Gene expression and clinical data from 374 metastatic samples including brain, bone, liver and lung and 348 normal samples from the same locations were collected. A total of 16 datasets for metastases (GSE100534, GSE101607, GSE10961, GSE44660, GSE11078, GSE12630, GSE14017, GSE14018, GSE43837, GSE46141, GSE14108.1, GSE14108.2, GSE40367, GSE41258, GSE50496.1, GSE85258) were downloaded from open repository Gene Expression Omnibus (GEO). The clinical information and description of datasets is summarized in online supplementary table 1. Metastatic samples were originated from six different primary locations (breast, colon, non-small-cell lung cancer (NSCLC), kidney, prostate and skin melanoma). Frequencies of location of metastases are in general in agreement with those previously reported,9 although a slight enrichment in brain metastatic samples existed due to over-representation in some of the datasets (online supplementary figure 1). In addition, a total of 348 samples from normal healthy tissue were collected from five GEO datasets (GSE7307, GSE45878, GSE803, GSE3526, GSE1133; online supplementary table 2).

Normalized gene expression data from the 16 datasets comprising six different microarray platforms were joined and transformed to log-2 scale. An adjustment for reduction of the batch effect was performed with ComBat function from R package sva. Online supplementary figure 2 shows a principal component analysis of the samples based on gene expression data, showing the efficacy of this method to generate a homogenized dataset. Data were analyzed separately for metastatic and normal samples.

Gene expression data were used to characterize the immune microenvironment of samples, using a variety of bioinformatics tools. The immunophenoscore (IPS) function was used to measure the immune state of the samples. IPS uses a number of markers of immune response or immune toleration to quantify and visualize four different immunophenotypes in a tumor sample (antigen presentation, effector cells, suppressor cells and checkpoint markers). It also generates a z-score summarizing these four categories. The higher the z-score of IPS, the more immunogenic the sample.14 To estimate the presence of immune cell populations in the metastatic tissues, two different tools were used. First, R package ESTIMATE was used.15 ESTIMATE (Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data) is a tool that predicts the tumor purity from gene signatures and calculates three scores: (1) stromal score—predicts the presence of stromal cell types in tumor bulk, (2) immune score—infers the infiltration of immune cells in tumor tissue and (3) estimate score—estimation of the tumor purity. Then, to obtain a more detailed picture of immune cell type infiltration, R package MCPcounter was used.16 MCPcounter (Microenvironment Cell Populations-counter) is a method for quantification of immune cell’s relative abundances in heterogeneous tissues using marker genes optimized for interrogating microarray data. Nine different cell types were interrogated (T cells, cytotoxic T cells, NK cells, B lineage, monocytic lineage, myeloid dendritic cells, neutrophils, endothelial cells and fibroblasts). Data were also interrogated with QuantiSeq (absolute method)17 and xCell (relative method)18 tools that also estimate immune cell infiltration.

For all the obtained scores, assumptions of normality and homoscedasticity were interrogated through statistical tests Levene and Shapiro, respectively. All comparisons between variables were analyzed using non-parametric tests (Kruskal-Wallis and Wilcoxon tests), for homogenization of methods, as variables were either not normally distributed, or variances were not equally distributed between groups. For all tests applied, differences were considered significant when p<0.05. To probe the lack of correlation of the scores between each metastatic sample and its related primary site of origin, a comparison between the different primary sites of origin was performed, for each metastatic site.

To make a cluster analysis separating samples on the basis of their immune status, a total of 25 immunity-related gene sets covering both innate and adaptive responses were manually selected from pathways’ databases and publications (detailed in online supplementary table 3). Gene Set Variation Analysis from R package GSVA was performed to obtain the immune profile of the metastatic samples. This function performs a non-parametric, unsupervised analysis for estimating variation of the given gene sets through the samples in the expression matrix, returning an enrichment score for each sample. GSVA function was performed with 1000 bootstraps and arguments as default.

