Predictive functional assay-based classification of PMS2 variants in Lynch syndrome

2.1 Cell culture and generation of Pms2-hemizygous mouse embryonic stem cells using CRISPR-Cas9

Hemizygous Pms2 mouse embryonic stem cells (mESCs) for use in the genetic screen were generated using CRISPR-Cas9. Briefly, two pairs of single-stranded oligonucleotides with BbsI overhangs targeting the C-terminal and N-terminal end of Pms2 were annealed and cloned into the Cas9 vector PX330-Puro (Cong et al., 2013). mESCs were transfected with the two PX330-Puro plasmids containing the gRNA's using Lipofectamine 2000 (Thermo scientific) and after 24 h, positively transfected cells were selected with 1 μg/ml Puromycin. After selection, cells were subcloned, and Pms2 hemizygosity was confirmed using allele-specific PCR. Pms2 protein expression was checked by immunoblotting.

2.2 Generation of a set of pathogenic PMS2 missense variants using a genetic screen

To compile a set of inactivating amino acid substitutions in Pms2, as a proxy for cancer-predisposing variants, a genetic screen was performed, essentially as described for Msh2 and Msh6 (Drost, Lützen, et al., 2013). In brief, mESCs hemizygous for Pms2 were exposed to the strong mutagen N-ethyl-N-nitrosourea (ENU; Sigma-Aldrich) with the aim of introducing random missense substitutions in the genome (Figure 2). Cells that had become MMR-deficient, presumably by an inactivating mutation at the monoallelic Pms2 gene, were selected using the Guanine analog 6-Thioguanine (6-TG, Sigma-Aldrich). 6-TG is cytotoxic via the lethal processing of methyl-6-thioGuanine:Thymidine mismatches by MMR. Consequently, a defect in MMR results in survival to the drug and 6-TG is a widely used agent used to select for MMR-deficient cells and to test for functional inactivation and cancer predisposition caused by a variant (Drost et al., 2020; Houlleberghs et al., 2016; Houlleberghs et al., 2020). Surviving clones were treated for 1 week in hypoxanthine-aminopterin-thymidine (HAT)-supplemented medium (Gibco) to kill inadvertent Hprt-mutant cells that also survive 6-TG treatment. Next, surviving clones were picked and expanded in 96-well plates. These clones were screened against loss of heterozygosity (LOH) at Pms2, an alternative way of losing the wild-type Pms2 allele, by intragenic allele-specific PCR. From remaining clones, Pms2 cDNA was generated and the critical, conserved, domains were sequenced using Sanger sequencing to identify the inactivating amino acid substitution. All primer sequences and PCR protocols are available upon request.

2.3 Selection of missense substitutions for CIMRA assay calibration and validation

We derived a set of PMS2 missense variants from the ClinVar and InSiGHT databases: 20 had been classified as IARC Class 1/2 [(likely) benign], and 11 as Class 4/5 [(likely) pathogenic], based on only clinical data (Accessed, October 28, 2019) (Supporting Information: Table S1). A total of 31 unique inactivating missense substitutions were derived from the genetic screen (Supporting Information: Table S2). A part of the variants identified in the databases and screen were excluded from the calibration and validation set based on the following criteria: missense variants in the start codon of PMS2; proven spliceogenic variants; variants with an unavailable in silico prior probability (Prior-P); nonconserved amino acid variants (if derived from the murine genetic screen); one variant used as a predisposing control in the CIMRA assay (Supporting Information: Tables S1 and S2).

2.4 Complete in vitro MMR activity (CIMRA) assay

CIMRA assays of the PMS2 missense variants were performed essentially as previously described (Drost, Koppejan, et al., 2013). Briefly, the variant PMS2 cDNA is synthesized by mutagenic PCR followed by in vitro transcription and translation of the variant PMS2 protein as well as of the wild-type MLH1 dimerization partner. Nuclear extract from MLH1-deficient human HeLa cells (which is also deficient for PMS2 protein due to protein instability without its partner) generated by CRISPR-Cas9 gene targeting is complemented with in vitro produced human (variant) PMS2/MLH1 protein and a mismatch-containing fluorescent substrate. Repair of the mismatch recreates a Hin1II restriction site, which is quantified using capillary fluorescent gelectrophoresis. Repair deficiency resulting from an inactivating VUS causes the absence of the repair-diagnostic fluorescent fragment, indicating that the VUS is cancer-predisposing.

2.5 Regression for CIMRA assay calibration and validation

Regression of CIMRA assay results against the known or assumed statistical probability of pathogenicity of the variants from the test set was performed following methods similar to those published for MLH1, MSH2, and MSH6 variants (Drost et al., 2019). Given the relatively small sample size of the PMS2 variant set (N = 51), 10-fold cross-validation was chosen as the preferred method to assess the fit of the regression model; 80% of the data are considered the training set and 20% of the data is considered the validation set in 10 different versions of the model. Analysis was performed using the Caret package in RStudio version 1.0.153. Only five of these variants had information on pathogenicity in the InSiGHT database based on multifactorial likelihood analysis (Supporting Information: Table S1) (B. A. Thompson et al., 2014). For this reason, variants assigned to Class 4 (likely pathogenic) using clinical criteria and the presumed inactivating variants derived from the genetic screen were conservatively assigned a probability of pathogenicity of 0.95 (Class 4), while variants in Classes 1 and 2 [termed (likely) benign]) were conservatively assigned a probability of pathogenicity of 0.05 (Class 2). These probabilities were then transformed to OddsPath using the following formula (B. Thompson et al., 2021): (Probability)/(1-Probability), and used as the dependent variable in the form of log10 in the regression formula. The relative average of the CIMRA assay results for each variant was used as the independent variable. To assess the fit of the regression model, a 10-fold cross-validation procedure was applied. A leave-one-out procedure was then used to calculate the OddsPath for each variant used for calibration of the regression formula. OddsPath for individual variants was aligned to ACMG/AMP evidence code strengths as per Tavtigian et al. (2018), shown in Supporting Information: Table S3.

