Large scale multifactorial likelihood quantitative analysis of BRCA1 and BRCA2 variants: An ENIGMA resource to support clinical variant classification

2.1 Variant selection for data collection

At the time of joining ENIGMA, members were asked to submit all variants in BRCA1 and BRCA2 that they considered to be of uncertain clinical significance, together with the number of families carrying each variant. We followed the rationale that high-risk variants will not occur commonly in the general population, as indicated by frequency measured in outbred reference datasets representative of subpopulations. Variants were thus classified as Class 1 Not Pathogenic if they were identified to occur at minor allele frequency >0.01 in one or more of the following datasets: South Asian, Latino, African, East Asian, Non-Finnish European subpopulations from the Exome Aggregation Consortium (ExAC) dataset (after excluding cancer-related information from The Cancer Genome Atlas (TCGA; http://exac.broadinstitute.org); European, African, Admixed American, East Asian, or South Asian sample sets from the 1000 Genomes Project (http://www.1000genomes.org). This exercise provided a baseline variant list for subsequent ENIGMA studies. For the analysis presented in this study, a subset of variants were prioritized for collection of segregation and breast tumor pathology data, based on number of observations/families in the initial ENIGMA variant list and/or bioinformatic score indicative of pathogenicity. Information for co-segregation analysis was provided in the form of a deidentified pedigree for families with known carrier status in more than one individual. Pedigree details included sex, cancer status, and age at cancer diagnosis, or age at interview if unaffected. Unaffected individuals known to have undergone prophylactic surgery (mastectomy or oophorectomy) were censored at age of earliest surgery. Breast tumor pathology information collected for known variant carriers included hormone receptor status (estrogen receptor [ER], progesterone receptor [PR], human epidermal growth factor receptor 2 [HER2]), and/or tumor grade. In addition, clinical queries to the Spurdle laboratory led to the collation of additional segregation and pathology information of potential value for multifactorial likelihood analysis of individual variants. Further, genotype data generated as part of the iCOGS project was available via collaboration with the Breast Cancer Association Consortium (BCAC) for a subset of variants, from up to 41,141 breast cancer cases and 38,694 controls of European ancestry, and 6,185 breast cancer cases and 6,614 controls of Asian ancestry (Michailidou et al., 2013). Variants included in the iCOGS project were prioritized for genotyping using the same approach as for the baseline ENIGMA variant list, with additional variants selected due to laboratory/bioinformatic evidence for effect on mRNA splicing. For each variant, a positive control DNA from a variant carrier was submitted for genotyping to facilitate calling of these rare variants. Lastly, we included a subset of variants for which multifactorial likelihood analysis results had previously been published, but the final classification reported was not “Class 5 Pathogenic” or “Class 1 Not Pathogenic”, and/or LRs were not all visible in the original publication (Easton et al., 2007; Farrugia et al., 2008; Lindor et al., 2012). Information from all these sources was collated for a total of 3,295 variants. This amalgamated list of variants was then circulated by email to ENIGMA consortium members to invite them to provide additional segregation or pathology information for inclusion in the analysis. Where relevant, additional nonoverlapping pathology or segregation information was sourced directly from publications for inclusion in the analysis. Overall, there were 1,008 informative pathology data points and 895 informative families for segregation analysis from ENIGMA collaborators, clinical enquiries, and nonoverlapping publications (see below for further explanation about LR assignment). After combining all information, at least one data point was available for 1,395 variants.

Variant descriptions are in accordance to HGVS recommendations. Nucleotide numbering corresponds to reference transcripts NM_007294.3 (BRCA1) and NM_000059.3 (BRCA2). Legacy description is also provided to assist comparison with historical records in the literature; the nucleotide numbering is from nucleotide one of the full gene sequence (Genbank: U14680.1/BRCA1; U43746/BRCA2) not the ATG initiator codon, and BRCA1 exon boundaries are from GenBank U14680.1 with exon four missing due to a correction made after the initial description of the gene.

