Comparison of SP142 and 22C3 PD-L1 assays in a population-based cohort of triple-negative breast cancer patients in the context of their clinically established scoring algorithms

The origin of our TNBC cohort has been previously described by Staaf et al. [30]. Briefly, a total of 408 TNBC patients were identified in Region Skåne between 2010/09 and 2015/03 by the Swedish National Breast Cancer Quality (NKBC) registry. Of those, 340 were enrolled in the Swedish Cancerome Analysis Network - Breast (SCAN-B) study (ClinicalTrials.gov ID NCT02306096), which is a population-based study in the southern health care region of Sweden and all patients with primary breast cancer are eligible (https://​www.​scan-b.​lu.​se/​) [31]. Eighty-four patients were thereafter excluded because of unclear TNBC status or insufficient tissue material. Of the 256 remaining patients included in our tissue microarray (TMA), 13 were excluded due to metastatic disease at diagnosis or prior to start of adjuvant chemotherapy (n = 8), bilateral breast cancer (n = 3), loss to follow-up before treatment start (n = 1), or non-TNBC status (n = 1). Clinicopathological characteristics and follow-up data was collected through clinical chart review and the last date of counting in events was 18th Oct 2019. Additional 11 patients were excluded since they only had TMA cores from residual disease after neoadjuvant chemotherapy (n = 6) or due to unevaluable TMA cores for the 22C3 staining (n = 5). Of the remaining 232 patients, who all underwent primary surgery (mastectomy or partial mastectomy), 166 received chemotherapy (CT-cohort) according to national guidelines, of which 155 received adjuvant and 11 neoadjuvant CT. Of these, 98.2% (163 of 166) received FEC or EC (5-fluorouracil, epirubicin, cyclophosphamide) based treatment with or without a taxane and three patients (1.8%) received less than 50% of planned CT. The remaining 66 patients did not receive any neo(adjuvant) CT, most often due to age or comorbidities (non-CT-cohort). Checkpoint inhibitors were not given to the patients in the cohort. Adjuvant radiotherapy was given according to national guidelines. All of the 166 CT patients were eligible for overall survival (OS) analysis, 165 for invasive disease-free survival (IDFS) and 163 for distant relapse-free interval (DRFI). In the non-CT-cohort, 64 were eligible for OS, 65 for IDFS and 63 for DRFI (Fig. 1, study flowchart). Clinicopathological characteristics in the CT-cohort (prior to eventual (neo)adjuvant CT) and the non-CT-cohort are presented in Table 1. RNA sequencing data for gene expression profiling (GEX) was available for 84% of the patients (194 out of 232 patients) through the SCAN-B consortium [31].

Scoring of PD-L1 expression by immunohistochemical testing was assessed in formalin-fixed, paraffin-embedded tumor samples in a TMA, using two different PD-L1 antibody clones: SP142 with Ventana BenchMark Ultra platform (Ventana Medical Systems, Inc., AZ, U.S) and 22C3 with Dako Autostainer Link 48 platform (Agilent, Inc., CA, U.S) IHC assays. Preparation and staining were done according to the manufacturer´s instructions. The TMA images were assessed in PathXL Philips Xplore (Koninklijke Philips N.V., NL). Each sample was represented by two TMA cores, each of 1.0 mm in diameter. PD-L1 in the adjuvant treated patients and the non-CT-cohort was evaluated on TMA cores from the surgical specimen. For the neoadjuvant patients, PD-L1 was evaluated on TMA cores from core needle biopsies taken prior to neoadjuvant treatment.

We evaluated PD-L1 staining according to two scoring methods: CPS and staining in ICs. CPS was defined as the combined number of PD-L1 stained TCs, tumor infiltrating lymphocytes (TILs) and macrophages (intratumorally and in adjacent stroma) divided by the total number of TCs, multiplied by 100. We evaluated CPS at a threshold of 1 and 10 according to PD-L1 evaluation in clinical phase III TNBC studies with pembrolizumab and the 22C3 assay [7, 8, 10]. The IC+ score was defined as percentage of the tumor area (non-necrotic, non-sclerotic area) covered by PD-L1 stained tumor infiltrating ICs and evaluated at a threshold of 1% as performed in phase III TNBC trials with atezolizumab and the SP142 assay [6, 9, 11, 13]. The score from the TMA core with highest value was set as the respective CPS and IC+ score for the tumor. PD-L1 expression in TCs (in CPS) included partial or complete membranous staining and in ICs (in CPS and IC+) membranous and/or cytoplasmic staining. Scoring of SP142 PD-L1 expression was done by a physician and a board-certified breast cancer pathologist where consensus in non-matching scoring had to be reached for 4,7% of the tumors (Additional file 1: Table S1). The 22C3 scoring was performed by a physician and in cases that were not clearly obvious, a board-certified breast cancer pathologist was consulted and consensus reached. IHC staining examples are illustrated in Fig. 2A. We scored CPS using the 22C3 assay and IC+ with both SP142 and 22C3 (note it is experimental scoring of IC+ with 22C3 since the IC+ scoring is not clinically applied for 22C3). We did not evaluate CPS for SP142 since it has been shown to have impaired sensitivity for PD-L1 staining in TCs in TNBC [20, 21]. When investigating the concordance between the assays, the SP142 IC+ of ≥ 1% and 22C3 CPS of ≥ 10 scores were compared as they are the only clinically established predictive cut-offs in TNBC. Moreover, since 22C3 CPS ≥ 1 also has been investigated in clinical trials, the concordance between SP142 IC+  ≥ 1% and 22C3 CPS ≥ 1 was evaluated. In addition to this, to compare under more similar, but explorative, scoring conditions, the concordance between SP142 IC+ and 22C3 IC+ was evaluated.