The resulting GSVA enrichment scores were then used to cluster the samples by agglomerative hierarchical clustering. First, samples distances were computed via the R function dist, with Euclidean distance. Next, hclust function generated a clustering from the distances, with “Ward-D2” linkage method, where dissimilarities are squared for obtaining more accurate clustering. The same process was performed for gene set distances. Three categories, characterized by different immune and inflammatory enrichments, were defined and named as “Low ImmuneCluster (LIC),” “Medium ImmuneCluster (MIC)” and “High ImmuneCluster (HIC).” Function cutree (dendextend package) was used to divide the dedrogram tree in the three groups. Finally, for visualization purposes, a heatmap was plotted with the representation of the 374 samples and 25 gene sets scores, previously scaled and centered. A dendrogram was drawn to visualize the distance tree for samples and gene sets. Proportions of the different tumor metastases among the three ImmuneClusters were tested by χ² test of proportions and plotted as barplot of percentages. Figure 1 shows a summary of the analysis performed from sample collection to immune microenvironment characterization.

Gene expression data from samples in Liu et al 19 were downloaded for validation purposes. Metastatic samples (n=111) were classified into the three ImmuneClusters. Overall survival (OS) after immunotherapy treatment was plotted using a Kaplan-Meier curve stratifying by HIC, MIC and LIC.

To calculate the Normal Tissue Signatures, normalized and ComBat-adjusted expression data for the 348 healthy samples were used. A differential expression analysis among the four healthy tissues was performed with R package Limma to generate a gene profile exclusive for each tissue type (brain, bone, liver and lung). All metastatic samples were scored with these signatures and used to explore the presence of healthy tissue contamination. Also, Normal Tissue Signatures were added as a continuous covariate in the model matrix for the Combat function. The resulting adjusted expression matrix values were used to recalculate immune scores in all metastatic samples.

To identify enrichment in specific cellular functions and pathways, a Gene Set Enrichment Analysis (GSEA) was performed comparing samples belonging to the extreme phenotypes HIC and LIC.20 Gene sets from MsigDB were interrogated (Hallmarks, Gene Ontologies, Oncogenic Pathways, Immunologic Pathways and Canonical Pathway that include the datasets KEGG, Reactome and Biocarta).

Also, to predict their putative response to anti PD-L1 drug, metastatic samples were scored with the GSVA method using the T-cell inflammatory (TIS) signature. This is a genetic profile reported as a good predictor of clinical response to pembrolizumab across a wide variety of tumor types.21

A classification algorithm was performed to find genes classifying samples into extreme phenotypes HIC and LIC. Data were divided into training (75% of samples, n=137) and test dataset (25% of samples, n=46). The Training dataset was used for predictor discovery and supervised classification to generate a plausible model. This division was carried out randomly and respecting the LIC/HIC proportions. The training datasets was used to identify differentially expressed genes (DEGs) between LIC and HIC groups. For this, an empirical Bayes analysis with R package Limma was performed using Benjamini and Hochberg’s method for false discovery rate correction. DEGs between these two extreme phenotypes were selected as those with log2 fold change (log2FC)> abs(2) and adjusted p-value <0.01. The DEGs identified using the training set were used in a binary decision tree with cross-validation (k=5) to identify an optimal classification model for LIC/HIC. The classification was made with the R package caret. The classification accuracy was evaluated by calculating the sensitivity, specificity, likelihood ratio (LR) and area under the curve (AUC). Finally, the predictive power of the selected decision tree was validated in an independent dataset (GSE51244). This dataset is compressed by 94 metastatic samples (lung and liver) from colorectal cancer (CRC). First, samples were classified into the three ImmuneClusters by our algorithm. Then, the samples were classified as CD74 high/CD74 low, by the median value of expression. Finally, the agreement between the CD74 high and HIC category was evaluated.

All data and R code used for the analysis are freely available at GitHub repository https://github.com/odap-ubs/mets-immunecluster.