2.2 Multifactorial likelihood analysis

A Bayesian model was used to combine evidence as previously described (Goldgar et al., 2008). In brief, the prior probability of pathogenicity was assigned based on a combination of Align-GVGD score and MaxEntScan splicing predictions, overlaid with expert knowledge incorporating prediction of variant effect on critical functional protein domains (Tavtigian et al., 2008; Vallee et al., 2016). Applicable prior probability of pathogenicity predictions for single nucleotide variants in BRCA1 and BRCA2 are available from the HCI Database of Prior Probabilities of Pathogenicity for Single Nucleotide Substitutions (http://priors.hci.utah.edu/PRIORS/). Align-GVGD does not score in-frame insertions and deletions, therefore to estimate prior probabilities for in-frame deletion variants studied, we took the highest Align-GVGD prior of the deleted bases. There were no in-frame exonic insertions included in this study. The higher of the two priors (missense vs. splicing) was assigned for analysis. Co-segregation analysis was performed as described by Thompson et al. (2003) for each family with more than one individual genotyped for the variant. Hazard ratio estimates were taken from Antoniou et al. (2003) for <30, 30–39, 40–49, 50–59, 60–69, 70–79, and 80+ age brackets. For individuals affected with ovarian cancer at 20–29 years, penetrance for ovarian cancer at age 30–39 was applied due to a lack of information in the younger penetrance class (Antoniou et al., 2003). If no age at last update was provided by the submitting center for unaffected individuals, test date was used to infer current age. Breast tumor pathology LRs were assigned based on the estimates in Spurdle et al. (2014), and considered age at diagnosis. When multiple tumors were present in one individual, the first diagnosed tumor with information available was taken, and only a single LR was assigned according to the extent of information available (out of the variables tumor grade, ER, PR, and HER2), following previous recommendations (Spurdle et al., 2014). Likelihood ratios for co-occurrence with a pathogenic variant (in trans), and reported family history analysis, were drawn from a previous publication (Easton et al., 2007). Case-control data from the iCOGS project were used to estimate LRs following methods described previously (de la Hoya et al., 2016).

Table S1 summarizes the LRs assigned for each component for each variant with at least one data point. Prior probabilities and LRs were combined to calculate posterior probabilities using Bayes rule: (Prior Probability x Combined LR)/(Prior Probability x [Combined LR + {1 − Prior Probability}]). Where multiple data points were available for a single data type, for example, segregation, LRs were combined multiplicatively. Using variant BRCA1 c.131G>T as an example: Prior Probability is 0.81; Combined LR is 6,440.7 (based on LR Segregation (156.17) x LR Pathology (41.24) x LR Co-occurrence (a) x LR Family History (a) x LR Case-Control (a)). Posterior probability is 0.99996, calculated as (0.81 × 6440.7)/(0.81 × [6440.7 + {1-0.81}]). Breakdown of clinical data type contributed, and the data sources (submitter, publication source), are shown in Table S2.

It has previously been proposed that a combined LR between 0.5 and 2, in particular if derived from a limited number of data points, provides insufficient observational data to perform a valid integrated analysis (Vallee et al., 2016). This is in accord with the idea that ACMG guidelines intrinsically include an indeterminate zone, between supporting benign and supporting pathogenic, for variants with insufficient or conflicting evidence for pathogenicity. Following this rationale, posterior probabilities of pathogenicity were not calculated for any variant with a combined LR between 0.5 and 2. Posterior probability of pathogenicity was calculated for a total of 734 variants, and classification assigned based on previously published cut-offs proposed for the International Agency for Research into Cancer (IARC) five tier classification scheme (Plon et al., 2008), with some modification of terms used to describe tiers (Spurdle et al., 2019), namely: Class 5 Pathogenic, >0.99; Class 4 Likely Pathogenic, 0.95–0.99; Class 3 Uncertain, 0.05–0.949; Class 2 Likely Benign, 0.001–0.049; and Class 1 Benign, <0.001. The variant classifications, with breakdown of LR components and sources, have been submitted to the following databases for public display:

