We inferred the immune abundance from RNAseq data using single-sample gene set enrichment analysis (ssGSEA, i.e. immune score in ESTIMATE) [10], gene set variation analysis (GSVA) [11] and deconvolution methods --- CIBERSORT [12] and TIMER [13]. In addition to the samples in Pan-MET dataset, we also evaluated immune level in normal tissue samples achived from Genotype-Tissue Expression (GTEx) Project. H&E stained sections in BRM-sTIL dataset were manually counted for percent sTILs using standard criteria developed by the international TILs working group [14]. Each slide was independently reviewed by two study personnel (JLN and CL) to minimize inter-observer variability. When the sTILs differed by 10% or more, the study pathologist (AH) made the final determination.

We estimated total immune abundance using RNAseq from 50 pairs of patient-matched PBTs and METs. For multiple METs that were matched to the same PBT, we first took the average. In general, METs showed a significantly lower total immune score compared to patient-matched PBTs (Fig. 1a; p < 0.001). The decrease in immune score was observed in METs collected from various sites, but was especially apparent in BRMs (p < 0.0001, Fig. 1b). Removing BRMs and combing all other METs, we noted a non-significant trend to decreased immune score in METs (p = 0.12, Fig. 1c). However, it should be noted that the small number of samples makes conclusions in the non-brain METs challenging. Validating the finding of decreased immune cells in brain METs, pathologic assessment of sTILs in an additional cohort of 49 patient-matched PBTs and METs revealed that BRMs also showed a significant decrease in the percentage of sTILs compared to patient-matched PBTs (p < 0.001, Fig. 1d). When grouping PBT/MET pairs by hormone receptor (HR) status and HER2 status, both datasets revealed a trend of decreased immune abundance in all subtypes, with TNBC subtype having the most significant decrease (p < 0.01, Additional file 2: Figure S1). Similar results were observed when we treated those METs matched to the same PBT as METs in different pairs (Additional file 2: Figure S2). While the total immune score only estimates the overall immune abundance in the bulk sample from RNAseq, and the sTILs percentage was carefully counted as the immune cell percentage in the stroma, the two measurements of immune abundance were significantly correlated (p < 0.001) for the 15 pairs of PBT/BRMs within both data sets (Fig. 1e). A lesser degree of agreement was only observed in tumors with extreme low sTILs (5%), possibly due to unstable estimates by both methods when the immune component is limited.Fig1Fig. 1

Lower immune abundance in metastatic breast tumors (METs) compared to primary breast tumors (PBTs) (a) Total immune score in PBT/MET pairs in Pan-MET dataset, together with the paired changes (MET-PBT). b Paired changes of total immune score removing BRMs in (a). c Total immune score grouped by MET sites. d Stromal tumor infiltrating lymphocytes (sTILs) percentages of 49 pairs of PBT/BRMs in BRM-sTIL dataset. e Spearman’s correlation between sTILs percentages and total immune score for 15 pairs of PBT/BRMs overlapped by Pan-MET and BRM-sTIL. ****p < 0.0001, ***p < 0.001, **p < 0.01, *p < 0.05 from two-sided Wilcoxon signed rank test in (a-d) and correlation test in (e)

We inferred the abundance of each immune cell population by two types of methods — enrichment analysis and deconvolution method. To validate those approaches, we first compared the GSVA scores of four common immune cell populations defined by both Davoli et al. [15] and Tamborero et al. [16]. The correlations ranged from 0.4 to 0.85 (Additional file 2: Figure S4), indicating overall high consistency. For further validation, we applied four methods; namely GSVA using the immune signatures from Davoli and Tomborero, and two methods of deconvolution (CIBERSORT and TIMER) to a publicly available single cell RNA-seq dataset [17], in which immune cell percentages were available using cell markers. Based on the correlations, the estimated levels of B cell, T cell, and macrophages by immune signatures from Davoli and Tamborero, and deconvolution method TIMER, were in general most highly correlated with actual abundance of corresponding cell types, although some signatures were not quite specific, such as CD4+ mature T cell and CD8+ effector T cell in Davoli signatures. CIBERSORT estimates showed lower correlations as expected, because the actual percentages were calculated based on three cell types, while CIBERSORT considered 22 cell types (Additional file 2: Figure S5).

