Distinct tumor architectures and microenvironments for the initiation of breast cancer metastasis in the brain

Summary

Brain metastasis, a serious complication of cancer, hinges on the initial survival, microenvironment adaptation, and outgrowth of disseminated cancer cells. To understand the early stages of brain colonization, we investigated two prevalent sources of cerebral relapse, triple-negative (TNBC) and HER2+ (HER2BC) breast cancers. Using mouse models and human tissue samples, we found that these tumor types colonize the brain, with a preference for distinctive tumor architectures, stromal interfaces, and autocrine programs. TNBC models tend to form perivascular sheaths with diffusive contact with astrocytes and microglia. In contrast, HER2BC models tend to form compact spheroids driven by autonomous tenascin C production, segregating stromal cells to the periphery. Single-cell transcriptomics of the tumor microenvironment revealed that these architectures evoke differential Alzheimer’s disease-associated microglia (DAM) responses and engagement of the GAS6 receptor AXL. The spatial features of the two modes of brain colonization have relevance for leveraging the stroma to treat brain metastasis.

Graphical Abstract

In brief

Gan et al. reveal distinct tumor architectures, stromal interfaces, and tumor-intrinsic mechanisms that are preferentially adopted by prevalent brain metastases during the initiation of colonization. These findings highlight critical spatial and microenvironmental features that can be leveraged to target minimal residual disease in the brain.

Introduction

Brain metastasis is an ominous form of cancer progression, with severe neurological complications, dismal survival rates, and limited treatment options^1,2. It is the most common malignancy in the central nervous system (CNS)³, and frequently occurs in patients with breast cancer, lung cancer, and melanoma¹. The risk of brain metastasis depends on the specific tumor type. For example, 20%-30% of patients with the HER2+ (HER2BC) or triple-negative (TNBC) basal breast cancer subtypes develop brain metastasis, whereas patients with luminal breast cancer subtypes do so infrequently (less than 10% of cases)^4,5. Therapeutic approaches based on leveraging immune components of the cerebral parenchyma have been hampered by a lack of knowledge about the brain metastatic stroma, particularly in the context of minimal residual disease.

Despite the prevalence and poor prognosis in patients, brain metastasis is a highly inefficient process at the cellular level, with the majority of disseminated cancer cells succumbing to physical, metabolic, or immunologic challenges⁶. To elucidate how disseminated cancer cells overcome these challenges, studies on mouse models and clinical tissue samples have identified metabolic adaptation of metastatic cells in the brain^7–9 and the molecular mediators underlying their interactions with vasculature^10–13, astrocytes^14–18, microglia^19–24, neurons²⁵, and infiltrated immune cells^26–28, focusing predominantly on advanced macrometastatic lesions^29–33. However, the crucial early stages of brain metastatic colonization, when elimination is the predominant fate of disseminated cancer cells and their survival is in the balance, remain obscure^19,21,34. In particular, the spatial features of brain metastatic colony formation and their role in disease progression are unknown.

Here we focus on the initiation of brain colonization by the two breast cancer subtypes with highest brain metastasis incidence, TNBC and HER2BC. We report two forms of brain colony architecture – perivascular versus spheroidal – that are adopted by TNBC and HER2BC with varying preferences. These two colonization patterns create distinct spatial interaction with the brain parenchyma, distinguished by infiltrative and restricted tumor microenvironment (TME) interfaces, respectively. Focusing on microglia as a prominent, highly reactive immune component of the TME, we show that TNBC and HER2BC both activate microglial responses characteristic of Alzheimer’s disease, yet with differences tied to the tumor-derived extracellular matrix (ECM) composition that promotes spheroidal brain colonization in HER2BC. Our findings illuminate distinctive strategies of brain colonization, microglial engagement, and intrinsic drivers in two major breast cancer subtypes, highlighting the importance of tumor spatial considerations in future efforts to eliminate metastatic disease in the brain.

Results

Perivascular and spheroidal metastasis initiation patterns in the brain

Upon extravasating from blood capillaries, metastasis-initiating cells from various carcinoma types occupy perivascular niches to establish metastatic colonies, which is particularly apparent in brain metastasis³⁵. The brain metastasis (BrM) models that we previously developed from human H2030-BrM lung adenocarcinoma (LUAD) and MDA-MB-231-BrM (MDA231-BrM for short) TNBC cells in athymic mice, and mouse E0771-BrM TNBC cells in immunocompetent mice exemplified the predominately vascular-cooptive growth pattern (Figures 1A, 1B, S1A, S1B, and Video S1). As reported by us^11,12 and others^19,35–37, extravasated cells migrate over the abluminal surface of capillaries and spread on their basement membrane to initiate proliferation, forming sheaths over the vessels before eventually transitioning to a multi-layered colony structure. The binding of cell adhesion molecules L1CAM and β1-integrins to basement membrane laminins mediates perivascular spreading^11,12,38.

Figure 1. Perivascular and spheroidal brain colonization patterns and stromal interfaces.

(A) Representative immunofluorescence (IF) staining of vasculature (CD31) and tumor lesions formed by brain metastatic derivatives (BrM) of indicated cell lines. BrM cells were selected by inoculating primary cancer cells into the arterial circulation of mice (human cells into athymic mice, and mouse cells into syngeneic immunocompetent mice) and isolating subpopulations that preferentially metastasize to the brain parenchyma^10,100. Tumor, GFP. Scale bars, 50 μm. (B) Percent of spheroidal colonies in indicated brain metastasis models (n = 5-6 mice/group). Box plot shows the median, first and third quartiles, and whiskers extend to the minimum and maximum. Two-way ANOVA test with Tukey’s multiple comparison test correction. (C) Representative IF staining showing the distribution of astrocytes (GFAP) and microglia and macrophages (IBA1) in indicated models. Tumor, GFP. Scale bars, 50 μm. (D-E) (D) Quantification and (E) and representative IF staining showing variable degrees of brain vasculature engagement in minuscule foci of patient metastatic breast cancer cells. See Table 1 for clinical information and STAR Methods for quantification details. (E) Cancer cells surround (top panel) or form clusters (bottom panel) adjacent to blood vessels (CD31), captured in longitudinal or transverse orientations as indicated. Tumor, cytokeratin. Scale bars, 25 μm. (D) Box plot summarizes the quantification of relative distances of metastatic cancer cells to their closest blood vessels in individual foci of cancer cells. Each data point represents the median relative distance among cells from a focus of cancer cells, with all corresponding values shown in Figure S1E. The tumor foci in (E) are highlighted with matching index numbers in Figure S1E. Box plot shows the median, first and third quartiles, and whiskers extend to the minimum and maximum. Two-way ANOVA and Tukey’s test. (F) Quantification of microglia and macrophage infiltration scores in surgically resected brain metastasis tissue samples derived from TNBC patients (n = 13) and HER2BC patients (n = 18). See Table S1 for scoring of individual samples and Figure S1F for corresponding representative staining. See also Figure S1, Table S1, and Video S1.

In contrast to this frequently observed pattern, metastatic HER2BC cells mainly grew as spheroidal colonies, as shown with human HCC1954-BrM and mouse MMTV-ErbB2-BrM models (Figures 1A, 1B, and S1A–C). Although not specifically reported by the authors, a spheroidal growth pattern was also apparent in brain micrometastases from human HER2BC cell lines JIMT-1 and SUM190^39,40, as opposed to a perivascular growth pattern for the mouse TNBC cell line 4T1-BrM³¹. In HCC1954-BrM cells, the spheroidal growth pattern emerged in small clusters consisting of as few as 4 cells (Figure S1C) and remained manifest as clusters grew larger (Figures S1A–C). Despite this pattern in incipient and established colonies, suppressing L1CAM expression inhibited HCC1954-BrM metastatic growth (Figure S1D), in line with our previous finding¹² of L1CAM being functionally important for a transient stage of vascular-cooptive survival in extravasated HER2BC cells. Taken together, these results suggested that the metastatic colonization of brain parenchyma followed distinct characteristic patterns, including a diffuse, perivascular pattern frequent in LUAD and TNBC and a previously unreported tight, spheroidal pattern enriched in HER2BC.

Infiltrative and segregated TME in perivascular and spheroidal colony patterns

These results raised interconnected questions about how the brain microenvironment reacts to the perivascular versus spheroidal incipient colonies and what drives these different modes of metastatic colonization.

To address the first of these two questions, we investigated how the distinctive brain colonization patterns spatially interact with astrocytes, microglia, and macrophages, major components of the brain TME⁴¹ (Figure 1C). In the predominantly vascular-cooptive colonies formed by MDA231-BrM and E0771-BrM TNBC cells, cancer cells were exposed to and co-mingled with GFAP⁺ astrocytes and IBA1⁺ microglia and macrophages, in agreement with previous reports¹⁴. In contrast to this infiltrative interface, both astrocytes and microglia/macrophages were largely limited from compact HCC1954-BrM and MMTV-ErbB2-BrM spheroidal colonies to their periphery. Astrocytes accumulated around the colonies without infiltrating cancer cell mass, and a dense layer of microglia/macrophages enwrapped the colonies.

To explore the clinical relevance of our findings, we analyzed brain tissue specimens harboring microscopic metastatic foci obtained from rapid autopsy of two TNBC patients and two HER2BC patients (Table 1). We observed a tumor type-associated continuum of spatial engagement with the vasculature, and specifically, a notable presence of vessel-enwrapping cancer cells in the brain parenchyma from the two TNBC patients examined and a higher occurrence of vessel-touching cell clusters in that of the two HER2BC patients. These observations suggested different predisposition for the two growth patterns by human breast cancer cells in initiating brain colonization (Figures 1D, 1E, and S1E).

Table 1.

Clinical information of rapid autopsy specimens.

Patient index	Type		Grade	Number of analyzed brain tissue blocks
	Molecular	Histological
RA-1	TNBC	Invasive ductal carcinoma, NOS	3	17
RA-2	TNBC	Invasive ductal carcinoma, NOS	3	5
RA-3	HER2+ ER− PR−	Invasive ductal carcinoma, NOS	3	2
RA-4	HER2+ ER+ PR+	Invasive ductal carcinoma, NOS	3	2

We also examined surgically resected brain metastasis tissues from patients harboring large, symptomatic brain metastases. Although representing an advanced stage of brain colonization, these specimens showed a high degree of intermingling of cancer cells with IBA1⁺ microglia and macrophages in the TNBC cases (Figures 1F and S1F; Table S1) and a more segregated interaction in HER2BC cases, where large lesions manifested as aggregated carcinoma cell clusters lacking infiltrating microglia and macrophages (Figure 1C).

High microglia reactivity in both brain colonization patterns

To further characterize the TME of these brain metastatic colonies, we adopted a metastatic niche labeling system⁴² that spatially enriches for parenchymal cells situated near the lesions (Figure 2A). We engineered MDA231-BrM cells and HCC1954-BrM cells to constitutively express, in addition to the cell-autonomous GFP, a cell membrane-permeable mCherry protein (sLP-mCherry) secreted outside of the cancer cells and taken up by adjacent cells, acting as a proximity TME label (Figures 2A, 2B, and S2A). The niche labeling system allowed us to dissociate an entire mouse brain harboring tens to hundreds of MDA231-BrM or HCC1954-BrM micrometastatic lesions and to use fluorescence-activated cell sorting (FACS) to isolate GFP⁺ mCherry⁺ cancer cells, GFP⁻ mCherry⁺ TME cells, and rest unlabeled cells from all lesions pooled together, without needing to individually resect them. We performed two independent single-cell RNA sequencing (scRNA-seq) experiments to analyze FACS-isolated cells (Figure 2C), and in the second one, included transcription and translation inhibitors during tissue harvesting and homogenizing steps to mitigate potential ex vivo perturbation to glial transcriptome⁴³.

Figure 2. Delineating TME cellular components by metastatic niche labeling and scRNA-seq analysis.

(A) Schematic of the metastatic niche labeling system⁴². (B) Representative IF staining showing HCC1954-BrM HER2BC cells (GFP⁺ mCherry⁺), proximal labeled TME cells (mCherry⁺), and distal unlabeled cells (mCherry⁻). Scale bar, 50 μm. See Figure S2A for corresponding staining in the MDA231-BrM model. (C) Schematic of the workflow for profiling single-cell transcriptome of the three indicated FACS-sorted groups of cells from dissociated metastases-bearing mouse brains. MDA231-BrM TNBC and HCC1954-BrM HER2BC samples were processed in parallel with cell hashing¹⁰¹. See STAR Methods for details. (D-F) Calling the cell types of all non-cancer cells from scRNA-seq experiments 1 and 2. (D) UMAP plot of the identified cell populations from 31657 cells. All macrophages, including BAM, BMDM, and MG, are highlighted by dashed gray lines. (E) Clustering showing the population average (indicated by color) for the marker genes of reference cell type and state clusters (columns) in each annotated cell population, denoted by the same abbreviated terms used in (D) (rows, with each row standardized to between 0 and 1). (F) Heatmap of global expression correlation (indicated by color) between annotated cell populations (rows) and the cell type and state clusters of reference single-cell transcriptome atlas of the mouse nervous system⁴⁴ (columns), z-normalized per row. Rows and columns are organized in identical order with clustermap in (E) to assist visual inspection. Cell types (BMDM, NK, NEUT, BAM, BC) absent in the reference atlas⁴⁴ are grayed out.(E, F) See STAR Methods and Table S2 for cell type annotation details. (G) Tissue dissociation-associated ex vivo activation scores (violin plot) and percent of the indicated subsets of macrophages (bar plot) among all (labeled and unlabeled) or only labeled macrophages, collected in two independent experiments (1, 2) from whole brain tissues bearing MDA231-BrM or HCC1954-BrM metastases. MG, microglia computationally identified to be in the G1 phase or not cycling. MG (cycling), microglia inferred to be cycling, given high scores of the S and G2/M phases. BAM, BMDM, abbreviated as in (D). All signature genes are listed in Table S2. See also Figures S2 and S3 and Table S2.

Using a reference single-cell transcriptome atlas of the mouse nervous system⁴⁴ and additional immune cell type markers complementing this atlas (Table S2), we identified 21 distinct cell populations among non-cancer cells (Figures 2D–F, S2B, and S3A). Given the phenotypic continuums observed in the data, we adopted the Milo framework⁴⁵ to systematically quantify phenotypic shifts between sample conditions. Milo constructed a cell-cell neighbor-graph and performed a statistical comparison of the density among different conditions across phenotypic neighborhoods in the graph (Figure S2C see STAR Methods). This analysis identified microglia, border associated macrophages, vascular leptomeningeal cells, and oligodendrocyte precursors present as reactive components within the TME (Figure S3A).