<http://hci-exlovd.hci.utah.edu/home.php?select_db=BRCA1>

<http://hci-exlovd.hci.utah.edu/home.php?select_db=BRCA2>

2.3 Datasets providing information for comparison and calibration of qualitative classification criteria

3.1 Correlation of multifactorial likelihood classifications with splicing assay data

Current ENIGMA BRCA1/2 classification criteria for spliceogenic variants, consistent with those of the InSiGHT Consortium developed for classification of mismatch repair gene variants (Thompson et al., 2014), present stringent recommendations for use of mRNA splicing data for variant interpretation (https://enigmaconsortium.org). Namely, a variant is only considered pathogenic on the basis on mRNA splicing data if there is no predicted functional transcript produced from the variant allele, as determined by assays of patient-derived mRNA that have assessed allele-specific expression of alternate transcripts. This stipulation is not specified for ACMG/AMP classification codes using splicing data (PS3, well-established in vitro or in vivo functional studies supportive of a damaging effect on the gene or gene product).

We undertook a comparison of multifactorial model classifications against published splicing assays results (including assays of patient material and construct-based assays) to calibrate use of splicing assay data for use as weighted information for qualitative classification, based on the LR ranges recently proposed as consistent with ACMG/AMP qualitative rule strengths for future classification in a Bayesian framework (Tavtigian et al., 2018). Of the variants falling outside of Class 3 Uncertain in this analysis, 99 had mRNA splicing data available, 25 of which had been assessed using allele-specific assays of mRNA from patient tissue. By comparing splicing effect to classifications derived from this study, we estimated a LR towards pathogenicity based on effect on mRNA splicing (Table 2; Table S3 for additional details). The very limited number of allele-specific assays did not allow for robust estimates of LRs, with the confidence intervals for LRs estimated for both partial and complete effect on splicing including unity. Nevertheless, results support the hypothesis that partial effect on splicing will not be as strongly predictive of pathogenicity as is complete effect on splicing (LR 3.82 vs. LR 6.36 from this analysis). Including all assay results, no effect on splicing provided strong evidence against pathogenicity (LR 0.02), whereas any impact on splicing (without measurement of allele-specific effects, or consideration of in-frame transcripts) provided moderate evidence for pathogenicity (LR 12.24). Recognizing the small sample sizes, and consequently large confidence limits, these results nevertheless demonstrate the value of mRNA splicing assays as a component in qualitative variant classification. We also highlight the possibility that there is likely to be considerable bias in variants selected for mRNA assays, with over-representation of variants at the highly conserved donor and acceptor dinucleotides, positions that, when altered, are likely to impact splicing more severely than spliceogenic variants located at other positions. We thus stress the importance of incorporating allele-specific expression assays into variant evaluation processes wherever possible, and to revisit such analysis with larger datasets in the future. We also recommend that future larger-scale comparisons to derive LRs for splicing assay data should consider in greater detail the predicted impact of the aberrant mRNA profiles on protein function, and in particular consider the relevance of in-frame isoforms that could be translated to result in (partially) functional protein. Further, as both partial mRNA splicing and in-frame transcripts may be associated with reduced cancer penetrance that is not inherently captured by the design of the multifactorial model, family-based and case-control studies may be necessary to tease out which such spliceogenic variants are indeed risk-associated, and if this level of risk is clinically actionable.

