both tumor and normal samples (>30X) and each repeat type for an MS locus should have at least three qualified reads to support that type. A repeat length that was too long (>3× the standard deviation from mean) was defined as an outlier and was not considered in the subsequent steps.

Qualified supporting reads were then normalized by the average sequencing depth for the tumor and normal samples, prior to determining the instability for each MS locus. In NovoPM-MSI, the length distribution at each MS locus was characterized by the normalized supporting reads in tumor and normal samples. For example, for one MS locus with X number of repeat types, two vectors were obtained by counting the supporting reads for each repeat type in the normal and tumor samples, given as Ni and Ti (i=1, 2, 3, …, X), respectively. This locus was determined as unstable if the two vectors were significantly different by using the non-parametric Mann-Whitney U test (P<0.05).

Finally, an instability score was calculated by the fraction of unstable loci within the target MS markers. An empirical cut-off was set as 0.2 based on the guidelines for defining MSI positivity (15,19).

MSI status determination by PCR

All the paired tumor-normal CRC samples were tested by the MSI-PCR method to obtain the standard MSI status. The PCR panel included six mononucleotide repeat markers (NR21, BAT26, NR27, BAT25, NR24 and MONO27) for MSI testing and three other markers (Penta C, Penta D and Amelogenin) for the sample contamination control. A fluorescence profile of amplified microsatellite DNA from paired samples was produced by capillary fluorescence electrophoresis (ABI 3730×l Genetic Analyzer, Applied Biosystems). If fluorescence peaks were present at any marker in the tumor sample but absent in the normal sample, then that marker was determined as unstable. MSI-H samples had two or more unstable markers, while MSI-L and MSS samples had one or zero unstable markers. Thus, the cut-off for the MSI-PCR method was 33.3% (2/6).

MSI detection by mSINGS and MANTIS

For mSINGS, a baseline control was first constructed by using 26 WBC samples from randomly selected MSI-negative cases. All the tumor samples were then tested by mSINGS, with defaults. The software determined an unstable locus by using the Z-test if the number of the types of repeat lengths was larger than (reference mean + 3*SD), where the reference mean and SD were from the baseline-control. The software reported the percentage of unstable loci, according to which MSI-H samples were determined if the percentage was higher than 20% (15).

Additionally, all the paired samples were tested by MANTIS, which required paired tumor-healthy samples and a target marker set. The input files for MANTIS were the same as those required for NovoPM-MSI. MANTIS calculates the absolute value of the stepwise difference between tumor and healthy samples at each target MS locus using supporting read counts. Then, an average distance value across all the targeted loci was obtained to present the tumor instability. The samples were considered MSI-H when the average distance value was >0.4 (16).

Performance comparison among three NGS-based tools

Three computational tools, NovoPM-MSI, MANTIS and mSINGS, were tested, with all samples having a standard MSI status assigned by MSI-PCR. Four performance indices, sensitivity (SN), specificity (SP), accuracy (ACC) and Mathew's correlation coefficient (MCC), were used to evaluate the detection performance among these tools. All the indices were calculated by convention.

Runtime efficiency was also tested by evaluating the processing time for each sample. Runtime data were compared by tools using pair-wise comparisons and the significance was adjusted by Bonferroni's correction (Welch's t-test, adjusted α=0.017).

Results

Target MS marker of NovoPM-MSI

A total of 19 MS loci of mononucleotide repeats located in non-coding regions were finally selected as the target marker set. These sensitive loci were sequenced after targeted gene enrichment during the NovoPM assay. Detailed information regarding these targets is listed in Table I.

Table I. Microsatellite markers used in NovoMSI detection.

Table I.

Microsatellite markers used in NovoMSI detection.

Chr	Start	End	Function	Gene	Type
1	161309452	161309480	Intronic	SDHC	(T)28
3	12649448	12649472	Intronic	RAF1	(A)24
3	37067099	37067120	Intronic	MLH1	(T)21
3	185802071	185802094	Intronic	ETV5	(A)23
4	55598211	55598236	Intronic	KIT	(T)25
6	135527343	135527365	Intronic	MYB	(A)22
7	140498359	140498380	Intronic	BRAF	(T)21
7	140508151	140508177	Intronic	BRAF	(A)26
8	38303723	38303746	Intronic	FGFR1	(T)23
8	48732074	48732095	Intronic	PRKDC	(A)21
9	5062500	5062531	Intronic	JAK2	(A)31
9	87357704	87357731	Intronic	NTRK2	(T)27
9	87479651	87479673	Intronic	NTRK2	(T)22
12	11962746	11962772	ncRNA_intronic	ETV6	(T)26
13	49039094	49039118	Intronic	RB1	(T)24
15	99192754	99192778	ncRNA_exonic	IRAIN/IGF1R	(T)24
17	29559061	29559087	Intronic	NF1	(T)26
21	42863078	42863102	Intronic	TMPRSS2	(T)24
X	123195593	123195618	Intronic	STAG2	(T)25

Validation of three computational methods against conventional MSI-PCR

Among the 113 paired CRC tumor-normal cases, nine samples were assigned MSI-H and 104 samples were assigned MSS according to the gold standard of MSI-PCR. The fraction of MSI-H cases in the present study collection was 7.9%, which was slightly lower than the number of MSI-H cases in the larger CRC population (15-20%) (20,21).

