Functional analysis of rare variants in mismatch repair proteins augments results from computation-based predictive methods

ABSTRACT

The cancer-predisposing Lynch Syndrome (LS) arises from germline mutations in DNA mismatch repair (MMR) genes, predominantly MLH1, MSH2, MSH6, and PMS2. A major challenge for clinical diagnosis of LS is the frequent identification of variants of uncertain significance (VUS) in these genes, as it is often difficult to determine variant pathogenicity, particularly for missense variants. Generic programs such as SIFT and PolyPhen-2, and MMR gene-specific programs such as PON-MMR and MAPP-MMR, are often used to predict deleterious or neutral effects of VUS in MMR genes. We evaluated the performance of multiple predictive programs in the context of functional biologic data for 15 VUS in MLH1, MSH2, and PMS2. Using cell line models, we characterized VUS predicted to range from neutral to pathogenic on mRNA and protein expression, basal cellular viability, viability following treatment with a panel of DNA-damaging agents, and functionality in DNA damage response (DDR) signaling, benchmarking to wild-type MMR proteins. Our results suggest that the MMR gene-specific classifiers do not always align with the experimental phenotypes related to DDR. Our study highlights the importance of complementary experimental and computational assessment to develop future predictors for the assessment of VUS.

KEYWORDS:

• Colon cancer
• colorectal cancer predisposition
• mismatch repair
• rare variant
• variants of uncertain significance

Abbreviations
ATM	=	Ataxia telangiectasia mutated
ATR	=	ATM- and Rad3-related
CDDP	=	cisplatin; also known as cis-Diammineplatinum(II) dichloride
DDR	=	DNA damage response
HCPS	=	hereditary cancer predisposing syndrome
LS	=	Lynch syndrome
MMR	=	mismatch repair
MMS	=	methyl methane sulfonate
VUS	=	variants of uncertain significance
wt	=	wild-type
5-FU	=	5-Flourouracil

Introduction

Inherited mutations in DNA mismatch repair (MMR) genes result in autosomal-dominant inheritance of hereditary nonpolyposis colorectal cancer (HNPCC) or Lynch Syndrome (LS). These mutations also strongly predispose individuals to endometrial cancer (EDC), colorectal cancer (CRC), and other tumors, forming a spectrum of hereditary cancer-predisposing syndromes (HCPS). These predispositions are largely due to mutations in 4 MMR genes, MSH2, MSH6, MLH1, and PMS2.^1-3 MMR is a post-replicative DNA repair mechanism in which 2 protein complexes, MutSα (formed of a heterodimer of MSH2 and MSH6) and MutSβ (a heterodimer of MSH2 and MSH3) wrap around replicated DNA, scan the sequence, and recognize errors in DNA replication. MutSα recognizes mismatched bases and small loops, while MutSβ recognizes both small and large loops (∼10 nucleotides), associated with insertions and deletions.⁴ When a MutS complex (either α or β) recognizes a nucleic aberration(s), it recruits an MLH1-PMS2 heterodimer (MutLα), a weak endonuclease that introduces nicks in the damaged DNA.^5,6 Subsequent recruitment of exonuclease 1 (Exo1) excises mismatched single-stranded DNA from the newly synthesized strand. Finally, the DNA replication machinery resynthesizes the DNA and fills the gap.⁴ Importantly, MMR proteins signal to induce cell cycle arrest, providing time to complete DNA repairs, or alternatively signal to induce apoptosis in the context of DNA damage too extensive to be repaired.⁷

Testing for mutations in MMR genes is routinely considered in clinic in suspected cases of LS. The current National Comprehensive Cancer Network (NCCN) guidelines for gastrointestinal cancers indicate that a pathogenic MSH2 or MLH1 mutation confers a 52–82% lifetime risk of CRC 52–82% and 25–60% for EDC. Pathogenic MSH6 mutations result in lifetime risk of 10–22% for CRC and 16–26% for EDC. Pathogenic PMS2 mutations are associated with risks of 15–20% for CRC and 15% for EDC.⁸ Patients with such mutations receive guidance and preventive care, and are advised to undergo intensive screening to consider prophylactic surgery. Complicating diagnosis and genetic counseling, ∼20–30% of MMR variants identified in high-throughput genomic studies, or in panel testing of individuals without access to family pedigree analysis, are rare variants of uncertain significance (VUS). Most of these cause single amino acid changes which may or may not impact protein function, and which have not yet been functionally linked to cancer risk.⁹ Understanding the significance of MMR variants would not only enhance risk assessment for families and individuals, but may also impact treatment decisions.¹⁰ Recent studies have shown that MMR-defective tumors are particularly susceptible to immunotherapy drugs targeting the programmed death 1 (PD-1) (Le, et al., 2015) and its ligand (PD-L1).¹¹

 Current methods for assessment of VUS rely on epidemiological and pathological data when available, but otherwise rely on a variety of computational tools.¹² Generic mutation phenotype prediction servers, such as SIFT,¹³ PolyPhen-2,¹⁴ and MutationAssessor,¹⁵ are often used (reviewed in ¹⁶). Such generic predictors make phenotypic predictions for missense variants in any protein. Their accuracies on independent benchmarks depend strongly on the nature of the testing set and its overlap with the training set.¹⁷ In an effort to increase the accuracy of phenotypic prediction for potential LS patients, several specific structural predictors have been developed for MMR proteins. MAPP-MMR¹⁸ makes predictions for MLH1 and MSH2, CoDP¹⁹ makes predictions for MSH6, and PON-MMR2²⁰ makes predictions for MLH1, PMS2, MSH2, and MSH6. These computational tools are primarily based on analyses of sequence conservation and the severity of the substitution based on the range of variation at the assessed position.¹²

Importantly, if a VUS is predicted to have a moderate or minimal effect on protein structure, but disrupts critical protein interactions that influence specific functions, the resulting biologic functionality or disease penetrance may be complex. For example, a specific VUS might impede apoptotic signaling, while leaving MMR fully functional.^21,22 However, in contrast to the common use of a growing number of computational prediction methods, direct experimental approaches to assess protein function are rarely used to explore and validate predictions, although some recent studies have suggested the value of this approach.^23,24 In this study, we have integrated phenotype predictions and functional biologic assessment for 15 missense VUS in 3 MMR genes (MSH2, MLH1, and PMS2). As part of this work, we have analyzed our results in the context of predictions made by 8 commonly used generic and specific computational predictors. Based on this work, we propose that a panel of cell-based in vitro methods can contribute to the assessment of the impact of VUS on DDR functionality.

