A multifactorial likelihood model for MMR gene variant classification incorporating probabilities based on sequence bioinformatics and tumor characteristics: a report from the Colon Cancer Family Registry

Abstract

Mismatch repair (MMR) gene sequence variants of uncertain clinical significance are often identified in suspected Lynch syndrome families, and this constitutes a challenge for both researchers and clinicians. Multifactorial likelihood model approaches provide a quantitative measure of MMR variant pathogenicity, but first require input of likelihood ratios (LRs) for different MMR variation-associated characteristics from appropriate, well-characterized reference datasets. Microsatellite instability (MSI) and somatic BRAF tumor data for unselected colorectal cancer probands of known pathogenic variant status were used to derive LRs for tumor characteristics using the Colon Cancer Family Registry (CFR) resource. These tumor LRs were combined with variant segregation within families, and estimates of prior probability of pathogenicity based on sequence conservation and position, to analyze 44 unclassified variants identified initially in Australasian Colon CFR families. In addition, in vitro splicing analyses were conducted on the subset of variants based on bioinformatic splicing predictions. The LR in favor of pathogenicity was estimated to be ~12-fold for a colorectal tumor with a BRAF mutation-negative MSI-H phenotype. For 31 of the 44 variants, the posterior probabilities of pathogenicity were such that altered clinical management would be indicated. Our findings provide a working multifactorial likelihood model for classification that carefully considers mode of ascertainment for gene testing.

Introduction

Lynch syndrome is an autosomal dominantly inherited disorder of predominantly colorectal and endometrial cancer predisposition caused by germline pathogenic variants in the DNA mismatch repair (MMR) genes MLH1 (MIM# 120436), MSH2 (MIM# 609309), MSH6 (MIM# 600678), and to a lesser degree PMS2 (MIM# 600259) (Lynch, et al., 2009; Viel, et al., 1998; Wang, et al., 1999). These pathogenic variants cause a functional defect in the DNA MMR complex, which leads to high levels of DNA microsatellite instability (MSI). Tumor MSI testing, and/or MMR protein immunostaining, is commonly used to identify potential Lynch syndrome cases (Aaltonen, et al., 1998; Lindor, et al., 2002; Salovaara, et al., 2000; Terdiman, et al., 2001). However, 10 to 15% of colorectal tumors detected in the general population have an MSI-H phenotype (Cunningham, et al., 1998; Herman, et al., 1998; Miyakura, et al., 2001; Toyota, et al., 1999), with molecular evidence indicating that these are largely caused by epigenetic silencing of the MLH1 promoter (Deng, et al., 2004; Domingo, et al., 2004; Domingo, et al., 2005; Lubomierski, et al., 2005; McGivern, et al., 2004; Wang, et al., 2003; Young, et al., 2005). A single somatic mutation in the oncogene BRAF (MIM# 164757, BRAF c.1799T>A p.(Val600Glu)) is frequently present in MSI-H cancers with MLH1 methylation (Davies, et al., 2002; Rajagopalan, et al., 2002), but not in MSI-H cancers arising in patients with a germline MMR gene pathogenic variant (Deng, et al., 2004; Domingo, et al., 2005; Lubomierski, et al., 2005; McGivern, et al., 2004; Wang, et al., 2003). Therefore BRAF p.Val600Glu has potential as a molecular marker to help distinguish MSI-H colorectal cancers caused by germline MMR gene pathogenic variant.

Between 20–50% of MMR gene sequence variants found by germline genetic testing in suspected Lynch syndrome colorectal and endometrial cancer cases are reported to be of uncertain clinical significance (Peltomaki and Vasen, 2004; Woods, et al., 2007). Establishing the pathogenicity of unclassified variants has direct clinical relevance, since pathogenic variant status determines clinical management with respect to genetic counselling approaches, presymptomatic screening, and choice/timing of possible prophylactic surgery for carriers.

Multifactorial likelihood analysis (also termed Bayesian Integrated evaluation) was established to address the issue of unclassified variants in the breast cancer susceptibility genes BRCA1 (MIM# 113705) and BRCA2 (MIM# 600185) (Goldgar, et al., 2008; Goldgar, et al., 2004). This approach provides a quantitative measure of variant pathogenicity, but first requires input of likelihood ratios (LRs) for different pathogenic variant-associated characteristics from appropriate, well-characterized reference datasets, with each LR comparing the probability of the observed data under the hypothesis that the variant is pathogenic compared to the corresponding probability that the variant is benign with respect to cancer risk. This method has since been proposed to assess cancer risk associated with rare variants of uncertain clinical significance for other cancer gene syndromes (Goldgar, et al., 2008; Tavtigian, et al., 2008), particularly since it provides a measure of causality in the form of the estimated posterior probability that a variant is pathogenic that can then be categorised into a qualitative five class scheme that is linked to clinical recommendations (Plon, et al., 2008).

We and others have proposed approaches to apply the multifactorial likelihood model to classification of MMR gene variants (Couch, et al., 2008; Hofstra, et al., 2008; Spurdle, 2010), and have developed methodology to estimate the likelihood of pathogenicity based on variant segregation with disease in families (Arnold, et al., 2009). The extension of the multifactorial approach to include other features was recently reported in a study assessing 35 unclassified variants identified in Italian clinic-based families (Pastrello, et al., 2011). However, the stringent selection criteria for MMR gene sequencing of the reference sets used for derivation of LR estimates for tumor characteristics by Pastrello et al suggests that the LR estimates for tumor characteristics assigned in this Italian study were likely upwardly biased.