Abundance of stromal TILs was evaluated by a board-certified breast cancer pathologist on hematoxylin–eosin stained whole slides from surgical specimen before eventual adjuvant chemotherapy and from pre-treatment core needle biopsies for the neoadjuvant treated patients. Abundance was calculated as percentage of TILs occupying the tumoral stromal area according to the international TILs working group (https://​www.​tilsinbreastcanc​er.​org/​) [32]. If more than one slide was available per patient, the average score was applied. Threshold for high versus low TILs (as binary variable) was set to 30% as performed in a previous pooled analysis of the prognostic value of TILs in early-stage TNBC patients [33], which also was near the mean value of TIL abundance in our cohort (27% in the overall cohort, 29% in the CT-cohort).

In the survival analyses, OS, IDFS and DRFI were defined as endpoints with support of the STEEP criteria [34]. OS was the time from diagnosis of primary breast cancer to death of any cause. IDFS was the time from primary diagnosis to the diagnosis of a breast cancer related invasive event (locoregional or distant) or, if no relapse had occurred, to death of any cause. In the absence of event in OS and IDFS the case was censored at last follow-up. DRFI was defined as the time from diagnosis to the diagnosis of a distant relapse of breast cancer or breast cancer related death, the case was censored at death of any other cause or at last follow-up if no DRFI event had occurred. Contralateral breast cancer and distant recurrences with uncertain origin were not included in DRFI but were included in IDFS. Follow-up time was defined as the time from diagnosis to date of death or to last follow-up.

Analyses of RNAseq data were performed in R (v 3.6.1), all remaining statistical analyses with SPSS (v 26.0). Concordance rate (expressed as percentage) was calculated to evaluate IHC inter-test reliability and kappa statistic applied as a measurement of the level of agreement. A kappa coefficient of ≥ 0.80 was interpreted as strong agreement, 0.60–0.79 as good, 0.40–0.59 as weak, 0.21–0.39 as minimal and < 0.20 as none agreement [35]. Area-proportional Venn diagrams were drawn with https://​www.​biovenn.​nl/​ [36]. Chi-square test was applied when comparing categorical values between groups (chi-square test for trend if more than two groups were compared). Nonparametric Mann–Whitney test was applied to compare non-categorical values between two groups. Survival data were analyzed by Kaplan–Meier estimates along with log-rank test and with Cox regression, reporting hazard ratio (HRs) and 95% confidence intervals (CIs). Multivariable Cox regression analyses were performed by including, aside from PD-L1 status, other traditional and prognostic factors: age at diagnosis, tumor size, lymph node status, Nottingham histologic grade (NHG) and TIL abundance as binary covariates. Four multivariable regression analyses were performed, i.e., one for each PD-L1 scoring method: SP142 IC+ , 22C3 CPS 10, 22C3 CPS 1 and 22C3 IC+. RNA sequencing data was matched against patient data generating a list of 16,258 genes across 194 samples. FKPM values were Log2-transformed, imputed (missing data to 0), mean-centered and scaled (samples and genes). The correlation between PD-L1 gene expression (by RNAseq) and PD-L1 protein expression (analyzed by IHC) was estimated using the Spearman method and visualized with boxplots (the median is indicated by the central line, upper and lower limits of the box represent the upper and lower quartiles and whiskers the × 1.5 interquartile range). A p-value less than 0.05 was considered statistically significant and all p-tests were two sided.

A higher positive detection rate (Table 1) was observed for SP142 IC ≥ 1% than for 22C3 CPS ≥ 10, 50.9% (118/232) versus 27.2% (63/232), when using these clinically applied predictive cut-offs (from the advanced TNBC setting).

Since 22C3 CPS ≥ 1 has also been investigated in clinical trials, we analyzed the percentage of PD-L1 positivity using this lower cut-off for 22C3. As expected, this resulted in a higher positive detection rate (53.9% (125/232)) compared to 22C3 CPS ≥ 10.

