Application of a five-tiered scheme for standardized classification of 2,360 unique mismatch repair gene variants lodged on the InSiGHT locus-specific database

Abstract

Clinical classification of sequence variants identified in hereditary disease genes directly affects clinical management of patients and their relatives. The International Society for Gastrointestinal Hereditary Tumours (InSiGHT) undertook a collaborative effort to develop, test and apply a standardized classification scheme to constitutional variants in the Lynch Syndrome genes MLH1, MSH2, MSH6 and PMS2. Unpublished data submission was encouraged to assist variant classification, and recognized by microattribution. The scheme was refined by multidisciplinary expert committee review of clinical and functional data available for variants, applied to 2,360 sequence alterations, and disseminated online. Assessment using validated criteria altered classifications for 66% of 12,006 database entries. Clinical recommendations based on transparent evaluation are now possible for 1,370 variants not obviously protein-truncating from nomenclature. This large-scale endeavor will facilitate consistent management of suspected Lynch Syndrome families, and demonstrates the value of multidisciplinary collaboration for curation and classification of variants in public locus-specific databases.

Identification of a high-risk disease-causing constitutional mutation in a cancer patient guides the clinical management of the whole family, with implications for counselling, cancer treatment options, pre-symptomatic surveillance, and consideration of risk-reducing surgery and/or medication regimes¹. Carriers of mutations in the mismatch repair (MMR) genes MLH1, MSH2, MSH6 and PMS2 causing Lynch Syndrome (LS)¹ have a substantially increased risk of colorectal and endometrial cancer, along with increased risk of ovarian, gastric, small bowel, urothelial, brain, hepatobiliary, pancreatic, bladder, kidney, prostate, and breast cancers^1-8. However, intensive management reduces mortality⁹.

Sequence variants of uncertain functional and clinical significance are common in genetic test reports. Although several lines of evidence can be evaluated to assess their significance, usually none of them can be used on its own to obtain clinically useful variant interpretation, and for many variants comprehensive data are lacking. Laboratories are generally conservative in designating pathogenic variants, assigning variants as “uncertain significance” unless overwhelming evidence of pathogenicity exists. Several schemes for classifying variants in genes associated with Mendelian conditions have been proposed for use in the clinical setting. Since clinically useful actions are currently only considered for high-penetrant mutations, all of these systems are aimed at differentiating high-penetrant from low-penetrant/neutral variants and do not consider intermediate risk variants. They differ in the range and format of data used for classification, and the number of variant classes^10-12. The International Agency for Research on Cancer (IARC) classification system, endorsed by the Human Variome Project (HVP), facilitates standardized categorization by defining classes that can be linked to validated quantitative measures of causality/pathogenicity from statistical models^13-16, or from validated interpretation of qualitative data¹⁷. Importantly, only the 5-class IARC system has been linked to clinical recommendations for all classes: clinical testing and full high-risk surveillance guidelines for Class 5 “pathogenic” and Class 4 “likely pathogenic”; advice to treat as “no mutation detected for this disorder” for Class 1 “not pathogenic” and Class 2 “likely not pathogenic”; and acquisition of additional data to provide more robust classifications for Class 2, Class 4 and Class 3 “uncertain”.

Locus-specific databases (LSDBs) are an important source of information for clinicians and researchers to assess data as well as opinion on the clinical relevance of disease gene sequence variants, and have a fundamental role in variant classification due to the added value of aggregated data. Consistent and normalized data curation is critical to the value of databases for categorizing the relationship between genetic variation and disease – especially for clinical application. It has previously been recommended by the IARC Working Group that a panel covering a range of expertise in variant classification provide consensus opinion on variant pathogenicity prior to publicly accessible display of such information¹⁸. Another important component of the classifications provided by LSDBs is transparency regarding the criteria and supporting information used for classification, so that LSDB users can consider the information for their own application in the research and clinical setting¹⁸.

The International Society for Gastrointestinal Hereditary Tumours (InSiGHT) has merged multiple gene mutation/variant repositories to create the InSiGHT Colon Cancer Gene Variant Database for MMR and other colon cancer susceptibility genes^19-23, hosted by Leiden Open Variation Database (LOVD). Following recommendations for LSDB curation¹⁸, InSiGHT formed an international panel of researchers and clinicians to review MMR gene variants submitted to the database. To encourage submission of unpublished clinical and research data to further facilitate variant classification, the microattribution approach²⁴ was implemented using Open Researcher and Contributor Identifications (ORCID). Here we present the results of the InSiGHT Variant Interpretation Committee (VIC) effort to develop, test and apply a five-tiered scheme to classify 2,360 unique constitutional MMR gene variants.

Curation of MMR gene variants submitted to the InSiGHT Colon Cancer Gene Variant Databases

Through December 2012, after 3,458 alterations to standardize nomenclature, there were 12,635 submissions of 2,730 unique MMR gene variants lodged in the InSiGHT database. Furthermore, 370 (13.6%) unique variants were not identified in constitutional (germline) DNA (see Supplementary Fig. 1 and Supplementary Table 1 for details), and were excluded from further analyses since: (i) no evidence exists that these occur as constitutional variants, and (ii) no clinical information was available to assess their potential role in hereditary disease. The 2,360 constitutional variants included: 932 MLH1 (39%), 842 MSH2 (36%), 449 MSH6 (19%), and 137 PMS2 (6%). Most variants were nonsense/frameshift predicted to cause protein truncation (800, 34%), followed by “not obviously truncating” non-synonymous variants as the next largest group, including missense substitutions, small in-frame insertions/deletions (indels) and read-through alterations of the translation termination codon (746, 32%).