In both MDA231-BrM TNBC and HCC1954-BrM HER2BC models, microglia accounted for more than 90% of all labeled or unlabeled macrophages detected, which also encompassed infiltrating bone marrow-derived macrophages (BMDM) and border associated macrophages (Figures 2G and S3B). This result held in both scRNA-seq experiments (Figure 2G), was validated in situ by IF staining of the putative microglia-specific marker TMEM119 (Figures S3C and S3D), and quantitatively agreed with published FACS analysis of tumor-associated macrophages in MDA231 TNBC and 99LN ER+/HER2+ models²⁴. Taken together, the sLP-mCherry system combined with scRNA-seq analysis allowed us to characterize the micrometastatic TME at a spatial resolution hardly achievable with other methods. Of note, in contrast to the initial brain colonization stages focused on here²², the TME composition of surgically resected, often pretreated, brain metastases from patients with advanced breast cancer, lung cancer or melanoma included BMDM constituting more than 50% of tumor-associated macrophages^24,29.

Brain metastases trigger Alzheimer’s disease-associated microglia (DAM) responses

To delineate how microglia react to the initiation of brain metastatic colonization, we analyzed the majority non-cycling microglia (i.e., computationally assigned to the G0 or G1 cell cycle phase, Figure S2D; see STAR Methods) of each experiment (Figures 3A and S4A), excluding cell cycle variation as a potential confounder. We first grouped the interconnected phenotypic neighborhoods concordantly enriched in or depleted of both TNBC- and HER2BC-labeled microglia (Figures 3A and 3B) to determine which gene programs were differentially expressed in tumor-proximal (labeled) or distal (unlabeled) microglia. Interestingly, along with diminishing the expression of basal microglia homeostatic genes (Hexb, Cx3cr1, Tmem119, Cst3, P2ry12; blue, Figures 3C and S4B; Tables S2 and S3), brain metastases triggered Alzheimer’s disease-associated microglial (DAM) responses, as shown by the induction of both global DAM signature genes defined by scRNA-seq transcriptome⁴⁶ (black, Figures 3C and S4B) and a smaller set of canonical DAM markers⁴⁷ (red, Figures 3C and S4B). The DAM phenotype was originally identified in microglia associated with amyloid-β plaques⁴⁶, and later found connected, both transcriptionally and functionally, to subsets of microglia from various other developmental and pathological contexts^47–51.

Figure 3. Tumor-associated microglia activate canonical DAM programs.

All results computed on the non-cycling microglia from experiment 1. See Figure S4 for replication by experiment 2. (A) (Left panel) UMAP embedding of the four indicated sources of altogether 11407 cells, and (right panel) contour plots of each source in the embedding. (B) Differential abundance of TNBC-labeled and HER2BC-labeled cells compared to unlabeled cells in all phenotypic neighborhoods, overlayed on UMAP embedding of the index cells of phenotypic neighborhoods. Color and dot size represent the log fold change (logFC) and Benjamini-Hochberg (BH)-adjusted P values of differential abundance, respectively. UMAP embedding of all cells are shown in the background to facilitate visually locating the index cells. (C) Volcano plot of the logFC in gene expression against corresponding BH-adjusted P values, comparing the groups of phenotypic neighborhoods that are concordantly enriched in versus depleted of breast cancer (BC) cell-labeled non-cycling microglia, which identified 838 differentially expressed genes (DEGs, Table S3). logFC thresholds used for calling DEGs are indicated by dashed lines. (D) Normalized enrichment scores (NES) of the top five positively or negatively enriched gene sets among 838 DEGs. BH-adjusted P values < 0.005. See STAR Methods and Tables S4 for full GSEA results. (E) (Left panel) UMAP plot showing the first diffusion component (DC1) values computed on all cells by color map; and (right panel) heatmap showing fitted trends (rows, z-normalized per row across neighborhoods) for indicated neighborhood-level signature scores and labeled cell enrichment (quantified by the logFC values shown in (B)) along the DC1 values of neighborhood index cells (shown in color, left panel). (C-E) All marker genes⁴⁷ and signatures are listed in Table S2. (F) Similar to (E), heatmap showing fitted expression trends of neighborhood-level signatures and their constituent genes (indicated by left parentheses) along the DC1 values of neighborhood index cells (rows, z-normalized per row across neighborhoods). See also Figure S4 and Tables S2–S4.

We next combined a differential abundance test with diffusion component analysis^52,53 to trace shifts in the transcriptional activities of the microglia along a continuous gradient of variation. Gene set enrichment analysis (GSEA) showed that the top diffusion component (DC1), characteristic of the major phenotypic variation, correlated with the up- and down-regulation of DAM and homeostasis signatures, respectively (Figure 3D; Tables S3 and S4), indicating the homeostasis-to-DAM transition as the primary variation source. As manifest from the trends along DC1 in both experiments, the expression levels of DAM signature genes rose, those of homeostasis fell together with a drop in microglial transforming growth factor β (TGF-β)-dependent and LRRC33 (TGF-β activation)-dependent response programs – corroborating TGF-β as a key enforcer of anti-inflammatory microglial homeostasis^54,55 (Figures 3E and S4D). Concomitant with this transition, the relative abundance in TNBC-labeled microglia peaked where the expression of all well-established stage 1 DAM marker genes⁴⁷ (Apoe, Tyrobp, B2m, Trem2) and gene signature started to rise, whereas the relative abundance in HER2BC-labeled microglia showed expression of the phenotypically more advanced stage 2 DAM signature and marker genes (Csf1, Ccl6, Axl, Spp1, Cd9, Cst7, Itgax, Lpl) (Figures 3F and S4E).

Differential engagement of two DAM stages

In mouse models of Alzheimer’s disease, microglia transition from homeostasis through an intermediate stage 1 DAM state, marked by induction of stage 1 DAM markers without stage 2 DAM markers, then to stage 2 DAM with additional upregulation of lysosomal, phagocytic, and lipid metabolism genes^46,47. The synchrony in the transitions of identities and transcriptional activities of the microglia along DC1 indicated that, in our models, TNBC metastasis-associated microglia were mostly restricted to the stage 1 DAM state, whereas HER2BC metastasis-associated microglia largely progressed towards the stage 2 DAM state. Despite potential cell cycle influences, this trend was also observed in the minority of metastasis-associated microglia in S and G2/M phases (Figures S2D and S4F–H; see STAR Methods for analysis details).

The fully developed stage 2 DAM state also displayed enhanced expression of MHC class I (H2-k1, H2-d1, H2-q7) and class II (H2-ab1) genes, and pathways related to actin remodeling (e.g., Arpc3, Brk1, Cotl1), oxidative phosphorylation (Cox5b, Uqcrq, Cox7c, Cox6b1) and glycolysis (Mif, Pkm, Ldha), congruent with the activation of phagocytosis (Itgax, Cyba, Fcer1g, Axl) and lipid metabolism (Lpl, Cst7, Npc2) in stage 2 DAM⁵¹ (Figures 3E, 3F, S4D, and S4E). Additionally, by clustering all genes differentially expressed in labeled microglia based on their expression trends⁵³, we detected a subset of genes (cluster 4 in Table S3) whose expression tracked the differential abundance in TNBC-labeled microglia along DC1, distinguishing stage 1 from stage 2 DAM states in brain metastasis. Among them, Tnf, Il1b, and Ccl4 (Figure 3F) are related to NF-κB-activating inflammatory signals observed in aging⁴⁹, brain malignancies^20,56,57, and neurodegenerative disorders^58–60. Tnf and Il1b in particular are enriched in the stage 1 DAM in Alzheimer’s disease⁶¹. These cytokines are known to promote tumor growth by enhancing cancer cell survival¹⁴, angiogenesis, and vascular permeability^62,63, suggesting pro-tumorigenic effects of microglia in TNBC metastasis.

Varied roles of GAS6-AXL signaling

IF staining of tyrosine receptor kinase AXL, a functionally important marker, validated the enrichment of stage 2 DAM in HER2BC brain metastases (Figures 4A and 4B). Macrophages⁶⁴ and microglia⁶⁵ can increase the level of AXL in response to pathological stimuli and use it to phagocytose apoptotic cells (Figure S5A) and debris⁶⁶. AXL binds to the bispecific ligand GAS6^66,67, triggering the actin remodeling required for phagocytic engulfment⁶⁸. The interaction avidity intensifies, when GAS6 concurrently engages the externalized phosphatidylserine (PtdSer) on apoptotic cell membranes via its opposing domain⁶⁷. In Alzheimer’s disease, AXL expression is increased specifically in microglia having direct contact with amyloid β plaques, where it mediates the detection and engulfment of amyloid β plaques decorated with GAS6 and PtdSer⁶⁹. Notably, HCC1954-BrM colonies were demarcated by a rim of largely AXL⁺ CD68⁺ microglia (Figure 4A) that exhibited a correlatively high expression level of the actin remodeling pathway (Figures 3E, S4D, and S5B; Table S2). FACS analysis (Figures S5C–E) and IF staining (Figure S5F) of cell engulfment events⁷⁰ confirmed the presence of cancer cell phagocytosis in incipient HER2BC brain metastases in situ. In contrast, microglia associated with MDA231-BrM colonies were mostly negative for AXL (Figure 4A), in agreement with the scRNA-seq analysis showing limited abundance of stage 2 DAM in TNBC (Figures 3E and S4D). The syngeneic models presented a similar pattern (Figure 4B). Of note, AXL⁺ microglia were also detected in advanced E0771-BrM lesions (Figure S5G). Overexpressing GAS6 in both models of HER2BC cancer cells to increase the concentration of AXL ligand in the TME (Figures 4C and 4D) reduced their brain metastatic activity (Figure 4E). This suggested that the enforced surge of GAS6 potentiated the capacity of AXL⁺ microglia for eliminating cancer cells in HER2BC colonies. A similar tumor-suppressive role was recently reported for the AXL signaling-mediated microglial phagocytosis in glioblastoma⁷¹.

Figure 4. Diverse roles of GAS6-AXL signaling in MDA231 TNBC and HCC1954 HER2BC brain metastases.

(A-B) Quantification and representative IF staining of AXL in IBA1⁺ microglia/macrophages associated with brain metastatic colonies formed by indicated TNBC and HER2BC cells. (Left panels) (A) Tumor, GFP. (B) Tumor, luciferase. Scale bars, 50 μm. (Right panels) Average intensity of AXL in IBA1⁺ cells associated with GFP⁺ or luciferase⁺ cancer cells in micrometastatic colonies (< 2.5 × 10⁵ μm² in xenograft models, < 1.0 × 10⁵ μm² in syngeneic models). RFU, relative fluorescence units. n = 30-70 colonies in 2-3 mice/group. Two-tailed Mann-Whitney U test. See STAR Methods for IF staining and quantification details. (C) Relative mRNA levels of GAS6 in indicated HER2BC cells overexpressing (OE) GAS6 or control vector, measured by qRT-PCR. Mean ± SEM of technical replicates of the assay, representative of two independent experiments. (D) Representative IF staining of GAS6 in brain metastatic colonies formed by indicated HER2BC cells overexpressing (OE) GAS6 or control vector. Tumor, luciferase. Scale bar, 50 μm. (E) Effect of GAS6 OE in HCC1954-BrM cells (n = 9 mice/group, 26 days post-intracardiac inoculation) and MMTV-ErbB2-BrM cells (n = 12-13 mice/group, 16 days post-intracardiac inoculation) on brain colonization measured by ex vivo bioluminescence imaging (BLI) of the brain. Box plot shows the median, first and third quartiles, and the whiskers extend to the minimum and maximum. Two-tailed Mann-Whitney U test. (F) AXL and GAS6 expression levels in indicated human cancer cells in vivo. Log-transformed, normalized UMI counts in single cells, computed on the cancer cells from both scRNA-seq experiments. Log fold change (logFC) comparing MDA231-BrM cells to HCC1954-BrM cells: AXL, logFC = 1.264, GAS6, logFC = 2.336. BH-adjusted P values < 0.01 for both genes. (G) Effect of AXL and GAS6 shRNA knockdown in MDA231-BrM cells (two shRNAs per target gene) on brain colonization measured by ex vivo BLI of the brain (n = 6-7 mice/group, 26 days post-intracardiac inoculation). Box plot shows the median, first and third quartiles, and whiskers extend to the minimum and maximum. Two-tailed Mann-Whitney U test. See also Figures S5–S7.

Aside from being expressed and mediating phagocytosis in macrophages and microglia, AXL is also expressed in human TNBC cells⁷² (Figure S6A), where its activation stimulates pro-survival PI3K-AKT-mTOR pathway (Figure S6B), enhances cell growth and proliferation signaling (Figures S6B and S6C), and promotes migration and invasion (Figure S6D)^73,74. Although increased AXL expression has been noted in HER2BC cells that undergo EMT⁷⁵, pan-human breast cancer cell line analysis of transcriptome and proteome (Figure S6A, DepMap database) revealed a specific enrichment of AXL and a significantly higher level of its ligand GAS6 in TNBC cells. Moreover, AXL and GAS6 transcript levels were positively correlated in TNBC cells (Figure S6A, top panel). Aligned with the pan-cell line analysis, MDA231-BrM cells showed abundant expression of AXL and GAS6 in vitro and in vivo (Figures 4F and S7A), in contrast to significantly lower levels in HCC1954-BrM at both transcript and protein levels (Figure S7A). Knocking down the expression of either gene in MDA231-BrM cells (Figure S7B) inhibited their brain colonization activity (Figures 4G) without altering the architecture of surviving colonies (Figure S7C), implying that AXL and its ligand GAS6 promoted TNBC brain metastasis by forming an autocrine loop sustaining the growth in a perivascular mode. In contrast, further reducing the low endogenous expression levels of AXL and GAS6 in HCC1954-BrM cells did not significantly affect brain colonization (Figure S7D). Thus, GAS6 in the TME activated pro-tumorigenic AXL signaling in MDA231-BrM colonies, while in stage 2 DAM in HER2BC models, GAS6 overexpression activated anti-tumorigenic AXL signaling, implying varied roles of GAS6-AXL signaling in the initiation of brain metastatic colonization by human breast cancer cells.

Heightened expression of ECM components in brain metastatic HER2BC

Next we investigated the relationship between the architecture of the nascent metastatic colonies and the microglia response elicited by these colonies. To this end, we sought to identify molecular drivers of colony morphology and the effect of these drivers on microglia, in particular the more advanced stage 2 DAM associated with HER2BC colonies.

We first compared the in situ transcriptomes of MDA231-BrM cells and HCC1954-BrM cells, using the Flura-seq technique to tag nascent cancer cell transcripts for rapid purification and sequencing⁷⁶ (Figure S8A; see STAR Methods). The analysis identified high activity of the AKT pathway, a central effector of HER2 signaling, and mesenchymal and ECM assembly signatures enriched in HCC1954-BrM cells (Figure S8B). Profiling in vitro transcriptomes¹⁰ detected ECM organization as a differentially expressed gene set in the BrM derivatives compared to parental counterparts in both HCC1954 and MMTV-ErbB2 HER2BC models, and not in the MDA231 TNBC model (Figures 5A and 5B). Among the 32 genes concordantly upregulated in the BrM derivatives of both HER2BC models (Table S5), COL4A1, CST6, TNC, SEMA7A, IL24, and S100A7A encode components of the matrisome, an ensemble of ECM proteins (core matrisome, including TNC, an ECM component of stem cell niches⁷⁷), ECM-modifying enzymes, ECM-binding growth factors, and other ECM-associated proteins⁷⁸. The expression levels of these matrisome genes displayed a trend of association with relapse in HER2BC patients⁷⁹ (Figure 5C). Taken together, these results suggested that ECM assembly was correlated with an inherent ability of HER2BC cells to metastasize to the brain.