In an explorative analysis, to evaluate 22C3 under more similar conditions as SP142, we applied the IC+ scoring method to 22C3. The positive detection rate for 22C3 IC ≥ 1% was 41.8% (97/232).

When comparing SP142 and 22C3 with the clinically applied scoring methods and cut-offs (IC ≥ 1% and CPS ≥ 10, respectively), a kappa value of 0.48 was obtained (interpreted as week agreement). Approximately half of the tumors (47.8%; 111/232) were negative with both antibodies, whereas 60 tumors (25.9%) were positive with both antibodies, resulting in a concordance rate of 73.7%. Fifty-eight tumors (25%) were positive with SP142, but negative with 22C3, whereas three tumors (1.3%) showed the opposite pattern (Fig. 2B). Taken together, almost half of the tumors (49.2%; 58/118) that stained PD-L1 positive with SP142 were considered to be negative with 22C3, when using these clinically established predictive cut-offs.

The kappa value increased to 0.63 (interpreted as good agreement) and the concordance rate to 81.5% when a threshold of ≥ 1 for CPS was applied for 22C3 where 189 tumors (out of 232) showed concordant PD-L1 status (89 tumors negative with both antibodies and 100 tumors positive with both; Fig. 2C). A lower number of tumors that stained positive with SP142 but negative with 22C3 was found than when using the ≥ 10 cut-off for CPS (n = 18 vs. n = 58). The number of tumors with the opposite pattern (i.e. negative with SP142 but positive with 22C3) was increased from 3 to 25.

Next, we evaluated the concordance between the two antibodies when scored with the same scoring method and cut-off, i.e. IC ≥ 1% (Fig. 2D, note that IC+ is not normally employed for the 22C3 antibody). This comparison resulted in the best concordance rate of 86.6% (201 concordant tumors: 109 negative with both and 92 positive with both) and a kappa-value of 0.73 (interpreted as good agreement). Five tumors were negative with SP142 but positive with 22C3 and 26 showed the opposite pattern.

We detected a significant positive association between PD-L1 IHC expression and PD-L1 (CD274) gene expression (mRNA) in the overall cohort. The Spearman correlation coefficients were similar between PD-L1 gene expression and all the two-categorical IHC scorings (r_s = 0.59 for SP142 IC+, r_s = 0.60 for both 22C3 CPS 1 and CPS 10, r_s = 0.62 for 22C3 IC+; all p-values < 0.001; Fig. 3A–D). When stratifying the 22C3 CPS into three categories (i.e. < 1, 1–9 and ≥ 10), a positive stepwise association between PD-L1 (CD274) gene expression and PD-L1 protein expression was observed (Fig. 3E; r_s = 0.67), establishing a good degree of association between transcript and protein measurements.

We also investigated PD-L1 gene expression levels in SP142 IC and 22C3 CPS concordant and discordant groups, respectively (Fig. 3F, G). Here, transcript levels in the discordant groups (i.e. 22C3 CPS < 10 and SP142 IC ≥ 1% or 22C3 CPS ≥ 10 and SP142 IC < 1%) were found to be at an intermediate level between the concordant positive group and the concordant negative group. No significant difference in PD-L1 mRNA expression was found between the two discordant groups.

Clinicopathological characteristics differed in patients receiving (neo)adjuvant CT and in those not receiving CT. The patients in the CT-cohort were younger (p < 0.001), had higher median TIL abundance, more proliferative (p = 0.007) and higher-grade tumors (p = 0.004), higher rate of PD-L1 expressing tumors (p = 0.006 for SP142 IC and p = 0.028 for 22C3 CPS status) and tended to have fewer deaths (p = 0.059) but had similar rate of relapses as compared to the non-CT-cohort (Table 1). Due to these differences, we chose to evaluate clinicopathological features in relation to PD-L1 status and perform outcome analyses separately in the CT-cohort and the non-CT-cohort.

In the CT-cohort, tumors with SP142 IC ≥ 1% were significantly associated with higher NHG (p = 0.004), higher Ki-67 proliferation index (p = 0.005), histological medullary features (p = 0.001) and increased stromal TIL abundance (p < 0.001), whereas age at diagnosis, tumor size and lymph node status were not significantly associated with PD-L1 status (Table 2). When using CPS ≥ 10 for 22C3, only medullary features and TIL abundance were significantly associated with PD-L1 status (both p values < 0.001) and the association between PD-L1 and NHG and Ki-67 did not reach statistical significance (Table 2). With the other cut-offs for 22C3 (CPS ≥ 1 and IC ≥ 1%; Additional file 2: Table S2), the results were similar to those obtained for SP142, with significant associations to NHG (p < 0.001 and p = 0.012, respectively), Ki-67 level (p = 0.009 and p = 0.006, respectively), medullary features (p = 0.002 for both 22C3 CPS 1 and 22C3 IC+) and TIL abundance (p < 0.001 for both CPS 1 and IC+).

