Evaluation of invasive breast cancer samples using a 12-chemokine gene expression score: correlation with clinical outcomes

A complete description of the TCC biobanking program has been previously published [17]. A retrospective review was performed on selected female patients with stage I to III breast cancer who were diagnosed between 1988 and 2012 and received primary surgery at the Moffitt Cancer Center. Patients may have received adjuvant therapies at Moffitt or at other locations. Snap-frozen tumor specimens from initially resected primary breast tumors were used for the gene expression profiles. Of the 813 unique gene expression files available, we included those from breast primary tumors only for which full clinical information was available. Patients chosen for study had available clinical and follow-up data within Moffitt Cancer Center’s electronic medical record system along with a genomic expression profile of their primary tumor. We excluded patients who had received any form of neoadjuvant therapy and with de novo metastatic disease, resulting in 366 patients in total.

The 12-chemokine score for these samples was extracted from the TCC database. The 12-chemokine score for TCC was calculated, in brief, as follows. Tumors from patients treated at the Moffitt Cancer Center were arrayed on modified Affymetrix HuRSTA-2a520709 GeneChips (Affymetrix, Santa Clara, CA, USA). Chips were normalized using iterative rank-order normalization (IRON) [18]. An RNA quality-related batch effect was identified in the resulting normalized data, which was removed by training a partial least squares (PLS) model [19] to the RNA integrity number (RIN) for each sample and then subtracting the first PLS component. The final 12-chemokine score across all tumors in TCC was calculated using the first component from a principal component analysis (PCA) model based on the 12-chemokine genes. The 13 following probe sets was used; CCL18: merck-NM_002988_at, CCL19: merck-NM_006274_at, CCL2: merck-NM_002982_at, CCL21: merck-NM_002989_at, CCL3: merck-D63785_x_at, CCL4: merck-NM_002984_at, CCL5: merck-NM_002985_at, CCL8: merck-NM_005623_at, CXCL10: merck-NM_001565_at, CXCL11: merck-NM_005409_at & merck2-NM_005409_at, CXCL13: merck-NM_006419_at, CXCL9: merck-NM_002416_at.

A two-sided Student t test and Bonferroni correction were used to test for differences in high (n = 183) and low (n = 183) gene expression in the 366 fully annotated specimens across the 12 genes in the chemokine gene expression signature. Principal component analysis was performed and visualized using Evince™. We also studied immune gene expression levels of B and T lymphocyte attenuator (BTLA), cluster of differentiation (CD)14, CD31, CD274 [programmed death ligand 1 or (PD-L1)], CD56, CD69, CTLA4, CXCL12, fibroblast activation protein (FAP), granzyme B, T cell immunoglobulin and mucin domain containing 3 (TIM3), indoleamine 2,3 dioxygenase (IDO1), interferon gamma (IFN-γ), interleukin (IL)2, IL10, IL4, IL6, Janus-associated kinase 1 (JAK1), lymphocyte activation gene 3 (LAG3), lymphocyte expansion molecule (LEM), homologous to lymphotoxin exhibits inducible expression and competes with HSV glycoprotein D for binding to herpesvirus entry mediator, a receptor expressed on T lymphocytes (LIGHT), MHC class I polypeptide-related sequence A (MICA), neural cell adhesion molecule 1 (NCAM1), nuclear factor kappa light chain enhancer of activated B cells (NFKB), nitrogen oxide synthase 1 (NOS1), programmed death 1 (PD1), perforin 1 (PRF1), signal transducer and activator of transcription 1 (STAT1), signal transducer and activator of transcription 3 (STAT3), and vascular endothelial growth factor A (VEGFA). We performed a two tailed Student’s t test with a Bonferroni correction for multiple testing, comparing mean expression levels of the above-noted immune genes between groups with high and low 12-chemokine gene expression scores.