Figure 5. Heightened gene expression of ECM components in HER2BC brain metastases.

(A) Schematic of the workflow of isolating, culturing, and comparing the transcriptome of BrM derivatives to parental breast cancer cell lines. (B) GSEA comparing indicated BrM derivatives to corresponding parental cell lines. Color shades indicate BH-adjusted P values of normalized enrichment scores. (C) Forest plots showing the hazard ratio of the expression of matrisome genes for relapse-free survival of HER2BC breast cancer patients. P values from left (S100A7A) to right (COL4A1) are as follows: 0.0889, 0.1649, 0.0641, 0.0894, 0.0135, 0.0012. (D) Schematic of the MetMap workflow using barcoded cancer cell line pools for high-throughput metastatic potential profiling (adapted from Ref.⁷). Relative metastatic potential was quantified by deep sequencing of barcode abundance from tissue. Comparing the transcriptome of in vivo brain metastases to in vitro cell culture per multiplexed cell line pool yielded the log2 fold change (log2FC) of gene expression shown in (E). (E) Relative in vivo expression, visualized by the log2FC values shown in a descending order, of the top ECM component genes differentially upregulated in brain metastasis samples composed predominantly of HER2BC cells (pink) than of TNBC cells (blue). Wilcoxon rank sum test (see STAR Methods for details). (F) GSEA showing the enrichment of ECM-related pathways in multiplexed brain metastasis samples composed predominantly of HER2BC cells (denoted in pink in (E)) than of TNBC cells (blue in (E)). Top four positively enriched gene sets: Gene Ontology (GO) terms of 1, collagen containing ECM (GO:0062023), 2, ECM (GO:0031012), 3, extracellular structure organization (GO:0043062), and 4, ECM structural constituent (GO:0005201). See also Figure S8 and Table S5.

As an orthogonal approach, we queried the metastasis map (MetMap) dataset that combines systematically mapped brain metastatic potential and expression patterns of 21 breast cancer cell lines from the Cancer Cell Line Encyclopedia, including three HER2BC cell lines (HCC1954, HCC1569, JIMT-1) and 18 TNBC cell lines⁷. The cell lines were engineered to express unique barcodes and inoculated into NOD-SCID-gamma (NSG) mice as multiplexed pools⁷ (Figure 5D). Most of the brain metastases formed by a pool of cells were found to consist primarily of one subtype of breast cancer as determined by barcode sequencing (Figures S8C and S8D). We could therefore attribute the in vivo expression changes from the in vitro cell culture, quantified by the log fold change in gene expression (log2FC in Figure 5E), to a dominant cell subtype (Figure S8D) to test whether the changes were associated with TNBC or HER2BC. The expression of multiple core matrisome genes was upregulated in vivo in HER2BC-dominant brain metastases but downregulated in the TNBC-dominant ones, with TNC being the top differentially HER2BC-associated one (Figure 5E). Moreover, GSEA uncovered ECM-related signatures to be highly enriched in HER2BC-brain metastases (Figure 5F). Overall, this pan-cancer cell line investigation corroborated that elevated expression of certain ECM components was closely associated with the brain metastatic potential of HER2BC.

Cancer cell TNC drives spheroidal colonization and stage 2 DAM in HER2BC brain metastasis

Focusing on TNC, we confirmed strong accumulation in both intercellular and peritumoral spaces in HCC1954 and MMTV-ErbB2 incipient colonies, enclosing the microglia in direct contact with, yet not penetrating, dense HER2BC spheroids to the interior of a net of TNC aggregates (Figures 6A and 6B). Analysis of patient-derived advanced metastasis tissues also showed higher TNC accumulation in HER2BC lesions than in TNBC lesions (Figure 6C).

Figure 6. Impact of peritumoral TNC deposition on spheroidal HER2BC brain colonization.

(A-B) Quantification and representative IF staining comparing peritumoral TNC deposition in indicated HER2BC or TNBC brain metastases (colonies of human cancer cells and mouse cancer cells formed in athymic mice and corresponding syngeneic mice, respectively). (A) Tumor, GFP. (B) Tumor, luciferase. Scale bars, 50 μm. (Right panels) Tumor-associated TNC signal per colony (see STAR Methods for quantification). RFU, relative fluorescence units. n = 10-40 colonies in 2-3 mice/group. Two-tailed Mann-Whitney U test. (C) Scoring and representative IHC staining of TNC levels in surgically resected brain metastasis tissue samples derived from TNBC patients (n = 13) and HER2BC patients (n = 18). Corresponding H&E staining shown in Figure S1F. T, tumor regions (encircled by white dotted lines) were annotated by a pathologist (T.A.B.). Scale bars, 200 μm. Two-tailed Mann-Whitney U test. (D) Effect of suppressing TNC expression on oncosphere formation by shRNA in HCC1954-BrM cells and by CRISPRi in MMTV-ErbB2-BrM, measured by the size of colonies after five days of growth in vitro. Two shRNAs or two sgRNAs. n = 100-175 colonies/group. Mean ± SEM. Unpaired t test. (E) Effect of the suppression of TNC expression on brain colonization by shRNA in HCC1954-BrM cells (n = 5 mice/group, 22 days after intracardiac inoculation into athymic mice, two shRNAs) and by CRISPRi in MMTV-ErbB2-BrM cells (n = 5-9 mice/group, 21 days post-intracardiac inoculation into athymic mice, two sgRNAs), both measured by the normalized ex vivo BLI signal of the brain. Box plot shows the median, first and third quartiles, and whiskers extend to the minimum and maximum. Unpaired t test. (F-H) (F, G) Quantification and (H) representative IF staining of brain metastatic colonies formed by indicated HER2BC cells expressing either a control vector or the shRNA or sgRNA that depletes TNC expression with highest efficiency. The number of colonies was normalized to the mean of the control. (F) The number of colonies and (G) percent of vascular-cooptive cells were determined 21 and 7 days post-intracardiac inoculation into athymic mice, respectively. n = 3-4 mice/group, including all colonies from 1/10 of the brain. Mean ± SEM. Unpaired t test. (H) Tumor, GFP. Scale bars, 25 μm. See also Figure S8 and Table S1.

Suppressing TNC expression (Figures S8E and S8F) attenuated the growth of HCC1954-BrM cells and MMTV-ErbB2-BrM cells as oncospheres in 3D culture in vitro (Figure 6D), and reduced brain metastatic burden (Figure 6E) and the number of incipient colonies (Figure 6F) in both models in immunodeficient mice. These results demonstrated a tumor-intrinsic pro-metastatic role of TNC in vivo. Although HCC1954-BrM cells (but not MMTV-ErbB2-BrM cells) formed colonies in the lungs when inoculated via the tail vein, the knockdown of TNC expression did not decrease their lung colonization activity (Figure S8G), suggesting that the pro-metastatic function of TNC in HER2BC cells was most critical in the brain. We previously reported that TNC is important for lung metastasis but not brain metastasis of MDA231 cells⁸⁰, indicating that TNC may promote different organotropic metastases depending on the subtype of breast cancer. Moreover, we observed a preponderance of vascular cooption in HER2BC cells with suppressed TNC expression (Figures 6G and 6H), which preceded significant reduction in the total number of colonies detected (Figure 6F). Collectively, these data corroborated the role for TNC as a key mediator of brain-tropic metastasis and spheroidal colony formation in HER2BC.

In exploring potential links between tumor architecture and microglial reactivity, we observed markedly lower AXL levels in microglia neighboring TNC-depleted HCC1954-BrM and MMTV-ErbB2-BrM colonies (Figures 7A and 7B). Additionally, treating primary microglia in vitro with recombinant TNC protein for an extended period of 18 hours blunted the expression of all canonical homeostasis markers (Hexb, Cx3cr1, Tmem119, Cst3, P2ry12), downregulated the expression of majority stage 1 DAM markers (Apoe, Tyrobp, Trem2), and enhanced the expression of multiple stage 2 DAM markers (Ccl6, Axl, Cst7, Itgax, Lpl) as well as MHC class genes (H2-k1, H2-ab1, H2-q7) (Figure 7C, left panel). The transcriptional responses depended on Toll-like receptor 4 (TLR4) (Figure 7C, left panel), which is thought to directly recognize TNC as a ligand activating downstream NF-κB signaling⁸¹. Additionally, the responses initiated with an immediate plunge in homeostatic marker expression (exemplified by Tmem119) and a concurrent surge in interferon β transcript (Ifnb1) and protein (IFN-β) production (Figure 7D). The induction of type I IFN response genes (Isg15, Ifitm3) ensued, along with a rise in DAM (B2m, Axl) and MHC class I (H2-q7) expression levels, which peaked at 12 hours before moderately diminishing during the TNC treatment (Figure 7D). Analysis of the two scRNA-seq experiments consistently identified a distinct subset of microglia with moderate Axl level and marked upregulation of type I IFN-responsive genes (Figures 7E and S9A, bottom panels; Table S6), situated as an intermediate state along the primary homeostasis-to-DAM transition axis (Figures 7E and S9A, top panels; see STAR Methods for analysis details). Of interest, similar microglia were recently identified as key players in neuronal development⁸² and pathologies^58,71, indicative of a conserved role of type I IFN signaling. Building on the established knowledge of AXL transcriptional regulation⁸³, the in vitro real-time trajectories (Figure 7D) and in vivo pseudo-time inferences (Figures 7E and S9A) together suggested the plausible model that TNC/TLR4 signaling-provoked IFN-β served as one of the direct signals for upregulating specific stage 2 DAM genes, notably Axl, as supported by the expression changes following a 6-hour recombinant IFN-β treatment (Figure 7C, right panel). IF analysis of both incipient and advanced metastases in rapid autopsy brain tissues revealed that the IBA1⁺ cells in niches enriched in TNC deposits, which were more abundant in HER2BC lesions, exhibited enhanced upregulation of AXL (Figures S9B–D), consistent with the findings in mouse models. Collectively, these observations indicated that tumor-derived TNC acted not only as an intrinsic driver of spheroidal colonization but also as a trigger of stage 2 DAM, which represented a chronic response to prolonged TNC exposure in establishing HER2BC colonies (Figure 7F).

Figure 7. Cancer cell-derived TNC induces stage 2 DAM in HER2BC brain metastases.

(A-B) (B) Quantification and (A) representative IF staining of AXL in IBA1⁺ microglia/macrophages associated with brain metastatic colonies formed by HCC1954-BrM cells (21 days post-intracardiac inoculation into athymic mice) and MMTV-ErbB2-BrM cells (14 days post-intracardiac inoculation into FVB/N mice), expressing either a control vector or the shRNA or sgRNA that depletes TNC expression with highest efficiency. Two-tailed Mann-Whitney U test. (A) Tumor, GFP. Scale bars, 50 μm. (B) n = 40-70 colonies in 2-3 mice/group. (C) Relative mRNA levels, measured by qRT-PCR, of the indicated homeostasis and DAM marker genes in primary mouse microglia treated (left panel) with recombinant TNC protein (2 μg/mL) and/or TLR4 inhibitor (TLR4i, 5 μM) for 18 hours or (right panel) with recombinant IFN-β (20 ng/mL) for 6 hours. For each in vitro treatment, the experimental group (right panel) or groups (left panel) were compared to the untreated control group normalized to 1. Mean ± SEM of technical replicates of the assay, representative of two independent experiments. (D) Relative mRNA levels of specified genes and IFN-β concentrations in culture supernatants, measured by qRT-PCR and ELISA respectively, in primary mouse microglia treated with recombinant TNC protein (2 μg/mL) at various time points. For Ifnb1, as the transcript level of the untreated group was below the detection limit of qRT-PCR, the 3, 6, 12, and 18-hour time points were compared to the 24-hour time point set to 1, with the untreated group assigned a value of 0. For all other genes, each time point was compared to the untreated group normalized to 1. Mean ± SEM of technical replicates of the assay, representative of two independent experiments. Error bars are not visible due to their small size. Original data available at Mendeley. (E) A type I IFN-responsive cluster identified in non-cycling microglia from scRNA-seq experiment 1. See Figure S9A for replication in experiment 2. (Top panel) Heatmap of cluster-averaged Axl expression (log-transformed, normalized UMI counts) and homeostasis, DAM, and type I IFN response signature scores (all signature genes listed in Table S2). Phonograph clusters were organized along increasing DAM scores (columns). Expression of each feature was standardized across clusters to between 0 and 1 (rows). (Bottom panel) Volcano plot of the logFC in gene expression against corresponding BH-adjusted P values, comparing the IFN-responsive Phenograph cluster 7 to the rest (see Table S6 for results). Gray dashed lines indicate the absolute logFC threshold values set to 0.322. (F) Schematic illustrates distinctive modes of colonization and stromal interface, various cancer cell-intrinsic mediators of colonization, and the induction of distinct DAM responses during the initiation of brain colonization in TNBC and HER2BC. See also Figure S9 and Tables S2 and S6.

Discussion

The results of this work illuminate two modes of tumor-stroma interplay employed by prevalent subtypes of breast cancer and other cancers to commence metastatic colonization of the brain. The two modes present distinct tumor architectures in mouse models and human clinical samples of micrometastatic disease, and are driven by different tumor-ECM interactions, autocrine growth mechanisms, and engagement of a highly reactive stroma to gain advantage (Figure 7F). Microglia segregation and TNC accumulation traits that distinguish these two brain metastasis initiation modes are also histologically discernable in brain macrometastatic lesions. As advances in magnetic resonance imaging for detecting small lesions open the potential for early treatment of brain metastases^84,85, our findings imply that subtype-associated colony architecture and interface with the TME are important determinants in the progression and potentially the treatment of brain metastasis.

The perivascular colony architecture characteristic of TNBC brain metastases, which allows better access to oxygen, nutrients, and basement membrane anchorage for survival and outgrowth⁸⁶, has been considered the representative mode of brain metastases initiation across multiple types of cancer^11,12,35,36. Real-time in vivo imaging of lung adenocarcinoma cells showed how simultaneous growth and fusion of proximal perivascular micrometatases gives rise to dense, globular macrometastatic lesions³⁸. Less commonly observed in the brain is the tight, spheroidal growth that HER2BC cells assume soon after infiltration, which is an archetypal growth pattern in primary tumors^87,88. The predominantly infiltrative or segregated interfaces with the TME likely stem from the inherent spatial configuration of tumor architectures. While these colony phenotypes are respectively prevalent in the models of TNBC and HER2BC brain metastasis we studied, the full range of brain metastatic cancers likely includes intermediate forms between these phenotypes as well as heterogenous and continuous distribution of the forms within each cancer.

