Sufficient tumor tissue for MMR status assessment was available for 508 patients. Apart from higher FIGO stage, the clinicopathological characteristics of the study cohort represent the full population-based cohort (n=865) (Table 1). Median follow-up time was 62 months. Maximum magnetic resonance imaging (MRI)-assessed tumor diameter was available for 243 patients by reassessment of pelvic MRIs performed as part of the primary workup.14

Clinicopathological characteristics of the study cohorts

MMR-D was initially detected in 8 of 508 cervical tumors on tissue microarrays (1.6%). To reduce the risk of falsely classifying the tissue microarray tumor sections as MMR-D, full sections of these tumors were re-stained resulting in three tumors being reclassified as MMR-P. Within the remaining five MMR-D tumors, all had loss of PMS-2, and three had combined loss of MLH-1 and PMS-2 (Figure 1, online supplemental table 4). No significant association between MMR-D and disease-specific survival was found (p=0.17; online supplemental figure 1A). Rescoring using <10% as cut-off for MMR-D was also not prognostic (online supplemental figure 1B). The prevalence of MMR-D was highest within the rare and very aggressive neuroendocrine carcinomas. Three of ten (30%) neuroendocrine carcinomas were MMR-D, representing 60% of all MMR-D cases (online supplemental table 5). Proficient and deficient neuroendocrine carcinomas had similar 5-year disease-specific survival (online supplemental figure 2).

Full section staining of PMS2 and MLH1 validates mismatch repair deficiency (MMR-D) in five patients. Representative images of full sections stained for MMR proteins with negative nuclear staining of tumor cells and positive staining of stroma or immune cells. Although some tumor cells exhibited cytoplasmic staining, no nuclear staining was found. This was confirmed by an expert pathologist. *Tissue block missing, photographs of tissue microarray (TMA).

To evaluate the prognostic value of individual MMR proteins, differential expression of MSH-6, MSH-2, PMS-2, and MLH-1 was scored from tissue microarrays (n=508) using the staining index method (Figure 2A). Weighted kappa scores were calculated from independently scoring (MCB and MKH) of a subset of cases demonstrating overall good concordance: k=0.826 for MLH-1 (n=65), k=0.839 for MSH-2 (n=63), k=0.771 for MSH-6 (n=66), and k=0.703 for PMS-2 (n=60).

Differential expression of MSH-2, but not MSH-6, MLH-1, and PMS-2, associates with disease-specific survival in cervical cancer. (A): Tumors were defined as ‘low’ or ‘high’ for each of the mismatch repair (MMR) proteins. Low and high MMR protein expression was defined based on staining index (SI) 0–9. (B–E) Disease-specific survival relative to MSH-2 (B), MSH-6 (C), MLH-1 (D), and PMS-2 (E) protein levels. Best cut-off for predicting disease-specific survival (Youden index) was applied (MSH-2: ‘low’ SI 0–4 and ‘high’ SI 6–9, MLH-1: ‘low’ SI 0–6 and ‘high’ SI 9, MSH-6: ‘low’ SI 0–6 and ‘high’ SI 9, PMS-2: ‘low’ SI 0–3 and ‘high’ SI 4–9). P-values are given by log-rank (Mantel–Cox) test. Numbers in parentheses indicate total number of patients/events.

High and low expression were defined from the Youden index as described in the Methods section. MSH-2 low (staining index 0–4) associated with poor disease-specific survival (p=0.018) (Figure 2B), whereas differential expression of MSH-6, MLH-1, or PMS-2 did not associate with survival (all p>0.05) (Figure 2C–E). MSH-2 protein level did not associate with any clinicopathological variables except for p53 status (online supplemental table 6). However, MSH-2 low independently predicted poor disease-specific survival after adjusting for age, FIGO stage, and histological type (HR 1.77, 95% CI 1.00 to 3.17, p=0.049) (Table 2). To account for possible interactions between age, FIGO stage, and histological type, interaction terms age*FIGO stage, FIGO stage*histological type, and histological type*age were explored. None influenced the effect. To evaluate the reproducibility and applicability of low MSH-2 tumor expression as a potential biomarker in future clinical settings, interobserver agreement for scoring MSH-2 low versus MSH-2 high was analyzed. Kappa agreement for categorizing MSH-2 in high or low was almost perfect (k=0.924).

Multivariate survival analysis of patients (n=457*) with high MSH-2 versus low MSH-2 staining index according to Cox’s proportional hazard regression method.

RNA sequencing and whole exome sequencing data were available for 72 and 75 patients, respectively. MSH-2 mRNA levels were significantly correlated with MSH-2 protein expression in overlapping samples (p=0.046, n=68) (Figure 3A). MSH-2 low tumors (n=9) had a higher mutational burden compared with MSH-2 high tumors (p=0.003, n=74) (Figure 3B). In gene set enrichment analyses, gene sets related to immune activation were enriched in MSH-2 low tumors. Among the 20 top-ranked gene sets in MSH-2 low tumors, 90% in the Ontology gene sets (C5) and 70% of Hallmark gene sets were related to immune response (GO ‘adaptive immune response’, ‘T-cell activation’ and Hallmark ‘interferon gamma response’, ‘inflammatory response’ (Figure 3C, online supplemental table 7). In MSH-2 high tumors, none of the enriched gene sets were related to immune response.

Transcriptomic and genomic characterization reveal high mutational burden and immune cell signaling in MSH-2 low tumors. (A/B) Immunohistochemical protein expression (MSH-2: SI 0–4 and SI 6–9) in relation to corresponding MSH-2 mRNA expression (A) and mutational load (B). (C) Gene set enrichment analysis showing distribution of top 20 ranked enriched gene sets within the C5 and Hallmark gene set collections (MSigDB) for tumors with low MSH-2 expression. Other Hallmark gene sets: ‘coagulation’, ‘KRAS signaling’, ‘p53 pathway’. Other C5 gene sets: ‘cornification’. (D) Top eight differentially mutated genes of MSH-2 low tumors compared with MSH-2 high tumors. Multiple mutations per case per gene is indicated as ‘multi-hit’. Numbers indicate frequencies of patients with mutations and percentages are indicated on the bar below. (E) RAD50 co-lollipop plot illustrating type and location of mutations. Three of four RAD50 mutations in MSH-2 low tumors were frameshift mutations at location p.Q975Lfs*6. In MSH-2 high all four missense mutations were detected within the same patient tumor. P-values are given by Mann–Whitney U test. SI, staining index.

Mutational analyses grouped for MSH-2 low (n=7) versus MSH-2 high (n=66) revealed a significantly higher mutational frequency in MSH-2 low tumors (p<0.001) (Figure 3C). Eight genes (KRT2, TRBV7-7, IGKV1-16, ELMO2, PHC2, METTL2B, ALDH2, and RAD50) had significantly higher mutation frequency in MSH-2 low tumors (p>0.01) (Figure 3D). RAD50 mutations were previously correlated to survival in other cancer types23 24 and were therefore further explored, revealing a recurrent (n=4) frameshift insertion mutation in MSH-2 low tumors (Figure 3E).