We then considered qualitative classification based on mRNA splicing results for all variants with a multifactorial likelihood calculation that is including Class 3 Uncertain variants (See Table S1, columns Splicing Results/s and Allele-Specific Assay). Following ENIGMA BRCA1/2 qualitative classification criteria (http://enigmaconsortium.org/), there were 15 variants that could be interpreted as Pathogenic based on splicing, that is no predicted functional transcript produced from the variant allele; of these, multifactorial data classified five as (Likely) Pathogenic, five as Uncertain, whereas four had insufficient data to perform a calculation. BRCA1 c.5453A>G was the only variant with truly discordant classification between splicing results (Class 5 Pathogenic) and multifactorial data analysis (Class 2 Likely Benign); the multifactorial classification was based on low prior probability of 0.03 for the presumed missense substitution Asp1818Gly, one pathology data point (LR 0.34), and relatively uninformative co-occurrence (LR 1.12) and family history (LR 0.91) data. This variant highlights a recognized limitation of current bioinformatic predictions used in the multifactorial analysis; the variant alters splicing by modifying an exonic splice enhancer (ESE; Rouleau et al., 2010). There are currently no bioinformatic prediction tools with adequate sensitivity and specificity to predict ESE loss or gain with any reliability, and this mechanism has thus not yet been incorporated into bioinformatic prior probability estimation. Although results from splicing assays can obviously add value for such examples, the poor predictability of ESEs and effects of variation on ESE function, hinders the prioritization of ESE-altering variants for splicing assays. We reiterate that the Class 2 Likely Benign tier implicitly allows for a 5% error rate in classification, and resources permitting, we would encourage additional data collection for all variants falling in this tier. Moreover, future inclusion of a LR derived for splicing impact as a component of the multifactorial likelihood analysis, where such splicing information is available, would likely shift the posterior probability for such variants into the range of Class 3 Uncertain and so prevent overt misclassification driven by bioinformatic prediction deficiencies. At this point in time, we would encourage additional clinical data collection for BRCA1 c.5453A>G to confirm that the clinical phenotype is consistent with a Class 5 Pathogenic assertion based on splicing data only.

3.2 Correlation of multifactorial likelihood classifications with protein functional assay data

Functional assays are considered strong evidence for or against pathogenicity using ACMG/AMP codes PS3 and BS3 (well-established in vitro or in vivo functional studies show (damaging/no damaging) effect on gene or gene product). A range of different assays have been used to assess effect of BRCA1 and BRCA2 variants on protein function, some limited to measuring impact on function of variants within a specific domain, and others measuring output relevant to a variant located anywhere in the coding region. To assess the strength of this evidence as a predictor of the clinical significance of anticipated missense BRCA1 and BRCA2 variations, it is important to consider several factors. Sensitivity and specificity of assays should be determined using missense variants that have previously been determined to be pathogenic or benign (Guidugli et al., 2014; Millot et al., 2012); that is assay profiles for truncating variants may not be appropriate to measure loss of function displayed by pathogenic missense variants. To prevent circularity, functional assay results should not have contributed to the classification of these “control” missense variants, as may be the situation for variants submitted to ClinVar as pathogenic. An additional factor to consider, but not addressable at this point in time, is that there are few BRCA1/2 variants robustly proven to be associated with moderate risk of cancer. There is thus a paucity of controls to calibrate assay results to detect moderate-risk variants. Moderate-risk variants are intuitively expected to have less impact on function than variants associated with a high cancer risk comparable to that of the average truncating allele, and severity of their impact on function may differ depending on the specific protein effects measured (Lovelock et al., 2007).