Variants had originally been assigned a classification by submitters according to the following classes: pathogenic; probably pathogenic; no known pathogenicity; probably no pathogenicity; effect unknown. No information was recorded to document the rationale for classification, or standards to classify variants. Considering 1,382 constitutional variants with multiple entries in the InSiGHT database, there were discordances in classification between submitters for 869 variants. Some of these discordances arose because of classification based on single data points or references, such as single functional assay results²², or inferences from individual publications originally lodged in the Mismatch Repair Genes Variant Database²³ (see example in Supplementary Table 2).

Development of a five-tiered system for consistent classification of MMR gene variants

The InSiGHT VIC (see Methods) was established in 2007 to address discrepancy in classification of MMR gene variants lodged on the InSiGHT database. Since March 2011 the VIC has made a concerted effort to develop standardized criteria for variant classification, employing a modified “Delphi consensus process”²⁵ to evaluate current scientific evidence and reach consensus. In line with the HVP²⁶, the IARC classification system¹⁰ for variant categorization (see Table 1) was adopted by InSiGHT for MMR variant classification. Briefly, multiple lines of evidence are required for classification, and evidence for each variant must include data associating the variant with both clinical and functional consequences (see Methods).

Table 1. InSiGHT variant classification scheme with accompanying recommendations for family management, adapted from the IARC 5-tiered classification system^*.

InSiGHT MMR variant class definition for Lynch syndrome**	Predictive testing of at-risk relatives	Surveillance for positive at-risk relatives	Research testing of relatives
5: Pathogenic	Yes	Full high-risk guidelines	Not indicated
4: Likely pathogenic	Yes***	Full high-risk guidelines	Yes
3: Uncertain	No***	Based on family history & other risk factors	Yes
2: Likely not pathogenic	No***	Based on family history & other risk factors. Treat as “no mutation detected” in this gene for this disorder	Yes
1: Not pathogenic	No***	Based on family history & other risk factors. Treat as “no mutation detected” in this gene for this disorder	Not indicated

^*Adapted from Plon et al¹⁰. Full high-risk surveillance guidelines for cancers in the Lynch spectrum are outlined in Vasen et al¹. Research testing entails cascade testing for the variant in affected and unaffected family members to facilitate segregation analysis, and is indicated for variants in classes 2-4 to refine classification. Consent from subjects through a protocol approved by a human subjects committee should be obtained.

^**Class definition is described in detail in the Supplementary Note and Supplementary Table 4, and is based on quantitative evidence defined by multifactorial likelihood posterior probability (with cut points >0.99 for Class 5; 0.95-0.99 for Class 4; 0.05-0.949 for Class 3; 0.001-0.049 for Class 2; <0.001 for Class 1) or combined qualitative evidence determined by consensus opinion as defined by the InSiGHT Variant Interpretation Committee. “Pathogenic” is defined as “clinically relevant in a genetic counseling setting such that germline variant status will be used to inform patient and family management.”

^***Recommend continued testing of proband for any additional available testing modalities available e.g. rearrangements.

The scheme was first tested on a subset of 117 MMR gene variants, and the criteria evolved and were refined by consensus to accommodate new data and inconsistencies over multiple classification teleconferences and face-to-face meetings. Final criteria were then applied retrospectively and to all remaining unique variants listed in the database (see Supplementary Table 3). Figure 1 shows an overview of the InSiGHT classification criteria (see Supplementary Note, Supplementary Table 4 for detailed criteria and justifications). At the close of each VIC teleconference/meeting, consensus classifications were noted. Where necessary, action items to improve or clarify classifications included:

1. Calls for missing clinical and functional information for specific variants to committee members and the general InSiGHT membership.

2. Requests for more detailed data or data clarifications from the authors of original publications.

3. Re-assessment of classification after additional data were obtained.

At the end of the process, the InSiGHT database was updated with the final consensus classifications and the supporting data, to ensure transparency.

Figure 1.

Overview of 5-tiered InSiGHT classification guidelines.

(a) Simplified guidelines describing levels and types of evidence required to reach different classes. See the supplementary information for the full guidelines (Supplementary Note) and detailed rationale behind each criterion (Supplementary Table 3). The Lynch Syndrome molecular phenotype described in Classes 5 and 4 includes microsatellite instability and/or loss of expression of relevant protein(s) as determined by immunohistochemistry. In this study, variants resulting in a premature termination codon or large genomic deletions of functionally important domains, generally considered pathogenic on the basis of DNA sequence alone, are referred to as Class 5a “assumed pathogenic” variants. All other variants reaching class 5 are termed Class 5b. (b) Flowchart used to assist in interpretation of available functional assay data. Assays reviewed for classification are shown in Supplementary Table 4, and the values used to define abrogated or normal function are shown in Supplementary Table 5. The cut-offs <25% and >75% set for protein expression, as used in previous publications^47,48, are very conservative given reported abrogated function associated with MLH1 expression defects of ~50% or lower⁴⁹. For variants that had normal/inconclusive/intermediate MMR activity in 2 independent assays, but deficient protein function in 2 independent assays, abrogated function was assigned. AF – allele frequency; PP – posterior probability of pathogenicity derived by multifactorial likelihood analysis; CMMRD – constitutional mismatch repair deficiency (MIM 276300); LR – likelihood ratio; LS – Lynch Syndrome; MSS – microsatellite stable; CRC – colorectal cancer; IHC – immunohistochemistry; NMD – nonsense mediated decay.