In this study, we have used a detailed set of qualitative criteria to consistently assign pathogenicity of MMR gene variants for a large set of colorectal cancer (CRC) cases (probands) with well-documented selection criteria for ascertainment and molecular characterization. We then compared the tumor MSI and BRAF mutation characteristics of probands with and without pathogenic MMR gene variants to estimate tumor LRs. Separate LRs were derived based on ascertainment criteria (i.e. clinic or population-based) to provide unbiased predictors of pathogenicity for inclusion in the multifactorial model, and were only applied to the relevant group. These tumor LRs were used together with existing approaches to assess variant segregation with disease in families, and estimates of prior probability of pathogenicity based on sequence conservation and position (see accompanying manuscript), to analyze 54 sequence variants in MLH1, MSH2 and MSH6 from 162 families. In addition, in vitro splicing analysis was performed for the subset of variants for which in silico analyses predicted a potential to cause splicing aberrations. We demonstrate the utility of both multifactorial likelihood and splicing analyses to classify exonic and intronic variants of MMR genes.

Methods and Materials

Subjects used in this study were recruited as colorectal cancer probands into the Colon Cancer Family Registry (Colon CFR) (Newcomb, et al., 2007). This includes probands providing data for reference sets (outlined below), and families analyzed for unclassified variant evaluation. Written informed consent was obtained from all participants. Recruitment of subjects and this laboratory-based study have been approved by the Human Research Ethics Committee of participating Institutions.

Selection criteria and testing for germline MMR gene sequence variation in the Colon CFR probands is described in Figure 1. All MMR gene sequence variants lodged in the Colon CFR database as of August 2009 were converted to HGVS nomenclature using the following cDNA reference sequences: NM_000249.3 for MLH1, NM_000251.1 for MSH2 and NM_000179.2 for MSH6. Nucleotide numbering reflects cDNA numbering with +1 corresponding to the A of the ATG translation initiation codon in the reference sequence, with the initiation codon as codon 1. All variants were re-evaluated based on standardised qualitative classification criteria (Supp. Table S1). Probands with Class 5 pathogenic variants, or probands with either Class 1 not pathogenic/no clinical significance sequence alterations or no known pathogenic variant after extensive gene screening were included in reference datasets used to estimate the likelihood ratios. Supp. Table S2 lists all the Class 5 and Class 1 MMR gene sequence variants and the rationale or publications providing supportive information for their classification.

Figure 1.

Colon Cancer Family Registry molecular testing schema. *somatic BRAF p.Val600Glu mutation testing. Note: BRAF results were not used in MMR gene testing selection criteria. USC: University of Southern California; QIMR: Queensland Institute of Medical Research; MSI-L/H: low/high microsatellite instability; MSS: microsatellite stable; MLPA: multiplex ligase-dependent probe amplification; DHPLC: denaturing high performance liquid chromatography.

Supp. Table S3 lists all variants selected for evaluation using multifactorial and splicing approaches. Variants denoted as Class 2 (likely not pathogenic), Class 3 (uncertain), or Class 4 (likely pathogenic) from patients ascertained from the Australasian Colorectal Cancer Family Registry (ACCFR) were selected for evaluation in this study. Additionally, variants reported inArnold et al. (2009) classified on segregation and splicing data only were also re-evaluated using the multifactorial approach. Clinical information was collated for all relevant ACCFR families, and when data became available for this project, also for Colon CFR families ascertained from other sites of the Colon CFR (Mayo Clinic, Ontario Familial Colorectal Cancer Registry and Fred Hutchinson Cancer Research Center) that were reported to carry at least one of the selected variants. Additional tumor and segregation data was extracted from publications to aid in the classification of some variants. The final sample set encompassed a total of 54 variants in 162 Colon CFR families, plus information from 26 families identified from the literature (188 families in total). All variants have been submitted to the InSiGHT locus-specific database (www.insight-group.org).

Splicing analyses were conducted on RNA extracted from lymphoblastoid cell lines (LCLs) established from available lymphocytes using EBV transformation based on the principles of Sugden and Mark (Sugden and Mark, 1977). cDNAs from LCLs established from nine healthy Red Cross donors were used as controls for in vitro splicing analyses.

Tumor Characteristics and Correlation with Pathogenic Variant Status

Characterization of colorectal tumors from Colon CFR probands was carried out previously as part of core Colon CFR activities (Newcomb, et al., 2007). Briefly, formalin fixed, paraffin embedded tissue sections were stained for MMR proteins MLH1, MSH2, MSH6, and PMS2 (Lindor, et al., 2002). Tumors were analyzed for MSI status using 10 microsatellite markers (BAT25, BAT26, BAT40, BAT34, D5S346, D17S250, ACTC, D18S55, D10S197, and MYCL), comparing to normal tissue as reference where possible, and classified according to number of markers demonstrating instability: MSI-H for ≥3; MSI-L for 1–2; MSS for 0 unstable markers (Lindor, et al., 2002). MSI data was only extracted from peer-reviewed journal articles if at least five of the above markers were tested. Somatic BRAF p.Val600Glu mutation testing was done in all available CRCs as described previously (Buchanan, et al., 2010). Relevant information up to August 2009 was extracted from the Colon CFR database for all CRC-affected probands and used to generate cross-tabulations by pathogenic variant status, and by proband ascertainment, as shown in Table 1. Statistical significance of differences in distribution between pathogenic variant carriers and non-carriers was assessed using a Pearson’s chi-squared test using STATA. A likelihood ratio (LR) for tumor characteristics was derived for MSI status and BRAF mutation status data using the formula:

• LR of causality for tumor category = (% pathogenic variant carriers)/(% non-carriers)

• The likelihood ratios for all the tumor categories are shown in Table 1.

Table 1.