The initial cut-offs to determine the MSI-H status for NovoPM-MSI, MANTIS and mSINGS were set at 0.2, 0.2 and 0.4, respectively. The validation of these methods against the MSI-PCR method is listed in Table II. All three computational methods correctly detected eight positive cases and one false negative case, which were not same for each tool, thus reaching the same sensitivity of 88.9%. These analyses differed in specificity with descending order of mSINGS (103/104, 99%), NovoPM-MSI (101/104, 97.1%) and MANTIS (90/104, 86.5%). According to Matthew's correlation coefficient, a more balanced measure, the performance of NovoPM-MSI (0.786) was slightly inferior to that of mSINGS (0.879), with MANTIS (0.516) showing the worst performance.

Table II. Performance comparison among NovoMSI, MANTIS and mSINGS, using microsatellite instability-PCR results as the gold standard.

Table II.

Performance comparison among NovoMSI, MANTIS and mSINGS, using microsatellite instability-PCR results as the gold standard.

Method	TP	TN	FP	FN	SN (%)	SP (%)	ACC (%)	MCC
NovoPM-MSI	8	101	3	1	88.90	97.10	96.50	0.786
MANTIS	8	90	14	1	88.90	86.50	86.70	0.516
mSINGS	8	103	1	1	88.90	99.00	98.20	0.879

[i] TP, true-positive; TN, true-negative; FP, false-positive; FN, false-negative; SN, sensitivity; SP, specificity; ACC, accuracy; MCC, Matthew's correlation coefficient.

Detection comparison among NovoPM-MSI, MANTIS and mSINGS

By plotting the instability score against samples, the detection performance among these computational tools was further assessed. NovoPM-MSI showed the least fluctuations within either the MSI-H (Fig. 2, left side of vertical line) or MSS category (Fig. 2, right side of vertical line). Although mSINGS had the highest MCC (0.879, Table II), the two misclassified samples were far from the default cut-off (0.2).

Figure 2. Microsatellite instability (MSI) scores from NovoPM-MSI (diamond), MANTIS (up triangle) and mSINGS (left triangle) for 113 CRC cases. The vertical line delimits all samples into MSI-high (left part) and MSS (right part) groups. The nine MSI-high samples determined by MSI-PCR are circled in black. The two horizontal lines indicate default cut-offs as references, with 0.2 for NovoPM-MSI and mSINGS and 0.4 for MANTIS. NovoPM-MSI shows the least fluctuations across the whole range of MSI status.

MANTIS showed the worst performance, with an accuracy of 86.7% (98/113) for the 14 false-positive samples that were reported (default cut-off: 0.4) and an acute fluctuate instability score produced.

In addition, the steady decrease in the instability score reported by NovoPM-MSI across MSI-H and MSS samples suggested an intermediate status for the MSI phenotype. We then modified a calling strategy of NovoPM-MSI by setting an ambiguous interval for the instability score. The small range was optimized by current observations set from 0.17 to 0.23 to include four samples with ‘false’ result. The samples were explicitly assigned MSI-H or MSS if the instability scores reported by NovoPM-MSI were >0.23 or <0.17. The samples were assigned MSI-ambiguous if the instability score fell into the ambiguous interval and these samples were recommended for further MSI-PCR testing. By setting the ambiguous interval, NovoPM-MSI reached 100% sensitivity and specificity. However, such intervals could not be properly utilized to improve the sensitivity and specificity of mSINGS.

Runtime comparison among NovoPM-MSI, MANTIS and mSINGS

In addition to the accuracy of detection, the runtime for each sample is also important when evaluating tool performance. The ranking of runtime efficiency was MANTS (16.3 sec)>NovoPM-MSI (18.3 sec)>mSINGS (109.0 sec). All the pair-wise comparisons were highly significant among the three tools (Welch's t-test, P<0.01, Fig. 3).

Figure 3. Comparison of the runtime efficiency among mSINGS, MANTIS and NovoPM-MSI (n=113). Data show the mean and standard error of the mean. Letters a, b, and c above the error bars represent three comparison groups. Data with different letters are significantly different (Welch's t-test, P<0.01). The significance level after Bonferroni's correction is 0.017.

Discussion

In the present study, we described an NGS-based strategy, NovoPM-MSI, for MSI detection in paired tumor-healthy samples. We selected 19 mononucleotide MS repeats as target loci (Table I) since several studies have shown that mononucleotide repeats are the most sensitive in MSI detection (22–24). This new method evaluates each of the target MS loci by comparing the length distribution between tumor and normal samples and then calculates the percentage of unstable MS loci and reports the instability score as well as the MSI status according to empirical cut-offs. NovoPM-MSI has comparable reliability to the gold standard MSI-PCR, with robust performance (Table II).