The major issues faced by the committee in the review process were determining data redundancy across multiple sources (resolved through discussion with authors), paucity of information, incomplete/inaccurate data, and difficulties in interpretation of functional assays. To facilitate functional assay interpretation, supporting information and flowcharts were developed (Fig. 1b, Supplementary Tables 5-6), and multiple meetings dedicated to review variants with apparently discordant functional assay results were coordinated (Supplementary Table 3).

Validation of the InSiGHT qualitative classification criteria

Nonsense or frameshift alterations, or large genomic deletions interrupting functionally important domains are generally considered pathogenic on the basis of DNA sequence alone; here these are referred to as “assumed pathogenic” variants (termed Class 5a in figures). There were 990 “assumed pathogenic” variants in the database, 640 of which were private mutations. In order to demonstrate the robustness of the qualitative classification criteria, 170 “assumed pathogenic” variants (68 MLH1, 75 MSH2, 13 MSH6, 14 PMS2) were reviewed as a validation set against the Class 5 (pathogenic) qualitative criteria required for the variants termed Class 5b in figures (see Methods, Supplementary Table 7). Class 5b designation required: evidence of abrogated protein function; at least two tumors with microsatellite instability (MSI) or appropriate loss of MMR protein expression; and a segregation likelihood ratio >10:1 (incorporating gene-specific cumulative risks²⁷) or variant co-segregation with disease reported in at least two Amsterdam criteria positive families. Class 5b was attained by all 60 validation set variants that had sufficient clinical data to assess these required criteria. The other 110 validation set variants could not be assigned to Class 5b largely because family co-segregation and tumor data were scarce or unobtainable - presumably because these variants are accepted as disease-causing and routinely used for clinical presymptomatic testing in families (see Implementing Microattribution). Of these, 72 were assigned to Class 4 due to lack of only one point of evidence, and 38 variants fell in Class 3 due to insufficient data. However, only 2/13 MSH6 and 2/14 PMS2 variants fulfilled Class 5, reflecting the lower penetrances and later ages of onset of PMS2²⁸ and MSH6²⁹ deleterious variants. Together these results indicate that the criteria for classification using qualitative data were sufficiently stringent to ensure conservative classification.

Application of InSiGHT classification guidelines to 2,360 individual constitutional MMR variants

Of the 12,006 eligible variant entries in the InSIGHT database, submitter versus final classification differed for 7,935 (66%), including changes from “no known pathogenicity” to Class 5 (pathogenic) and vice versa (Fig. 2a). The overall breakdown of final classifications is shown in Figure 2b. In addition to the 990 “assumed pathogenic” truncating/large deletion variants (Class 5a), consistent medical management is now also possible for the remaining 1,370 “not obviously truncating” variants; these include 167 (12%) Class 5 (pathogenic) variants (Class 5b), 183 (14%) Class 4 (likely pathogenic) variants, 86 (6%) Class 2 (likely not pathogenic) variants and 169 (12%) Class 1 (not pathogenic) variants.

Figure 2. Outcome of standardized 5-tiered InSiGHT classification of constitutional MMR gene variants.

a) The plot represents the proportion of the 5-tiered classifications for all documented constitutional variants in the database, against the original LOVD database classifications assigned by submitters for each entry. Class 5a is a subset of Class 5 containing the assumed pathogenic nonsense mutations, small frameshift indels, and large deletions. Class 5b includes not-obviously truncating variants considered to be pathogenic on the basis of combined evidence (See Supplementary Note). Results show that standardized classification led to altered classifications for a considerable proportion of variant entries, including downgrading for variants submitted as pathogenic (24%), and upgrading of variants with unknown pathogenicity to likely pathogenic (5.6%) or pathogenic (48%). In addition, clinically important misclassifications were identified for unique variants initially submitted as not pathogenic (54 unique variants reclassified as Class 5b, and 25 reclassified as Class 4) and unique variants submitted as pathogenic (28 unique variants reclassified as Class 1, 16 reclassified as Class 2, and 218 reclassified as Class 3). b) Pie chart showing distribution of final InSiGHT VIC classifications.

As shown in Figure 3 and Supplementary Figure 2, non-synonymous variants (see Fig. 3 footnote) made up the majority of Class 3 variants (524/765; 68%) and newly assigned Class 5b variants (91/167; 54%; see Supplementary Table 8 for detailed information supporting classifications). Substitutions of canonical dinucleotide splice sites fell predominantly in Class 4, due to lack of functional RNA analyses; however if experimentally tested these would likely move to Class 5b. Intronic variants outside conserved splice sites were the most prevalent variant type in Class 1.

Figure 3. Classifications of all documented unique variants by variant type.

