Immune gene expression profiling reveals heterogeneity in luminal breast tumors

We analyzed data and biospecimens collected from a hospital-based BC case-control study in Hong Kong as previously described [20]. In brief, fresh frozen breast tumors and paired normal tissues were collected from newly diagnosed BC patients in two HK hospitals between 2013 and 2016. Patients with pre-surgery treatment were excluded from the study. Clinical characteristics and BC risk factors were obtained from medical records and questionnaires. The study protocol was approved by ethics committees of the Joint Chinese University of Hong Kong-New Territories East Cluster, the Kowloon West Cluster, and the National Cancer Institute (NCI). Written informed consent was obtained prior to the surgery for all participants.

Paired tumor and histologically normal breast tissue samples were processed for pathology review at the Biospecimen Core Resource (BCR), Nationwide Children’s Hospital, using modified TCGA criteria [21]. Specifically, only tumors with > 50% tumor cells and normal tissue with 0% tumor cells were included for DNA/RNA extraction.

RNA sequencing (RNA-Seq) data were generated in 139 tumors and 92 histologically normal breast tissue samples that passed on all quality control metrics at Macrogen Corporation on Illumina HiSeq4000 using TruSeq stranded RNA kit with Ribo-Zero for rRNA depletion and 100-bp paired-end method. Gene expression was quantified as TPM (transcript per million) using RSEM [22], and log₂TPM was used for statistical analyses. PAM50 subtype was defined by an absolute intrinsic subtyping (AIMS) method [23]. Three computational algorithms were used to characterize immune cell composition in both tumor and paired normal breast tissue: ESTIMATE [24], CIBERSORT [25], and MCP-counter [26]. While ESTIMATE (for overall infiltration of immune cells) and MCP-counter (for eight immune cell subpopulations) both measure the abundance of immune cells in a given sample, CIBERSORT estimates intra-sample proportions of 23 immune cell subpopulations.

Whole-exome sequencing (WES) was performed on 104 paired tumor and normal samples (59 of them also had RNA-Seq data) at the Cancer Genomics Research Laboratory (CGR), NCI, using SeqCAP EZ Human Exome Library v3.0 (Roche NimbleGen, Madison, WI) for exome sequence capture. The captured DNA was then subjected to paired-end sequencing utilizing Illumina HiSeq2000. The average sequencing depth was 106.2x for tumors and 47.6x for the paired normal tissues. Somatic mutations were called using four callers, and the analyses were based on mutations called by three or more of four established callers (MuTect [27], MuTect2 (GATK tool), Strelka [28], and TNScope by Sentieon [29]).

SNP rs12628403, which is a proxy for the APOBEC3B deletion (r² = 1.00 in Chinese from Beijing (CHB) in HapMap samples), was genotyped in germline DNA with a custom TaqMan assay as previously described [30].

We assembled hemotoxylin and eosin (H&E)-stained frozen sections from the same frozen tumors used for DNA/RNA extraction and formalin-fixed paraffin-embedded (FFPE) sections from the same set of HK patients. Using the Halo image analysis platform (Indica Labs, Albuquerque, New Mexico), we developed a multi-step approach for the quantification of TILs which was based on supervised machine learning analysis of histological images (Additional file 1: Figure S1). In the first step, we trained an algorithm to segment the tumor into epithelial, stromal, and adipose tissue regions (panel B, Additional file 1: Figure S1). Next, we trained a cell-detection algorithm to identify TILs based on contexture (nuclear detection weight = 0.35; nuclear contrast threshold = 0.54), size (5–20 μm) and shape (minimum nuclear roundness = 0.45) (panel E, Additional file 1: Figure S1) within well-defined regions of interest. By focusing on the stroma (intra-tumoral and peri-tumoral; panels E and F Additional file 1: Figure S1), we then applied this algorithm to the centralized evaluation of TILs in all images.

We analyzed two available, independent datasets to replicate our findings: 564 luminal patients in TCGA [3] and 112 luminal patients in a Korean BC genomic study (KBC) [31]. We analyzed TCGA Asians (n = 29, mean age: 51 years), African Americans (AA, n = 72, mean age 58 years), and European ancestry (EA, n = 463, mean age 60 years) separately. PAM50 was called using the same AIMS method for each TCGA sample as it was used in HKBC. KBC patients were much younger, with a mean age at diagnosis of 40 years. PAM50 subtype and mutation calling for KBC were previously detailed [31]. Immune classification and composition across all datasets (HKBC, TCGA, and KBC) were analyzed using the same methods.