We also compared NovoPM-MSI to two other NGS-based MSI detection tools, mSINGS (15) and MANTIS (16). All three tools exploit information from the read count distribution but differ in the instability evaluation. MANTIS assesses step-wise differences between tumor and matched normal samples at the same MS locus and calculates an average difference across all the MS loci to reflect the general instability of a sample but not a single locus. Notably, the aggregate averaged distance may be biased when using relatively fewer markers. One advantage of MANTIS is that this approach is less susceptible to sequencing errors or poorly performing loci, but this advantage becomes increasingly apparent when using more target loci, which may be limited in most gene panels. In the present study, MANTIS produced the most false-positive calls (14 in 104) and showed the most fluctuations in the instability score (Table II and Fig. 2). mSINGS utilizes incidentally sequenced MS loci and evaluates the instability of each MS locus by comparison with a reference distribution generated from pooled normal samples. This approach assesses single locus instability using the Z-score test and reports the percentage of unstable loci as an instability score. mSINGS performed well for the 113 cases in the present study, with both high sensitivity (88.9%) and specificity (99%). However, its disadvantage in runtime, which was nearly six times longer than that of MANTIS, was also evident (Fig. 3). Additionally, the two misclassified samples could not be saved through setting an ambiguous interval. By contrast, NovoPM-MSI combines the advantages of the above tools. First, this approach examined an intermediate marker set (Table I), which was customized for most gene panels. Second, NovoPM-MSI runs much faster than mSINGS (18.3 vs. 109.0 sec) and has a runtime comparable to that of MANTIS (16.3 sec). Finally, by setting an ambiguous interval (0.17-0.23) from empirical observations, the detection performance of NovoPM-MSI was greatly improved compared with that of the other tools. In short, NovoPM-MSI efficiently detected MSI status in a robust and rapid manner.

The advantages of NGS-based methods for MSI detection over traditional methods are obvious. MSI status can be simultaneously available with other genomic alterations, including SNV, gene fusion and indels, by NGS pipelines, without dedicated MSI detection. In addition, NGS-based methods enable the examination of more MS loci than that of the MSI-PCR method, providing a comprehensive assessment of MSI status. Although considered the gold standard, MSI-PCR reportedly shows 97% sensitivity and 95% specificity (25). In the present study, the three NGS-based methods showed performances comparable to those of conventional PCR (Table II). Recent studies have suggested that the MSI-H phenotype should be subdivided to guide important clinical applications (15,26–28). The rough binary classification of MSI only provides qualitative results and cannot meet this requirement. Alternatively, NGS-based methods can fill the gap by reporting both qualitative status and quantitative instability score. The NovoPM-MSI method reported in the present study showed reasonable classification (Fig. 2).

The present study may have some limitations. The NovoPM-MSI method may be limited by the small number of MSI-H cases (9/113), primarily due to the low frequency of MSI-H phenotypes in CRCs driven by MMR deficiency (~15%) (20,21). The accuracy of NovoPM-MSI would be increased if more MSI-H cases were accumulated, regardless of the one or two cut-offs used. We are now endeavoring to collect more CRC cases with different MSI status to build large validation cohort for NovoPM-MSI. In addition, using white blood cells as a normal control is relatively easy, but in most clinical settings, only tumor samples are available (e.g., retrospective FFPE without matched normal specimen). Under such conditions, tumor-only methods, including mSINGS, are required. In addition, MSI status is most closely studied in CRCs; it is found to be present in endometrial, ovarian, cervical, gastric and prostate cancers (27). The current NovoPM-MSI method has been developed for CRCs, but its applicability to other cancers may be tested in the future if cases from more cancer types are accumulated.

In conclusion, findings of the present study show that NovoPM-MSI can rapidly detect MSI status with high accuracy and robustness compared to conventional MSI-PCR and other NGS methods based on read count distributions. As the genetic characteristics of tumors become increasingly more critical in the clinical management of patients, we hope that the NovoPM-MSI method will provide more valuable information in clinical service when conducting tumor sequencing.

Acknowledgements

The authors would like to thank Ms Wanchun Zang, an experienced analyst for clinical data (Beijing Novogene Bioinformatics Technology Co. Ltd.), for her helpful suggestions on the manuscript.

Funding

This work was supported by grants from the research program of Novogene Bioinformatics Technology Co., Ltd. (grant no. P101ZY17010010).

Availability of data and materials

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Authors' contributions

LZ, GS, LL, GC and XZ designed the study. LZ and GS improved the algorithm, analyzed and interpreted all the data. LZ, GS, LL and GC wrote the manuscript. YY contributed to sample collection and project coordination. GC and XZ provided the research materials, supervised the study and edited the manuscript. All authors read and approved the final version of the manuscript.

Ethics approval and consent to participate

The present study was approved by the Ethics Committee of Shengjing Hospital of China Medical University. Written informed consent was obtained from all the patients.

Patient consent for publication

Not applicable.

Competing interests

LZ, GS, LL, YY and GC are employees of Beijing Novogene Bioinformatics Technology Co., Ltd. The authors declare that they have no competing interests.

References