The plot represents the proportion of the different variant types within the 5 classes. Class 5a is a subset of Class 5 containing the assumed pathogenic mutations (nonsense mutations, small frameshift indels, and large deletions). All other variants reaching class 5 are termed Class 5b (see Supplementary Note). The different variant types are: PTC – variants that introduce premature terminating codons, i.e. nonsense mutations and small frameshift indels; LGDel – large genomic deletions or disrupting inversions; LGDup – large genomic duplications; SS – variants in the canonical splice site dinucleotides; NS – not obviously truncating non-synonymous variants outside the Kozak consensus sequence i.e. missense, small in-frame insertion/deletions, and read-through translation termination codon alterations; S – synonymous variants; I – intronic variants outside the canonical splice site dinucleotides; ATG/UTR – variants in the initiation codon, and the 5′ or 3′ untranslated regions. See Supplementary Figure 2 for further details of variant types, by MMR gene.

Final categorization (see methods) of “not obviously truncating” variants as Class 1, 2, 4 or 5 was achieved by applying qualitative criteria for 391 variants, using quantitative multifactorial likelihood analysis methodology for 192 variants (based on bioinformatic prior probability plus evidence from segregation and/or tumour data, see Thompson et al¹⁶), and either quantitative or qualitative criteria for 26 variants. Where classifications derived using quantitative versus qualitative criteria differed, this reflected the amount of data available rather than deficiencies in the classification criteria, with no variants considered Class 1/2 using one approach and Class 4/5 using the other. Six synonymous variants reached Class 5b due to their effects on splicing. For substitutions occurring in initiation codons (often assumed to be pathogenic^30-32), only 1/9 had sufficient evidence to determine pathogenicity.

Implementing Microattribution

Microattribution is a means to incentivize placement of unpublished data in the public domain, by assigning scholarly contribution to authors similar to citations conventions afforded to journal articles³³. Retrospective and prospective microattribution was implemented to acknowledge and encourage the submission of unpublished data to the InSiGHT database, including submission of additional detailed clinical information from authors of published reports. Microattribution was assigned for initial variant submission, segregation and family history data, pathology (MSI, immunohistochemistry) information, in vitro functional assays (mainly RNA splicing), and variant frequency in normal individuals. As of July 2013, a total of 6,015 microattributions were conferred, including 3,763 for variant submission, 2,111 for family and tumor pathology data, 97 for in vitro assays, and 25 for frequency data. Notably, 19% of the clinical and functional microattributions contributed additional information critical to classification of the “assumed pathogenic” validation subset (Class 5a). These data also highlighted that clinical testing for “assumed pathogenic” variants is mostly undertaken in the presymptomatic setting (see above). The contribution of microattribution to final classification of “not obviously truncating” variants is shown in Figure 4. Importantly, classification was altered for 57/169 (34%) variants for which novel unpublished data were obtained. Moreover, implementation of microattribution stimulated submission of 128 novel MMR variants, yet to be classified.

Figure 4. Contribution of microattribution to classification of “not obviously truncating” variants.

Dark shading (YES) indicates the proportion of variants for each class, where the additional data obtained through microattribution contributed to their final classification.

Preliminary Analysis of Class 3 Variants of Uncertain Significance

Missense variants in MMR genes are abundant among Class 3 (uncertain) variants and present a considerable clinical problem. Quantitative multifactorial likelihood analysis is an effective approach to missense variant classification, since validated bioinformatic predictions³⁴ based on amino acid conservation and physicochemical properties can be used as a surrogate for in vitro variant effect on protein function. In silico analyses previously shown to have high accuracy (area under receiving operator characteristic [ROC] curve 0.93)³⁴ were used to estimate the prior probability of pathogenicity for all 481 Class 3 (uncertain) missense variants (see Fig. 5), to prioritize requests for data to facilitate future multifactorial analysis. The distribution of prior probabilities for MLH1 and MSH2 Class 3 variants is clearly bimodal, suggesting that ~50% of MLH1/MSH2 missense variants may be classified as pathogenic after further investigation. In total, 401 missense variants had extreme prior probabilities of 20% or ≥80%, 270 of which were <10% or >90%, indicating that Class 1 or Class 5 could be easily reached by incorporating segregation or tumor information. It is also possible that some Class 3 variants with low prior probability of pathogenicity based on predicted missense alteration will cause splicing aberrations, as already observed for 42/746 of not obviously truncating non-synonymous variants. Incorporation of validated bioinformatic splicing prediction tools into the MMR gene multifactorial model, as is under development for BRCA1/2³⁵, will assist prioritization of such likely spliceogenic variants.

Figure 5. Probabilities of pathogenicity for 481 Class 3 “uncertain” missense variants, derived by in silico analysis.

Distribution of probabilities of pathogenicity as estimated from a calibrated algorithm based on customized MAPP and PolyPhen2 scores³⁴, for (a) MLH1, n=186; (b) MSH2, n=169; (c) MSH6, n=145; (d) PMS2, n=24; (e) all four MMR genes, showing stratification of variants with prior probabilities ≤20% or ≥80% to prioritize variants for further investigation and classification.

Investigating potential effects of Class 3 regulatory variants (see methods) showed that all 15 5′UTR variants fell within multiple transcription factor binding sites, but no evidence for miRNA binding interruption for six 3′UTR variants was found (data not shown). Multifactorial analyses and transcription assays would help elucidate if these variants affect gene function.

Discussion

The InSiGHT VIC has successfully undertaken a collaborative effort to establish standardized variant interpretation guidelines using a modified Delphi process, encourage data submission, and provide objective assessment of MMR gene variants involved in Lynch Syndrome. The criteria developed provide a basis for standardised clinical classification of variants to inform patient and family management by genetic counselling¹⁰. This initiative has achieved the systematic evaluation of 2,360 constitutional variants, which will benefit thousands of families internationally. Importantly, 605 variants not resulting in premature termination codons, including 217 non-synonymous substitutions, have now been assigned to Class 5 (pathogenic) and Class 4 (likely pathogenic), or Class 1 (not pathogenic) and Class 2 (likely not pathogenic). These can now also be used as standards for the calibration of functional assays^36,37.