In searching for cancer cell-intrinsic drivers for brain metastasis of HER2BC, we detected a robust ECM deposition program, comprising TNC, periostin, fibronectin, and multiple collagens, with TNC emerging as a component of particular interest. TNC is highly expressed during embryonic development and in regions of active neurogenesis in the adult CNS, where neuronal development is tightly orchestrated by ECM components⁸⁹. TNC, fibronectin, and periostin directly bind to each other and with integrin cell adhesion receptors^90–92 to induce stemness-related signaling pathways in development and wound healing^80,93. We show that TNC knockdown shifts incipient HER2BC brain lesions into the perivascular mode and reduces their metastatic growth. These concurrent changes imply that prolonged post-extravasation spreading on the vasculature, while favorable for TNBC expansion in the brain, is detrimental to the survival of HER2BC cells, and that the deposition of ECM components enables these cells to exit the transient vascular cooption and adopt a spheroidal growth mode more advantageous to colonize the brain.

TNBC and HER2BC incipient brain metastases activated non-identical, albeit partly overlapping, DAM responses originally defined in Alzheimer’s disease, with inflammatory IL1⁺ stage 1 DAM prevailing in TNBC and phagocytic AXL⁺ stage 2 DAM in HER2BC. Expression of selected DAM marker genes has been noted in bulk-averaged transcriptome of microglia from mouse brain tissues harboring LUAD metastatases²². Additionally, our analysis of a dataset from surgically resected pan-cancer human brain macrometastases uncovered heterogenous enrichment of various DAM genes in tumor-associated macrophages, including genes annotated as canonical markers (like APOE, TREM2, SPP1, and AXL), and also those that have been reported by multiple DAM studies^94,95 (such as APOC1, C1AQ, and IL1B), all of which are significantly upregulated in the metastasis-associated microglia in our BrM models. Such expression patterns support the DAM phenotype as a prominent, common feature of the TME in brain metastases in patients and mouse models. Moreover, the TNC-dependent induction of stage 2 DAM we found in HER2BC brain metastasis aligns with previous reports of TNC serving as an endogenous ligand to the pattern recognition receptors on microglia and macrophages in other diseases, including stroke⁹⁶, rheumatoid arthritis⁹⁷, and TNBC lung metastasis⁹⁸. These instances of TNC activation may relate to the conserved macrophage reactivity to ECM remodeling during tissue regeneration⁹⁹.

Our GAS6-AXL findings point at tumor-microglia interactions with therapeutic implications. The diffuse contact of TME cells with perivascular TNBC colonies may facilitate access to stroma-derived GAS6 supplementing the cancer cell-derived GAS6 to trigger pro-survival GAS6-AXL signaling in cancer cells. Inhibiting this signaling presents an option for targeting brain metastases in TNBC-like cases. In contrast, the stromal restriction in spheroidal HER2BC colonies may serve as a protective barrier limiting the exposure of cancer cell mass to AXL⁺ phagocytic stage 2 DAM, a potential liability introduced by tumor-derived TNC. Here, enhancing the GAS6-AXL signaling in microglia may suppress tumor growth in cases that resemble these HER2BC brain metastases, as recently shown in the context of glioblastoma immunotherapy⁷¹.

Limitations of the study

It should be noted that the sLP-mCherry niche labeling system provided a practical, yet possibly cell type-biased, approach for enriching the TME cells of micrometastases, as other cell types may be less efficient than microglia in taking up sLP-mCherry to be gated as mCherry⁺ in FACS. The scarcity of rapid autopsy tissue samples underscores the importance of sustained biobanking initiatives to enable identifying the largely unknown determinants of incipient metastatic colonization in the clinical context. The proposed model of cancer cell TNC deposit-orchestrated DAM program awaits systematic dissection in the endogenous context. Two additional questions warranting future work are the epitranscriptomic landscapes responsible for the heightened expression of ECM components in HER2BC epithelial progenitors and the molecular mechanisms by which TNC and other ECM components identified drive spheroidal growth for the initiation of metastatic colonization. The present insights on DAM programs in incipient colonization prompt future systematic studies on how different cancer types and DAM states influence each other in various disease stages and clinically relevant contexts.

STAR★METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact Joan Massagué (MassaguJ@mskcc.org).

Materials availability

All unique reagents generated in this study, including plasmids and cancer cell lines, are available from the lead contact with a completed Materials Transfer Agreement.