Materials & methods (See also Supplementary Methods)

Computational analysis of VUS, protein, and mRNA structure

MMR variants were analyzed by using multiple computational tools for assessment of SNP impact on protein function. We used several generic predictors such as: PolyPhen-2.1¹⁴ http://genetics.bwh.harvard.edu/pph2/, SIFT¹³ http://sift.jcvi.org/, PROVEAN²⁵ http://provean.jcvi.org/index.php, MutationAssessor¹⁵ http://mutationassessor.org/r3/ (see supplementary methods for extended predictor and computational methods). Additionally, we analyzed each variant with the Biological Assembly Support Vector Machine (BA-SVM),²⁶ a generic phenotype predictor developed by our group which incorporates sequence conservation, structural disorder, physiochemical properties, and interface data from the structures of oligomeric protein complexes (biologic assemblies) into a support vector machine. We also used 2 MMR-specific predictors: MAPP-MMR¹⁸ (http://mappmmr.blueankh.com) and PON-MMR2 ²⁰ (http://structure.bmc.lu.se/PON-MMR2). All the predictive tools provide a score from which the prediction on the protein variant can be categorized as benign/neutral, pathogenic/disease-causing, or intermediate/uncertain. For each variant, we also calculated the DDG of the mutation using FoldX²⁷ to evaluate protein stability. Variants were also submitted to the Human Splicing Finder (HSF),²⁸ SNPfold,²⁹ and Mfold³⁰ servers to evaluate potential alterations in mRNA splicing or structure.

Molecular modeling

A complete 3-dimensional model of the MutSα heterodimer was constructed from the crystal structure of the human MSH2/MSH6 heterodimer (PDB: 2O8B³¹) by building loops with the program YASARA³² for several loop regions missing coordinates in the experimental structure. The MutLα C-terminal heterodimer model was created with HHpred,³³ using the human C-terminal MLH1 crystal structure (PDB: 3RBN (Dombrovsky et al., unpublished, DOI: 10.2210/pdb3rbn/pdb) and S. cerevisiae MutLα C-terminal heterodimer crystal structure (PDB 4FMO³⁴) as templates. Energy minimization was performed with YASARA, and side chains were rebuilt with SCWRL4.³⁵ The N-terminal heterodimer was constructed by superposing the human N-terminal domain crystal structures of PMS2 (PDB: 1EA6³⁶) and MLH1 (PDB 4P7A,³⁷ using the MLH1 homodimer structure (PDB: 4P7A) as a template.³⁷

Results

MMR variants and computational predictions of pathogenicity

Fig. 1A outlines the activity of the MutSα/β and MutLα complexes during MMR repair. The MMR variants in this study (Table 1) were selected based on the following criteria: 1) Identified in patients with LS and/or HCPS; 2) Rare, missense germline variants; 3) Clinically classified as VUS; 4) Not previously analyzed. Five VUS each in MSH2, MLH1, and PMS2 (Fig. 1B,C) that met the above criteria were analyzed using 8 generic and 2 MMR protein-specific predictors (Tables 2 and S2). The scores from each predictor were translated into a 3-tiered classification system (Neutral (N), Moderate (M), Deleterious (D)) to facilitate comparison between predictor programs (Table S1). Results in Tables 2 and S2 highlight the variability in the results of these computational programs. While there is some consensus for 6 variants (PMS2: V475E, I853M; MLH1: R423T, H264L, V213L; MSH2: R534P, for the remaining 9 VUS there was significant disagreement among the predictions. We also incorporated analysis of codon usage bias in the initial assessment, which exhibited differences that correlated with the observed phenotype of some of the variants as discussed below (summarized in Table S2).

Figure 1. (A) Basic schematic of the MMR pathway in replication-coupled repair. During DNA replication, a misincorporation of nucleotide(s) results in the creation of mismatched base(s) in the daughter strand of the DNA. If DNA polymerase fails to recognize and correct the misincorporated base pair(s), the post -replicative MMR pathway is activated. A MutS heterodimer, incorporating MSH2 with MSH6 (MutSα) or MSH3 (MutSβ), recognizes the mismatch, and recruits MutLα (a MLH1 heterodimer PMS2) to nick the daughter strand at either the 3′- or -5′ end, depending on the location of the mismatch. Nicks in DNA are extended by Exo1, after which the DNA replication machinery fills in the correct base(s) and DNA ligase connects the new sections of the DNA. (B) Mapping of variants/mutants on MMR protein structures. Panels: A. Cartoon representation of MutSα. Purple: MSH2. Orange: MSH6. Magenta spheres: mutation sites. Cyan: dsDNA. Gray sphere and blue sticks: Mg²⁺ and ADP. (B) Cartoon representation of MutLα, with N-terminal domain on top and C-terminal domain below. Unstructured linker arms are represented as dashed lines and are not to scale. Green: MLH1. Cyan: PMS2. Magenta spheres: mutation sites. Cyan: dsDNA. Gray sphere and orange sticks in N-terminal domain: Mg²⁺ and ADP. Orange sticks on C-terminal domain: EXO1 fragment. C-G. Zoomed-in MSH2 mutations. H-K. Zoomed-in MLH1 mutations. L-O. Zoomed-in PMS2 mutations.

Table 1. Variant information, clinical classification and ExAC summary. Key: AA, amino acid; CMMR-D, constitutional MMR deficiency; CRC, colorectal cancer; HCPS, Hereditary Cancer Predisposing Syndrome; Hx, History; LS, Lynch Syndrome; MMR, Mismatch Repair; MSS, Microsatellite Stable; NF, Not Found; Pt, Patient. ExAC estimates were based on May 2016 data release (http://exac.broadinstitute.org/about).