The consensus clustering was conducted using ConsensusClusterPlus [32], based on expression of 130 immune-related genes (within 13 previously reported metagenes, including T cell signatures, activated CD8/NK cells, interferon-stimulated genes, and etc., Additional file 2: Table S2) [33]. Expression levels of these metagenes correspond to activities of various types of immune cells and reflect various immune functions. The prognostic and predictive values of these metagenes have been previously assessed in TCGA and other independent datasets [34, 35]. For each of the 500 resampling of subjects, we sampled 80% of the subjects and clustered them using agglomerative hierarchical clustering with Pearson correlation as the distance metric. We evaluated up to 10 clusters and chose 3 clusters (k = 3) as they fit the data best.

A comprehensive characterization of immune cell composition in both tumor and paired normal breast tissue was achieved by using three computational algorithms: ESTIMATE [24], CIBERSORT [25], and MCP-counter [26]. The ANOVA test was used to compare mean differences across the luminal immune subtypes for immune cell populations and their immune scores. Logistic regression was used to assess the associations between the immune subtypes (outcome) and genomic alterations, patient characteristics, and BC risk factors, with the adjustment for age at diagnosis and body mass index (BMI). The Kaplan–Meier method was used to assess overall survival among patients, stratified by immune subtypes. A multivariable Cox proportional hazards model was also used to test the differences in survival across immune subtypes with the adjustment of age at diagnosis and tumor stage. Since most of our analyses were exploratory, we did not adjust for multiple testing. All statistical tests were two-sided and performed using SAS version 9.3 (SAS Institute, Cary, NC, USA) or R version 3.4.4 (R Foundation for Statistical Computing, Vienna, Austria).

We conducted unsupervised consensus clustering of 92 luminal tumors using expression of 130 immune-related genes. The best separation was achieved by dividing the luminal patients into three subtypes (lum1: n = 40; lum2: n = 36; lum3; n = 16; Fig. 1a); lum1 and lum3 were enriched with luminal-A tumors and lum2 enriched with luminal-B tumors (Additional file 2: Table S3). Lum1 expressed low levels of most immune genes (Fig. 1b) and therefore was designated as low-TIL. Lum2 had high expression of STAT1 and other interferon-stimulated genes (ISGs), but low expression of other immune genes (Fig. 1b), designated as high-ISG. Lum3 (defined as high-TIL) showed the highest expression level of most immune genes (Fig. 1b) such as immune checkpoint genes (e.g., PD-L1 and CTLA-4), chemokine genes and their receptors (e.g., CXCL9 and CXCL10), and effectors (e.g., GZMK and PRF1) (Additional file 1: Figure S2), reflecting a T cell-inflamed phenotype. Compared to low-TIL and high-ISG tumors, high-TIL tumors had higher abundance of most immune subpopulations (estimated by MCP-counter, Fig. 2a), except for neutrophils and cells of monocytic lineage. The abundance score for each immune subpopulation in high-TIL luminal tumors was comparable to that of HER2-enriched and basal-like tumors (Fig. 2a; Additional file 1: Figure S3; P values see Additional file 2: Table S4). Adjusting for tumor purity, which was inferred using ESTIMATE purity score, did not change the results (Additional file 1: Figure S4). The results from TIL assessment based on H&E staining of frozen and FFPE sections were consistent, confirming that TILs were more abundant in tumor stroma in high-TIL versus low-TIL patients (Fig. 2b).

We also inferred the fractions of 23 immune cell subpopulations in these patients using CIBERSORT, which estimates the relative fraction of each cell population in a sample rather than the absolute abundance. Figure 2c shows the fractions of seven subpopulations with the average fraction > 10% across all samples. We found that high-TIL tumors showed higher fractions of CD8+ T cells and tumor-killing M1 macrophages [36] than those of low-TIL and high-ISG tumors, while they had lower frequencies of tumor-promoting M2 and undifferentiated M0 macrophages (P values see Additional file 2: Table S5).

Based on expression levels of the same 130 immune genes used in HKBC, luminal tumors in each TCGA population (Asian, African American, and White) and KBC were similarly assigned to three subtypes using consensus clustering, with the presence of a high-TIL luminal subtype seen in all populations (Fig. 3). The pattern was more similar in the three Asian populations, with a more pronounced separation of the high-TIL subtype from the other two subtypes. Consistent with HKBC results, high-TIL tumors in all replication datasets showed higher overall immune score (by ESTIMATE, Fig. 3), higher abundance of most immune subpopulations (by MCP-counter, Additional file 1: Figure S5a), and higher fractions of CD8+ T cells and M1 macrophages (by CIBERSORT, Additional file 1: Figure S5b). Like HKBC, high-TIL tumors showed upregulation of genes in immune activation and regulation activities (Additional file 1: Figure S5c), while high-ISG tumors expressed higher levels of ISGs (e.g., DDX58) than tumors in the other two luminal immune subtypes (Additional file 1: Figure S5d).