Estimation of likelihood ratios for tumor microsatellite instability and BRAF mutation status

Tumor Category	Pathogenic Variant Carriers	Non-carriers	Likelihood Ratio of Causality: % MC/% NC	P-value: MC vs NC
Clinic-based	n=90	n=190		P-trend <10−6
MSS	3.33% (n=3)	80.00% (n=152)	0.04 (CI 95% 0.01, 0.13)	<10−6
MSI-L	5.56% (n=5)	9.47% (n=18)	0.59 (CI 95% 0.23, 1.5)	0.2648
MSI-H	91.11% (n=82)	10.53% (n=20)	8.66 (CI 95% 5.69, 13.2)	<10−6
Population-based*	n=137	n=3893		P-trend <10−6
MSS	7.30% (n=10)	76.03% (n=2960)	0.10 (CI 95% 0.05, 0.17)	<10−6
MSI-L	3.65% (n=5)	11.17% (n=435)	0.33 (CI 95% 0.14, 0.78)	0.0055
MSI-H	89.05% (n=122)	12.8% (n=498)	6.96 (CI 95% 6.29, 7.70)	<10−6
Clinic-based	n=69	n=14
MSI-H BRAF +	2.90% (n=2)	28.57% (n=4)	0.10 (CI 95% 0.02, 0.50)	0.005
MSI-H BRAF −	97.10% (n=67)	71.43% (n=10)	1.36 (CI 95% 0.97, 1.90)
Population-based*	n=52	n=189
MSI-H BRAF +	3.85% (n=2)	47.62% (n=90)	0.08 (CI 95% 0.02, 0.32)	0.0007
MSI-H BRAF −	96.15% (n=50)	52.38% (n=99)	1.84 (CI 95% 1.59, 2.13)

^*As noted in Figure 1, tumor MSI-H and MSI-L status was used to select population-based cases for MMR sequencing, with only a subset of population-based MSS cases selected for sequencing. Therefore, the number of pathogenic variant carriers in the total population-based sample set was estimated separately for each MSI stratum based on the data available for probands that had undergone sequencing. See Methods and Supp. Table S4 for more detailed explanation. Frequency of BRAF mutation in MSI-H cases differs between population-based and clinic-based cases (Mantel-Haenszel χ² 13.83, p=0.0002). BRAF +: positive for BRAF p.Val600Glu mutation; BRAF −: negative for BRAF p.Val600Glu mutation; MC: mutation carrier; NC: non-carrier; CI: confidence interval.

Note: Since MSI status was a criterion for MMR gene variant testing of population-based probands in the Colon CFR database, but NOT for clinic-based probands (see Figure 1) (Newcomb, et al., 2007), the MSI likelihood ratios for population-based probands were based on estimations of the number of pathogenic variant carriers in all population-based cases with tumor MSI data, as detailed in Supp. Table S4.

BRAF mutation status provided no additional predictive capacity over MSI status for MSS and MSI-L probands, and categories were collapsed accordingly, to calculate a BRAF likelihood ratio for the subset of probands with an MSI-H tumor. The distribution of BRAF mutations by MMR-mutation status differed between the clinic- and population-based sample sets, and separate BRAF estimates were thus generated and used for clinic-based versus population-based probands.

Estimation of Prior Probability of Pathogenicity

Missense substitutions

The derivation of prior probabilities for missense substitutions are described in detail elsewhere (accompanying manuscript). Briefly, sequence conservation at relevant positions in protein multiple sequence alignments of MLH1, MSH2 and MSH6 were assessed using a combination of MAPP (Stone and Sidow, 2005) and a custom version of PolyPhen-2.1 (Adzhubei, et al., 2010) (http://hci-lovd.hci.utah.edu/). The MAPP impact score and custom PolyPhen-2.1 output “pph2 probability” were used as continuous variables. A minimum and maximum prior probability of pathogenicity of 0.10 and 0.90, respectively, was set to ensure that additional sources of information were required to reach posterior probabilities that alter clinical management of patients with variants (i.e. Class 1, Class 2, Class 4, and Class 5) i.e. direct classification based on the prior probability alone was not possible. All synonymous substitutions were assigned a prior of 0.10. Prior probabilities were derived from calibration of the above programs based on analysis of a consistently classified and carefully annotated reference set of MMR missense alterations that excluded all missense alterations known to alter splicing.

Intronic substitutions

There is a relative paucity of published information on splicing assays for intronic MMR gene variants (Supp. Table S5). All MMR gene variants located in the donor or acceptor dinucleotides assayed in vitro have been reported to be associated with splicing aberrations. However, the majority of variants reported in the literature appear to have been selected for in vitro analysis on the basis of bioinformatic prediction and the potential for negative publication bias for variants with no associated splicing aberration cannot be discounted. In the absence of a large study that has assessed the pathogenicity of intronic MMR gene variants without selection bias, we thus conservatively assumed prior probabilities of pathogenicity previously calculated for BRCA1 and BRCA2 variants based on analysis of a large dataset (Easton, et al., 2007). Namely, variants located in canonical GT-AG dinucleotides have a prior probability of 0.96 and other intronic variants and alterations in the 3’UTR were assigned a prior probability of 0.26.