The clinical significance of 32% of the variants investigated remains uncertain. A large proportion (71%) of these were “private” variants occurring in only one family, which are problematic to classify due to the paucity of available clinical information. Clinicians play a fundamental role in promoting collection of segregation and other information relevant for classification. We anticipate that development of this interpretation scheme, plus the implementation of microattribution, will create incentive to assist in accumulating clinical data. The value of microattribution for data accrual has previously been demonstrated for hemoglobinopathies²⁴, and the InSiGHT initiative now demonstrates the clinical utility of data collection. The promotion of standardised data formats will assist transition into fully quantitative unbiased classification, eventually incorporating other components of the qualitative guidelines. In addition, the difficulties experienced in interpreting apparently discordant functional assays emphasize the importance of assay validation and standardization^38,39. Such experience will be directly applicable to functional analysis of deep intronic and regulatory variants, which are increasingly detected with the advancement of DNA sequencing technologies.

To accommodate the lower penetrance and reported lesser degrees of tumor MSI associated with MSH6 and PMS2 mutations^28,29,40-44, gene-specific criteria should also be considered for future iterations of the classification guidelines e.g., stipulating inclusion of segregation odds for MSH6 and PMS2 variants for classification, and use of modified panels to detect MSI status.

Another challenging issue to contemplate will be incorporating intermediate risk alleles⁴⁵ into classification schemes, including the clinical recommendations that might be linked to such variants. The identification of a subset of MMR gene alleles with apparently discordant clinical and functional features that renders them “resistant” to classification will provide the basis for future studies to define the most appropriate methodology and criteria to identify such variants. Further studies will also be required to assess if variants with abrogated DNA damage response but normal mismatch repair⁴⁶ are associated with the same clinical features as classical pathogenic mutations in MMR genes.

The InSiGHT database is a well-recognized resource for the clinical and research community, receiving over 20,000 hits/month. The development and adoption of standard templates allows transparency in the review process. Database users can view relevant information and sources in relation to guideline interpretation when considering the classification provided by the committee. The guidelines must evolve to accommodate additional kinds of evidence, but we anticipate no clinical issues as long as the variant classifications are dated and linked to a dated set of guidelines with the supporting information used to derive the classification. The final classifications have also been submitted to NCBI’s ClinVar for higher exposure, but expert classifications and underlying data rest with InSiGHT.

This is the first large-scale comprehensive classification effort demonstrating the value of expert panel evaluation to curation of an LSDB, and providing summary information used to assign variant pathogenicity. It also shows how classification may be assisted by promoting standardized data submission from stake-holders in the clinical and research setting, in order to access unpublished clinical and functional information to facilitate variant classification. Therefore, the InSiGHT initiative provides an important model of LSDB-centric multidisciplinary collaboration for transparent DNA variant interpretation.

Online Methods

InSiGHT Variant Interpretation Committee expertise

The InSiGHT Variant Interpretation Committee (VIC) (current chair author MG), includes 40 multidisciplinary experts from 5 continents (See Supplementary Table 9 for disciplines covered by VIC members). The Committee is responsible to its Governance Committee, which in turn is responsible to the InSiGHT Council. InSiGHT has recently joined the Human Variome Project and is a founding member of its Gene and Disease Specific Council. The InSiGHT Council specifically considered the need and responsibility associated with classification assignment on its database, and took all reasonable steps to both invite the highest possible expertise to contribute to the classification process and to ensure its processes and legal standing are robust.

InSiGHT database curation

Mutalyzer⁵⁰ was used to standardize the nomenclature of all variants present on the database as of December 2012. Variants with multiple submissions that were originally assigned a classification of pathogenic/probably pathogenic, and no known pathogenicity/probably no pathogenicity, were included in the group of “discordant” variants used to test the classification criteria. All unique variants identified in the database were assigned to one of the following sources: constitutional, somatic, artificial, and unknown.

Development of 5-tier InSiGHT classification criteria

The InSiGHT classification criteria were developed using the Delphi method²⁵. A 5-tiered classification system originally developed for consistent classification of MMR gene variants identified amongst Colon Cancer Family Registry participants^16,34 was selected as a baseline for the InSiGHT classification criteria. This system included the option of classification based on posterior probabilities arising from multifactorial likelihood analysis^15,16,51,52, and also multiple combinations of qualitative data not yet calibrated for inclusion in quantitative analyses but which are often reported in the literature or available from clinical sources. These baseline classification criteria were critically reviewed by InSiGHT members attending the InSiGHT San Antonio meeting in April 2011, and by VIC members via email. In response to comment, the rules were amended for clarity, to apply more stringent interpretation of functional assay data, and to consider additional points of evidence. These InSiGHT rules were used for variant classification over a series of 11 meetings (10 teleconferences and one face-to-face meeting), with further changes incorporated after each meeting to include additional points of evidence identified to be relevant during the review process as the committee encountered different combinations of useful data from published and unpublished sources. For example, after discussion, co-occurrence of a variant with a pathogenic mutation in the same gene with clinical information regarding constitutional MMR deficiency phenotype⁵³ was included as an in vivo test of MMR function, and the 1000 genomes data⁵⁴ was accepted as a test for population frequency. Consistency of the accumulative evidence required for a given class was reviewed by presentation of the rules at a face-to-face meeting of committee members. Supporting documentation was developed to assist the interpretation of splicing and functional assay results, by author BAT in consultation with a subset of committee members with specific expertise in this field (see Fig. 1b, Supplementary Tables 4-5). Where necessary, rule alterations were applied retrospectively to variants evaluated in previous meetings. The finalised rules (shown in simplified format in Fig. 1, detailed in the Supplementary Note) were then used to assess all remaining variants lodged in the InSiGHT database.