GENE	AA	DNA	AA (as used in text)	DNA (as used in text)	Domain	Clinical condition(ClinVar)	ExAC (allele count)	ExAC (allele freq.)	ClinVar additional data
MSH2	p. Pro5Gln	NM_000251.2: c.^14C>A	P5Q	^14C>A	MutS domain I /DNA binding	LS & HCPS	3	0.0000704(< 1/10,000)
MSH2	p. Arg55Gly	NM_000251.2:c.163C>G	R55G	163C>G	MutS domain I /DNA binding	HCPS	2	0.00002445(< 1/10,000)
MSH2	p. Arg534His	NM_000251.2:c.1601G>A	R534P	1601G>C	MutS domain III /DNA repair, recombination, MSH3/MSH6 interaction	HCPS		NF	R534H:One) Pt CRC; tumor MSS and normal IHC (ClinVar).
									Two) Patient with suggestive LS hx.
									PMID: 24728327
MSH2	p.Met813Ile	NM_000251.2:c.2439G>A	M813I	2439G>A	MutS domain V /ATPase, MuTL homolog interaction	LS & HCPS	1	0.000008241(Singleton)
MSH2	p.His839Arg	NM_000251.2:c 2516A>G	H839R	2516A>G	MutS domain V /ATPase, MuTL homolog interaction	HCPS	Two	0.00001648(< 1/10,000)	2 families with CRC (PMID: 26053027), CRC family- Pt also had variant in MLH1 (c.793C>T) (PMID: 19419416). CRC family- Pt had variant, but inconclusive segregation data (PMID: 15613555)
MLH1	p.Thr116Arg	NM_000249.3:c.347C>G	T116R	347C>G	Nudix domain/ Nucleoside hydrolase	LS		NF
MLH1	p.Val213Leu	NM_000249.3:c.637G>T	V213L	637G>T	MuTL Transducer domain	LS		NF
MLH1	p.His264Leu	NM_000249.3:c.791A>T	H264L	791A>T	MuTL Transducer domain	HCPS		NF
MLH1	p.Arg423Thr	NM_000249.3:c.1268G>C	R423T	1268G>C	Exonuclease Interaction	NF		NF	NF
MLH1	p.Arg725His	NM_000249.3:c.2174G>A	R725H	2174G>A	MutL C-terminal dimerization / Endonuclease domain	LS & HCPS	18	0.0001484(< 1/10,000–0.001)	Pt with CRC and another Pt with Prostate cancer and polyps (ClinVar).
PMS2	p.Ser36Arg	NM_000535.6:c.106A>C	S36R	106A>C	N-terminal/ ATPase and DNA binding	HCPS		NF
PMS2	p.Val475Glu	NM_000535.6:c.1424T>A	V475E	1424T>A	Disordered region	HCPS		NF
PMS2	p.Asp699His	NM_000535.6:c.2095G>C	D699H	2095G>C	MutL C-terminal dimerization / Endonuclease domain	HCPS		NF	Homozygous variant in CMMR-D (PMID:23729388)
									Pt tumor- loss of PMS2 protein (PMID:20205264)
PMS2	p.Asp792Asn	NM_000535.6:c.2374G>A	D792N	2374G>A	MutL C-terminal dimerization / Endonuclease domain	HCPS		NF
PMS2	p.Ile853Met	NM_000535.6:c.2559C>G	I853M	2559C>G	MutL C-terminal dimerization / Endonuclease domain	LS & HCPS		NF

Key: AA, amino acid; CMMR-D, constitutional MMR deficiency; CRC, colorectal cancer; HCPS, Hereditary Cancer Predisposing Syndrome; Hx, History; LS, Lynch Syndrome; MMR, Mismatch Repair; MSS, Microsatellite Stable; NF, Not Found; Pt, Patient

Table 2. Summary of generic and gene-specific computational predictors. Key: AA, mutant amino acid; PP2, PolyPhen-2 score; MUTA, MutationAssessor score; BA, Biological Assembly SVM score; PROV, Provean score; N, neutral; M, moderate; D, damaging; n/a, not available. These predictive tools provide a variety of scores from which protein variants can be categorized as benign/neutral, pathogenic/disease causing, or intermediate /uncertain. For comparative purposes, raw prediction scores were translated into 3 classes (neutral, moderate, damaging) using the expanded criteria given in Table S2.

GENE	AA	DNA	SIFT	PP2	MUTA	BA	PROV	PON-MMR2	MAPP-MMR	D	M	N
MSH2	P5Q	^14C>A	N	M	N	D	M	N	—	1	2	3
MSH2	R55G	163C>G	D	M	M	D	D	N	M	3	3	1
MSH2	R534P	1601G>C	D	D	M	D	D	D	D	6	1	0
MSH2	M813I	2439G>A	N	N	M	D	N	D	D	3	1	3
MSH2	H839R	2516A>G	N	M	N	M	N	N	N	0	2	5
MLH1	T116R	347C>G	D	D	M	M	D	D	M	4	3	0
MLH1	V213L	637G>T	N	N	N	N	N	N	N	0	0	7
MLH1	H264L	791A>T	D	D	D	M	D	D	D	6	1	0
MLH1	R423T	1268G>C	N	N	N	M	N	N	—	0	1	5
MLH1	R725H	2174G>A	D	D	M	D	D	M	N	4	2	1
PMS2	S36R	106A>C	M	D	N	M	M	M	—	1	4	1
PMS2	V475E	1424T>A	N	N	N	N	N	N	—	0	0	6
PMS2	D699H	2095G>C	D	D	D	M	D	N	—	4	1	1
PMS2	D792N	2374G>A	M	M	M	M	D	N	—	1	4	1
PMS2	I853M	2559C>G	N	M	N	N	N	N	—	0	1	5

Key: AA = mutant amino acid; PP2 = PolyPhen-2 score; MUTA = MutationAssessor score; BA = Biological Assembly SVM score; PROV = Provean score; N = neutral; M = moderate; D = damaging; “–“ = not available. All these predictive tools provide a score from which the protein variant can be categorized as benign/neutral, pathogenic/disease causing, or intermediate /uncertain. For comparative purposes, raw prediction scores were translated into 3 classes (neutral, moderate, damaging) using the expanded criteria given in Table S2. The text color correspond to color coding in Table S1 and S2.