Multifactorial Likelihood Classification

Likelihood ratios for segregation were derived by Bayes factor analysis adapted from the method of Thompson et al (2003), as described previously (Arnold, et al., 2009). Penetrance estimates for MLH1 and MSH2 variants were derived fromQuehenberger et al. (2005) and those for MSH6 variants were derived fromBaglietto et al. (2010). The variants MLH1 c.790+10A>G, MLH1 c.2146G>A (p.Val716Met), MSH2 c.339G>A, MSH2 c.*221G>T, MSH2 c.*226A>G and MSH6 c.1186C>G (p.Leu396Val) were present in multiple families and co-occurred with pathogenic variants in some probands. The bi-allelic families concerned were excluded from multifactorial analysis to ensure the segregation likelihood ratio for estimating causality of the variant would not be biased by the presence of a pathogenic variant in the family. Bayesian scores were derived twice for those families in which two different variants of unknown clinical significance were observed in trans, analysing risk associated with each variant separately, under the assumption that only one variant was causal.

Likelihood ratios (LR) for tumor characteristics shown in Table 1 were applied as follows: clinic-based probands and all relatives with known MSI status were assigned a LR according to MSI phenotype and recruitment location. All probands and relatives with MSI-H status and known BRAF mutation status were assigned an LR according to BRAF mutation status and site of recruitment. The MSI and BRAF likelihood ratios were used for CRCs only.

For each variant under study, the individual likelihood ratios available from segregation and tumor characteristics were multiplied to calculate an overall multifactorial likelihood ratio, under the assumption that each factor was statistically independent. Bayes rule was then used to calculate a posterior probability that the variant was pathogenic from the multifactorial likelihood ratios and the prior probability. Variants were classified according to the 5 class IARC quantitative scheme (Plon, et al., 2008), based on the posterior probability.

Splicing assays and bioinformatic prediction of splicing aberrations

Variants were selected for splicing analysis based on lymphoblastoid cell line (LCL) or lymphocyte availability. LCL cycloheximide splicing assays, PCR sample preparation and sample sequencing were done using methods described previously (Whiley, et al., 2011). cDNA (synthesized using random hexamers – Sigma, St Louis MO USA) was used as template in PCR reactions with specific primers targeting the potential splice-sites (Supp. Table S6). PCR for variants was performed under the conditions presented in Supp. Table S6. Each RT-PCR analysis included a set of 9 healthy Red Cross control LCLs (see methods). The long-range DNA polymerase Elongase (Invitrogen, Carlsbad CA USA) was used to test for intron inclusion in variant carriers. PCR products from MLH1 c.588+1G>T, MLH1 c.678-1G>C, MLH1 c.791-1G>C, MLH1 c.884G>A p.(Ser295Asn), MLH1 c.1559-2A>T, MSH2 c.645+1G>A and MSH2 c.1387-9T>A were cloned using pGEM®-T Vector (Promega Madison, WI, USA) and verified by sequencing.

Results from splicing assays were compared to bioinformatic predictions of splicing aberrations. We utilized Human Splicing Finder version 2.4 (www.umd.be/HSF/), which evaluates splicing signals present in any human gene by using matrices to predict 5’ and 3’ splice sites and splice regulatory sites using different algorithms, including Human Splicing Finder matrices and MaxEntScan (Desmet, et al., 2009; Yeo and Burge, 2004). The coding DNA HGVS nomenclature was input into HSF for each variant. We determined the difference between variant and wild-type output scores as a proportion of wild-type scores for HSF matrices and MaxEntScan. NNSplice (http://www.fruitfly.org/seq_tools/splice.html) (Reese, et al., 1997) was also used to assess the effect of variants on mRNA splicing, with the exact sequence input consecutively for wild-type and variant sequences, and a default minimum score of 0.4 was set for both 5’ and 3’ splice sites. The qualitative classification criteria inSpurdle et al. (2008) for splicing assay interpretation was applied in some instances to classify variants if it improved on the multifactorial likelihood classification.

Results

In this study we assessed 54 variants using multifactorial analysis, with the addition of prior probabilities for missense variants and a likelihood ratio for tumor characteristics. We also carried out splicing assays on all probands with LCLs available for in vitro analysis and no previous reported splicing assay results, including alterations in the splice consensus dinucleotides (MLH1 c.588+1G>T, MLH1 c.791-1G>C, MLH1 c.1559-2A>T and MSH2 c.645+1G>A).

Multifactorial Likelihood Analysis

We were able to classify 31 variants as either not pathogenic (9 variants) or pathogenic (22 variants) in MLH1, MSH2 and MSH6 using the multifactorial likelihood model (Table 2; detailed information used to derive individual LRs is provided in Supp. Table S7). In most instances, the likelihood ratios derived from tumor characteristics provided more information towards the final classification than segregation likelihood ratios, with obvious discordance between tumor results and segregation for only one variant MSH2 c.488T>G. There were an additional 6 variants considered likely not pathogenic and 4 likely pathogenic. Tumors from 3 cases that are carriers of an MSH2 variant in the 3’ untranslated region (MSH2 c.*129T>C and MSH2 c.*226A>G) demonstrated a tumor phenotype indicative of sporadic CRC (a positive mutation status for BRAF p.Val600Glu, MSI-H and MLH1 protein loss), which is consistent with the neutral and likely neutral classifications observed.

Table 2.

