CD8⁺ lymphocyte infiltration is an independent favorable prognostic indicator in basal-like breast cancer

The study population consists of 3,992 female patients with invasive breast cancer diagnosed between 1986 and 1992 in the province of British Columbia. This cohort was collected from the Breast Cancer Outcomes Unit database maintained by the British Columbia Cancer Agency (BCCA). During the study period, 75% of patients with breast cancer in the province were referred to the BCCA; non-referred patients were generally older or had no indications for adjuvant therapy [37, 38]. Of the patients referred to the BCCA, approximately 25% had available formalin-fixed paraffin-embedded blocks with sufficient tumor tissues for tissue microarray (TMA) construction. Thus, the study cohort represents about 20% of all of the patients with breast cancer diagnosed in the province during the study period. The mean age of the cohort at diagnosis was 58.9 years (23 to 95 years), and the median follow-up was 12.6 years. Baseline clinical information of the study population includes age at diagnosis, histology, grade, tumor size, number of involved axillary nodes, lymphovascular invasion (LVI), and dates of diagnosis, recurrence, death, and cause of death (breast cancer versus other). As shown in Table 1 among the study cases, approximately half (51.1%, 2,040/3,992) were poorly differentiated tumors (grade 3), 47.3% (1,888/3,992) had breast tumors over 2 cm, 43.1% (1,719/3,992) were node-positive, and 42.8% (1,710/3,992) had LVI. Histological categorization on these cases, including assignment to the medullary subtype, was determined by a central review of full sections which was performed at the time of referral to the BCCA. During the time period of this study cohort, most patients with breast cancer were treated according to the provincial guidelines developed by the BCCA on the basis of patient age, tumor size, nodal status, and LVI. Patients were defined as high-risk if their lymph nodes were positive, if there was evidence of LVI, or if the tumor was both greater than 2 cm and ER^- at the time of diagnosis. High-risk patients were treated with AST according to their age and menopausal status. Low-risk patients were not given any AST. This study and the use of de-identified data were approved by the Clinical Research Ethics Board of the BCCA and the University of British Columbia. We were permitted access to the de-identified patient outcome information from the Breast Cancer Outcomes Unit database, maintained by the BCCA. In compliance with the Canadian Tri-Council Policy Statement for ethical research involving human subjects, the requirement for informed consent was waived as this study was limited to anonymous archival specimens.

The centralized provincial laboratory of Vancouver General Hospital retained single archival blocks for each of the 3,992 patients. One 0.6-mm core per patient was used, 17 TMAs representing these samples were constructed, and immunohistochemistry and scoring for ER, PR, HER2, the Ki67 proliferation marker, EGFR, and CK5/6 were performed as previously described [31, 37, 39‐42]. Immunohistochemistry for CD8⁺ TILs was performed by using the antibody against human CD8 (clone C8/144B, dilution 1:100) in accordance with the protocol of the manufacturer (DakoCytomation, Glostrup, Denmark). Intrinsic breast cancer subtypes were determined by the immunohistochemical expressions of ER, PR, HER2, Ki67, EGFR, and CK5/6. Luminal A was defined as ER⁺ or PR⁺, HER2^-, and low Ki67 (< 14%); luminal B was defined as ER⁺ (or PR⁺) and HER2^- with high Ki67 (≥ 14%); luminal/HER2 subgroup was defined as ER⁺ (or PR⁺) and HER2⁺; HER2⁺/ER^- was defined as HER2⁺ with ER^- and PR^- [42]; and triple-negative subgroup (TNP) was defined as ER^-, PR^-, and HER2^-. The core basal subgroup was defined as triple-negative with either EGFR⁺ or CK5/6⁺, and the five negative phenotype (5NP) was defined as triple-negative as well as EGFR^- and CK5/6^- [31]. The 3,992 patients with breast cancer were thereby categorized as follows: 38.0% (1,518/3,992) luminal A, 20.8% (829/3,992) luminal B, 5.6% (223/3,992) luminal/HER2, 6.3% (250/3,992) HER2⁺/ER^-, and 15.8% (630/3,992) triple-negative, of which 8.3% (330/3,992) could be categorized as core basal and 4.1% (162/3,992) as 5NP; the remainder had a partial or unassignable subtype because of missing or ambiguous biomarker data (Table 1).