Specific VUS affect mRNA and protein expression of MMR genes

To establish consequences of analyzed VUS for mRNA and protein expression, providing a baseline for interpreting functional results, we transiently introduced plasmids expressing wild-type (wt) or VUS forms of MSH2, MLH1, or PMS2 into the highly transfectable HEK293 cell line (Fig 2 and Fig. S1).³⁸ For MSH2, expression of VUS R534P and M813I was equivalent to wt while H839R was variable. Unexpectedly, MSH2 variants P5Q and R55G were not detectable. For MLH1, variants T116R, V213L, H264L, and R423T, were expressed equivalently to wt, while R725H was reduced by ∼60%. For PMS2, all VUS (S36R, V475E, D699H, D792N, and I853M) were expressed equivalently to wt. For each overexpressed protein, we also examined expression of the binding partner in each heterodimer (based on MSH2-MSH6 and MLH1-PMS2 pairing, Fig 1B and C): this was not significantly affected for any VUS (Fig. 2).

Figure 2. Analysis of MMR variant protein expression in HEK293 cell lines. (A-B) represents expression of MSH2 variants (P5Q, R55G, R534P, M813I, H839R) (dark gray bars) with corresponding MSH6 expression (light gray bars). (C-D) represents expression of MLH1 variants ((T116R, V213L, H264L, R423T, R725H) (dark gray bars) with corresponding PMS2 expression (light gray bars). (E-F) represents expression of PMS2 variants ((S36R, V475E, D699H, D792N, I853M) (dark gray bars) with corresponding PMS2 expression (light gray bars). For all experiments, E, empty vector transfection and WT, wild type transfection. (A, C, E) Quantification from 3 independent experiments ± SEM. Statistics were performed using a one-sample T-test, where * represents p < 0.05, ** represents p < 0.01 and *** represents p < 0.001. Protein was analyzed 48 hr post transfection. The data are analyzed relative to wt expression as % expression ± SEM. (B, D, F) Top image represents MSH2/MLH1/PMS2 with GAPDH and bottom image represents corresponding MSH6/PMS2/MLH1 with GAPDH

We also assessed mRNA levels for each VUS (Fig. S1). Surprisingly, the mRNA expression of MSH2 VUS P5Q and R55G was barely detectable above the endogenous expression in HEK293 cells, while the level of MSH2 VUS H839R was reduced by ∼77% versus wt (Fig S1A), which was comparable to the observed protein expression differences for these variants (Fig 2A, B). For MLH1 and PMS2, wt and VUS mRNAs were expressed at equivalent levels with the exception of MLH1 T116R, which exhibited reduced (∼40%) RNA expression (Fig. S1B), even though no differences in steady-state protein expression were noted (Fig 2C, D).

Impact of MMR VUS on basal cellular viability

The expression of MMR proteins has been associated with the activation of cell cycle checkpoints and apoptosis, especially in response to certain DNA damaging agents.³⁹ Several studies have used in vitro biochemical methods for assessment of VUS.^40,41 A limitation of these methods is that they do not fully recapitulate the intracellular environment: they may minimize the impact of VUS that impact protein interactions with diverse partners, or alternatively, fail to capture compensating mechanisms that reduce the impact of some damaging VUS.³⁹ In vivo studies have shown that re-expression of MSH2 or MLH1 can induce apoptosis in both repair-proficient and repair-deficient cells, with loss of such pro-apoptotic activity a reliable indication of damaged MMR signaling function. Hence, several studies have used cell culture-based methods to address MMR functionality in regulating cellular viability and cell death under basal conditions, and during DNA damage response.^42-46

Using these approaches, we analyzed the consequences of VUS function in CRC cell line models in which no wt copy of each affected gene was present. We transfected plasmids expressing the MSH2 series VUS into LoVo cells,⁴⁷ the MLH1 series into HCT116 cells,⁴⁸ and the PMS2 series into Vaco481 cells⁴⁹ (Fig S2 and S3). In the absence of endogenous wt forms of each protein, the protein expression patterns for each VUS were generally similar to those observed in the HEK293 cells (Fig. 2), with the exception that MSH2 H839R and PMS2 I853M were now somewhat reduced compared with wt (∼50%, Fig S2A, B; ∼35%, Fig S2E, F). A complete summary of expression results is presented in Table 3.

Table 3. Summary of experimental and structural analysis. AA, amino acid change; +, proficient; -, deficient; +(-), reduced or between empty and wt; wt, as wt; +(-)*; response likely different with different DNA damaging agent.