Data and code availability

Bulk and single-cell RNA-seq data have been deposited in the Gene Expression Omnibus database. Accession numbers are listed in the key resources table. Other original data presented in main and supplemental figures are publicly accessible at Mendeley (https://doi.org/10.17632/4nhh9cp6xw.1). Codes for conducting the scRNA-seq analysis are accessible at GitHub (https://github.com/dpeerlab/Brain-metastasis-TME). All software programs used for analyses are publicly available and listed in the key resources table.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies and conjugates
Chicken anti-Green Fluorescent Protein antibody	Aves Labs	Cat #: GFP-1010 RRID:AB_2307313
Chicken anti-Green Fluorescent Protein antibody	Abcam	Cat #: ab13970 RRID:AB_300798
Rabbit anti-Green Fluorescent Protein antibody	Thermo Fisher	Cat #: A11122 RRID:AB_221569
Chicken anti-mCherry antibody	Abcam	Cat #: ab205402 RRID:AB_2722769
Rabbit anti-RFP antibody	Rockland	Cat #: 600-401-379 RRID:AB_2209751
Rat anti-CD68 antibody	BioRad	Cat #: MCA1957 RRID:AB_322219
Rabbit anti-CD68 antibody	Boster	Cat #: PA1518 RRID:AB_2813855
Rabbit anti-IBA1 antibody	Abcam	Cat #: ab178847 RRID:AB 2832244
Rabbit anti-IBA1 antibody	Wako Chemicals	Cat #: 01919741 RRID:AB_839504
Goat anti-IBA1 antibody	Invitrogen	Cat #: PA518039 RRID:AB_10982846
Rat anti-CD31 antibody	BD Biosciences	Cat #: BDB550274 RRID:AB_393571
Goat anti-CD31 antibody	R&D Systems	Cat #: AF3628 RRID:AB_2161028
Goat anti-mAxl antibody	R&D Systems	Cat #:AF854 RRID:AB_355663
Goat anti-mGas6 antibody	R&D Systems	Cat #: AF986-SP RRID:AB_3076301
Goat anti-hGas6 antibody	R&D Systems	Cat #: AF885-SP RRID:AB_3076302
Rabbit anti-Tenascin C antibody	Millipore	Cat #: AB19011 RRID:AB_2203804
Rabbit anti-Tenascin C antibody	Abcam	Cat #: ab108930 RRID:AB_10865908
Rat anti-Tenascin C antibody	R&D Systems	Cat #: MAB2138 RRID:AB_2203818
Rabbit anti-Collagen type IV antibody	Serotec	Cat #: 21501470 RRID:AB_2082660
Rat anti-GFAP antibody	Thermo Fisher	Cat #: 130300 RRID:AB_2532994
Rat anti-BrdU antibody	Abcam	Cat #: ab6326 RRID:AB_305426
Mouse anti-Human Cytokeratin antibody	Agilent Technologies	Cat #: M351501-2 RRID:AB_2631307
Rabbit anti-Firefly Luciferase antibody	Abcam	Cat #: ab185924 RRID:AB_2938620
Mouse anti-V5 antibody	Thermo Fisher	Cat #: R960-25 RRID:AB_2556564
Rabbit anti-TMEM119 antibody	Abcam	Cat #: ab209064 RRID:AB_2800343
Rat anti-P2RY12 antibody	BioLegend	Cat #: 848001 RRID:AB_2650633
Leica Bond Polymer anti-Rabbit HRP secondary antibody	Leica Biosystems	Cat #: DS9800 RRID:AB_2891238
CF® Dye 430	Biotium	Cat #:96053
CF® Dye 594	Biotium	Cat #: 92174
Alexa-Fluor 488 Donkey anti-Chicken	Jackson ImmunoResearch	Cat #: 703545155 RRID:AB_2340375
Alexa-Fluor 488 Donkey anti-Goat	Thermo Fisher	Cat #: A32814 RRID:AB_2762838
Alexa-Fluor 546 Donkey anti-Goat	Thermo Fisher	Cat #: A11056 RRID:AB_2534103
Alexa-Fluor 546 Donkey anti-Mouse	Thermo Fisher	Cat #: A10036 RRID:AB_2534012
Alexa-Fluor 568 Donkey anti-Rabbit	Thermo Fisher	Cat #: A10042 RRID:AB_2534017
Alexa-Fluor 647 Donkey anti-Chicken	Jackson ImmunoResearch	Cat #: 703605155 RRID:AB_2340379
Alexa-Fluor 647 Donkey anti-Rat	Thermo Fisher	Cat #: A48272 RRID:AB_2893138
Alexa-Fluor 647 Donkey anti-Rabbit	Thermo Fisher	Cat #: A31573 RRID:AB_2538183
Alexa-Fluor 750 Donkey anti-Goat	Abcam	Cat #: ab175745 RRID:AB_2924800
Alexa-Fluor 750 Donkey anti-Rabbit	Abcam	Cat #: ab175728 RRID:AB_2924801
Human TruStain FcX (Fc receptor blocking solution)	BioLegend	Cat #: 422301 RRID:AB_2818986
Anti-mouse CD16/32	BioLegend	Cat #: 101319 RRID:AB_1574973
TotalSeq-A0253 anti-human Hashtag 3 antibody	BioLegend	Cat #: 394605 RRID:AB_2750017
TotalSeq-A0255 anti-human Hashtag 5 antibody	BioLegend	Cat #: 394609 RRID:AB_2750019
TotalSeq-A0256 anti-human Hashtag 6 antibody	BioLegend	Cat #: 394611 RRID:AB_2750020
TotalSeq-A0258 anti-human Hashtag 8 antibody	BioLegend	Cat #: 394615 RRID:AB_2750022
TotalSeq-A0306 anti-mouse Hashtag 6 antibody	BioLegend	Cat #: 155811 RRID:AB_2750037
TotalSeq-A0307 anti-mouse Hashtag 7 antibody	BioLegend	Cat #: 155813 RRID:AB_2750039
TotalSeq-A0308 anti-mouse Hashtag 8 antibody	BioLegend	Cat #: 155815 RRID:AB_2750040
FITC Annexin V	BioLegend	Cat #: 649020
APC/Fire™ 750 anti-mouse/human CD11b [M1/70]	BioLegend	Cat #: 101261 RRID:AB_2572121
BB700 Rat anti-mouse CD45 Clone 30-F11 (RUO)	BD Biosciences	Cat #: 566439 RRID:AB_2744406
Brilliant Violet 605™ anti-mouse Ly-6G/Ly-6C (Gr-1)	BioLegend	Cat #: 108440 RRID:AB_2563311
PerCP/Cyanine5.5 anti-mouse CD45 Clone 30-F11	BioLegend	Cat #: 103131 RRID:AB_893344
Brilliant Violet 711™ anti-mouse Ly-6C Clone HK1.4	BioLegend	Cat #: 128037 RRID:AB_2562630
APC-Cyanine7 Anti-Human/Mouse CD11b [M1/70]	Tonbo Biosciences	Cat #: 25-0112-U100 RRID:AB_2621625
Bacterial and virus strains
Stbl3 competent E. coli	ThermoFisher	Cat #: C737303
endA competent E. coli	New England Biolabs	Cat #: C3040H
Biological samples
Brain metastasis tissue samples derived from TNBC patients (n = 13) and HER2BC patients (n = 18)	Department of Pathology, MSKCC
Chemicals, peptides, and recombinant proteins
Dulbecco’s Modified Eagle’s high glucose medium	Media Preparation Core, MSKCC	Powder Cat #: 52100047
Dulbecco’s Phosphate-Buffered Saline, no calcium, no magnesium	Media Preparation Core, MSKCC	Powder Cat #: 21600044
Roswell Park Memorial Institute 1640 medium	Media Preparation Core, MSKCC	Powder Cat #: 31800105
Fetal Bovine Serum	Sigma Aldrich	Cat #: F2442
L-glutamine	Thermo Fisher	Cat #: 25030081
Penicillin-Streptomycin	Thermo Fisher	Cat #: 15140163
Amphotericin B	Gemini Bio-Products	Cat #: 400104
DiD’ solid; DiIC18(5) solid (1,1’-Dioctadecyl-3,3,3’,3’-Tetramethylindodicarbocyanine, 4-Chlorobenzenesulfonate Salt)	Thermo Fisher	Cat #: D7757
B-27 Supplement, serum free	Thermo Fisher	Cat #: 17504001
Human Recombinant bFGF, ACF	StemCell Technologies	Cat #: 02634
Human EGF Recombinant Protein	Thermo Fisher	Cat #: PHG0311
D-luciferin, Potassium Salt	GoldBio	Cat #: LUCK-10G
Heparin sodium salt	Sigma Aldrich	Cat #: H3393
Paraformaldehyde Aqueous Solution	Electron Microscopy Sciences	Cat #: 15710S
Tissue-Tek O.C.T Compound	Sakura	Cat #: 4583
Isoflurane Solution	Covetrus	Cat #: 029405
Sucrose	Fisher Scientific	Cat #: S53
Polyethylene glycol	Millipore Sigma	Cat #: 8170025000
Glycerol	Thermo Fisher	Cat #: BP2291
β-Mercaptoethanol	Sigma Aldrich	Cat #: M3148100ML
Ethanol	Fisher Scientific	Cat #: 04-355-222
Lipofectamine 2000	Thermo Fisher	Cat #: 11668019
Opti-MEM	Thermo Fisher	Cat #: 31985062
Lenti-X Concentrator	Clontech	Cat #: 631231
Polybrene	Santa Cruz Biotechnology	Cat #: sc134220
G 418 disulfate salt solution	Thermo Fisher	Cat #: 10131035
Puromycin dihydrochloride	Sigma Aldrich	Cat #: P962010ML
Actinomycin D	Sigma Aldrich	Cat #: A1410
Triptolide	Sigma Aldrich	Cat #: T3652
Anisomycin from Streptomyces griseolus	Sigma Aldrich	Cat #: A9789
5-Fluorocytosine	Sigma Aldrich	Cat #: F71291G
SeaPlaque Agarose	Lonza	Cat #: 50100
BsmBI-v2	New England Biolabs	Cat #: R07395
CellTrace Calcein Violet, AM, for 405 nm	Thermo Fisher	Cat #: C34858
CellTracker™ Green CMFDA Dye	Thermo Fisher	Cat #: C7025
CypHer5E NHS Ester	Cytiva	Cat #: PA15401
DAPI (4’,6-Diamidino-2-Phenylindole Dilactate)	Thermo Fisher	Cat #: D3571
DAPI (4’,6-Diamidino-2-Phenylindole Dilactate)	Sigma Aldrich	Cat #: D9542
Hoechst 33342 solution, 20 mM	Thermo Fisher	Cat #: 62249
Cultrex 3-D culture matrix reduced factor basement membrane extract (RGF BME)	Fisher Scientific	Cat #: 344500501
Fisher BioReagents™ Bovine Serum Albumin (BSA) DNase- and Protease-free Powder	Thermo Fisher Scientific	Cat #: BP9706100
Advanced DMEM/F-12	Fisher Scientific	Cat #: 12634028
Gibco™ GlutaMAX™ Supplement	Fisher Scientific	Cat #: 35-050-061
R&D Systems™ Mouse M-CSF Recombinant Protein	Fisher Scientific	Cat #: 416ML010
Recombinant Mouse IFN-beta Protein	R&D Systems	Cat #: 8234-MB-010/CF
Human Tenascin-C Purified Protein	Millipore Sigma	Cat #: CC065
BOND Epitope Retrieval Solution 2	Leica Biosystems	Cat #: AR9640
Mowiol 4-88	Calbiochem	Cat #: 475904
Aprotinin	R&D Systems	Cat #: 4139/10
Leupeptin hemisulfate	Tocris	Cat #: 1167
cOmplete™, Mini, EDTA-free Protease Inhibitor Cocktail	Roche	Cat #: 11836170001
Halt™ phosphatase inhibitor	Thermo Fisher	Cat #: 78427
Staurosporine solution from Streptomyces sp.	Sigma-Aldrich	Cat #: S6942
Gibco™ HBSS, no calcium, no magnesium	Fisher Scientific	Cat #: 14170112
Sample Diluent Concentrate 2 (2X)	R&D Systems	Cat #: DYC002
Substrate Reagent Pack	R&D Systems	Cat #: DY999
Stop Solution, 2N Sulfuric Acid	R&D Systems	Cat #: DY994
R428	MedChemExpress	Cat #: HY-15150
Critical commercial assays
RNeasy Mini Kit	QIAGEN	Cat #: 74106
RNeasy Micro Kit	QIAGEN	Cat #: 74004
QIAshredder	QIAGEN	Cat #: 76956
Transcriptor First Strand cDNA Synthesis Kit	Roche	Cat #: 04897030001
SMARTer PCR cDNA Synthesis Kit	Clontech	Cat #: 634926
TruSeq RNA Sample Prep Kit v2	Illumina	Cat #: RS1222001
NEBNext Ultra RNA Library Prep Kit	New England Biolabs	Cat #: E7530S
Adult Brain Dissociation Kit	Miltenyi Biotec	Cat #: 130107677
Dead Cell Removal Kit	Miltenyi Biotec	Cat #: 130090101
Mouse IFN-beta Quantikine ELISA Kit	R&D Systems	Cat #: MIFNB0
Human Total Axl DuoSet IC ELISA Kit	R&D Systems	Cat #: DYC1643-2
Human Gas6 DuoSet ELISA	Novus Biologicals	Cat #: DY885B
Proteome Profiler-Human Phospho-Kinease Array Kit	R&D Systems	Cat #: ARY003C
CellTiter-Glo® 2.0 Cell Viability Assay	Promega	Cat #: G9242
Pierce™ BCA Protein Assay Kit	Thermo Fisher	Cat #: 23227
Deposited data
Patient Survival Datasets	Györffy et al.	PMID: 20020197
	Lánczky et al.	PMID: 27744485
Microarray Gene Expression Data	Bos et al.	GEO: GSE12237
Atlas of the Adolescent Mouse Brain	Zeisel et al.	mousebrain.org
DepMap	Broad Institute	depmap.org/portal/
MetMap	Jin et al.	GEO: GSE148283, GSE148372
Raw and processed data files for RNA-seq	This study	GEO: GSE223351
Raw and processed data files for Flura-seq	This study	GEO: GSE223247
Raw and processed data files for single-cell RNA-seq	This study	GEO: GSE223309
R and Python codes	This study	https://github.com/dpeerlab/Brain-metastasis-TME
Mendeley dataset	This study	DOI: 10.17632/snccvxy28g.1
Experimental models: Cell lines
HEK293T	ATCC	ATCC #: CRL-3216
HMC3	ATCC	ATCC #: CRL-3304
E0771-BrM	Derived in house	This paper
MDA231-BrM	Derived in house	This paper
HCC1954-BrM	Derived in house	This paper
MMTV-ErbB2-BrM	Derived in house	This paper
Experimental models: Organisms/strains
Mouse: Hsd:Athymic Nude-Fox1nu	ENVIGO	Order code: 069 RRID:ISMR_ENV:HS D-069
Mouse: Crl:NU(NCr)-Foxn1nu	Charles River	Strain Code: 490 RRID:IMSR_CRL:490
Mouse: C57BL/6J	The Jackson Laboratory	Strain #: 000664 RRID:IMSR_JAX:000664
Mouse: B6(Cg)-Tyrc-2J/J (B6 albino)	The Jackson Laboratory	Strain #: 000058 RRID: IMSR_JAX:000058
Mouse: FVB/NJ	The Jackson Laboratory	Strain #: 001800 RRID: IMSR_JAX:001800
Oligonucleotides
TaqMan human TNC (Hs01115665_m1)	Thermo Fisher	Cat #: 4453320
TaqMan human GAS6 (Hs01090305_m1)	Thermo Fisher	Cat #: 4331182
TaqMan human GAPDH (Hs02786624_g1)	Thermo Fisher	Cat #: 4331182
TaqMan human ACTB (Hs01060665_g1)	Thermo Fisher	Cat #: 4331182
TaqMan human AXL (Hs01064444_m1)	Thermo Fisher	Cat #: 4453320
TaqMan mouse Tnc (Mm00495662_m1)	Thermo Fisher	Cat #: 4453320
TaqMan mouse Gas6 (Mm00490378_m1)	Thermo Fisher	Cat #: 4453320
TaqMan mouse Gapdh (Mm99999915_g1)	Thermo Fisher	Cat #: 4331182
TaqMan mouse ActB (Mm0120567_g1)	Thermo Fisher	Cat #: 4453320
TaqMan mouse Axl (Mm00437221_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Hprt (Mm03024075_m1)	Thermo Fisher	Cat #: 4331182
Taqman mouse Apoe (Mm01307193_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse B2m (Mm00437762_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Ccl6 (Mm01302419_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Cd9 (Mm00514275_g1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Csf1 (Mm00432686_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Cst3 (Mm00438347_m1)	Thermo Fisher	Cat #:: 4448892
Taqman mouse Cst7 (Mm00438351_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Cx3cr1 (Mm02620111_s1)	Thermo Fisher	Cat #: 4453320
Taqman mouse H2-ab1 (Mm00439216_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse H2-k1 (Mm01612247_mH)	Thermo Fisher	Cat #: 4453320
Taqman mouse H2-q7 (Mm00843895_s1)	Thermo Fisher	Cat #: 4448892
Taqman mouse Hexb (Mm01282432_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Ifitm3 (Mm00847057_s1)	Thermo Fisher	Cat #:4453320
Taqman mouse Ifnb1 (Mm00439552_s1)
Taqman mouse Itgax (Mm00498701_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Isg15 (Mm01705338_s1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Lpl (Mm00434764_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse P2ry12 (Mm01950543_s1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Spp1 (Mm00436767_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Tmem119 (Mm00525305_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Tnf (Mm00443258_m1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Trem2 (Mm04209424_g1)	Thermo Fisher	Cat #: 4453320
Taqman mouse Tyrobp (Mm00449152_m1)	Thermo Fisher	Cat #: 4453320
97-mer oligonucleotide containing shTNC 1: TGCTGTTGACAGTGAGCGACAGAGGTGACATGT CAAGCAATAGTGAAGCCACAGATGTATTGCTTG ACATGTCACCTCTGCTGCCTACTGCCTCGGA	This paper	Synthesized by IDT
97-mer oligonucleotide containing shTNC 2: TGCTGTTGACAGTGAGCGACAGCTATTGACAGTT ACAGAATAGTGAAGCCACAGATGTATTCTGTAA CTGTCAATAGCTGCTGCCTACTGCCTCGGA	This paper	Synthesized by IDT
97-mer oligonucleotide containing shAXL 1: TGCTGTTGACAGTGAGCGAAAAGTCTCTAATTCT ATTAAATAGTGAAGCCACAGATGTATTTAATAG AATTAGAGACTTTGTGCCTACTGCCTCGGA	This paper	Synthesized by IDT
97-mer oligonucleotide containing shAXL 2: TGCTGTTGACAGTGAGCGCCCAAAGTCTCTAATT CTATTATAGTGAAGCCACAGATGTATAATAGAA TTAGAGACTTTGGATGCCTACTGCCTCGGA	This paper	Synthesized by IDT
97-mer oligonucleotide containing shGAS6 1: TGCTGTTGACAGTGAGCGCCCAGGAAACGGTGA AAGTGAATAGTGAAGCCACAGATGTATTCACTTT CACCGTTTCCTGGATGCCTACTGCCTCGGA	This paper	Synthesized by IDT
97-mer oligonucleotide containing shGAS6 2: TGCTGTTGACAGTGAGCGAAGCGAGGACTGTAT CATCTGATAGTGAAGCCACAGATGTATCAGATG ATACAGTCCTCGCTCTGCCTACTGCCTCGGA	This paper	Synthesized by IDT
Oligonucleotides for Tnc sgRNA 1:	This paper	Synthesized by IDT
Sense: CACCGCACACACCCTAGCCTCTGGT
Antisense: AAACACCAGAGGCTAGGGTGTGTGC
Oligonucleotides Tnc sgRNA 2:	This paper	Synthesized by IDT
Sense: CACCGACACACACCCTAGCCTCTGG
Antisense: AAACCCAGAGGCTAGGGTGTGTGTC
Oligonucleotides for Gal4 sgRNA:	This paper	Synthesized by IDT
Sense: CACCGAACGACTAGTTAGGCGTGTA
Antisense: AAACTACACGCCTAACTAGTCGTTC
Recombinant DNA
Plasmid: HSV1-TK/GFP/Fluc/	Ponomarev et al.
Plasmid: pLV[Exp]-BsdEF1A>hGAS6[NM_000820.4]	Vector Builder	Cat #: Ecoli(VB900126-3640pet)
Plasmid: pLV[Exp]-Bsd-EF1A>mGas6[NM_019521.2]	Vector Builder	Cat #: Ecoli(VB230330-1312wkh)
Plasmid: pLV[Exp]-Bsd-EF1A>ORF_Stuffer	Vector Builder	Cat #: Ecoli(VB900122-0480ezn)
Plasmid: sLP-mCherry-P2A-eGFP	Tammela Lab
Plasmid: IGI-P0492 pHR-dCas9-NLS-VPR-mCherry	Jacob Corn Lab	Addgene #102245
Plasmid: dCas9-KRAB-MECP2	Yeo et al.	Addgene #110821
Plasmid: lentiGuide-Hygro-mTagBFP2	Ho et al.	Addgene #99374
Plasmid: SGEN	Fellmann et al.	Addgene #111171
Plasmid: LENC	Fellmann et al.	Addgene #111163
Plasmid: LEPZ	Fellmann et al.	Addgene #111161
Software and algorithms
Living Image	PerkinElmer	Version 4.4
ImageJ	NIH	Version 2
FIJI	NIH	Version 2.3.0/1.53q
Imaris	Oxford Instruments Group	Version 9.5.0
STAR RNA-seq aligner	Dobin et al.	Version 2.5.3a
HTSeq	Anders et al.	Version 0.6.1p1
DESeq2	Bioconductor	Version 3.4.1
Xenome	Conway et al.	Version 1.0.1
DAVID	Huang et al.	Version 6.8
RStudio	RStudio	Version 1.2.5029
Bioconductor	RStudio	Version 3.15
Prism	GraphPad	Version 8.4.3
SEQC	https://github.com/dpeerlab/seqc	Version 0.2.1
Scanpy	Wolf et al.	Version 1.6.0
Python	Python	Version 3.8.5
CellBender	Fleming et al.	Version 0.2.0
DropletUtils	Lun et al.	Version 1.10.3
SHARP	https://github.com/hisplan/sharp	Version 0.2.1
DoubletDetection	https://github.com/dpeerlab/DoubletDetection	Version 2.5.2
PhenoGraph	Levine et al.	Version 1.5.6
Milo	Dann et al.	Version 0.1.0
EdgeR	Chen et al.	Version 3.32.1
MAST	Finak et al.	Version 1.16.0
GSEA	Mootha et al.	Version 4.0.3
Scran	Lun et al.	Version 1.14.6
Palantir	Setty et al.	Version 1.1
FIJI Wound Healing Size Tool	Suarez-Arnedo et al.	Version 2021.1.16
Slideviewer	3DHistech	Version 2.6.0.166179
FlowJo	BD Biosciences	Version 10.10.0
Gen5	Agilent	Version 2.09
Other
Clamp Lamp Light with 8.5 Inch Aluminum reflector	Simple Deluxe	Cat #: B08MZKQNP4
250W clear bulb	VWR	Cat #: 36548001
ProLong Diamond Antifade mountant	Thermo Fisher	Cat #: P36961
VWR Slides Micro Frosted	VWR	Cat #: 48312-003
Fisherbrand™ Superfrost™ Plus microscope slides	Fisher Scientific	Cat #: 12-550-15
Fisherbrand Premium cover glasses	Fisher Scientific	Cat #: 125485M
Corning Falcon Standard tissue culture dishes	Fisher Scientific	Cat #: 08772E
BioLite™ cell culture treated dishes	Thermo Fisher	Cat #: 130183
Falcon™ Polystyrene microplates	Fisher Scientific	Cat #: 07-772-1
Single-use Syringe/BD PrecisioGlide needle	VWR	Cat #: BD309625
Myelin Removal Beads II, human, mouse, rat	Miltenyi Biotec	Cat #: 130090101
Oligo (dT)25 magnetic beads	New England Biolabs	Cat #: S1419S
4-well Nunc Lab-Tek II chambers	Thermo Scientific	Cat #: 154453
Autoradiographic Film 5×7”	LabScientific	Cat #: VAR ALF 1318

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Cell lines

MDA231 and MMTV-ErbB2, including parental cells and BrM derivatives, as well as HEK293T cells were cultured in Dulbecco’s Modified Eagle’s (DME) high glucose medium (Media Preparation Core, MSKCC, Cat: 52100047). HCC1954 and E0771, including parental cells and BrM derivatives, were cultured in Roswell Park Memorial Institute (RPMI) 1640 medium (Media Preparation Core, MSKCC, Cat: 31800105). Both medium were supplemented with 10% fetal bovine serum (FBS) (Sigma Aldrich, Cat: F2442), unless otherwise specified, 2 mM L-glutamine (Gln) (Thermo Fisher Scientific, Cat: 25030081), 100 IU/mL penicillin/streptomycin (P/S) (Thermo Fisher Scientific, Cat: 15140163), and 1 μg/mL amphotericin B (Gemini Bio-Products, Cat: 400-104). All parental cancer cells were female in origin. Human microglial HMC3 cells were cultured in Advanced DMEM/F-12 (Thermo Fisher Scientific, Cat: 12634028) supplemented with 2 mM GlutaMAX (Fisher Scientific, Cat: 35-050-061), 10% FBS, and 100 IU/mL P/S. All cells were obtained from ATCC and grown in a humidified incubator at 37 °C with 5% CO₂ and verified to be mycoplasma-free on a monthly basis.