For this reason, we compared our multifactorial analysis results for missense variants classified outside of Class 3 Uncertain to results from selected published functional assays (also see Methods). Briefly, these included: (a) domain-specific or generic assays assessing variant effect on protein function, and calibrated against missense variants previously classified as pathogenic or benign using multifactorial likelihood analysis that is using bioinformatic and clinical information (Bouwman et al., 2013; Fernandes et al., 2019; Hart et al., 2019; Mesman et al., 2019; Petitalot et al., 2019); and (b) multiplex reporter assays (Findlay et al., 2018; Starita et al., 2018) reported to have reasonable to good sensitivity and specificity by comparison to ClinVar classifications (including truncating, splicing, and missense variants). There were 16 (Likely) Pathogenic and 61 (Likely) Benign variants with a protein functional assay result from at least one study (Table 3; Table S4 for additional details). All 56 variants reported to have no functional impact were classified as (Likely) Benign, as were the four of five variants demonstrating partial function in at least one assay. The fifth variant BRCA1 c.5216A>T p.(Asp1739Val) was classified as Likely Pathogenic based on posterior probability of 0.97. As outlined in Supp Table S4, this missense substitution variant was reported to have complete loss of function using transcription activation and cell survival assays, but partial activity by Petitalot et al. (2019); the latter categorization was based on the combination of somewhat decreased solubility, decreased BACH1 binding (reported in yet another publication, Lee et al. (2010)), and normal homologous recombination and localization (Petitalot et al., 2019). Of the 16 variants reported to impact function completely (and with no evidence otherwise by another of the functional studies selected), 15 were classified by multifactorial likelihood analysis as (Likely) Pathogenic, and the other as (Likely) Benign. Of note, the latter variant BRCA2 c.8351G>A p.(Arg2784Gln) did complement lethality in the mouse embryonic stem cell assay (Mesman et al., 2019), but was coded as impacting function based on homologous recombination assay results from the same study (Mesman et al., 2019), and was reported to impact homologous recombination in an independent study (Hart et al., 2019). We note that for the two exceptions highlighted (BRCA1 c.5216A>T, BRCA2 c.8351G>A), the results from survival assays were concordant with the multifactorial likelihood classification.

Considering results overall, we estimated a LR towards pathogenicity based on assays of protein function from at least one study (Table 3). Acknowledging the caveat of small sample sizes, and at least one observation (and thus liberal frequency estimates) assumed for cells without counts, our results support use of functional assay data as moderate or strong evidence in determining pathogenicity assertions for missense variants. Specifically, complete impact on function with no conflicting evidence is strongly predictive of missense variant pathogenicity (estimated LR 57.19, lower confidence bound 8.15). No impact on protein function provides moderate evidence against missense variant pathogenicity (LR 0.07 with upper bound 0.45, equating to an LR of 15.26 against pathogenicity). The results confirm the value of results from these selected protein functional assays as a component in qualitative classification of missense variants. They also stress the importance of considering discordances across different assay methods as an approach to select individual variants for further consideration as potential moderate-risk variants, variants that may not always be detectable as risk-associated using statistical models developed for high-risk variants. Further, as noted before, the BRCA1/2 multifactorial model is designed to capture clinical features of patients with the average high-risk pathogenic variant, and we cannot exclude the possibility that some variants demonstrating impact on function (partial, or even complete for at least one assay type) are moderate-risk alleles. It will thus be important to prioritize variants such as BRCA1 c.5216A>T p.(Asp1739Val) and BRCA2 c.8351G>A p.(Arg2784Gln), where some functional data conflict the clinical information data (thereby arguably considered Uncertain according to ACMG/AMP qualitative criteria) for further study as potential moderate-risk variants.

3.3 Correlation of multifactorial likelihood classifications with frequency in reference population datasets

Variant frequency in disease-free controls can be used to provide evidence against pathogenicity, and indeed minor allele frequency (MAF) > 1% in a nonfounder population is considered stand-alone evidence against pathogenicity for BRCA1/2 variants by the ENIGMA consortium. An algorithm to define a “maximum credible population allele frequency” (Whiffin et al., 2017) has been proposed as a method to select MAF cut-offs as evidence against pathogenicity, and indeed was used as a basis to select relevant minor allele frequency cut-offs for the PTEN and CDH1 adaptations of ACMG/AMP rule codes BA1 (stand-alone) and BS1 (strong) evidence against pathogenicity, described as “allele frequency is greater than expected for the disorder.” The output of this algorithm can vary widely depending on input assumptions for disease penetrance and prevalence of the disorder, and is complicated for multicancer syndromes where penetrance varies for cancer type and even cancer subtype. Further, absence from control datasets has been proposed as moderate evidence for variant pathogenicity (ACMG/AMP rule code PM2).