PROTEIN	AA	RNA expression	Protein expression	Viability	DDR ATM/ATR	DDR gH2AX	PON-MMR2	MAPP-MMR	Variant functionality	Structural Comments
MSH2	P5Q	–	–	+	–	+(–)	Neutral	N/A	Defective	Alters RNA splicing
										Compromises DNA binding
MSH2	R55G	–	–	+	–	+(–)	Neutral	Moderate	Defective	Alters RNA splicing
										Helix destabilization
MSH2	R534P	+	+	+	–	+(–)	Pathogenic	Pathogenic	Defective	Compromises MutSα dimerization
MSH2	M813I	+	+	WT	WT	WT	Pathogenic	Pathogenic	Functional	Not expected to alter function
MSH2	H839R	+(–)	+(–)	+(–)	WT	WT	Neutral	Neutral	Moderately	Alters RNA structure and alters MutSα dimerization
									Functional
MLH1	T116R	+(–)	+	WT	+(–)*	WT	Pathogenic	Moderate	Moderately	Alters MutLα – MutSα dimerization
									Functional
MLH1	V213L	+	+	WT	WT	+(–)*	Neutral	Neutral	Moderately	Structural effects unclear
									Functional
MLH1	H264L	+(–)	+	WT	+(–)*	+(–)*	Pathogenic	Pathogenic	Moderately	Introduces a solvent-exposed hydrophobic group
									Functional
MLH1	R423T	+	+	WT	WT	+(–)*	Neutral	N/A	Moderately	On unstructured linker region
									Functional
MLH1	R725H	+	+(–)	WT	+(–)*	+(–)*	Moderate	Neutral	Moderately	May alter association on N- and C-terminal domains around DNA
									Functional
PMS2	S36R	+	+	WT	WT	WT	Moderate	N/A	Functional	Not expected to alter function
PMS2	V475E	+	+	WT	+(–)*	WT	Neutral	N/A	Moderately	On unstructured linker region
									Functional
PMS2	D699H	+	+	+(–)	WT	+(–)*	Neutral	N/A	Moderately	Alters helix capping and MutLα dimerization
									Functional
PMS2	D792N	+	+	+(–)	WT	+(–)*	Neutral	N/A	Moderately	Structural effects unclear
									Functional
PMS2	I853M	+	+(–)	WT	WT	WT	Neutral	N/A	Functional	Not expected to alter function

In functional testing, introduction of plasmids for wt MSH2 and its VUS M813I reduced basal LoVo cellular viability by ∼70–75%, while VUS H839R reduced basal viability ∼56%, (Fig 3A), which may reflect the partially reduced expression of this protein. In contrast, the well-expressed variant R534P did not reduce viability, similar to the empty vector control and poorly expressed variants P5Q and R55G. Based on measurement of caspase 3/7 cleavage, the observed reduction of viability approximated the induction of apoptosis by wt MSH2 and MSH2 variants M813I and H839R (Fig 3B). For MLH1, no significant differences were observed between wt and VUS, which each reduced HCT116 viability by ∼25% in comparison to empty vector (Fig. 3C), and each increased apoptosis by ∼40% (Fig 3D). Similarly, expression of PMS2 and the VUS series in each case significantly reduced viability of Vaco481 cells (Fig. 3E), with wt and VUS S36R, V475E, and I853M reducing viability slightly more than by ∼52–61%, but PMS2 VUS D669H and D792N. These results paralleled capacity to induce apoptosis, with wt and VUS S36R, V475E, and I853M exhibiting slightly higher caspase3/7 cleavage than D669H and D792N (Fig 3F).

Figure 3. Expression of wt and MMR variants impacts cellular viability. LoVo (A, B), HCT116 (C, D), or Vaco481 (E, F) CRC cells were transfected with Empty vector (black bars) or wt (red bars) or VUS (dark gray bars) for MSH2 (A, B), MLH1 (C, D), or PMS2 (E, F). Cell viability (A, C, and E) was determined 72 hours post-transfection by CellTiter-Blue (CTB) assay. The data are represented as % viability relative to E and are from 3 independent experiments ± SEM. The statistical comparisons are with wt. Statistical analysis was performed using a generalized linear model, where * represents p < 0.05. B, D, and F) Post transfection (72 h) cells were assayed for caspase-3/7 activity using the Apo-ONE Homogeneous Caspase 3/7-activity kit according the manufacturer's protocols. Measurements were taken using a Perkin Elmer Envision Plate Reader and plotted as % of wt (red bars) from 3 independent experiments ± SEM. Statistics were performed using a generalized linear model, where * represents p < 0.05, ** represents p < 0.01 and *** represents p < 0.001.

Impact of VUS on viability in response to DNA-damaging treatments

Loss of MMR leads to paradoxical resistance to chemotherapeutic/DNA-damaging agents such as platinum-based compounds (e.g. cisplatin (CDDP)), anti-metabolites such as 5-flourouracil (5-FU)), and methylating agents such as methyl methanesulfonate (MMS)),^{39, 50, 51} as MMR-deficient cells are able to tolerate moderate DNA damage, whereas MMR responses induces cytostasis and apoptosis. To specifically evaluate the impact of the VUS on response to DNA-damaging agents, we analyzed viability of CRC cell lines expressing wt or variant forms of MSH2, MLH1, or PMS2 72 hours after treatment with a concentration series of CDDP, 5-FU, or MMS (Fig. 4, Table 3 and extended data in Table S3 with red text highlighting significant dose response and significant dose response relative to wt). Platinum agents and 5-FU are used clinically in the treatment and management of CRC.⁵²

Figure 4. Impact of VUS on IC50 for DNA damaging drugs. LoVo (A-C), HCT116 (C-E), and Vaco481 (F-H) CRC cells were transfected with empty vector (E, black curves) or plasmid expressing wt (red curves) or VUS (gray curves) MSH2 (A-C), MLH1 (D-F), or PMS2 (G-I). 48 hours after transfection cells were treated with vehicle (V), methyl methanesulphonate (MMS), or cisplatin (CDDP) or 5-Flurouracil (5-FU) for 2 h at the concentrations indicated. Cell viability was measured by short-term CellTiter-Blue (CTB) assay (72 hours after drug addition). Data are represented as % viability and are from 3 independent experiments ± SEM. For extended p-values for log dose effect (slope) and log dose slope compared with wt see Table S3 (p-values hihglighted in red are significant).

For MSH2 in LoVo cells, the 2 poorly expressed VUS (P5Q and R55G) and VUS R534P had IC50 curves for all 3 drugs similar to cells transfected with empty vector (Fig 4A). Strikingly, VUS M813I and H839R showed more subtle alterations in patterns of response, referenced to wt MSH2. For example, while expression of wt MSH2 had the greatest impact in viability in the absence of drug, IC50 curves were relatively flat as concentration of 5-FU was increased, indicating limited dose responsiveness. However increasing concentrations of CDDP and MMS indicated significantly increased dose responsiveness of wt MSH2 (Fig 4A and Table S3). In contrast, VUS H839R, expressed at a lower but detectable level had a weaker effect on basal viability, but was associated with a proportionally greater loss of cell viability following treatment with these drugs, while VUS M813I had an intermediate effect compared with wt.