Classification of MMR gene variants by literature review and data collation

Variants occurring in 1000 genomes⁵⁴ with allele frequency greater than 1% were automatically classified as Class 1. Committee members were invited to participate in at least one classification meeting. A core group participated in each meeting, with attendance invited from Committee membership to make up the balance. Before each meeting, participants were assigned, through randomisation, a subset of variants to be assessed. Each attendee was provided with literature pertaining to the list of variants to be discussed, and where relevant, additional unpublished clinical or research information submitted by committee members to InSiGHT curator JPP. Meeting attendees were requested to thoroughly review and summarize all information pertaining to the subset of variants in a spreadsheet template, and provide a class assignment based on their interpretation of the information accessed. All reviewer summaries, submitted clinical information, and causality analysis results were compiled into a single file to allow comparison of data and class assignments for each variant, and circulated to the teleconference participants. During committee meetings, variants were discussed one at a time, assessing the following: class assigned by each reviewer; rationale for classification according to the classification guidelines; difficulties in interpreting specific data sources; assessment of possible redundancy of information due to multiple publications including all or part of the same information pertaining to a variant; differences in interpretation of the guidelines as provided and adjustments required to improve their clarity; consensus view on variant class considering the preceding discussion; action required to obtain additional information for refining classification of variants that remained in Class 2, 3 or 4 at the close of discussion. Where classifications differed using qualitative versus quantitative criteria, this was due to differences in availability of specific data types for the two approaches, and the most extreme classification was assigned for relevant variants. Author BAT applied rules-based classification for variants that were truncating/large deletion from nomenclature, canonical splice site with no splicing data, or frequency >1% in a control reference group. Author BAT then collated all information for all unique “not obviously truncating” variants (including those reviewed in teleconferences previously), and determined which variants had sufficient information to allow classification outside of class 3. Summary information for these variants were circulated for independent class assignment by at least three reviewers from the VIC, and classification finalized at teleconferences or by email.

Validation of Qualitative Criteria

A subset of truncating variants and large genomic deletions were selected to validate the qualitative classification criteria. The variants were selected on the basis of availability of data from the first point of evidence in the qualitative Class 5 criterion, i.e. in vitro functional assay results (e.g. protein truncation test or genomic/mRNA confirmation of large deletions); Constitutional MMR Deficiency Syndrome phenotype; or different haplotypes across multiple families. Published and unpublished data for these variants were then used to validate the other points of evidence required for Class 5 “pathogenic”.

Preliminary Analysis of Class 3 “uncertain” Variants

In silico probabilities of pathogenicity were estimated for all Class 3 missense variants, as described elsewhere³⁴. Preliminary bioinformatic analysis of Class 3 regulatory variants was undertaken using the ENCODE data⁵⁵ on UCSC genome browser.

Implementation of the microattribution process

The variant interpretation process utilizes both published and unpublished data. For published literature the pubmed ID was used to reference the original work. Some unpublished data was recorded in the InSiGHT database at study initiation, and InSiGHT members were also requested by email to contribute information important for variant classification using a standardized submission template. Data submitters were requested to provide a permanent unique publicly searchable ID, preferably from the ORCID system to facilitate adoption of the microattribution approach. Microattribution was assigned for the different types of information corresponding to the points of evidence required for classification, namely submitters were allocated one credit of microattribution for each type of information received:

1. Variant (Mandatory)

2. Family History/Pedigree

3. MSI

4. Immunohistochemistry

5. In-vitro functional

6. RNA splicing assays

7. Population frequency

All unpublished data received by the VIC was recorded in microattribution tables for each element type, where each microattribution table lists a unique researcher ID along with submitted information. Microattribution counts for submitters are publically available on the InSiGHT website. Additionally, the data will be made available in nanopublication format.

Statistical Analysis

Multifactorial likelihood analysis was done for variants with appropriate tumor and segregation data available, using methods previously reported^16,34,51, described briefly as follows. Bayes factor analysis was conducted by author BAT to assess MLH1, MSH2, MSH6, and PMS2 variant causality from segregation data^16,51, for both published and unpublished pedigrees with sufficient relevant information on cancer and variant carrier status. Penetrance estimates from Senter et al²⁸ were used in the Bayes segregation analysis²⁷ of PMS2 variants. Where family relationship status was unknown, a conservative segregation likelihood ratio (LR) was derived i.e. setting affected carriers as first-degree relatives, which is less informative than segregation between second-degree relatives. Colorectal tumor MSI and somatic BRAF mutation status were used to assign LRs according to tumor phenotype, derived as previously reported from the ratio of these characteristics in known mutation carrier cases vs non-mutation carrier cases¹⁶. For each variant, the individual LRs (co-segregation, tumor) were multiplied to calculate the odds for causality. Then, a posterior probability was calculated from combining the prior probability (in silico for missense variants³⁴ or based on sequence location for all other variants¹³) and the odds for causality using Bayes rule: posterior = (prior × odds × (1/(1-prior)))/(prior × odds × (1/(1-prior))+1).