Animals

All animal experiments were conducted in accordance with a protocol approved by the Memorial Sloan Kettering Cancer Center (MSKCC) Institutional Animal Care and Use Committee (IACUC). Athymic nude (Hsd:Athymic Nude-Foxn1^nu, ENVIGO, order code: 069) female mice aged between 4-6 weeks were used for metastatic colonization assays of MDA231-BrM cells and HCC1954-BrM cells. B6(Cg)-Tyrc-2J/J (B6 albino mice, strain #: 000058) and FVB/NJ (strain #: 001800) female mice aged between 4-6 weeks, both from The Jackson Laboratory, were used for metastatic colonization assays of E0771-BrM cells and MMTV-ErbB2-BrM cells, respectively.

Surgically resected clinical samples and immunohistochemistry

All human tissues were obtained under MSKCC Institutional Review Board biospecimen research protocol 15-101. All patients provided pre-procedure informed consent. The archival formalin-fixed, paraffin-embedded (FFPE) brain metastasis of HER2BC (18 cases) and TNBC (13 cases) clinical tissue blocks used for immunostaining were identified by database search and chart review. IHC was performed by the MSKCC Molecular Cytology Core using antibodies against Tenascin C (Millipore, Cat: AB19011, 0.5 μg/mL) and IBA1 (Abcam, Cat: ab178847, 0.2 μg/mL). Tissue processing and histopathological data interpretation were overseen by an expert breast cancer pathologist (E.B.).

Rapid autopsy samples

All human tissues were obtained under the Last Wish Program (LWP), a MSKCC Institutional Review Board biospecimen research protocol 15-021 that coordinates the collection, annotation, and storage of high-quality postmortem tissue samples. All patient participants of LWP provided informed consent prior to their passing. The FFPE brain tissue blocks of two HER2BC patients and two TNBC patients that contained small tumor foci were identified by the database search and review of H&E slides from archived tissue blocks by an experienced pathologist (R.M., director of LWP), assisted by the analysis of MRI scans of the patients by an expert neurosurgeon (N.S.M.). IF staining of recut sections from the identified FFPE blocks was performed as described below (see IF staining and imaging of paraffin-embedded sections).

METHOD DETAILS

Brain metastatic cell isolation

Brain-tropic metastatic derivatives (BrM) of cell lines MDA231, HCC1954, and MMTV-ErbB2 were established as previously described^10,11,100. BrM derivatives of cell line E0771, expressing a V5-tagged firefly luciferase reporter, were generated and provided by J. Remsik in the Adrienne Biore Lab (MSKCC)¹⁰².

Animal studies

Brain metastatic colonization assays were performed as follows: mice were anaesthetized by 100 mg/kg ketamine and 10 mg/kg xylazine. 1.0 × 10⁵ BrM cells resuspended in 100 μL ice-cold phosphate buffered saline, no calcium, no magnesium solution (PBS, Media Preparation Core, MSKCC, Cat: 21600044) supplemented with 2% FBS were injected into the left ventricle of the mice with a 26G × 3/8” needle attached to tuberculin syringe (VWR, Cat: BD309625). During the course of colonization assays, metastatic burden was monitored by non-invasive bioluminescence imaging (BLI) in vivo using the Xenogen IVIS-200 system. Specifically, mice were anaesthetized in an induction chamber connected to isoflurane (Covetrus, Cat: 029405), and imaged within 2-4 minutes after retro-orbital injection of 1.5 mg D-luciferin (GoldBio, Cat: LUCK-10G) dissolved in 100 μL PBS. After a designated period of time post-injection indicated in the figure legends for individual experiments, mouse brains were collected for histological analysis. Briefly, mice were imaged by BLI in vivo as described above (see Brain metastatic cell isolation), euthanized by CO₂, and transcardially perfused with 10 mL PBS containing 1.5 mg D-luciferin potassium salt and 10 mg/L heparin sodium salt (Sigma Aldrich, Cat: H3393). Whole brains were immediately isolated, imaged by BLI ex vivo, and subsequently incubated with rotation at 4 °C first in 4% paraformaldehyde (PFA) (Electron Microscopy Sciences, Cat: 15710-S) for 24 hours and then in 70% (v/v) ethanol (EtOH, Fisher Scientific, Cat: 04-355-222) or 30% (w/v) sucrose (Fisher Scientific, Cat: S53) in PBS for 48 hours, with three PBS washes in between. EtOH-preserved brains were processed and embedded in paraffin by HistoWiz. Subsequently, the embedded tissues were coronally sectioned into 5 μm-thick slices and mounted on microscope slides (Fisher Scientific, Cat: 12-550-15), which were then stored at 4 °C. Sucrose-preserved brains were embedded into Tissue-Tek O.C.T. compound (Sakura, Cat: 4583), mounted onto the platform of a sliding microtome (Thermo Fisher Scientific, Cat: HM450), frozen to −30 °C. 80 μm-thick slices were coronally sectioned, and serially stored in ten 2 mL volume centrifuge tubes (USA Scientific, Cat: 14209704) at −20 °C in anti-freezing solution, containing 30% polyethylene glycol (Millipore Sigma, Cat: 8170025000) and 30% glycerol (Thermo Fisher, Cat: BP2291) in PBS. Lung colonization assays were similarly performed except that BrM cells were injected into the lateral tail vein of the mice to initiate lung colonization. In both brain and lung colonization assays, BLI signal was measured using the ROI tool in Living Image software (PerkinElmer, version 4.4). To accommodate for moderate variation in the initial number of BrM cells inoculated per mouse in the experiments that investigated the impact of TNC on brain and lung colonization, the endpoint ex vivo BLI signal of each mouse was adjusted by a normalization factor. This factor was determined by dividing the whole-body BLI signal of a mouse measured upon intracardiac inoculation to that averaged among all mice in certain experiment.

Oncosphere culture

4-well Nunc Lab-Tek II chambers (Thermo Scientific, Cat: 154453) were coated with 200 μL Cultrex 3-D culture matrix reduced factor basement membrane extract (RGF BME) (Fisher Scientific, Cat: 344500501) per well, and incubated at 37 °C for 30 minutes to promote gelling of the matrix. Single-cell suspensions of HCC1954-BrM cells and MMTV-ErbB2-BrM cells were plated in the coated 4-well chambers at a density of 1.0 × 10⁴ cells/mL per well in corresponding culture media supplement with 2% 3-D Culture Matrix RGF BME solution, 1X B-27 (Life Technologies, Cat: 17504-001), 20 ng/mL human EGF recombinant protein (Life Technologies, Cat: PHG0311), and 20 ng/mL human recombinant bFGF (StemCell, Cat: 02634). Cells were cultured for five days and imaged with EVOS M5000 imaging system (Invitrogen, Cat: AMF5000). FIJI, an implementation of ImageJ (NIH, version 2.3.0/1.53q) was used to quantify the area of the oncospheres.

In vitro cell proliferation assay

Cell proliferation was assessed using the CellTiter-Glo 2.0 Cell Viability Assay (Promega, Cat: G9242). MDA231-BrM cells were seeded at a density of 4.0 × 10³ cells per well in 96-well plates in full-serum (10% FBS) medium and incubated overnight for attachment. Subsequently, the medium was replaced with low-serum (2% FBS) medium supplemented with 1 μM AXL inhibitor R428 (MedChemExpress, Cat: HY-15150) or an equivalent volume of DMSO (R428 solvent, Fisher Scientific, Cat: D128500). Each experimental condition was replicated in eight wells. The number of viable cells was measured at 0, 24, 48, and 72 hours, following the manufacturer’s instructions of the assay. The luminescent signal, proportional to the ATP amount and thereby indicative of the viable cell number per well, was detected by the Synergy H1 Hybrid microplate reader (Agilent) with the Agilent Gen5 software (version 2.09).

In vitro wound scratch assay

A total of 3.0 × 10⁵ MDA231-BrM cells were plated per well of a 6-well plate in full-serum (10% FBS) medium and cultured to approximately 90% confluency. A straight uniform wound was then created across the center of each well by scratching the cell monolayer using a 200 μL pipette tip. The wells were then washed twice with PBS to remove detached cells and replenished with serum-free medium, supplemented with 1 μM R428 or an equivalent volume of DMSO. Three wells were used per experimental condition. Whitefield images of wound scratch areas were captured at 0, 24, and 48 hours post-wounding on EVOS M5000 imaging system, focusing on two randomly selected areas per well for a total of six areas per condition. The fraction of open area remaining in the scratch was quantified using the FIJI Wound Healing Size Tool¹⁰³ plugin (version 2021.1.16), with settings of Variance window size 50, Threshold value 50, and Percentage of saturated pixels 0.001.

Treatment of primary adult mouse microglia culture with IFN-β or with Tenascin C and/or TLR4 inhibitor

As previously reported⁵⁴, adult primary microglia were isolated from the brains of C57BL/6 mice (The Jackson Laboratory, strain #: 000664) aged between 6-10 weeks. For each experiment, 3-4 mouse brains were harvested and processed in parallel by enzymatic dissociation using the Adult Brain Dissociation Kit, mouse and rat (Miltenyi Biotec, Cat: 130-107-677). The dissociated tissues were cleared of debris using the debris removal solution as part of the Kit. The resulting cell suspension was incubated on ice for 20 minutes with APC/Fire^™ 750 anti-mouse/human CD11b [M1/70] antibody (BioLegend, Cat: 101261, 1:200), BB700 rat anti-mouse CD45 clone 30-F11 (RUO) antibody (BD Biosciences, Cat: 566439, 1:200), and 10 μg/mL DAPI (Thermo Fisher, Cat: D3571) in staining buffer containing 0.5% (w/v) bovine serum albumin (Thermo Fisher Scientific, Cat: BP9706100) in PBS. Following two washes of staining buffer, CD11b⁺ CD45^low cells were FACS-sorted into microglia culture medium of Advanced DMEM/F-12 supplemented with GlutaMAX, 10% FBS, 100 IU/mL P/S, and mouse recombinant M-CSF (Fisher Scientific, Cat: 416ML010). Sorted primary microglia were plated into 24-well polystyrene microplates (Fisher Scientific, Cat: 08-772-1) at a density of 0.8-1.0 × 10⁵ cells per well in 1 mL of medium and cultured at 37 °C, 5% CO₂. After replacing with fresh medium on the following day, cells were cultured for at least five days without perturbation and subsequently subject to treatment of 20 ng/mL recombinant mouse IFN-β protein (R&D Systems, Cat: 8234-MB-010/CF) for 6 hours. Alternatively, cells received 2 μg/mL purified human tenascin C (TNC) protein (Millipore Sigma, Cat: CC065, 1:50 from 100 μg/mL stock in PBS) for specified durations (Figures 7C and 7D) and/or 5 μM TLR4 inhibitor (TLR4i) TAK-242 (Millipore Sigma, Cat: 614316, 1:2000 from 1 mM stock reconstituted in DMSO). Corresponding controls included equivalent volumes of PBS or DMSO. Post-treatment, cells were analyzed for gene expression as described in qRT-PCR. For the TNC treatment time-course experiment (Figure 7D), IFN-β protein production was also measured as outlined in ELISA assays. The control group received PBS for 24 hours, matching the longest duration of TNC treatment. To minimize technical variations during sample collection, TNC administration was staggered, allowing the simultaneous collection of supernatant and RNA from all time-point samples at a uniform endpoint.

In vitro engulfment of apoptotic HER2BC cells by human microglial HMC3 cells

To probe the role of AXL signaling in microglial phagocytosis, we employed FACS analysis to assess the engulfment of CypHer5E-labeled apoptotic HER2BC cells following 1-hour incubation with human microglial HMC3 cells, treated with or without the AXL inhibitor R428. The CypHer5E dye (Cytiva, Cat: PA15401) exhibits heightened fluorescence in the acidic environment typical of the endosome/lysosome pathway. Such pH sensitivity effectively minimized misidentification of HMC3-HER2BC cell doublets as phagocytic HMC3 cells, thereby facilitating robust detection of the phagocytic uptake of HER2BC cells and cellular contents. Additionally, we utilized the HMC3 model, characterized by elevated AXL expression under homeostatic conditions^61,104, to allow specifically perturbing AXL signaling using R428.

3.5 × 10⁴ HMC3 cells were plated per well in 24-well polystyrene microplates two days prior to the engulfment assay, reaching approximately 1.0 × 10⁵ at the time of the assay. After overnight incubation to allow for cell attachment, the medium was replaced with fresh medium containing 1 μM R428 or an equivalent volume of DMSO for 24 hours. Four wells were used for each condition as biological replicates. Immediately preceding incubation with apoptotic HER2BC cells, HMC3 cells were stained for 45 minutes in medium supplemented with 10 μM Hoechst 33342 (Thermo Fisher, Cat: 62249) and 10 μM CellTracker^™ Green CMFDA (Thermo Fisher, Cat: C7025), along with either R428 or DMSO. The stained HMC3 cells were then washed twice with medium to remove excess dye. To induce apoptosis in HER2BC cells as required for their effective engulfment by HMC3 cells in vitro, HCC1954-BrM cells or MMTV-ErbB2-BrM cells were treated with 100 μM staurosporine (STS; Sigma-Aldrich, Cat: S6942) for 18 hours. Both adherent and suspended HER2BC cells were collected post-treatment, stained with 1 μM CypHer5E (Cytiva, Cat: PA15401) in serum-free HBSS (Fisher Scientific, Cat: 14170112) for 45 minutes, and then incubated in full-serum HMC3 medium for 25 minutes to remove excess dye. For the engulfment assay, 5.0 × 10⁵ apoptotic HER2BC cells were introduced to each well of HMC3 cells in 400 μL of medium correspondingly supplemented with either 1 μM R428 or DMSO, matching prior treatments. After 1-hour incubation, residual HER2BC cells were removed with three 4 °C PBS washes. HMC3 cells were then harvested and analyzed on LSR Fortessa (BD Biosciences), and data were processed using FlowJo (BD Biosciences, version 10.10.0). Engulfment of HER2BC cells by HMC3 cells was evaluated by quantifying the percentage of CypHer5E⁺ events within the Hoechst⁺ CellTracker Green⁺ HMC3 population and the median CypHer5E fluorescence intensity among CypHer5E⁺ events. The CypHer5E gate was set based on a negative control consisting of HMC3 cells that did not undergo incubation with apoptotic HER2BC cells.