For MLH1/HCT116 and PMS2/Vaco481, cells transfected with empty vector had almost no reduction in viability following treatment with any of the drugs, whereas introduction of either wt or VUS forms of MLH1 or PMS2 resulted in a dose-dependent loss of viability. Surprisingly, even though the expression of MLH1 R725H was significantly reduced compared with wt (Fig. S2C and D), this difference did not impair drug response phenotype (Fig. 4B). Also surprisingly, several MLH1 VUS had a non-statistically significant trend toward greater drug-induced loss of viability than did wt (Fig 4B). For PMS2, I853M CDDP showed a significant dose response difference from wt, while VUS D699H and D792N had a near-significant difference from wt specifically following CDDP and MMS treatment (Fig 4C).

Expression of VUS impacts DDR signaling

The kinases ataxia telangiectasia mutated (ATM) and ATM- and Rad3-related (ATR) are the key regulators of DNA damage response (DDR) signaling,⁵³ activated by MMR-dependent detection of specific DNA lesions.⁵⁴ MSH2 directly recruits ATR to the sites of DNA damage,⁵⁴ while ATM activation may stabilize the MutL complex in response to DNA damage.⁵⁵ ATM-dependent phosphorylation of histone H2AX phosphorylation at Ser139 (gH2AX) in the presence of DNA damage marks chromatin for recruitment of DNA repair proteins.⁵⁶ Certain MSH2 missense mutations have been suggested to uncouple repair and DDR signaling functions,^21,22 making it of interest to assess how the VUS affect ATM/ATR signaling activation and other indicators of DDR signaling.

We used antibodies directed at canonical ATM/ATR substrate phosphorylation sites (phospho-S*/T*Q),^57,58 and monitored appearance of gH2AX foci (as in ²³) in cells overexpressing wt or VUS proteins, either under basal conditions or 16 hours after treatment with 15mM of CDDP or 5-FU (determined as responsive dose in Fig 4) (Fig 5, Table 3 and Table S4 (red text highlights significance)). Overall, VUS phenotypes paralleled those obtained with viability and apoptosis assays, with VUS capable of reducing viability and inducing apoptosis performing similarly to wt MMR proteins in inducing ATM/ATR phosphorylation and gH2AX induction under both basal and drug treatment conditions, with some interesting exceptions. Comparably increased ATM/ATR-dependent and gH2AX phosphorylation (Fig 5A, B) was induced by MSH2 wt, VUS M813I and H839R, but not by VUS P5Q, R55G and R534P. In contrast, several of the MLH1 VUS and all of the PMS2 VUS showed reduced induction of ATM/ATR-dependent and gH2AX phosphorylation compared with wt MLH1 and PMS2 (Fig. 5C–F).

Figure 5. Impact of VUS on wt DNA damage signaling responses. ImageExpress automated quantitation of signal for ATM/ATR substrate phosphorylation (A, C, E.) or gH2AX (B, D, F) in Lovo (A, B), HCT116 (C, D) or Vaco481 (E, F) CRC cells. Imaging is performed 16 hours after drug treatment with vehicle (V), with 15 mM cisplatin (CDDP) or 5-flourouracil (5-FU) of cells transfected with empty vector (E, black bars) or plasmids expressing wt (red bars) or VUS for indicated proteins. Numbers reflect mean positive fluorescence (for ATM/ATR phosphorylation) or mean foci count (for gH2AX). Data are represented as % positive mean relative to wt and are from 3/4 independent experiments ± SEM. Statistical analysis was performed using a generalized linear model, where where * represents p < 0.05, ** represents p < 0.01 and *** represents p < 0.001.

Structural and computational analysis of VUS

Based on all the data described above, several VUS reduce or eliminate mRNA expression (MSH2-P5Q, R55G, H839R; MLH1-T116R, H264L) and/or protein expression (MSH2-P5Q, R55G, H839R; MLH1-R725H; PMS2-I853M), while several VUS impair function but not expression (MSH2-R534P; MLH1-V213L, R423T, R725H; PMS2-V475E, D699H, D792N). The experimental results may be compared with the output of programs designed to predict changes in protein stability or function or association with disease .^13-15,18-20 For 2 mutations with consistent experimental functionality, MSH2-R534P with impaired function and PMS2-I853M with wt functionality, the computational predictions indicate agreement with the experimental results (6/7 deleterious (D) and 5/6 neutral (N) respectively; see Table 3). In 5 cases, there was strong consensus among the predictors: MSH2-H839R (5/7 N), MLH1-V213L (7/7 N), MLH1-H264L (6/7 D), MLH1-R423T (5/6 N), and PMS2-V475E (6/6 N), while the experimental results indicated some wt functionality and some impaired functionality (columns 4–6, Table 3). For the remaining 8 cases, the predictors produced mixed results. Of these, 2 mutations were completely defective due to lack of mRNA expression (MSH2-P5Q and MSH2-R55G), which the predictors are generally not designed to detect since they are based on protein sequence and structural information. Two mutations were similar to wild type in all 3 assays (MSH2-M183I and PMS2-S36R), while the remaining 4 were moderately function, having impaired function in one or 2 of the 3 assays (MLH1-T116R, MLH1-R725H, PMS2-D699H, PMS2-D792N)

To further guide interpretation of biologic data, we mapped each VUS onto the coordinates of the 3-dimensional structures of the MMR heterodimers to examine local interactions for each affected residue. We also calculated the DDG of mutation with the program FoldX ,²⁷ and examined potential alterations to transcription and translation using online prediction tools. Here we highlight the most relevant results for several variants (Fig 1B and C), with additional text and data for the entire test set noted in supplementary text and Table S1 respectively.