Gene expression data were used to categorize metastatic samples according to their immune status. First, IPS scores were used as a general indicator of immune system activation across samples. Metastases in lung showed a higher IPS z-score than those in bone, brain or liver (p=0.00006), suggesting a different immune microenvironment modulation (figure 2A). Specifically, lung metastases showed higher scores for antigen presentation (p=0.00002) and effector cells (p=0.0002), whereas they showed the lower score for suppressor cells (p=0.002). Interestingly, no differences across metastatic locations were found in the immune checkpoint category (figure 2B).

Next, ESTIMATE software was used to interrogate samples about their stromal and immune cell infiltration. Metastasis in lung and bone showed more abundance of stromal cells. In agreement with the IPS results, lung metastases scored better in the immune category (p=0.0002; figure 3A). To explore this issue in detail, MCPcounter, a tool for the quantification of different immune cell populations was used. As a result, lung metastases scored high in B lineage (p=0.0001), cytotoxic lymphocytes (p=0.001), myeloid dendritic cells (p=0.004), endothelial cells (p=0.006) and T-cell categories (p=0.03). No differences were found in NK cells, monocytic lineage and neutrophil categories. Interestingly, both lung and bone metastases showed an enrichment in fibroblast in comparison with brain and liver samples (p=0.00001; figure 3B). To validate this results, other tools apart from MCPcounter were used. In agreement with our results, absolute method QuantiSeq validates B-cell and T-cell abundance (specifically CD8+ T cells) in lung metastases, whereas relative xCell algorithm validates lung metastases enrichment in dendritic and endothelial cells. Neither QuantiSeq nor xCell validates differences in cancer associated fibroblasts (CAF) infiltration. QuantiSeq detected monocytes as higher infiltrated in lung versus liver metastases (online supplementary table 4).

To exclude the possibility that normal tissue from surgical margins were biasing our results, profiles for healthy tissue gene expression in bone, brain, liver and lung using transcriptomic data were constructed. Strong differences were found when comparing metastatic and normal tissues suggesting that little contamination existed, if any (online supplementary figure 3). Only lung metastases showed some resemblance with lung normal tissue. However, no correlation was found between the IPS score and the normal tissue score in lung metastatic samples thus excluding the possibility that samples showing higher immune scores had more normal tissue contamination (online supplementary figure 4A). Moreover, immune characterization was repeated adjusting all samples’ gene expression by their normal tissue score and almost the same results were obtained (online supplementary figure 4B and C).

Finally, we wondered if differences between the seeding organs existed. Within metastatic samples in brain and liver, no differences existed in IPS scores neither in infiltration composition regardless of their primary site or origin (online supplementary figure 5A and B). This suggested that, to some extent, immune adaptation to the new environment is a mechanism shared across tumors metastasizing in brain and liver. However, differences existed in lung and bone metastases. In lung, only slight differences were found in B lineage infiltration score (p=0.02; online supplementary figure 5C). In bone metastases, significant and stronger differences existed since samples from colorectal and kidney primary tumors have higher levels of antigen presentation and effector cells score but lower suppressor cells score. Also, they have more lymphocytes and myeloid dendritic cell infiltration (online supplementary figure 5D). This result suggested that the primary origin of bone metastases is indeed affecting the immune phenotype of their subsequent metastases.

To further explore the existence of metastases with a hot immune phenotype, manually curated gene sets related to both adaptive and innate immune responses were used to perform a hierarchical clustering for the 374 samples. Three groups emerged categorized as “HIC” (19%), “MIC” (51%) and “LIC” (30%) (figure 4A). Online supplementary figure 6A summarizes the GSVA scores for each metastatic site, lung being the most immunogenic one in agreement with previous results. Indeed, when proportion of metastases for each ImmuneCluster were represented, the HIC was enriched in lung metastases, whereas MIC was in brain and LIC in liver (figure 4B). When primary sites of origin were compared with the three clusters, kidney cancer type showed an enrichment of HIC samples (probably reflecting the enrichment of lung metastatic ones) whereas prostate was the one with the lower. Finally, normal tissues were also scored using this approach but pairs of normal-metastatic tissues did not cluster together (online supplementary figure 6B).