Histological sections corresponding to 28 cases (prepared from the mirror image of the portion of tumor submitted for the mRNA microarray analysis) were retrieved from the Moffitt Cancer Center Anatomic Pathology Division’s repository as a pilot analysis to study correlation between the CS and tumor-localized, ectopic lymphoid node-like structures (TL-ELNs). Half of the specimens were from the top 10^th percentile 12-chemokine gene expression scores and half were from the bottom 10^th percentile. All of the specimens were 10% formalin fixed and paraffin embedded. Random representative hematoxylin and eosin (H&E)-stained sections through all selected tissue blocks were evaluated for the presence or absence of TL-ELNs. To further characterize the TL-ELNs, tissue sections were stained using the avidin-biotin complex method with retrieval under high pH. Pre-diluted monoclonal antibodies to CD3, CD4, CD8, and CD20 (Ventana Medical Systems, Tucson, AZ, USA) were used for the manual morphometric analysis of TL-ELNs by brightfield microscopy. To ensure pathologic concordance, two pathologists at our institution reviewed the tissue sections. Scores of 0 to 3 were assigned based on the following features: 0 = no lymphoid infiltrate noted in slide, 1 = 1 group lymphoid infiltrate, 2 = 2 groups of lymphoid infiltrate, and 3 = 3 or more groups of lymphoid infiltrate. Both pathologists were blinded as to the 12-chemokine gene expression scores of the individual samples. Clinical information was accessible only to the principal investigator and authorized collaborators, and all samples were anonymously coded before analysis. The Fisher’s exact test was used to test the association of 12-chemokine expression scores and H&E staining scores. The McNemar test was used to analyze the strength of agreement between the scoring methods of the two pathologists.

We compared clinical and pathological factors of patients with low versus high 12-chemokine gene expression scores calculated by principal component analysis (determined by median split). Correlation between clinicopathologic factors and the 12-chemokine gene expression score was tested using the chi-square test with the exact method using Monte Carlo estimation. Kaplan-Meier curves were created for both overall survival (OS) and recurrence-free survival (RFS), and log-rank tests were used to compare 12-chemokine gene expression scores and 12-chemokine gene expression scores stratified by antibody status [estrogen receptor (ER) or progesterone receptor (PR) positive], ER and PR negative (with HER2 negative or missing), and HER2 positive. Multivariable survival models were fit using Cox proportional hazards model. Final models were chosen using backward selection, with a removal alpha of 0.05. All p values were two-sided unless otherwise stated and considered statistically significant at the 0.05 level. The final multivariate survival model incorporated age, pathologic stage, and ER status based on this criteria. All statistical analyses were performed using SAS (version 9.4; SAS Institute; Cary, NC, USA). A second log rank Kaplan Meier OS and RFS analysis was done on the larger KM Plot breast cancer dataset [20] due to the small number of events within the TCC cohort using the corresponding 12-chemokine gene probeset (216598_s_at (CCL2), 205114_s_at (MGC12815), 204103_at (CCL4), 204655_at (TCP228), 214038_at (CCL8), 210072_at (CCL19), 204606_at (CCL21), 203915_at (Humig), 204533_at (CXCL10), 211122_s_at (SCYB9B), 205242_at (CXCL13)) on the available Affymetrix arrays. In all KM Plot analyses the high versus low chemokine patients were split by mean upper tertile expression of the 12 chemokine genes computed across the entire dataset with equal weighting. Intrinsic subtypes in the KM Plot dataset are based on St. Gallen criteria using estrogen receptor 1 (ESR1), HER2, and antigen KI-67 (KI67). Prediction analysis of microarray 50 (PAM50) was used to classify breast cancer molecular subtypes within the TCC dataset (11). The estrogen, progesterone, and HER2 receptor status of the samples were classified by TCC pathologists using College of American Pathologists (CAP) criteria and obtained from the TCC database. Estrogen and progesterone receptors were positive if >1% immunohistochemical (IHC) staining was noted, and HER2 status was positive if 3+ by IHC or with a positive ratio by in situ hybridization assay. GraphPad Prism software (GraphPad Software, San Diego, CA, USA) was used to compare the mean chemokine gene expression scores with a Tukey test to obtain 95% confidence intervals between the molecular subtypes and receptor statuses by immunohistochemistry.