Neutral variants

Three variants were found to be experimentally functional in all 3 assays compared with wt: MSH2-M813I, PMS2-S36R, PMS2-I853M. The conservative, hydrophobic mutations MSH2-M813I and PMS2-I853M are both in heterodimeric interfaces (with MSH2 and MLH1 respectively) but molecular modeling suggested no significant steric conflicts upon mutation and preservation of hydrophobic contacts. PMS2 S36R is solvent exposed on the N-terminal domain of PMS2, easily accommodating a larger, polar side chain at the same site. The computational results are further detailed in supplementary text.

Interface-disrupting VUS

Several more drastic variants were located at the interfaces of biologic assemblies, and have the potential to disrupt dimerization events. Here we highlight VUS MSH2-R534P from Fig. 1B, while details on MSH2-H839R, MLH1-T116R, MLH1-R725H and MLH1-D699H are in supplementary text. Functional assays showed the greatest impairment of activity for a VUS affecting MSH2-R534, which is at the MutSα heterodimer interface. The R534 side chain is in close proximity to backbone carbonyls of R988 and N989 on MSH6, with N-O distances of 3.1 and 3.7 Angstroms, respectively. Additionally, R534 backbone amide hydrogen stabilizes a β turn of residues 532–535 through a hydrogen bond to the backbone carbonyl oxygen of residue 535. The R534P mutation would likely disrupt the β turn at the interface by eliminating the amide hydrogen (due to proline ring closure) and the backbone hydrogen bond. Additionally, any favorable intermolecular electrostatic interactions to residues 988 and 989 of MSH6 would be lost.

Ambiguous variants

The effects for some variants were more difficult to rationalize with a high level of confidence from a structural perspective. Both MLH1-R423 and PMS2-V475 are located in the disordered linker regions between the N- and C-terminal domains of MLH1 and PMS2, respectively (Fig. 1C). Although neither of these sites were included in the critical linker arm regions identified by Plys et al.⁵⁹ the biophysical consequences of mutations in this region are difficult to predict. Other ambiguous variants such as PMS2-D792N, MLH1 V213L and H264L are summarized in supplementary text.

RNA expression-deficient variants

Five variants showed partial or severe deficiencies in RNA expression: MSH2-P5Q (severe), MSH2-R55G (severe), MSH2-H839R (partial), MLH1-T116R (partial), and MLH1-H264L (partial). Alternative splicing leading to premature truncation of protein coding capacity can trigger nonsense-mediated decay,⁶⁰ providing one possible explanation for reduced steady-state mRNA levels. The Human Splicing Finder (HSF) server²⁸ suggests that mRNA splicing can potentially be altered through several mechanisms, including the creation of an exonic splicing silencer (ESS) site, creation of a new donor or acceptor site, or alteration of an exonic splicing enhancer (ESE) site. For the variants affecting RNA expression, HSF predicts the following, MSH2-P5Q (ESS), MSH2-R55G (ESS, donor), MLH1-T116R (ESS, ESE, and acceptor), and MLH1-H264L (ESS). Since we have transfected DNA sequences without introns, these changes are not necessarily relevant to the expression of RNAs unless they create new splice sites. Of the remaining 11 VUS, 7 of them also have predicted changes by the Human Splicing Finder (Table S2). For individuals carrying a VUS in an MMR gene, changes in RNA splicing may be a mechanism for downregulating expression of the gene product, even though this was not assessed in our experiments.

Alternatively, gross changes in mRNA secondary structure can also destabilize gene transcripts.⁶¹ SNPfold²⁹ predicted statistically significant structural alterations to RNA structure for MSH2-P5Q (p = 0.0226) and MLH1-T116R (p = 0.0115), which showed impaired RNA expression, and for MLH1-V213L (p = 0.0106) and PMS2-V475E (p = 0.0436) which did not. Mfold³⁰ was used to generate the predicted mRNA 2° structures of the reference and variant sequences (Fig. S4). Mfold predicts that the P5Q RNA structure gains a circular loop in the middle of the stem (Figs. S4A, B). The R55G RNA structure was predicted to form a circular loop instead of 2 short arms (Figs. S4C, D). H839R is predicted to eliminate a significant loop in the RNA structure. Such changes, if validated, could significantly affect mRNA stability.

Discussion

This study examined the relation between computational prediction of VUS severity and MMR functionality. We systematically assessed the role of 15 rare germline VUS in MSH2, MLH1 and PMS2 (Fig. 1B,C and Table 1), using multiple assays to address the consequences for RNA and protein expression, viability, apoptosis induction, and DNA damage signaling. Of the VUS analyzed, most were detected either in patients with features selective of LS or HCPS, or with definitive LS, while some of the VUS (MSH2-R534P, MSH2-H839R, MLH1-R725H, and PMS2-D699H) (Table 1) were also detected in patients with CRC. Major conclusions of the study were: 1) existing generic and gene-specific predictor programs failed to correctly classify many novel VUS, and care should be taken when using these methods in a clinical setting; 2) for at least 3 VUS, single residue changes were associated with significant changes at the steady-state mRNA level, which may reflect differences in mRNA secondary structure; 3) the use of multiple function-testing assays was useful to distinguish different categories of potential defect associated with individual VUS; and 4) mapping the variants onto protein structure provided additional insights suggesting the mechanistic basis for their effect, with many VUS localizing in important functional domains (Fig. 1B,C and Table 1).

Among the VUS assessed, most were expressed at normal protein levels, but several exhibited moderate impairment of activity in one or more assays, with a small number associated with more profound defects. Most of the MLH1 and PMS2 variants were classified as moderately functional (Table 3). Interestingly, VUS affecting these genes sometimes varied significantly in their consequences for viability and DDR signaling in response to cisplatin and 5-FU DNA damaging treatments (Fig. 5, Table 3, and Table S4). Three of 5 MSH2 variants assessed had moderate to severe defects in activity. For 2, P5Q and R55G, MSH2 bearing these variants were detectable neither at the mRNA nor the protein level (Fig. 2 and Figs. S1 and S2), which may reflect impact of the variants at the level of RNA structure, with variant-associated changes potentially influencing interaction of affected transcripts to RNA binding proteins that regulate stability (Fig. S4). In contrast, although MSH2 variant R534P was expressed at wt levels, this VUS was associated with complete loss of function in regulation of basal or drug treatment-associated cell viability or DDR signaling (Figs. 4 and 5).