STATA 11 was used to calculate sample size of truncating variant validation set: H₀: p=0.01 with the following assumptions =0.05 (one-sided) and power=0.95.

All other analyses were completed using the statistical package R and GraphPad Prism 6. For meta-analysis of population frequency data, the proportions were combined using an inverse variance random effects model, to account for heterogeneity between studies.

InSiGHT Collaborators

Adela Castillejo⁴⁷, Adrienne Sexton⁴⁸, A.K.W. Chan²⁸, Alessandra Viel⁴⁹, Amie Blanco⁵⁰, Amy French⁵¹, Andreas Laner²², Anja Wagner⁵², Ans van den Ouweland⁵², Arjen Mensenkamp⁵³, Artemio Payá⁵⁴, Beate Betz³⁷, Bert Redeker⁵⁵, Betsy Smith⁵⁶, Carin Espenschied⁵⁷, Carole Cummings⁵⁸, Christoph Engel⁵⁹, Claudia Fornes⁶⁰, Cristian Valenzuela⁶¹, Cristina Alenda⁵⁴, Daniel Buchanan⁶², Daniela Barana⁶³, Darina Konstantinova⁶⁴, Dianne Cairns⁶⁵, Elizabeth Glaser⁶⁶, Felipe Silva⁶⁷, Fiona Lalloo⁶⁸, Francesca Crucianelli⁴⁴, Frans Hogervorst⁶⁹, Graham Casey⁷⁰, Ian Tomlinson⁷¹, Ignacio Blanco¹⁰, Isabel López Villar⁷², Javier Garcia-Planells⁷³, Jeanette Bigler⁷⁴, Jinru Shia⁷⁵, Joaquin Martinez-Lopez⁷⁶, Johan J.P. Gille⁷⁷, John Hopper⁷⁸, John Potter⁷⁹, José Luis Soto⁴⁷, Jukka Kantelinen³¹, Kate Ellis⁸⁰, Kirsty Mann⁴⁸, Liliana Varesco⁸¹, Liying Zhang⁸², Loic Le Marchand⁸³, Makia J. Marafie⁸⁴, Margareta Nordling⁸⁵, Maria Grazia Tibiletti⁸⁶, Mariano Ariel Kahan⁸⁷, Marjolijn Ligtenberg⁵³, Mark Clendenning⁶², Mark Jenkins⁷⁸, Marsha Speevak⁸⁸, Martin Digweed⁸⁹, Matthias Kloor⁹⁰, Megan Hitchins⁹¹, Megan Myers⁵⁰, Melyssa Aronson⁹², Mev Dominguez Valentin⁹³, Michael Kutsche⁹⁴, Michael Parsons¹, Michael Walsh⁶², Minttu Kansikas³¹, Mohd Nizam Zahary⁹⁵, Monica Pedroni⁹⁶, Nao Heider⁹⁷, Nicola Poplawski⁹⁸, Nils Rahner⁹⁹, Noralane M. Lindor¹⁰⁰, Paola Sala¹⁰¹, Peng Nan¹⁰², Peter Propping¹⁰³, Polly Newcomb⁷⁹, Rajiv Sarin¹⁰⁴, Robert Haile⁷⁰, Robert Hofstra⁵², Robyn Ward⁹¹, Rossella Tricarico⁴⁴, Ruben Bacares⁷⁵, Sean Young¹⁰⁵, Sergio Chialina⁶⁰, Serguei Kovalenko¹⁰⁶, Shanaka R. Gunawardena⁵¹, Sira Moreno¹⁰⁷, S.L. Ho²⁸, S.T. Yuen²⁸, Stephen N. Thibodeau⁵¹, Steve Gallinger¹⁰⁸, Terrilea Burnett⁸³, Therese Teitsch¹⁰⁹, T.L. Chan²⁸, Tom Smyrk⁵¹, Treena Cranston¹¹⁰, Vasiliki Psofaki¹¹¹, Verena Steinke-Lange⁹⁹, Victor-Manuel Barbera¹¹²