Stained TMA slides were digitally scanned and CD8⁺ TILs were visually scored by a pathologist who was blinded to the clinical characteristics and outcomes of the patients. Scoring and quantification of CD8⁺ TILs were carried out as described in a recent study [24]. In brief, intratumoral CD8⁺ TILs (iTILs) were defined as CD8⁺ lymphocytes located within tumor cell nests or in direct contact with the breast carcinoma malignant epithelial cells, whereas stromal CD8⁺ TILs (sTILs) were defined as CD8⁺ lymphocytes in the adjacent peritumoral stroma without direct contact with the carcinoma cells. Total CD8⁺ TILs (tTILs) were measured by combining the counts of iTILs and sTILs for each tissue core. To assess the reproducibility and reliability of the scoring, 490 cases were repeatedly scored by the same pathologist after a period of time (4 weeks), and 200 cases were randomly selected from the whole cohort and iTILs were re-scored by a second pathologist. Pearson correlation analysis was used to check the reliability of the repeated scoring by the same scorer, and the intraclass correlation coefficient (ICC) was used to assess the reliability of re-scoring by the two scorers. High correlation coefficients were obtained (Pearson r was at least 0.94, and ICC was 0.74).

The outcome variable in this study was breast cancer-specific survival (BCSS). Optimal cutoff points for TILs counts against BCSS were chosen on the basis of recently published findings from an independent series [24] and checked by receiver operating characteristic curve analysis by using 10-year BCSS as the endpoints, as described in the Supplemental method section (Additional file 1). The optimal cutoff points for iTIL, sTIL, and tTIL used in this study were 1, 3, and 2, respectively. To specify, CD8⁺ iTIL expression was categorized as low when iTIL was 0 (no CD8⁺ iTIL counted) and high when iTIL was at least 1 (1 or more CD8⁺ iTILs in the assessed tissue core); sTIL low means fewer than 3 CD8⁺ sTILs per core, and tTIL low means fewer than 2 CD8⁺ tTILs were identified in a core.

Analysis of the association between TILs and clinicopathologic variables was performed by using SPSS version 19.0 and R 2.11.1. Because the distributions of the outcome variable (BCSS) were not normal in the study cohort, non-parametric Wilcoxon testing was used to check the bivariate relationship between BCSS and TILs and other potential confounding variables, including age at diagnosis, grade, tumor size, involvement of lymph nodes, LVI, and intrinsic subtypes. Chi-squared testing was used to check the relationship between TILs and those potential confounding variables. For survival analysis, the event under study was death from breast cancer. BCSS time was defined as the number of years between the date of diagnosis of breast cancer and the date of death attributable to breast cancer. Survival time was censored at the time a patient died from another cause or when the follow-up period ended. For univariate survival analyses, the Kaplan-Meier function analysis was performed to estimate probabilities of BCSS. Log-rank testing was used to assess differences in BCSS among different subgroups. For multivariate survival analyses, Cox proportional hazards regression models were built to estimate the TIL hazard ratio (HR), which was adjusted by the potential confounding variables on the basis of the partial maximum likelihood estimation. Smoothed, rescaled Schoenfeld residual plots were performed to test proportional hazards assumptions. Only cases with sufficient information for all covariates were included in the multivariate analysis. Wald statistics were used to test the significance of individual coefficients. Interactions between TILs and some covariables were checked by building Cox regression models for different levels of those variables and comparing HRs of TILs. All of the tests were two-sided at a significance level of 0.05. Supplementary analyses were also performed by using relapse-free survival as an outcome variable; relapse-free survival time was defined as the number of years between the date of diagnosis of breast cancer and the date of any type of relapse, including local, regional, and distant relapses of the disease.