Unclassified variants present a variety of unique challenges to the genetic counselor and the clinician as well as the patient and their family. Previous studies have advised application of standardized approaches to assess the pathogenicity of the VUS. Rasmussen et. al. suggested a 3-step diagnostic decision process,⁶² incorporating computational methods to predict effects on mRNA splicing, analysis of evolutionary sequence conservation, and assaying the activity of the VUS protein in vitro. They also suggested studying the biologic, biochemical and biophysical properties of the VUS to further aid in diagnosis. The International Society for Gastrointestinal Hereditary Tumors (InSiGHT) recently published a 5-tiered standardized scheme with methods to deal with MMR VUS.⁶³ They suggested that computational methods (MAPP-MMR and Polyphen-2.1) may exhibit high accuracy in variant prediction. By contrast, in our small test set of 15 rare or novel variants, all of the computational predictors performed rather poorly in assessing function of the genes in regulating viability, apoptosis, and DDR signaling. The best results came from BA-SVM,²⁶ which correctly classified 10 out of the 15 mutants, according to our 3-tiered system.

Gene-specific predictors are an alternative to generic predictors. However, for the 15 VUS in this study, the results were not promising. Using a 3-tiered classification system (Functional, Moderate/Intermediate, and Defective), MAPP-MMR correctly classified 2 out of 8 variants, but was not capable of making predictions on the 5 PMS2 variants and 2 of the MHS2/MLH1 variants. PON-MMR2 correctly classified 3 out of 15 variants, based on empirically determined functionality (summarized in Table 3). Additionally, these gene-specific methods have been trained on rather small variant sets (< 200), with few or no PMS2 variants. Moreover, the experimental data upon which these methods are trained may contain some errors, as a variant may affect a protein function that was not assayed. For example, a pathogenic MLH1 c.2041G>A (p. Ala681Thr) mutation exhibited MSI-high tumors but was functional in various in vitro assays of expression, MMR complementation and interaction with PMS2.⁶³ More refinement and training on larger data sets would be beneficial.

Both generic and gene-specific predictors are based on evolutionary conservation of protein sequences and amino acid properties. Most of them do not consider important biologically relevant details; such as RNA splice sites, codon usage bias, biologic assembly interfaces, post-translational modification sites, etc. Incorporating such data could improve results, as evidenced by the superior performance of the biologic assembly SVM predictor. Computational methods that predict protein dysfunction most often rely on amino acid substitution in relation to evolutionary conservation across species. In only a very few cases do these feature structural assessment of the protein.⁶² There is not only a need for protein specific computational predictors, but also programs that are trained on a large balanced set of (pathogenic and neutral) mutations per protein for machine learning and dysfunction prediction.^26,64 In case of PMS2 and MSH6 gene variants, limited variants are known as well as limited information is available on the disease spectrum.⁶⁵

Another possible confounder for structural prediction methods is that a given mutation may affect a specific function that is not standardly assayed. Most VUS studies focus on biologic assays in yeast and cell-free MMR assays which do not reflect the cellular environment.^39,62 Our alignment of VUS with structural features of MSH2, MLH1, and PMS2 (Fig 1B,C) indicated multiple domains were targeted, leading to suggestion of a wide range of potential mechanisms of action (summarized in Table 1 and 3). The approaches used here were specifically focused on defining in vivo effects of VUS, rather than providing a biochemical assessment of their mismatch repair capacity. While there are strengths to this approach, in simulating in vivo consequences, there are also potential caveats for surrogate assays based on viability and apoptosis. For example, it has been shown that MRE11, a component of the ATM-regulated double strand break signaling, is inactivated in several MMR-deficient cell lines, including the HCT116 and LoVo lines.^66,67 Hence, behavior of VUS may vary based on the genetic background of the models in the experimental study. While it is not possible to develop clinical tests for all possible activities, our data suggest that that MMR repair should not be the sole criterion upon which variant function should be judged. Rather, apoptotic signaling and therapeutic-dependent viability should be evaluated when considering how a novel MMR mutation will affect a patient.

Finally, our data imply that incorporating RNA and biologic assembly interface features could aid in the development of future predictive programs. Other studies have also emphasized the desirability of combining predictive analyses at the DNA, RNA (splicing and secondary structures) and protein level as well as incorporating additional features such as codon usage bias,^68,69 with one study suggesting that many VUS would be predicted to influence RNA splicing, stability or translation of encoded variant proteins.⁶² Although the incorporation of computational RNA analyses requires empirical confirmation, they provide workable hypotheses to explain experimental results, and potentially should be incorporated in next-generation prediction models.

Disclosure of potential conflicts of interest

No potential conflicts of interest were disclosed.

Author contributions

S.A. designed, coordinated, performed most studies and writing, P.J.H.: writing, structural modeling and discussion; R.S., M.S., SD, A.B., R.L.: mutagenesis, expression, and in vitro analyses, E.N.: RNA analysis, sequencing, M.J.H., R.D. & E.A.G.: discussion, designing, planning studies and writing.

Acknowledgments

The Fox Chase Cancer Center (FCCC) Cell Culture, Genomics, DNA Sequencing, Molecular Modeling, Biostatistics and High Throughput Facilities provided resources and expertise. The authors, facilities, and this work were supported by NCI Core Grant P30 CA06927 (CURE supplement, FCCC), Temple Genomic Research Funding, R01 DK108195 (to E.A.G.), T32 CA09035 (to S.A. and P.J.H.), R01 GM084453 (to R.L.D), and FCCC Summer Undergraduate Research Fellowships (to M.S. and R.L.). We thank Dr. Mark D. Andrake (FCCC, Molecular Modeling Facility) for assistance with modeling, and Dr. Sandy Markowitz (Case Western Reserve University) for providing Vaco481 CRC cells. All authors declare that there is no conflict of interest regarding this manuscript.