Detecting in vivo engulfment of HER2BC cells by metastasis-associated microglia and macrophages

To facilitate capturing HER2BC cell engulfment events in situ, we engineered HER2BC cells to express a histone 2B (H2B)-fused mCherry reporter. Specifically, HCC1954-BrM cells and MMTV-ErbB2-BrM cells were transduced with lentivirus expressing H2B-mCherry under the control of PGK promoter¹⁰⁵ Addgene #21217). The enhanced inherent stability of histone 2B leads to a slower degradation of H2B-mCherry compared to most cancer cell proteins after engulfment by IBA1⁺ microglia and macrophages. This property prolongs the timeframe during which a previous engulfment event remains detectable by IF staining (see IF staining and imaging of paraffin-embedded sections) after it has occurred.

As a complementary approach to IF staining (Figure S5F), we employed FACS analysis, which provides higher fluorescence sensitivity than microscopy, to effectively identify phagocytic microglia containing even minimal amounts of H2B-mCherry remnants from HER2BC cells. To account for autofluorescence variations arising from tumor presence when gating mCherry signal, we compared brain tissues bearing metastases from HCC1954-BrM cells or MMTV-ErbB2-BrM cells expressing the H2B-mCherry reporter, to time-matched control tissues harboring metastases from corresponding HER2BC cells without the reporter. The single-cell suspension obtained from brain tissue dissociation was incubated on ice for 30 minutes with 10 μg/mL DAPI and the antibodies listed below, using the staining buffer described alongside the dissociation process in Treatment of primary adult mouse microglia cultures with IFN-β, Tenascin C, and/or TLR4 inhibitor. After staining and subsequent washes, cells were analyzed on the LSR Fortessa with appropriate compensations applied. Data were processed using FlowJo. Microglia and the mCherry⁺ phagocytic cells within this population were gated according to the strategies outlined in Figures S5D and S5E, respectively.

Antigen	Cell population	Fluorophore	Source	Dilution
CD45	Immune cells	PerCP/Cyanine5.5	BioLegend	1:100
CD11b	Myeloid cells	APC-Cyanine7	Tonbo Biosciences	1:100
Gr-1	Neutrophils	Brilliant Violet 605™	BioLegend	1:400
Ly6C	Monocytes	Brilliant Violet 711™	BioLegend	1:400

Gene overexpression (OE)

To achieve GAS6 OE, HCC1954-BrM cells and MMTV-ErbB2-BrM cells were transduced with lentivirus expressing GAS6 under the control of EF1α promoter (for HCC1954-BrM, pLV[Exp]-BsdEF1A>hGAS6[NM_000820.4], VectorBuilder, Cat: Ecoli(VB900126-3640pet)); for MMTV-ErbB2-BrM, pLV[Exp]-Bsd-EF1A>mGas6[NM_019521.2], Cat: Ecoli(VB230330-1312wkh)). As the negative control, cells were transduced with lentivirus carrying a scrambled sequence instead (pLV[Exp]-Bsd-EF1A>ORF_Stuffer, Vector Builder, Cat: Ecoli(VB900122-0480ezn)).

Gene expression knockdown

RNA interference

shRNA-mediated knockdown of TNC, GAS6, and AXL expression in HCC1954-BrM cells and MDA231-BrM cells were performed by cloning mir-E-based shRNA sequences targeting these genes into corresponding lentiviral or retroviral vectors listed below to replace their existing shRNA sequences that target the Renilla luciferase gene as negative control¹⁰⁶. Five 97-mer oligonucleotides (IDT), each containing a different candidate shRNA sequence, were designed per gene using the web-based tool splashRNA (splashrna.mskcc.org)¹⁰⁷. qRT-PCR analysis was conducted for each engineered cell line to select the top two shRNAs with high knockdown efficiency to target particular gene as shown below.

shRNA	Plasmid backbone	Vector type	Selection marker	Cells
shTNC 1	SGEN106	Lentiviral	G 418 (500 μg/mL) (Thermo Fisher, Cat: 10131035)	HCC1954-BrM
shTNC 2	SGEN	Lentiviral	G 418 (500 μg/mL)	HCC1954-BrM
shAXL 1	LENC106	Retroviral	FACS sorting for mCherry+ cells	MDA231-BrM, HCC1954-BrM
shAXL 2	LENC	Retroviral	FACS sorting for mCherry+ cells	MDA231-BrM, HCC1954-BrM
shGAS6 1	LEPZ106	Retroviral	Puromycin dihydrochloride (5 μg/mL, Sigma Aldrich, Cat: P962010ML)	MDA231-BrM, HCC1954-BrM
shGAS6 2	LEPZ	Retroviral	Puromycin dihydrochloride (5 μg/mL)	MDA231-BrM, HCC1954-BrM

CRISPR interference (CRISPRi)

We used CRISPRi to suppress the expression of Tnc in MMTV-ErbB2-BrM cells, as shRNA knockdown (KD) of these genes was found to reduce the transcript level only by 50%. MMTV-ErbB2-BrM cells were engineered to express the two components of CRISPRi system – dCas9-KRAB and sgRNAs – via two steps of lentiviral transduction and selection. To first establish cells that stably express dCas9-KRAB, we cloned a lentiviral vector carrying transcription repressor sequence dCas9-KRAB. Specifically, we replaced the VP64-P65-Rta sequence (encoding transcription activator complex) in the lentiviral vector of dCas9-VPR-mCherry fusion protein (for CRISPRa, Addgene #102245) with the KRAB-MECP2 sequence from the transient expression vector of dCas9-KRAB-MECP2 (for CRISPRi, Addgene #110821)¹⁰⁸. After lentiviral transduction, we isolated mCherry⁺ cells by FACS sorting, which were then introduced with specific sgRNAs targeting Tnc as the second step as follows. The sgRNAs listed below, including the one targeting Gal4 DNA binding domain that was designed as a negative control, were cloned into the lentiGuide-Hygro-mTagBFP2 vector (Addgene #99374) as previously described¹⁰⁹. Specifically, in synthesizing the oligonucleotides (IDT) to clone sgRNAs, 5’CACC and 5’AAAC overhangs were added to sense and antisense oligonucleotides, respectively, to create cohesive ends in annealing. The annealed oligonucleotides were ligated into the backbone of lentiGuide-Hygro-mTagBFP2 vector, obtained by BsmBI-v2 (New England Biolabs, Cat: R07395) digestion. The mCherry⁺ tagBFP⁺ cells were FACS-sorted after the second round of lentiviral transduction, and tested for the efficiency in knocking down Tnc expression by qRT-PCR.

Target genes		sgRNA sequence
Tnc	1	5’-GCACACACCCTAGCCTCTGGT-3’
	2	5’-GACACACACCCTAGCCTCTGG-3’
Gal4		5’-GAACGACTAGTTAGGCGTGTA-3’

Immunoblotting

ELISA assays

To quantitatively compare AXL and GAS6 protein levels between human TNBC and HER2BC models, we employed ELISA assays to measure membrane-bound AXL receptor abundance in total cell lysates and secreted GAS6 ligand concentration in culture supernatant. MDA231-BrM cells and HCC1954-BrM cells, 1.0 × 10⁶ of each, were plated in 10 cm petri dishes (Fisher Scientific, Cat: 08772E) and cultured for 2.5 days to reach approximately 90% confluence without medium change. The post-culture supernatant was collected and its volume measured to account for any evaporation during culture. Cells were harvested, counted, yielding a total of approximately 4.5 × 10⁶ cells for each model, and then lysed in 0.5 mL of lysis buffer, which was prepared by appropriately diluting Sample Diluent Concentrate 2 (R&D Systems, Cat: DYC002) and supplementing with Halt^™ phosphatase inhibitor (Thermo Fisher, Cat: 78427) and cOmplete^™ protease inhibitor (Roche, Cat: 11836170001). The total protein concentration of the cell lysates was determined using the Pierce^™ BCA Protein Assay Kit (Thermo Fisher, Cat: 23227). AXL abundance was measured with the Human Total Axl DuoSet IC ELISA Kit (R&D Systems, Cat: DYC1643-2) and scaled to the total protein content in the lysates. GAS6 concentration was measured with the Human Gas6 DuoSet ELISA Kit (Novus Biologicals, Cat: DY885B). To assess the amount of GAS6 secreted per cell, the measured concentration was multiplied by the total supernatant volume, and then divided by the number of post-culture cells in the petri dish to derive the cell number-normalized value.

The production of IFN-β protein by primary microglia treated with TNC over various durations was detected using the Mouse IFN-beta Quantikine ELISA Kit (R&D Systems, Cat: MIFN80). Similar to GAS6, IFN-β levels were normalized against the count of primary microglial cells, which remained constant due to their minimally proliferative nature in culture and were quantified via FACS sorting prior to seeding (see Treatment of primary adult mouse microglia culture with IFN-β or with Tenascin C and/or TLR4 inhibitor). All ELISA assays were conducted according to the manufacturer’s instructions, with three technical replicates performed for each sample in all assays.

Phospho-kinase array assay

To efficiently probe the impact of AXL signaling on major kinase phosphorylation in TNBC, we employed the Proteome Profiler-Human Phospho-Kinase Array (R&D Systems, Cat: ARY003C), which allowed for simultaneous analysis of multiple kinases. 5.0 × 10⁶ MDA231-BrM cells were seeded in a 15 cm petri dish (Thermo Fisher, Cat: 130183) and cultured to approximately 70% confluence. The cells were then treated with either 1 μM R428 or an equivalent volume of DMSO, as the control, for 24 hours in serum-free medium. After treatment, cells were lysed directly in the dish using 1 mL of the lysis buffer described in ELISA assays, and the total protein concentration was also determined as outlined there. Lysates from the R428 treatment and control samples, each containing 360 μg of total protein, were applied to individual membranes pre-blotted with an array of phosphorylated kinase detection antibodies, with each antibody presented in technical duplicate spots. Following the incubation and wash steps per the manufacturer’s instructions, the membranes were exposed to a single X-ray film (LabScientific, Cat: XAR AFL 1318) for 10 minutes. The film was then developed using SRX-101A tabletop X-ray film processor (Konica Minolta) and scanned with Epson Perfection V600 photo scanner (Epson), and the resulting image analyzed with FIJI. The relative phosphorylation signals were quantified by measuring the total pixel densities within each spot and then the subtracting background densities of identical spot size.

Immunofluorescence (IF) staining and imaging of free-floating tissue sections

Tissue sections archived in anti-freezing solution were washed thoroughly in PBS three times to remove residual cryoprotectant. Sections were permeabilized with two washes of PBS-T, that is, PBS supplemented with 0.25% Triton X-100 (Fisher Scientific, Cat: AC215682500). To inactivate the fluorescence of GFP in cancer cells and background autofluorescence of brain tissues, sections were incubated with 30% H₂O₂ (Sigma Aldrich, Cat: 216763500ML) and 0.02 M HCl in PBS under a clamp lamp with 8.5 inch aluminum reflector (Simple Deluxe, Cat: B08MZKQNP4) and a 250W clear bulb (VWR, Cat: 36548001) for 1 hour at 4 °C. The sections were subsequently incubated in the blocking buffer of 5% normal donkey serum (Jackson ImmunoResearch, Cat: 017000121) or 10% normal goat serum (Life Technologies, Cat: 50-062Z), selected to match the host species of secondary antibodies, and 2% (w/v) bovine serum in PBS-T for 1 hour at room temperature. After blocking, sections were incubated in primary antibodies diluted in blocking buffer for overnight at 4 °C, washed in PBS-T six times, and then incubated in 1:500 secondary antibodies and 10 μg/mL DAPI diluted in blocking buffer for 2-3 hours, followed by three washes in PBS-T and three washes in PBS, all at room temperature. Each wash step was 5-10 minutes long. All incubation and wash steps were performed with gentle shaking. Antibodies and corresponding dilutions were detailed below. After the last PBS wash, sections were transferred onto 1.0 mm-thick glass slides (VWR, Cat: 48312003) and allowed to air dry until translucent, mounted with ProLong Gold Diamond antifade mountant (Thermo Fisher, Cat: P36961), and sealed with 0.13- to 0.17 mm-thick glass coverslips (Fisher Scientific, Cat: 125485M). The sealed slides were first cured at room temperature and placed at −20 °C for long-term storage. Images of the slides were taken with a TCS SP5 confocal microscope (Leica Microsystems) or a Ti2-E motorized microscope (Nikon) equipped with Crest X-Light V2 LFOV25 spinning disk confocal (Nikon) and processed as described below (see Imaging analysis).

Primary antibodies for IF	Source	Dilution
Chicken anti-Green Fluorescent Protein Antibody	Aves Labs	1:250
Rabbit anti-Green Fluorescent Protein Antibody	Thermo Fisher	1:1000
Rabbit anti-Firefly Luciferase Antibody	Abcam	1:50
Chicken anti-mCherry Antibody	Abcam	1:250
Rat anti-CD68 Antibody	BioRad	1:100
Rabbit anti-IBA1 Antibody	Wake Chemicals	1:200
Goat anti-IBA1 Antibody	Invitrogen	1:200
Rat anti-CD31 Antibody	BD Biosciences	1:100
Goat anti-CD31 Antibody	R&D Systems	1:100
Goat anti-mAxl Antibody	R&D Systems	1:50
Goat anti-mGas6 Antibody	R&D Systems	1:50
Goat anti-hGas6 Antibody	R&D Systems	1:50
Rat anti-Tenascin C Antibody	R&D Systems	1:50
Rabbit anti-Collagen type IV Antibody	Serotec	0.4 μg/mL
Rat anti-GFAP Antibody	Thermo Fisher	1:500

IF staining and imaging of paraffin-embedded sections

To facilitate robust, efficient analysis of IF signal based on a large number of brain metastatic colonies, we utilized automated IF staining and slide scanning at the Molecular Cytology Core, MSKCC. IF staining was performed on Bond RX research stainer (Leica Biosystems), in which paraffin-embedded 5 μm-thick tissue sections were first pretreated with EDTA-based epitope retrieval ER2 solution (Leica, Cat: AR9640) for 20 minutes at 95 °C. Multiplexed staining was conducted in a sequential order as indicated in the table below. For each round of staining, after 1-hour incubation of certain primary antibody, Leica Bond Polymer anti-rabbit HRP secondary antibody (Leica Biosystems, Cat: DS9800) was applied, followed by particular tyramide conjugate to detect the immunoreactivity with HRP-catalyzed signal amplification. The conjugates used include Alexa-Fluor tyramide conjugate 488 or 647 (Life Technologies, Cat: B40953, B40958) and CF^® Dye tyramide conjugate 430 or 594 (Biotium, Cat: 96053, 92174). Afterwards, epitope retrieval was performed to denature the primary and secondary antibodies used in the current round before the primary antibody of the following round was applied. After all rounds of IF staining were completed, slides were washed in PBS, incubated in 5 μg/mL DAPI (Sigma Aldrich, Cat: D9542) in PBS for 5 minutes, and rinsed in PBS and then mounted in Mowiol 4–88 (Calbiochem, Cat: 475904). In cases where multiplexed IF staining required using primary antibodies raised in the same host species, the Bond RX research stainer was also used, as its sequential staining protocol allowed staining and detecting the primary antibodies one at a time.