Among the 3,992 breast tumor cases, intact cores bearing infiltrating breast carcinoma sufficient for interpretation of immunohistochemical data for CD8 staining were available from 3,403 (85.2%) tumors. Median counts of CD8⁺ TILs per 0.6-mm TMA core were 0 for iTIL (interquartile range, or IQR, of 0 to 1), 2 for sTIL (IQR of 0 to 10), and 3 for tTIL (IQR of 0 to 12). Of the 3,403 interpretable cases, 32.4% had tumor infiltrated with at least one CD8⁺ iTIL and 60.6% by at least one CD8⁺ sTIL (Figure S1 of Additional file 2). The distributions of CD8⁺ iTILs and sTILs were both significantly and positively skewed (Figure S2 of Additional file 3). Because analytical results from all types of TILs interpretation were broadly similar, results presented in this paper are based primarily on iTILs analysis, which is the fastest and simplest to perform. As shown in Table 1 the presence of iTIL is significantly associated with young age, high grade, medullary histology, ER negativity, HER2 positivity, and the core basal intrinsic subgroup, the category that has the highest prevalence of cases displaying intratumoral lymphocytes.

To examine the prognosis of CD8⁺ TILs in the study population, we first applied univariate Kaplan-Meier function survival analysis in the whole cohort. The results did not show a significant difference in BCSS between breast cancer patients with an iTIL count of at least 1 and an iTIL of 0 (P = 0.761). Since the distribution of iTILs was associated with patient age at diagnosis, tumor grade, and ER and HER2 status, we next assessed the survival functions of iTIL associated with BCSS in groups with different age, tumor grade, and ER and HER2 status. Figure 1 showed that, in younger patients (< 50 years) and in those with ER^- tumors, cases with iTILs had significantly better BCSS than those without. Reversed associations were observed in patients who were at least 50 years old or who had ER⁺ breast cancer. No significant associations were detected in cases stratified by grade (grade 1 + 2 versus grade 3) or HER2 status (HER2⁺ versus HER2^-). These results indicated that age and ER status could have an interaction with the association between iTILs and patient survival in breast cancer.

We built Cox proportional hazards regression models to estimate the HR for iTILs. Smoothed, rescaled Schoenfeld residual plots showed that iTILs and most other covariables satisfied the proportional hazards assumptions well during the period of follow-up. Only iTILs in the luminal A subgroup varied slightly during longer follow-up.

Results from the univariate Cox regression model analysis showed that iTILs was not a significant prognostic factor associated with BCSS in the cohort as a whole: HR = 1.02, 95% confidence interval (CI) = 0.89 to 1.17. To take into consideration potential confounders, a multivariate Cox regression model was built to assess the association between iTILs and BCSS, including the covariates of age at diagnosis, tumor grade and size, lymph node status, LVI, and intrinsic subtype. Table 2 showed that the adjusted HR of iTIL was 0.79 (95% CI = 0.68 to 0.91), meaning that, in the whole cohort, the probability of BCSS among patients with an iTIL count of at least 1 was 21% (1 to 0.79) higher than among those with an iTIL count of 0 after adjustment for age, grade, tumor size, lymph node status, LVI, and intrinsic subtypes. Besides iTIL, tumor grade and size, nodal status, LVI, and intrinsic subtype, each had significant effects on BCSS. To examine the effect of interaction between age, ER status, and iTIL, we built multivariate Cox regression models for iTILs at different levels of age and ER status. These analyses showed that the adjusted HRs for iTILs were 0.65 (95% CI = 0.51 to 0.84) for those younger than 50 years old and 0.89 (95% CI = 0.74 to 1.06) for those at least 50 years old; the adjusted HRs were 0.61 (95% CI = 0.47 to 0.77) for those with ER^- tumors but 0.91 (95% CI = 0.77 to 1.11) for those with ER⁺ tumors. Therefore, interactions between iTIL and age and ER status might modify the effect size for iTILs in the unstratified whole cohort of patients with breast cancer.