Mismatch repair gene multifactorial likelihood analysis results

Variant	Prior probability of pathogenicityf	Tumor characteristics LR	Segregation LR	Odds for causality	Posterior probability of pathogenicity	In vitro splicing results	Classificationg
MLH1
c.112A>C p.(Asn38His)	0.750	3.73×108	1871.3	6.97×1011	1.000	Wild-type	Class 5 - Pathogenic1
c.113A>G p.(Asn38Ser)	0.723	6.960	27.47	191.2	0.998	Wild-type	Class 5 - Pathogenic
c.116+5G>C	0.260	75.00	21.68	1625.9	0.998	Intron inclusion	Class 5 - Pathogenic
c.198C>Ta	0.100	N/A	0.690	0.690	0.071	Wild-type	Class 3 - Uncertain
c.199G>A p.(Gly67Arg)	0.962 → 0.9	2210.3	2.148	4747.2	1.000	Wild-type	Class 5 - Pathogenic
c.230G>A p.(Cys77Tyr)	0.880	11.78	3.161	37.23	0.996	Not tested	Class 5 - Pathogenic
c.303T>G	0.100	N/A	0.531	0.531	0.056	Wild-type	Class 3 - Uncertain
c.307-29C>Aa	0.260	4.10×10−12	4.08×10−4	1.67×10−15	5.88×10−16	Wild-type	Class 1 - Not pathogenic
c.350C>T p.(Thr117Met)	0.893	1.53×1011	1.04×106	1.60×1017	1.000	Wild-type	Class 5 - Pathogenic
c.389A>G p.(Tyr130Cys)b	0.757	1.360	1.1782	1.602	0.833	Wild-type	Class 3 - Uncertain
c.588+1G>T	0.960	883.3	3.706	3273.4	1.000	Exon skipping	Class 5 - Pathogenic
c.589-2A>G	0.960	N/A	25.11	25.11	0.998	4 bp deletion	Class 5 - Pathogenic
c.678-1G>C	0.960	N/A	1.671	1.671	0.976	2 bp deletion	Class 4 - Likely pathogenic
c.702G>Ac	0.100	N/A	0.949	0.949	0.095	Wild-type	Class 3 - Uncertain
c.790+2dupT	0.960	6.960	312.8	2176.7	1.000	Exon skipping	Class 5 - Pathogenic
c.790+10A>Ga	0.260	4.00×10−4	N/A	4.00×10−4	1.41×10−4	Wild-type	Class 1 – Not pathogenic2
c.791-1G>C	0.960	8.660	122.2	1057.8	1.000	Exon skipping	Class 5 - Pathogenic
c.884G>A p.(Ser295Asn)	0.723	11.78	2.430	28.62	0.987	Exon skipping	Class 5 - Pathogenic
c.1040C>A p.(Thr347Asn)	0.007 → 0.1	N/A	1.61×10−3	1.61×10−3	1.79×10−4	Not tested	Class 1 - Not pathogenic
c.1321G>A p.(Ala441Thr)b	0.011 → 0.1	1.360	0.577	0.784	0.080	Not tested	Class 3 - Uncertain
c.1559-2A>T	0.960	102.0	1.762	179.7	1.000	Exon skipping	Class 5 - Pathogenic
c.1852_1853delAAinsGC p.(Lys618Ala)	0.899	2.78×10−13	1.17×10−1	3.26×10−14	2.89×10−13	Wild-type	Class 1 - Not pathogenic
c.1852_1854delAAG p.(Lys618del)	0.500	5624.3	579.9	3.26×106	1.000	Wild-type	Class 5 - Pathogenic
c.1990-1G>A	0.96	N/A	28.44	28.44	0.999	Exon skipping	Class 5 - Pathogenic3
c.2059C>T p.(Arg687Trp)	0.429	8.660	4.769	41.30	0.969	Wild-type	Class 4 - Likely pathogenic3,4
c.2146G>A p.(Val716Met)a	0.327	4.16×10−4	0.204	8.48×10−5	4.12×10−5	Wild-type	Class 1 - Not pathogenic2,5,6,7,8
c.2153A>C p.(His718Pro)	0.931 → 0.9	N/A	0.811	0.811	0.880	Wild-type	Class 3 - Uncertain
MSH2
c.6G>T	0.100	0.100	0.696	0.070	0.008	Not tested	Class 2 – Likely Not pathogenic
c.339G>Aa	0.100	0.004	0.119	4.75×10−4	5.27×10−5	Wild-type	Class 1 - Not pathogenic
c.484G>A p.(Gly162Arg)d	0.962 → 0.9	8.660	1.636	14.17	0.992	Wild-type	Class 5 - Pathogenic
c.488T>G p.(Val163Gly)e	0.422	0.087	25.56	2.213	0.618	Wild-type	Class 3 - Uncertain
c.571_573delCTC p.(Leu191del)	0.500	11.78	19.81	233.3	0.996	Wild-type	Class 5 - Pathogenic
c.645+1G>A	0.960	11.78	7.870	92.70	1.000	154 bp deletion	Class 5 - Pathogenic
c.913G>A p.(Ala305Thr)	0.929 → 0.9	N/A	0.520	0.520	0.824	Wild-type	Class 3 - Uncertain
c.942+3A>T	0.260	N/A	20871.32	20871.32	1.000	Exon skipping	Class 5 - Pathogenic
c.1013G>A p.(Gly338Glu)	0.959 → 0.9	11.78	1.350	15.90	0.993	Wild-type	Class 5 - Pathogenic
c.1387-9T>A	0.260	75.00	1.682	126.1	0.978	Intron inclusion	Class 4 – Likely pathogenic
c.1662-9G>A	0.26	2.80×10−12	N/A	2.80×10−12	9.85×10−13	Wild-type	Class 1 - Not Pathogenic
c.1666T>C	0.1	5.86×10−6	N/A	5.86×10−6	6.52×10−7	Wild-type	Class 1 - Not pathogenic
c.1865C>T p.(Pro622Leu)	0.927 → 0.9	75.00	67134.69	75.00	0.999	Wild-type	Class 5 - Pathogenic
c.1906G>C p.(Ala636Pro)	0.045 → 0.1	2.42×1011	4.142	1.00×1012	1.000	Wild-type	Class 5 - Pathogenic9,10
c.2005+8dupAa	0.260	1.840	N/A	1.840	0.393	Wild-type	Class 2 - Likely Not pathogenic
c.2006-36_2006- 33dupTTAA	0.260	N/A	1.000	1.000	0.260	Not tested	Class 3 - Uncertain
c.*129T>C	0.260	0.003	0.793	0.002	7.72×10−4	Not tested	Class 1 - Not pathogenic
c.*221G>Ta	0.260	0.040	1.351	0.054	0.019	Not tested	Class 2 - Likely Not pathogenic
c.*226A>Ga	0.260	0.080	N/A	0.080	0.027	Not tested	Class 2 - Likely Not pathogenic
MSH6
c.1019T>C p.(Phe340Ser)	0.002 → 0.1	0.100	1.000	0.100	0.011	Wild-type	Class 2 - Likely Not pathogenic
c.1164C>Td	0.100	N/A	1.07×10−4	1.07×10−4	1.19×10−5	Wild-type	Class 1 - Not pathogenic
c.1186C>G p.(Leu396Val)a	0.026 → 0.1	0.100	0.520	0.052	0.006	Wild-type	Class 2 - Likely Not pathogenic
c.1193T>A p.(Val398Glu)	0.913 → 0.9	11.78	1.523	17.93	0.994	Not tested	Class 5 - Pathogenic
c.2057G>A p.(Gly686Asp)	0.916 → 0.9	1.840	1.998	3.676	0.971	Wild-type	Class 4 - Likely pathogenic
c.2341C>T p.(Pro781Ser)c	0.826	N/A	1.040	1.040	0.831	Wild-type	Class 3 - Uncertain
c.3226C>T p.(Arg1076Cys)e	0.805	N/A	1.113	1.113	0.821	Wild-type	Class 3 - Uncertain
c.*24_28delGTTGA	0.260	N/A	0.907	0.907	0.242	Wild-type	Class 3 - Uncertain