The most commonly used “control” reference sets (ExAC, and more recently gnomAD) include males and females that were ascertained for noncancer related studies mostly at ages younger than the average age at onset of BRCA1/2-related breast or ovarian cancer, but individual-level information about cancer phenotypes is not available. Even assuming that these reference sets are largely cancer unaffected, it must be considered that penetrance in female pathogenic variant carriers for breast cancer is not complete, and much lower for male carriers. Indeed, known pathogenic BRCA1/2 variants have been identified in these population control sets, even after accounting for “founder” pathogenic variants (Maxwell, Domchek, Nathanson, & Robson, 2016). As a result, the current ENIGMA BRCA1/2 classification guidelines used empirical data to select frequency cut-offs for qualitative classification criteria (https://enigmaconsortium.org/). Specifically, allele frequency ≥ 0.001 and < 0.01 in large outbred control reference groups was selected as a component of evidence against pathogenicity, based on the upper 95% confidence interval (binomial Exact) of the frequency observed for the most common pathogenic allele in non-Finnish European and other population groups drawn from ExAC and gnomAD. The absence from controls has not yet been incorporated into the ENIGMA BRCA1/2 guidelines.

To determine the utility of variant frequency in (or absence from) reference population sets for future BRCA1 and BRCA2 variant classification, and to formally assess the strength of this evidence, we estimated a LR based on MAF in gnomAD v2.1 (noncancer), and considered them against LR cut-offs suggested for ACMG/AMP rules (Tavtigian et al., 2018). Results are shown in Table 4. We considered variants observed only once across all sample sets reviewed as a separate category (see Methods), and categorized the remaining variants into three MAF bins: 0.01 > MAF ≥ 0.0001; 0 < MAF < 0.0001; and not observed. The proportion of variants seen only once across all five sample sets was 13.32% for (Likely) Benign variants compared with 8.99% for (Likely) Pathogenic Variants, which equates to an LR of 0.67, considered uninformative for pathogenicity prediction. Variants classified as (Likely) Benign were spread relatively evenly across the frequency categories, whereas (Likely) Pathogenic variants were only seen at MAF < 0.0001, and the vast majority (88%) were not seen in population controls. Assuming conservatively a single (Likely) Pathogenic variant to fall in the category “ ≥ 0.0001 & < 0.01,” the LR estimate is 0.05, equating to an LR of 22.10 (3.12–156.21) against pathogenicity, considered strong evidence that a variant is benign. The estimated LR against pathogenicity for a variant seen in gnomAD at MAF < 0.0001 is 7.97 (2.59–24.50), which meets moderate evidence against pathogenicity. Last, the estimated LR towards pathogenicity for a variant not detected in gnomAD is 2.50 (2.16–2.91), corresponding to supporting evidence in favor of pathogenicity; these findings suggest that whereas “absence in controls” may be useful for BRCA1/2 variant classification, such evidence should carry less weight than the PM2 moderate code proposed by ACMG/AMP for generic use.

Overall, these findings, based on empirical data, have utility to inform the ongoing adjustment of the ENIGMA BRCA1/2 guidelines. Use of these frequency cut-offs for variants designated in this study (Table S1) as Class 3 Uncertain or with posterior probability not calculated, suggests strong evidence against pathogenicity for 98 variants, moderate evidence against pathogenicity for 147 variants, and supporting evidence in favor of pathogenicity for 468 variants. It is relevant to acknowledge that some variants detected in public databases - and also in clinical datasets - may be somatic rather than germline in origin, arising due to clonal hematopoiesis of indeterminate potential. However, the proportion of variants arising due to this mechanism is expected to be rare for BRCA1 and BRCA2 (estimated as < 0.2% of all pathogenic variants in one study of >200,000 cancer gene tests (Coffee et al., 2017)), and we would also anticipate that the majority of such variants would be filtered out by generic allele fraction cut-offs used by sequencing pipelines set up to detect true germline variants.