⁴⁷Department of Molecular Genetics, Elche University General Hospital, Elche, Spain. ⁴⁸Familial Cancer Centre, Royal Melbourne Hospital, Australia. ⁴⁹Oncological Referral Center, IRCCS, Aviano, Italy. ⁵⁰Hereditary GI Cancer Prevention Program, University of California, San Francisco, USA. ⁵¹Department of Laboratory Medicine and Pathology, Mayo Clinic, Rochester, Minnesota, USA. ⁵²Department of Clinical Genetics, Erasmus Medical Center, Rotterdam, The Netherlands. ⁵³Department of Human Genetics, Radboud University Medical Center, Nijmegen, The Netherlands. ⁵⁴Department of Pathology, Hospital Universitario Alicante, Spain. ⁵⁵Department of Clinical Genetics, Academic Medical Center, Amsterdam, The Netherlands. ⁵⁶Benefis Sletten Cancer Institute, Great Falls, MT, USA. ⁵⁷Division of Clinical Cancer Genetics, City of Hope, Duarte, CA, USA. ⁵⁸The Family Cancer Clinic, St Mark’s Hospital, Harrow, UK. ⁵⁹Institute for Medical Informatics, Statistics and Epidemiology, University of Leipzig, Germany. ⁶⁰Molecular genetics, STEM Lab, Rosario, Argentina. ⁶¹School of Medicine, New York University, USA. ⁶²Department of Population Health, QIMR Berghofer Medical Research Institute, Brisbane, Australia. ⁶³U.O.C. Oncologia ULSS5 Ovest Vicentino, Ospedale di Montecchio Maggiore (VI), Italy. ⁶⁴Molecular Medicine Center, Medical University of Sofia, Bulgaria. ⁶⁵Liverpool Women’s Hospital, Liverpool, UK. ⁶⁶International Society for Gastrointestinal Hereditary Tumours. ⁶⁷Laboratory of Genomics and Molecular Biology, A C Camargo Cancer Center, Brazil. ⁶⁸Manchester Centre for Genomic Medicine, Central Manchester University Hospitals NHS Foundation Trust, UK. ⁶⁹Family Cancer Clinic and Department of Pathology, The Netherlands Cancer Institute, The Netherlands. ⁷⁰Department of Preventive Medicine, University of Southern California, Los Angeles, CA, USA. ⁷¹Molecular and Population Genetics Laboratory, London Research Institute, Cancer Research UK, UK. ⁷²Hospital 12 de Octubre, Universidad Complutense, Madrid, Spain. ⁷³Institute of Genomic Medicine, University of Valencia, Spain. ⁷⁴Medical Sciences, Amgen Inc., Seattle, WA, USA. ⁷⁵Department of Pathology, Memorial Sloan-Kettering Cancer Center, USA. ⁷⁶Molecular Biology Laboratory, Hospital Universitario 12 de Octubre, Madrid, Spain. ⁷⁷Clinical Genetics, VU University Medical Centre, The Netherlands. ⁷⁸Centre for Molecular Environmental, Genetic and Analytic (MEGA) Epidemiology, University of Melbourne, Victoria, Australia. ⁷⁹Public Health Sciences Division, Fred Hutchinson Cancer Research Center, Seattle, WA, USA. ⁸⁰Hunter Family Cancer Service, Waratah, Australia. ⁸¹Center for Hereditary Tumours, National Institute for Cancer Research, Genoa, Italy. ⁸²Diagnostic Molecular Genetics Laboratory, Memorial Sloan-Kettering Cancer Center, USA. ⁸³University of Hawaii Cancer Center, Honolulu, Hawaii, USA. ⁸⁴Cancer Genetics Unit, Kuwait Medical Genetics Centre, Kuwait. ⁸⁵Department of Molecular and Clinical Genetics, Institute of Biomedicine, Sahlgrenska Academy, University of Gothenburg, Sweden. ⁸⁶Unit of Pathology, Varese Hospital, Varese, Italy. ⁸⁷Molecular Oncology ICBME, Hospital Italiano de Buenos Aires, Argentina. ⁸⁸Division of Genetics, Trillium Health Partners, Credit Valley Hospital, Mississauga, Ontario, Canada. ⁸⁹Charité Berlin, Institute of Human Genetics, Germany. ⁹⁰Department of Applied Tumor Biology, Institute of Pathology, University of Heidelberg, Germany. ⁹¹Lowy Cancer Research Centre, Prince of Wales Clinical School, Faculty of Medicine University of New South Wales, Australia. ⁹²Familial GI Cancers Unit, Mount Sinai Hospital, Toronto, Canada. ⁹³Department of Oncology, Clinical Science, Lund University, Sweden. ⁹⁴Laboratory for Molecular Medicine Genetic, Praenatalzentrum laboratories, Hamburg, Germany. ⁹⁵Human Genome Centre, School of Medical Sciences, Universiti Sains Malaysia, Malaysia. ⁹⁶Department of Medicine and Medical Specialties, Modena University Hospital, Italy. ⁹⁷RIKEN, Japan. ⁹⁸SA Pathology, Women’s and Children’s Hospital, Australia. ⁹⁹Medical Faculty, Institute of Human Genetics, University of Duesseldorf, Germany. ¹⁰⁰Department of Health Science Research, Mayo Clinic, Scottsdale, Arizona, USA. ¹⁰¹Hereditary Cancers of the Digestive Tract Unit, Predictive and Preventive Medicine, National Tumor Institute IRCCS Foundation, Milan, Italy. ¹⁰²School of Life Sciences, Fudan University, China. ¹⁰³Institute of Human Genetics, University of Bonn, Germany. ¹⁰⁴SARIN LAB, ACTREC, Tata Memorial Centre, Mumbai, India. ¹⁰⁵Cancer Genetics Laboratory, British Columbia Cancer Agency, Vancouver, Canada. ¹⁰⁶Genetic Technologies Limited, Australia. ¹⁰⁷Genetics Service, Hospital Virgen del Camino, Spain. ¹⁰⁸Zane Cohen Centre for Digestive Diseases, Toronto, Canada. ¹⁰⁹Dartmouth Medical School, Dartmouth College, Lebanon, New Hampshire, USA. ¹¹⁰Oxford Medical Genetics Laboratories, Oxford University Hospitals NHS Trust, The Churchill Hospital, Oxford, UK. ¹¹¹Biochemical Laboratory, University Hospital of Ioannina, Greece. ¹¹²Research Laboratory, University Hospital of Elche, Elche, Spain.