^aCo-occurrence or co-observation of variant with a pathogenic MMR gene variant in some families;

^bCo-occurrence of both variants in the proband (phase unknown);

^c,dCo-observation of both variants in the proband;

^eCo-observation of both variants in a single family;

^fPrior probability of pathogenicity <0.1 and >0.9 were set at 0.1 and 0.9 respectively. The prior was derived using the following equation Logit(Pr)= −9.20 + 2.27(M) + 4.26(P), where Pr is the Prior probability, M is ln(MAPP score), and P is the “pph2 probability” output by PolyPhen-2.1 (retrained without MMR gene data); Tumor characteristics LR based on colorectal tumors only; The Odds for Causality are calculated as the product of the individual statistically independent components (Tumor Characteristics LR X Segregation LR). The Posterior Probability = Posterior Odds/(Posterior Odds + 1), where the Posterior Odds = Prior Probability X Odds for causality X (1/1-prior probability);

^gBased on the criteria in Plon et al., 2008, additional data was extracted from peer-reviewed literature to achieve classification for six variants: 1 - Van Riel et al., 2010, 2 - Pastrello et al., 2011; 3 - Caldes et al., 2002; 4 - Gallinger et al., 2004; 5 - Genuardi et al., 1998; 6 - Muller-Koch et al., 2001; 7 - Kamory et al., 2006; 8 - Barnetson et al., 2008; 9 - Yuan et al., 1999; 10 - Foulkes et al., 2002;

N/A - Not applicable, LR - likelihood ratio

The multifactorial classification, for each variant assessed in this study is reported on a newly created LOVD based MMR database (http://hci-lovd.hci.utah.edu/), which details the prior probabilities, likelihood ratios for the individual components, and posterior probabilities.

As shown in Table 2, there were 9 variants that were observed with known pathogenic variants, and 8 variants that were observed with another variant of unknown clinical significance. The earliest age of onset of a Lynch cancer for these probands was 37 years old (range 37 to 64), and none had an abnormal clinical phenotype, suggesting that the carriers of these variants were not carriers of two different bona fide pathogenic variants. Indeed, the final classifications from multifactorial analysis (which at this point in time does not incorporate a LR for co-occurrence) were Class 1 (4/9 variants), Class 2 (4/9 variants) and uncertain (1/9 variants) for variants co-observed with a known pathogenic variant.

In vitro splicing assays

Several of the cases assayed expressed varying levels of naturally occurring alternative splicing isoforms, particularly MLH1 Δ exon 9/10, MLH1 Δ exon 3, a 227 base-pair insertion of intron 1 and 145 base-pair insertion of intron 2 both in MLH1 (Figure 2a). Splicing aberrations not identified in controls were observed for 7 out of the 26 variants assayed (Supp. Table S8, Figure 2). The splicing programs predicted the interruption of the intron-exon junction for 5 variants, all of which showed aberrant splicing in vitro. The variants MLH1 c.588+1G>T, MLH1 c.791-1G>C, MLH1 c.884G>A p.(Ser295Asn), and MLH1 c.1559-2A>T all result in exon skipping (Figure 2a, b, c). MSH2 c.645+1G>A produces multiple transcripts: an in-frame deletion of exon 3 and an out-of-frame product from cryptic donor site usage 154 base-pairs upstream of the wild-type donor site (Figure 2d). These splicing results were consistent with multifactorial likelihood classifications for these variants, Class 4 (MLH1 c.884G>A) or Class 5 (the remaining intron-exon junction variants). Moreover, analysis of cDNA sequence for the exonic variant MLH1 c.884G>A indicated that the variant allele did not express any full length transcript, justifying that the classification for this variant may be upgraded to Class 5 on the basis of the splicing data.

Figure 2.