One might expect metastases to be immune cold, as being very aggressive tumors. In agreement, the HIC was the one comprising the less number of samples. This was an interesting group of samples characterized by an elevated expression of the human leukocyte antigen class II (HLA-II) complex and genes involved in antigen presentation pathway. Also, they showed high levels of the therapeutic targets PD-L1 and CTLA4, thus suggesting a hypothetic treatment with immune checkpoint inhibitors. Supporting this hypothesis, all samples belonging to HIC scored very high in the TIS signature reporting to be correlated with response to anti PD-L1 checkpoint inhibitor pembrolizumab (p<0.001; figure 5). Moreover, independent dataset comprising metastatic samples from patients with melanoma treated with immune checkpoint inhibitors were classified into the ImmuneClusters and interrogated about survival. Interestingly, samples classified as HIC and MIC showed a tendency toward better OS than LIC samples (figure 6A). Half lung metastases were classified as HIC (figure 6B). Although number of samples was too small to reach a conclusion, it is worth that two out of the total four lung metastases experienced progressive disease, whereas one experienced stable disease and one achieved a complete response. On the contrary, six out of seven brain metastasis suffered progressive disease and only one exhibited a partial response. Although speculative, these results pave the way to a hypothetic treatment of the highly inflammatory metastatic tumors (figure 7).

Samples in the three clusters were interrogated about their level of immune infiltration. As expected, the HIC scored better in all categories being these tumors highly infiltrated in all categories both activating and inhibiting the immune system. It is interesting to note that all bone metastasis in the HIC scored very high in the T-cell category (online supplementary figure 7A). However, when markers of exhaustion were examined, samples belonging to HIC showed the higher levels of expression (online supplementary figure 7B). This result suggested that although highly infiltrated by T cells, these are not functional but exhausted.

Then, a functional analysis comparing samples belonging to the extreme phenotypes HIC and LIC was done using GSEA. As expected, HIC samples were highly enriched in immunity-related pathways and in inflammation ones such as interleukin (IL)-2–STAT5, IL-6-JAK-STAT3, interferon, tumor necrosis factor (TNF)-α, TNF-γ and nuclear factor kappa-light-chain-enhancer of activated B cells (NF-κB), among others. Also, HIC samples were enriched in antigen presentation, Toll like receptor 4 (TLR4) signaling and CTLA4 pathways. Regarding cellular functions, inflammatory response and adaptive immune response emerged as the most significant ones (online supplementary table 5 and online supplementary figure 8). Since those samples were initially clustered on the basis of their immune phenotypes, these results were not surprising at all. However, the functional analysis also reported interesting results not directly related to the immune system. For example, apoptosis and KRAS signaling pathways were upregulated pathways in HIC samples.

The more differentially expressed genes (DEG) among ImmuneClusters were selected (log2FC > abs(2), n=43; online supplementary table 6, online supplementary figure 9) and used to search for a biomarker or panel of biomarkers useful to identify samples belonging to the HIC. A decision tree algorithm with bootstrapping selected CD74 as the gene that best categorizes between the two groups (online supplementary figure 10A). The model was first validated on the test samples within our dataset, with good prediction values (sensitivity=1, specificity=0.96, LR(+)=9.8, LR(−)=0.2, AUC=0.98, receiver operating characteristic curve in online supplementary figure 10B). When tested in an external dataset, CD74 showed high accuracy classifying samples into LIC and HIC (classification error=0.1).

Interestingly, when samples belonging to MIC were reclassified, 76 out of 191 samples belonging to the MIC were classified as HIC, whereas 113 were classified as LIC in our dataset. In the validation dataset, 14 out of 18 samples belonging to this cluster were classified as HIC, whereas 4 were classified as LIC. This suggested that a percentage of metastasis showing and intermediate phenotype could resemble to the highly infiltrated ones. Indeed, when interrogated using the TIS score, about half of metastases classified as MIC scored very high suggesting a putative response to immunotherapy (figure 5). Thus, CD74 might be marker of inflammatory metastases.