We further assessed the association of CD8⁺ TILs with patient survival in different breast cancer intrinsic subtypes, first using univariate Kaplan-Meier function survival analysis. No difference in BCSS was detected between those with an iTIL count of at least 1 and an iTIL count of 0 within the luminal A and luminal B subgroups (Figure 2a, b). Although we observed an apparent difference between the two groups among HER2⁺/ER^- cases, this was not statistically significant (P = 0.064) (Figure 2c). However, as shown in Figure 2d, a large and significant difference in BCSS was found between cases with an iTIL count of at least 1 and those with an iTIL count of 0 among triple-negative breast cancers. By stratifying triple-negatives into core basal and 5NP subgroups, we observed a much larger difference in BCSS between cases with an iTIL count of at least 1 and those with an iTIL count of 0 in the core basal intrinsic subgroup. Patients with an iTIL count of at least 1 basal-like tumor had significantly better survival than those with an iTIL count of 0 (mean survival time of 14.5 vears versus 11.0 years, P < 0.001) (Figure 2e). No such association was found among triple-negative, non-basal (5NP) cases (Figure 2f). We also performed survival analysis in all patients with ER^- breast cancer, excluding the core basal cases, and found no significant difference in BCSS between cases with an iTIL count of at least 1 and those with an iTIL count of 0 (P = 0.434).

To confirm the association between iTIL and BCSS and to assess the independent prognostic effect size in different breast cancer intrinsic subgroups, multivariate Cox proportional hazards regression models were built to estimate the iTIL HRs, which were adjusted by the potential confounders. Results in Table 3 showed that the HRs of iTIL were not significant in the luminal A, luminal B, and HER2⁺/ER^- intrinsic subgroups. However, iTIL was demonstrated to be a significantly independent favorable factor for BCSS in triple-negative cases because of a strong effect in the core basal subgroup (Table 4). Among core basal cases, the presence of any intratumoral CD8⁺ lymphocytes (iTILs of at least 1) was associated with a 65% higher probability of BCSS than among those tumors lacking intratumoral CD8⁺ lymphocytes (iTIL of 0) and this was statistically significant even after adjusting for age at diagnosis, grade, tumor size, lymph node status, and LVI. Considering that medullary breast carcinoma, a histologically evident subtype known to carry a good prognosis, usually has a core basal immunophenotype and could be responsible for some of the observed effect, we repeated the multivariate Cox regression analysis for core basal cases by excluding those with medullary carcinoma (27 cases). The results still showed a similar and significant HR (HR = 0.38, 95% CI = 0.24 to 0.59), which therefore could not be attributed to medullary histology. In contrast, the multivariate analysis did not show any association between iTILs and BCSS in the 5NP subgroup (that is, triple-negative breast cancers that do not express basal markers). These results demonstrated that the prognostic effect of iTILs was significantly different in these two subgroups of triple-negative cases, indicating that the association of iTIL with BCSS exists primarily in only the core basal intrinsic subgroup.

To confirm the prognostic value of CD8⁺ TILs in breast cancer, we also evaluated the distributions of sTILs and tTILs in relation to patient and tumor characteristics and the associations of sTILs and tTILs with survival. Results similar to those from the analysis with iTILs were obtained. In brief, high expressions of sTILs and tTILs were significantly correlated with young age, high grade, larger tumor size, medullary histology, ER negativity, HER2 positivity, and the core basal phenotype (Table S1 of Additional file 4) and again were significantly associated with better BCSS in only the core basal intrinsic subgroup (Figure S3 of Additional file 5 and Tables S2 and S3 of Additional file 4).