Splicing aberrations arising from MLH1 and MSH2 variants detected using RT-PCR. Variant carrier and control cycloheximide treated (+) and untreated (-) lymphoblastoid cell lines. A) Alternative splicing detected for MLH1 c.588+1G>T (P1), MLH1 c.678-1G>C (P2), MLH1 c.791-1G>C (P3), MLH1 c.884G>A (P4) and two healthy controls (C1 and C2). The 227 base-pair insertion of MLH1 intron 1, 145 base-pair insertion of MLH1 intron 2, Δ exon 3, Δ exon 6, Δ exon 10, and Δ exon 9/10 are all naturally occurring isoforms. B) A schematic representing the splicing aberrations caused by MLH1 c.588+1G>T (P1), MLH1 c.791-1G>C (P3) and MLH1 c.884G>A (P4). MLH1 c.588+1G>T causes exon 7 skipping and both MLH1 c.791-1G>C and MLH1 c.884G>A lead to upregulation of Δ exon 10. C) MLH1 c.1559-2A>T results in two aberrant transcripts not present in controls: Δ exon 14 and Δ exon 14/15. D) MSH2 c.645+1G>A (P6 and P7 are both variant carriers) causes an upregulation of exon 3 skipping and 154 bp partial deletion of exon 3 as a minor transcript. E) MLH1 c.678-1G>C causes a 2 bp deletion at the start of exon 8. F) MSH2 c.1387-9T>A results in a 7 bp insertion of intron 8 caused by the creation of a de novo acceptor site. M: marker; bp: base-pair; fs: frameshift.

Bioinformatic prediction of aberrant splicing was consistent for the remaining two variants displaying aberrant splicing. MLH1 c.678-1G>C interrupts the wild-type acceptor splice site and creates a de novo acceptor leading to a 2 base-pair out-of-frame deletion identified by sequencing (Figure 2e), both of which were predicted by all 3 of the bioinformatic splicing programs (Supp. Table S8). MSH2 c.1387-9T>A creates a de novo acceptor site causing a 7 base-pair inclusion of intron 8 (Figure 2f), with de novo acceptor creation and interruption of the wild-type acceptor site predicted by MaxEntScan (Supp. Table S8). Although further quantitative splicing analysis would be required to determine if the splicing aberrations caused by MLH1 c.678-1G>C and MSH2 c.1387-9T>A are associated with a minor or major transcript, the Class 4 likely pathogenic classification from multifactorial analysis for these variants suggests that these splicing aberrations are causative.

In addition, splicing assay results showed no evidence of an aberration for the intronic variant MSH2 c.2005+8dupA, justifying that this variant can be altered from Class 3 uncertain (indicated by multifactorial analysis) to Class 2 likely not pathogenic. Splicing data did not alter the classifications of any other variants.

Discussion

In this study we derived likelihood ratios from the Colon CFR dataset for the CRC characteristics of MSI and BRAF mutation status for use in multifactorial analysis, carefully considering issues relating to ascertainment bias in this large and well-characterised dataset. We provide LRs that are applicable to variants discovered within clinic-based or population-based datasets, with the caveat that the reference pathogenic variants in this study are also representative of studies outside the Colon CFR dataset.

These LRs for tumor features were then used in a multifactorial model together with a segregation likelihood ratio and an estimated prior probability to assess clinical significance of 54 MMR gene variants, and contributed noticeably to classification. In total, application of the model altered the class of 31 variants under study (21 as Class 1 or 5, and another 10 Class 2 or 4), with implications for the future management of all families with these variants. Additionally, we were able to confirm the Class 5 classification for 10 variants considered pathogenic using largely qualitative information fromArnold et al. (2009). The priors based on in silico analysis of the 25 missense variants were in agreement with tumor and segregation data except for the variants MLH1 p.(Lys618Ala) and MSH2 p.(Ala636Pro). Underestimation of the prior for the latter Ashkenazi Jewish founder pathogenic variant traces to the observation that the position corresponding to Ala636 is quite variable in our protein multiple sequence alignments; moreover, mutS crystal structures do not place the residue in an alpha helix (Lamers, et al., 2000; Obmolova, et al., 2000). Consequently, the substitution to proline is not obviously destabilizing and neither MAPP nor PolyPhen gives the variant a high prior probability. Nonetheless, Foulkes et al. (2002) suggested that the substitution could cause steric hindrance that might interrupt ATP hydrolysis. Still, in spite of the relatively low prior probability for this variant, the data from tumor characteristics and co-segregation were sufficient to provide a posterior probability indicative of pathogenicity. The classification of MLH1 c.1852_1853delAAinsGC p.(Lys618Ala) is a contentious issue with some functional assays suggesting reduced protein function (Belvederesi, et al., 2006; Blasi, et al., 2006; Guerrette, et al., 1999; Kondo, et al., 2003; Perera and Bapat, 2008), but family studies (including this study) indicating it is not pathogenic (Castillejo, et al., 2011). Consistent with these inconclusive findings, a recent case-control study concluded that the variant was not fully penetrant, and was associated with a two-fold increase in risk of Lynch syndrome-associated tumors (Medeiros, et al., 2012).

There is currently no sufficiently large and well-characterised dataset to estimate prior probabilities of pathogenicity based on in silico splicing data for the MMR repair genes, and it is not feasible to estimate a splicing prior from the literature due to bias in selection of variants for in vitro mRNA assays (including this study with preselection by bioinformatic prediction splicing assays). Therefore, since the splicing machinery and regulatory elements are similar between genes (Baralle and Baralle, 2005), the BRCA1/2 splicing prior probabilities (Easton, et al., 2007) were used in the multifactorial analysis as a conservative estimate of prior probability for all intronic variants (within and outside acceptor/donor dinucleotides). Our analysis of possible splicing aberrations for a subset of variants supported the findings from multifactorial analysis, with 9/12 demonstrating splicing aberrations having a posterior probability Class 5. The remaining 3 variants were Class 4 from multifactorial analysis, and semi-quantitative allelic expression analysis for one exonic variant indicated that it could be considered Class 5 using qualitative interpretation {Spurdle, 2008 #587}.