In the non-CT-cohort, a significant positive association between TIL abundance and PD-L1 status was observed, irrespective of PD-L1 IHC evaluation method (all p-values < 0.001 for SP142 IC+ , 22C3 CPS 1 and 22C3 IC+ ; for 22C3 CPS 10: p = 0.006 for median TIL score and p = 0.026 for TILs as binary covariate). Associations between the other clinicopathological parameters and SP142 IC ≥ 1%, 22C3 CPS ≥ 10 or 22C3 IC ≥ 1% did not reach significance (Additional file 3: Table S3). When using 22C3 CPS ≥ 1 cut-off, NHG was significantly associated with PD-L1 IHC expression (p = 0.045) and Ki-67 borderline significant (p = 0.051; Additional file 3: Table S3).

When using the clinically established cut-offs for both SP142 (IC ≥ 1%) and 22C3 (CPS ≥ 10) in univariable Cox regression analyses, a positive PD-L1 status was significantly associated with a better DRFI (HR = 0.47, 95% CI 0.22–1.00, p = 0.049 for SP142 IC+ and HR = 0.18, 95% CI 0.04–0.76, p = 0.019 for 22C3 CPS 10; Table 3 and Fig. 4). The HRs for IDFS and OS also indicated a better prognosis for patients with PD-L1 positive tumors (HRs ranging from 0.46 to 0.53), but only reaching significant level for IDFS and SP142 IC status (95% CI 0.26–0.89, p = 0.02). The results for 22C3 CPS ≥ 1 and 22C3 IC ≥ 1% showed a similar pattern although only reaching significancy for 22C3 CPS 1 and IDFS (HR = 0.53, 95% CI 0.29–0.98, p = 0.043; Table 3 and Additional file 3: Fig. S1).

Next, we performed a subgroup analysis where we divided the 22C3 CPS 10 negative group (i.e., those with CPS < 10) into one group positive with SP142 (i.e. IC ≥ 1%; n = 47) and one group negative with SP142 (IC < 1%, n = 71). No significant difference in DRFI was observed between these two groups (log rank p = 0.562; Fig. 5). For the group with 22C3 CPS ≥ 10, a similar division was not meaningful since all the patients in the CT-cohort that had 22C3 CPS ≥ 10 also scored SP142 IC ≥ 1%. These results suggest that if information for PD-L1 status with 22C3 CPS 10 is available, SP142 does not add any further prognostic information for DRFI.

In multivariable Cox regression analysis, PD-L1 status was found not significantly associated to outcome for any of the clinical endpoints, irrespective of IHC assay and cut-off (Table 3). Of note though, a trend towards better DRFI was observed for 22C3 CPS ≥ 10 staining (HR = 0.26, 95% CI 0.06–1.20, p = 0.084). Stromal TIL abundance was the only covariate showing independent significant association to outcome in multivariable analyses, where it was positively associated with improved IDFS irrespective of PD-L1 assay and cut-off included in the analysis (HRs ranging from 0.24 to 0.27 and p-values from 0.003 to 0.007, Table 3A-D) and with a better DRFI in a multivariable model where SP142 IC+ was included (HR = 0.33, 95% CI 0.11–0.99, p = 0.047; Table 3A).

The scarcity of patients in the non-CT-cohort did not allow for robust multivariable Cox regression analyses. PD-L1 status was not significantly associated with DRFI in univariable analysis (HRs ranging from 0.56 to 0.77, p-values not significant). For IDFS, the HRs for PD-L1 status were similar as in the CT-cohort (HRs ranging from 0.53 to 0.65 compared 0.46 to 0.57 in the CT-cohort), but in this small group of TNBC patients not treated with (neo)adjuvant CT with few events, the p-values were not significant (Additional file 5: Table S4 and Additional file 6: Fig. S2). Stromal TIL abundance was not significantly associated with any of the clinical endpoints in univariable analyses (HRs ranging from 0.90 to 1.34). Age (as continuous variable) was negatively associated with OS (HR = 1.07, 95% CI 1.01–1.14, p = 0.027) and IDFS (HR = 1.05, 95% CI 1.00–1.11, p = 0.043), tumor size was negatively associated with all the endpoints (HR = 2.41, 95% CI 1.08–5.39, p = 0.003 for IDFS; HR = 2.54, 95% CI 1.02–6.32, p = 0.045 for OS; HR = 4.63, 95% CI 1.00–21.53, p = 0.051 for DRFI) and lymph node status negatively associated with DRFI (HR = 8.06, 95% CI 2.13–30.59, p = 0.002; Additional file 5: Table S4).