Comparing patient-matched PBTs and METs, the GSVA score and abundance estimate from deconvolution methods for most immune cell populations were significantly lower in METs (Fig. 2a-c). Adjusting for total immune abundance, most immune cell populations were still lower, but M2-like macrophages were significantly higher in METs (Fig. 2d). Since CIBERSORT provides an empirical p value testing the null hypothesis that a particular sample does not contain any of the 22 cell types, we removed 16 pairs with at least one sample with p > 0.05, M2-like macrophages were still higher in METs, but there was only a trend to significance (Additional file 2: Figure S6). Significant increment was also observed in the ratio of the relative percentages of M2 and M1, indicating dominant level of M2 over M1 (Fig. 2e). When separating PBT/MET pairs to different MET sites or HR/HER2 subtypes, the results were generally consistent (Additional file 2: Figure S7-S8). Due to the lack of adjacent normal tissues, it is impossible to fully eliminate the effect contributed by the different cellular composition of the normal tissues. However, when comparing the percentage of M2-like macrophages in normal tissues with RNAseq data downloaded from GTEx, we observed that M2 macrophages was lower in normal brain and small intestine and similar in ovary (normal bone tissue is not available in GTEx) compared to normal breast, suggesting that the increased M2 macrophage in METs was not due to the presence of normal tissues (Additional file 2: Figure S9).Fig2Fig. 2

Paired comparison of the abundance of immune cell population in PBT/MET pairs in Pan-MET. a-b GSVA score changes (MET-PBT) of (a) Davoli signature and (b) Tamborero signature. c Abundance changes estimated by deconvolution method TIMER. d Changes of percentages relative to total immune level estimated by deconvolution method CIBERSORT. e Changes of the ratio of relative percentages of M2 and M1. ****FDR < 0.0001, ***FDR < 0.001, **FDR < 0.01, *FDR < 0.05 by Benjamini-Hochberg correction. Two-sided Wilcoxon signed rank test

To further validate in silico results, we selected three pairs of PBT/BRMs and three pairs of PBT/OVMs, which were shown to have higher M2-like macrophages relative to the total immune abundance, for multispectral immunofluorescence (Fig. 3a). Three pairs of PBT/OVMs and two pairs of PBT/BRMs showed increased macrophages in METs, and the majority of METs had lower B cells and T cells (Fig. 3b), consistent with percentage estimated from CIBERSORT (Fig. 3c and Additional file 2: Figure S10).Fig3Fig. 3

Multispectral immunohistochemical (mIHC) staining of selective pairs in Pan-MET. a mIHC staining images of one pair of PBT/OVMs and PBT/BRMs. b Percentage (by cell) of each immune cell population denoted by markers using mIHC staining. c Relative percentages of corresponding immune cell populations estimated by CIBEROSRT

To examine the contribution of each clinical variable, we tested the association between immune level (at PBT, MET and their changes) and all clinical variables available (Additional file 7: Table S8, Additional file 8: Table S9). Both the RNAseq and the sTIL dataset revealed that HR+ PBTs have significantly lower immune scores than HR- PBTs (Fig. 4a). Further, HR+ METs tended to have a smaller decrease in immune abundance compared to PBTs, although this was only significant in the BRM-sTIL dataset. However, stratifying tumors by HR and HER2 status revealed that METs in all categories had lower immune level than paired PBTs (Additional file 2: Figure S1), indicating that decreased immune is not entirely due to HR status. On the other hand, therapies were also strongly associated with the immune level, but they were highly related to tumor subtypes − 94% of ER+ cases received endocrine therapy; 64% HER2+ cases and 6% HER2- patients received HER2 treatment; 87% of all cases received chemotherapy. Due to the heterogeneity of the treatments, and the association with subtype, it is not possible to correct for this confounding variable.Fig4Fig. 4

Association of immune abundance with clinical variables and survivals. a Association between immune score and sTILs with clinical variables. b Association between survivals and immune score of PBT/BRM pairs in (b) Pan-MET dataset and (c) BRM-sTIL dataset. ****p < 0.0001, ***p < 0.001, **p < 0.01, *p < 0.05 from Wilcoxon signed rank and Kruskal-Wallis test in (a) and log-rank test in (b)-(c)

We hypothesized that immune level of PBT may be associated with metastasis-free-survival (MFS), while immune level of MET and its change from PBT to MET are potentially associated with survival-post-metastasis (SPM). Combining all PBT/MET pairs into one cohort, immune score was not significantly associated with MFS or SPM (Additional file 2: Figure S11), likely due to the confounding effect of different MET sites on outcome. Considering PBT/BRM pairs had the largest sample size, we tested the potential association between immune score and survival specifically in PBT/BRMs. In the pan-MET dataset, there was a trend in association between higher immune levels in PBTs and longer time to development of BRMs (i.e. MFS) (Fig. 4b). However, such a trend was not observed between SPM with immune levels in BRM or immune level change between PBT and BRM (Fig. 4b). In the BRM-sTIL dataset, higher sTILs percentage in PBT was not associated with MFS. Instead, there was a trend toward an association between a higher sTILs percentage in MET and longer SPM (Fig. 4c). We did not observe significant associations between the relative level of M2-like macrophage and survivals (Additional file 2: Figure S12).