It was not appropriate to utilise other LRs derived using BRCA datasets such as co-occurrence and family history (Easton, et al., 2007; Goldgar, et al., 2004) due to the different phenotypes present for the mismatch repair genes compared to BRCA1 and BRCA2. Although it is known that co-occurrence or co-observation of pathogenic variants in the MMR genes (also termed compound heterozygosity or bi-allelic mutations) is associated with very early onset gastrointestinal tumors (under 20 years old), haematologic malignancies and neurofibromatosis type 1-like features (Bandipalliam, 2005), further study is required to assign with confidence the specific features that can be used to derive LRs for pathogenic variants that co-occur (same MMR gene) or are co-observed (different MMR genes). Similarly the family history profiles associated with MMR gene pathogenic variant status will have to be assessed from large well-characterised datasets to derive LRs for this component.

This is the first instance of incorporating BRAF mutation status in the multifactorial model. Mining of the Colon CFR MSI and somatic BRAF tumor data from unselected colon cancer patients of known pathogenic variant status indicates that a BRAF mutation negative MSI-H colon tumor is associated with a ~12-fold increase in likelihood of pathogenicity. This data also suggests that MSI in a tumor is a better predictor of pathogenic variant carrier status than absence of a BRAF mutation. Positive BRAF mutation status was most informative in identifying probable non-pathogenic variants, as was MSS tumors. These findings are reflected by observations in the literature (Lucci-Cordisco, et al., 2006; Parsons, et al., 2012; Zaanan, et al., 2011). However, the incidence of MSI in sporadic CRCs (Cunningham, et al., 1998; Herman, et al., 1998; Miyakura, et al., 2001; Toyota, et al., 1999) highlights the importance of a combinatorial tumor characteristic likelihood ratio. Using MSI alone can sometimes lead to over-estimation of pathogenicity. For the variant MLH1 c.2146G>A p.(Val716Met) there are two MSI-H tumors eligible for analysis that have not had BRAF mutation testing, which increase the odds of causality. However, additional evidence suggests that this variant is not pathogenic and that something else other than the missense substitution is causing the MSI. A likelihood ratio based on MLH1 promoter methylation would also be a beneficial addition to the multifactorial model as a negative indicator of Lynch syndrome.

BRAF mutation status is commonly associated with MLH1 promoter hypermethylation in MSI-H sporadic CRCs (Deng, et al., 2004; McGivern, et al., 2004). Therefore, the presence of BRAF p.Val600Glu in a CRC should be sufficient to exclude MMR gene variant screening for Lynch syndrome (Bessa, et al., 2008; Bouzourene, et al., 2010; Jensen, et al., 2008; Loughrey, et al., 2007). However, there are seven pathogenic MMR variants reported in probands from this Colon CFR dataset and additional pathogenic variants reported in the literature (Lagerstedt Robinson, et al., 2007; Walsh, et al., 2009) that have BRAF mutation positive tumors, suggesting that recognition of bona fide Lynch tumors may be missed using such triaging for gene variation screening.

Few studies have addressed classifying MMR gene variants using multifactorial analysis. To our knowledge, three papers have tried integrating tumor characteristics with multiple lines of evidence to classify variants. Arnold et al. (2009) used Bayesian segregation analysis, but used tumor pathology data qualitatively together with other observational data (functional data and family history) and in vitro splicing data to classify variants.Barnetson et al. (2008) devised an arbitrary scoring system for variant classification, that included IHC and MSI tumor status as points of evidence, but there was no attempt to derive an empirical LR. Furthermore, Pastrello et al. (2011) described a Bayesian approach to MMR variant classification that included a likelihood ratio for clinicopathologic features such as IHC and MSI based on a large tumor dataset (Engel, et al., 2006). Nevertheless, ascertainment bias in the dataset was not taken into account in derivation of the LRs, therefore leading to an overestimation in the likelihood of pathogenicity for variant carriers. In the estimation of the tumor characteristic LR we have taken into account ascertainment bias for the tumors, and chose not to estimate a likelihood ratio based on IHC data due to the high correlation between tumor protein loss and MSI, and because some missense substitutions can lead to stably expressed non-functional proteins not identified using IHC (Mangold, et al., 2005; van Riel, et al., 2010).

The tumor characteristics likelihood ratio based on MSI and BRAF mutation status was an integral component in the classification of nearly two thirds of the variants reported in this study. Therefore, this LR used in combination with segregation analysis and an in silico missense prior probability is a beneficial addition to the MMR multifactorial model for variant classification. While the strength of the multifactorial approach is the ability to draw on multiple data sources readily available in the clinical setting, we do acknowledge a number of assumptions that underlie the basic model. Namely, that the individual components are mutually independent, that all variants of clinical significance have similar cancer penetrances to known pathogenic variants, that estimates of penetrance used in segregation analysis have been accurately estimated, and that the phenotypic features of different types of pathogenic variants (missense, splicing, truncating) are sufficiently similar to allow likelihood ratios estimates from one type of pathogenic variant to be relevant to other classes.

The results of the multifactorial analyses will affect the clinical management of 76% (41/54) of the variants reported within. Additionally, a mismatch repair gene public LOVD-based database has been created to report these classifications and future multifactorial likelihood classifications.