In HKBC, most clinical characteristics or BC risk factors examined, such as tumor grade, nodal status, age at menarche, parity, age at first birth, breastfeeding, and age at menopause, did not vary significantly across immune subtypes (Additional file 2: Table S6). However, the average BMI was higher in high-TIL (mean = 27.9) than in low-TIL (mean = 24.1) and high-ISG patients (mean = 24.6). The differences remained significant after the adjustment of age, menopausal status, and tumor purity (P = 0.0018 for high-TIL vs. low-TIL and P = 0.0057 for high-TIL vs. high-ISG). In addition, high-TIL tumors had slightly lower ESR1 (estrogen receptor alpha) but higher ESR2 (estrogen receptor beta) expression levels, resulting in a lower ESR1/ESR2 ratio (P = 0.001) compared with low-TIL and high-ISG tumors (Fig. 4a, Additional file 1: Figure S6a). The association between the low ESR1/ESR2 ratio and the high-TIL subtype was consistently seen in all TCGA populations (Additional file 1: Figure S7a).

High-TIL patients tended to be younger than patients with low-TIL tumors in HKBC as well as in the replication datasets (Additional file 1: Figure S7b), although the significant difference was only seen among the TCGA Whites (P = 0.018). The short follow-up time in HKBC prohibited us from assessing the prognostic outcome in relation to the immune subtypes. We therefore conducted survival analysis using TCGA data of 905 BC patients. We combined all ethnicity groups because few deaths occurred among Asian or African American patients. As shown in Additional file 1: Figure S7c, the high-TIL subtype was associated with the best 10-year overall survival among all subtypes (P = 0.008), although the difference became non-significant after the adjustment for age at diagnosis and stage (hazard ratio [HR] = 0.6, 95% confidence interval [CI] = 0.26–1.4, P = 0.22). The attenuation of the significance was likely due to younger ages in the high-TIL subtype as stage did not differ across luminal immune subtypes (P = 0.72).

To evaluate the possible contribution of germline variation in APOBEC3B to immune profiles and mutational events, we genotyped a SNP (rs12628403) that is a proxy for the APOBEC3B deletion in germline DNA [30]. In HKBC, the frequency of the rs12628403-C allele that tags the 30-kb deletion (44.7% among 76 luminal patients and 40.4% among all 114 patients with genotyping data) was similar to what was reported in East Asian populations [17]. We found the expected associations between the APOBEC3B deletion and decreased levels of APOBEC3B expression in both tumor and normal tissue, validating SNP rs12628403 as a proxy for APOBEC3B deletion (Additional file 1: Figure S8). The frequency of the deletion allele did not vary significantly by immune subtypes, either in HKBC or in TCGA White (Table 1). In addition, the expression level of APOBEC3A_B, which is a hybrid transcript resulting from the APOBEC3B deletion, did not vary significantly by luminal immune subtypes (P = 0.36). Further, the ESTIMATE immune scores did not vary across different genotypes of SNP rs12628403 (P = 0.56). Similar results were obtained in the analysis based on all tumor subtypes. In TCGA Whites, the homozygous deletion of APOBEC3B was very rare; only 2 of 329 luminal patients with genotyping data were homozygote and neither of them was in the high-TIL subtype (Table 1).

In an exploratory analysis of a subset of luminal tumors with both RNA-Seq and WES data (n = 59), we found that, after age and BMI adjustment, high-TIL tumors were associated with a higher nonsynonymous mutation burden (P = 0.03 compared to low-TIL tumors, Fig. 4b, Additional file 1: Figure S6b) and a higher frequency of APOBEC-signature mutations (mean 23.6%) compared with low-TIL (7.6%, P = 0.045) and high-ISG (8.3%, P = 0.089) tumors. Notably, all TP53 mutations (n = 8, Table 2) observed among luminal patients occurred in high-ISG tumors. The similar enrichment of TP53 mutations in high-ISG tumors was also seen in TCGA Whites (P = 0.0064, Table 2). The frequency of PIK3CA mutations did not vary significantly by immune subtypes in HKBC but showed a slight increase in high-TIL tumors in TCGA Whites (P = 0.031 compared to low-TIL tumors).

In our HKBC data, neither abundance nor fractions of the examined immune cell populations in paired normal breast tissue varied significantly across the three luminal immune subtypes (Additional file 1: Figure S9), suggesting that the distinguishing TIL levels between high-TIL and other tumors were not driven by the differences in their systematic normal TIL levels. Compared to matched normal (N) tissue, low-TIL and high-ISG tumors showed either no change or lower abundance of immune cell populations (such as cytotoxic lymphocytes), whereas high-TIL, like non-luminal tumors, had higher abundance scores of CD3+ T cells, CD8+ T cells, and B lineage cells (T-N difference > 0, Fig. 5; P value of CD8+ T cells = 0.0002 and 0.025 for high-TIL and non-luminal patients; other P values see Additional file 2: Table S7). These observations indicate a tumor-derived activation of specific immune responses in high-TIL and non-luminal tumors but not in other luminal tumors.