3.1 Genetic screen for the identification of inactivating Pms2 missense variants

To enable calibration of CIMRA activity of PMS2 variants into an OddsPath and its validation, a set of clinically classified Class 1/2 [(likely)benign] and Class 4/5 [(likely) pathogenic] variants was derived from the ClinVar and InSiGHT databases (Supporting Information: Table S1). Of the variants classified as (likely) pathogenic in ClinVar, five disrupted the start codon and one was used as the inactivating control in the CIMRA assay (p. Ser46Ile), and therefore these could not be used for calibration of the assay. Two variants (p. Ile668Val, p. Lys301Asn) were excluded since they have previously been shown to impact splicing (Li et al., 2015; B. A. Thompson et al., 2020; van der Klift et al., 2015). After these exclusion criteria, as described in Section 2, were applied, 20 (likely) benign and 3 (likely) predisposing missense PMS2 variants were included from the databases (Table 1). Unfortunately, presumably owing to the low cancer penetrance of inactivating PMS2 defects, too few clinically classified (likely) predisposing variants were available to generate a reference set for calibration purposes. To obtain sufficient functionally inactivating PMS2 substitution variants for assay calibration and validation we performed a genetic screen in mESc, as previously described for Msh2 and Msh6 (Drost et al., 2020; Drost, Lützen, et al., 2013; Figure 2). A total of 34 MMR-deficient mESC lines were obtained from the genetic screen, containing 31 unique Pms2-inactivating amino acid substitutions, to serve as proxy pathogenic/predisposing variants for calibration and validation (Supporting Information: Figure S1 and Table S2). Of these unique amino acid substitutions, two residues were not conserved between mice and humans and for one no Prior-P could be derived as two nucleotides were altered. The human equivalents of the remaining 28 Pms2-inactivating substitution variants were included, as proxies for predisposing PMS2 variants, in the calibration and validation set of the CIMRA assay (Figure 3, Table 1). Of note, variant p.Asn749Lys, identified in the genetic screen, was also listed as an unclassified human VUS in the ClinVar database (Accessed, October 15, 2021), which supports its role in cancer predisposition.

3.2 CIMRA assay calibration and validation

We used the CIMRA assay to determine the MMR activity of the set of 51 human PMS2 substitution variants, comprising the 20 class 1/2 variants and 31 (likely) predisposing or genetic screen-derived inactivating variants (Table 1). The CIMRA assay results demonstrated that all clinically classified (likely) benign protein variants had a near-wild type MMR activity (range: 78%–146% of wild type), indicating that the CIMRA assay is robust in detecting benign variants with no predicted splice defect. (Likely) predisposing and genetic screen-derived variants, derived from the databases and the genetic screen, all showed low repair of, on average, 8% of wild type (range: 2.6%–42%; Figure 4 and Table 1). Four variants from the genetic screen, despite having associated moderate to low in silico predictions of pathogenicity (Prior-P), indicative of non-predisposing, nevertheless displayed loss of MMR function when tested in the CIMRA assay (p. Arg169His, p. Ile831Ser, p. Met850Arg, and p. Met850Lys; Table 1 and Figure 5). The resistance of the corresponding cell lines to 6-TG combined with the biochemical MMR deficiency strongly suggests that these are indeed predisposing and indicate false-negativity of the in silico-derived Prior-P.

We then performed linear regression of the CIMRA assay repair levels from the 51 PMS2 variants against the given log-transformed OddsPath to generate a final calibration equation: log₁₀(OddsPath) = −0.0237 * CIMRA assay activity + 1.4127. Statistical metrics derived during cross-validation showed a good performance of the regression model (R² = 0.95 and root mean squared error = 0.29).

For OddsPath estimation for each individual variant, we used a leave-one-out procedure to permit probability estimation. For all variants previously classified as (likely) pathogenic/predisposing by clinical criteria or as inactivating and therefore (proxy) pathogenic/predisposing by the genetic screen, the average calculated OddsPath was 17 (range: 2.5−23), overall considered as supporting to strong strength towards pathogenicity for application under ACMG/AMP code PS3 (Supporting Information: Table S3; https://www.insight-group.org/criteria) (Tavtigian et al., 2018). For all clinically classified (likely) benign variants, the average calculated OddsPath was 0.10 (range: 0.01–0.38), overall considered to be supporting to strong evidence against pathogenicity for application under ACMG/AMP code BS3 (Figure 5, Supporting Information: Table S3). Thus, the CIMRA assay-derived OddsPath displays a bimodal distribution, in support of its high predictive value for predisposing and for benign variants. The OddsPath for individual variants, and relevant ACMG/AMP codes are shown in Table 1.

As summarized in Table 2, none of the “benign” reference set variants were incorrectly assigned a high OddsPath (in favor of pathogenicity), and none of the “predisposing” reference set variants were incorrectly assigned a low OddsPath.