3.4 Caveats, considerations, and conclusions

This study is the largest presentation of multifactorial likelihood analysis to date. Although we mined existing public data and requested clinical data for calculations from more than 300 individuals on the ENIGMA mailing list, we recognize that the classifications assigned may alter with addition of data from other sources, and/or with the application of qualitative classification criteria. Efforts to collate information in a transparent manner that retains patient confidentiality, and within the bounds of ethical constraints, remain a challenge. As one step towards transparency, we present summary estimates of LRs for the individual components included in the analysis, and the sources of the different information types. Further discussion, and probably technical developments, will be required to determine how more detailed information, for example, segregation scores for individual families, may be presented for future large-scale studies.

A prepublication iteration of the classification dataset was used as a reference set for the Critical Assessment of Genome Interpretation (CAGI) 5 experiment, results from which are presented in this same Journal issue (Cline et al., ). Comparison of various different prediction methods from six different teams highlighted that prediction of mRNA splicing is an important inclusion in variant interpretation algorithms, and also indicated that variant interpretation may be improved by incorporating amino acid accessibility as a component of bioinformatic prediction of variant effect. It also showed that use of clinical information, when available, provides significant improvements to variant classification over purely bioinformatic approaches. In addition to the CAGI 5 experiment, we chose to use an updated dataset for calibration of isolated data types commonly used as components of qualitative classification approaches. Specifically, the variant classifications from multifactorial likelihood analysis were derived without use of laboratory splicing and functional data, or variant frequency in outbred reference populations. This allowed us to estimate independent LRs for or against pathogenicity for these evidence types, with several purposes: to assess the validity of ACMG/AMP code strengths proposed for these evidence types when applying them to classification of BRCA1/2 variants; to justify specific ENIGMA BRCA1/2 classification criteria incorporating population frequency or mRNA splicing data; to provide guidance on incorporation of BRCA1/2 protein functional assay data, to assess BRCA1/2 predicted missense variants specifically, in quantitative or qualitative (rules-based) classification models. We acknowledge that further analyses are necessary before such LR estimations be formally reviewed and incorporated into guidelines for BRCA1/2 Expert Panel variant interpretations. Namely, it will be important to investigate LR estimations with additional variants that have been previously assessed using the multifactorial likelihood approach, and to derive LRs separately for BRCA1 and BRCA2 given differences in the penetrance for truncating variants in these two genes and the potential for differences in sensitivity and specificity of different laboratory assays to detect impact on function. Nevertheless, we note that, despite the fact that known BRCA1/2 pathogenic variants are seen in reference population datasets, and acknowledging reservations that absence in control populations overall is not a predictor of variant pathogenicity, it appears that this feature does have value for examining the clinical relevance of rarer variants presenting for assessment at this point in time; that is, accounting for the fact that “known pathogenic” variants observed in control datasets are more likely to have already been observed in the clinical setting and already classified using other information types. This observation highlights issues around constancy of evidence strengths over time, and indicates that it will be important to reconsider such analysis periodically to re-estimate LRs and corresponding rule strengths as the pool of variants remaining to be classified alters over time. Following this line of thought, it will be important that estimates of prior probability based on bioinformatic predictions are re-estimated using updated datasets that reflect changed variant pools, and altered patient ascertainment in the era of multigene panel testing. Results arising from the CAGI 5 experiment (Cline et al., ), and other similar studies, are likely to inform development of such bioinformatic methods.

In summary, we have used the multifactorial likelihood analysis approach to generate 248 new or considerably altered BRCA1/2 variant classifications, information that is relevant for medical management – including determining patient eligibility for screening or PARPi treatment, and cascade testing of their relatives. We have also shown the value of this dataset for confirming existing ClinVar assertions, and for calibration of additional data types useful for variant interpretation. We have provided as supplementary information details regarding data sources and likelihood scores for all variants investigated, so providing a resource that will facilitate continued assessment of variants as additional information accrues, and further calibration of new lines of evidence relevant for variant interpretation.