Slides were scanned on a Pannoramic scanner (3DHistech) using a 20X/0.8NA objective. Whole-slide scans were viewed using SlideViewer (3DHistech, version 2.6.0.166179). Representative images of individual brain metastatic lesions were taken with higher resolution on an Axioplan 2 widefield microscope (Zeiss) using a 40X/0.95NA objective and an Axiocam 506 monochrome camera (Zeiss).

IF set	Order	Primary antibodies for automated IF			Tyramide conjugates
		Antibody	Source	Dilution
1	1	Goat anti-mAxl Antibody	R&D Systems	1:50	CF® Dye 594
	2	Mouse anti-Human Cytokeratin Antibody or	Agilent Technologies	1:250	Alexa-Fluor 488
		Chicken anti-Green Fluorescent Protein Antibody or	Abcam	2 μg/mL
		Mouse anti-V5 Antibody	Thermo Fisher	1:50
	3	Rabbit anti-CD68 Antibody or	Boster	1 μg/mL	CF® Dye 430
		Rabbit anti-Tenascin C Antibody	Millipore Sigma	1:500
	4	Rabbit anti-IBA1 Antibody	Abcam	0.025 μg/mL	Alexa-Fluor 647
2	1	Rabbit anti-TMEM119 Antibody	Abcam	1:200	CF® Dye 594
	2	Mouse anti-Human Cytokeratin Antibody or	Agilent Technologies	1:250	Alexa-Fluor 488
		Chicken anti-Green Fluorescent Protein Antibody or	Abcam	2 μg/mL
		Mouse anti-V5 Antibody	Thermo Fisher	1:50
	3	Rabbit anti-IBA1 Antibody	Abcam	0.025 μg/mL	Alexa-Fluor 647
3	1	Mouse anti-Human Cytokeratin Antibody or	Agilent Technologies	1:250	Alexa-Fluor 488
		Chicken anti-Green Fluorescent Protein Antibody or	Abcam	2 μg/mL
	2	Rabbit anti-RFP Antibody	Rockland	1.25 μg/mL	CF® Dye 594
	3	Rabbit anti-IBA1 Antibody	Abcam	0.025 μg/mL	Alexa-Fluor 647

Whole-mount immunostaining and volume imaging of iDISCO-cleared brain

Immunostaining and imaging of brain hemisphere samples were conducted following the standard iDISCO protocol (idisco.info/idisco-protocol/)¹¹⁰ over four major steps, 1) pretreatment with methanol (MeOH, Fisher Scientific, Cat: A4524), 2) immunolabeling, 3) tissue clearing, and 4) imaging. Briefly, during 1) pretreatment, harvested tissue samples were dehydrated through serial incubation in 20%, 40%, 60%, 80%, and 100% MeOH/H₂O solution, each solution for 1 hour and 100% MeOH twice. After incubation with 66% Dichloromethane (DCM, Sigma Aldrich, Cat: 270997-2L)/33% MeOH overnight at room temperature, the dehydrated samples were bleached with 5% H₂O₂ in MeOH overnight at 4 °C, and subsequently rehydrated by 1-hour incubation with 80%, 60%, 40%, 20% MeOH/H₂O and then PBS. Proceeding with 2) immunolabelling, the rehydrated samples were first permeabilized with 20% DMSO, 2.3% (w/v) Glycine (Fisher Scientific, Cat: G48212), 0.2% Triton-X in PBS, and then blocked in 10% DMSO, 6% normal donkey serum, 0.2% Triton-X in PBS, each step for two days at 37 °C. After blocking, samples were incubated with primary antibodies diluted in 5% DMSO, 3% normal donkey serum in PTwH buffer, consisting of PBS with 0.2% Tween 20 (Sigma Aldrich, Cat: P13791L) and 0.001% heparin, for seven days at 37 °C, followed by five washes with PTwH buffer over one day at room temperature, and then incubated with secondary antibodies diluted in 3% normal donkey serum in PTwH buffer for seven days at 37 °C, followed by another five washes with PTwH buffer over one day at room temperature. To 3) clear the tissue, immunolabeled samples were dehydrated as in 1) pretreatment by serial incubation in 20%, 40%, 60%, 80%, and 100% MeOH/H₂O solution, each solution for 1 hour and 100% MeOH twice, and then incubated with 66% DCM/33% MeOH for 3 hours, followed by two 15-minute washes with 100% DCM to remove residual MeOH, with all steps at room temperature. Finally, samples were stored in DiBenzyl Ether (DBE, Sigma, Cat: 108014-1KG) without shaking at 4 °C until imaging. All above steps were performed with shaking except for the final storage. The 5 mL microcentrifuge tubes (Thermo Scientific, Cat: 14568101) used throughout the process were filled with buffer to the top to prevent air from oxidizing the samples. Before 4) imaging, the tubes were inverted a couple times to mix the DBE solution, and cleared samples were transferred to the chamber of a Luxendo MuVi SPIM light sheet microscope (Bruker) to be scanned with a 10X objective, 30% laser power, and in line mode with the resolution of 50 px and light sheet thickness of 3 μm.

Primary antibodies for iDISCO	Source	Primary antibody dilution	Secondary antibodies for iDISCO	Secondary antibody dilution
Goat anti-CD31 Antibody	R&D Systems	1:100	Alexa-Fluor 488 Donkey anti-Goat	1:200
Rabbit anti-IBA1 Antibody	Wako Chemicals	1:200	Alexa-Fluor 568 Donkey anti-Rabbit	1:200
Chicken anti-GFP Antibody	Aves Labs	1:250	Alexa-Fluor 647 Donkey anti-Chicken	1:150

Image processing and analysis

Confocal microscopy images were minimally processed with FIJI. Representative images were displayed with linear adjustments of contrast and brightness, using identical adjustment settings where fluorescent signals were quantitatively compared (e.g., AXL and CD68 between MDA231-BrM lesions and HCC1954-BrM lesions in Figure 4A). Figure 1A images showed maximum intensity projection of z-stack images (11 planes, Δz = 0.3 μm), which captured the structure of blood vessels better than single z-plane images. To facilitate visualization of cancer cells immunolabeled by anti-GFP antibody in the 488 nm channel, where brain tissue exhibited a high level of autofluorescence, FIJI functions of Threshold and Analyze Particles were used to mask non-cancer cell regions in cases of high autofluorescence background. No other images were processed by such background removal step. Light sheet microscopy images of a brain hemisphere were registered and stitched by the built-in Image Processor function in Luxendo MuVi SPIM light sheet microscope. The 3D-reconstructed images obtained were adjusted for contrast and brightness in Imaris (Oxford Instruments Group, version 9.5.0). 1 mm³ cubic regions containing brain metastases were cropped and displayed as 3D images or videos with optimal rendering.

From whole-slide scans of 5 μm-thick paraffin-embedded sections, the regions of interest (ROIs) encompassing metastatic colonies were annotated and exported as TIFF images using SlideViewer. To mitigate potential confounding edge effects, the rims of tissue sections, where non-specific antibody immunoreactivity tends to aggregate, were excluded from annotation. For both ROIs extracted from whole-slide scans and individual confocal microscopy images, the IF signal quantification was carried out using FIJI batch analysis scripts, which enabled automation of the following steps in the analysis process.

Quantification of tumor-associated TNC signal

The quantification proceeded as follows. 1) The Threshold function was run to segment the area positive of particular cancer cell marker (e.g., GFP, luciferase, and cytokeratin) to create a binary mask of the tumor region:

$$M_{\text{tumor}}(x,y)\in (0,1),$$

where the mask value at position $(x,y)$ was set to 1 if it belonged to a cancer cell and 0 otherwise. The threshold value was determined through manual visual inspection to ensure proper alignment of the mask with the original cancer cell marker staining. 2) To include TNC deposits surrounding a metastatic colony, the Dilate function was applied to augment the tumor region mask $M_{\text{tumor}}(x,y)$ by 3.25 μm, approximately the distance of a layer of mouse microglia/macrophage cells, to obtain a mask of the larger tumor-associated area $M_{\text{tumor–associated}}(x,y)$. 3) The total TNC intensities in the expanded tumor-associated area mask was calculated by the Measure function as element-wise sum of the Hadamard product between $M_{\text{tumor–associated}}(x,y)$ and TNC intensities $I_{\mathrm{T}\mathrm{N}\mathrm{C}}(x,y)$ across all positions $(x,y)$:

$$s_{\mathrm{T}\mathrm{N}\mathrm{C}}={\sum }_{(x,y)}{M}_{\text{tumor–associated}}(x,y)\odot I_{\mathrm{T}\mathrm{N}\mathrm{C}}(x,y).$$

$s_{\mathrm{T}\mathrm{N}\mathrm{C}}$ was subsequently divided by the area of the tumor region mask $M_{\mathrm{t}\mathrm{u}\mathrm{m}\mathrm{o}\mathrm{r}}(x,y)$ to yield a size-normalized, colony-level metric of tumor-associated TNC abundance:

$$s_{\mathrm{T}\mathrm{N}\mathrm{C}}^{\prime }=\frac{s_{\mathrm{T}\mathrm{N}\mathrm{C}}}{{\sum }_{(x,y)}{M}_{\mathrm{t}\mathrm{u}\mathrm{m}\mathrm{o}\mathrm{r}}(x,y)}.$$

Quantification of AXL signal in metastasis-associated microglia and macrophages

As with the TNC signal quantification, 1) the Threshold function was run to obtain a binary mask of the tumor-associated area $M_{\text{tumor–associated}}(x,y)$ and a binary mask of IBA1⁺ microglia and macrophages $M_{\text{IBA1}}(x,y)$. 2) The region corresponding to the IBA1⁺ microglia and macrophages in contact with cancer cells was subsequently determined by the Hadamard product between these two masks:

$$M_{\text{tumor–associated IBA1}}(x,y)=M_{\text{tumor–associated}}(x,y)\odot M_{\text{IBA}1}(x,y),$$

such that the resulting mask value at position $(x,y)$ was set to 1 only when this position was both within the tumor-associated area and positive of IBA1 staining. 3) The average AXL IF signal in tumor-associated microglia and macrophages was quantified by summing up AXL pixel intensities $I_{\mathrm{A}\mathrm{X}\mathrm{L}}(x,y)$ inside the mask $M_{\text{tumor–associated IBA1}}(x,y)$ and subsequently dividing the sum by the area of the mask, using the Measure function:

$$S_{\text{IBA1}}=\frac{{\sum }_{(x,y)}{M}_{\text{tumor–associated IBA1}}(x,y)\odot I_{\mathrm{A}\mathrm{X}\mathrm{L}}(x,y)}{{\sum }_{(x,y)}{M}_{\text{tumor–associated IBA1}}(x,y)}.$$

Quantification of tumor-vasculature engagement in rapid autopsy specimens

To quantify the degree of tumor-vasculature engagement, we measured the shortest distance from the edge of cytokeratin⁺ cancer cell nuclei to CD31⁺ blood vessels as follows. 1) Considering both our focus on brain colonization initiation and the constraints posed by resolving 3D structures using 5 μm-thick sections for larger tumor foci, we focused on ROIs containing small groups of cells (i.e., 3-20 cells) that were captured in isolation within 2D sections. 2) The Threshold function was applied to DAPI images to segment individual nuclei and to cytokeratin and CD31 images to obtain the tumor region mask $M_{\text{tumor}}(x,y)$ and blood vessel mask $M_{\text{vasculature}}(x,y)$, respectively. The threshold values were carefully adjusted to ensure robust detection of the cancer cells and refined blood vessel structures, especially the capillaries, while effectively excluding background noise. 3) The area $A_{\text{cell}}$ of the nuclei of individual cancer cells, i.e., DAPI⁺ objects that overlapped with $M_{\text{tumor}}(x,y)$, was determined by Measure function. 4) Subsequently, the Euclidian Distance Map function was run to calculate the shortest distance from the edge of a cancer cell nucleus to the blood vessel mask $M_{\text{vasculature}}(x,y)$, which we denoted as $d_{\text{cell–vasculature}}$. 5) To circumvent the confounding effects of the heterogeneity in cancer cell sizes especially evident among different tumor foci from clinical samples, we calculated the projected area diameter of the mean nuclei area ${\bar{A}}_{\text{cell}}$ of cells in particular tumor focus, representing a robust metric for the size of these cells:

$$l_{\text{tumor focus}}=2{\left({\bar{A}}_{\text{cell}}/\pi \right)}^{1/2}.$$

6) We subsequently normalized $d_{\text{cell–vasculature}}$ by $l_{\text{tumor focus}}$ to derive a cell size-calibrated metric for the degree of separation from a cancer cell to its nearest vasculature, in approximate terms of how many cells away it was located:

$$d_{\text{cell–vasculture}}^{\prime }=\frac{d_{\text{cell–vasculature}}}{l_{\text{tumor focus}}}.$$

To test the dependence of vasculature engagement on tumor subtype, data comprising the median $d_{\text{cell–vasculture}}^{\prime }$ value among cancer cells in each tumor focus were grouped per subtype (RA-1 and RA-2 for TNBC, RA-3 and RA-4 for HER2BC) and compared between subtypes using a Wilcoxon signed-rank test.

Quantification of AXL level in microglia/macrophages and surrounding TNC abundance in rapid autopsy specimens

To explore the global association between these two properties in rapid autopsy specimens across patients, we annotated the ROIs containing tumor lesions of all sizes for the two TNBC patients and two HER2BC patients, and subsequently compiled the data from all ROIs across available brain tissue slides of each patient for downstream visualization. 1) Using a similar approach as described above (see Quantification of AXL signal in metastasis-associated microglia and macrophages), the Threshold function was applied to identify and isolate tumor-associated IBA1⁺ microglia and macrophages as individual masks, $m_{\text{tumor–associated IBA1}}(x,y)$. 2) The Analyze particles function was applied to obtain solid whole-cell masks for IBA1⁺ cells, ${m^{\prime }}_{\text{tumor–associated IBA1}}(x,y)$, which cover the “hollow” IBA1⁻ DAPI⁺ nuclei area inside $m_{\text{tumor–associated IBA1}}(x,y)$ masks. (3) The AXL level in these cells was subsequently calculated as the average of AXL intensities $I_{\mathrm{A}\mathrm{X}\mathrm{L}}(x,y)$ inside the objects of whole single-cell masks as:

$$s_{\mathrm{A}\mathrm{X}\mathrm{L}}=\frac{{\sum }_{(x,y)}{m^{\prime }}_{\text{tumor–associated IBA1}}(x,y)\odot I_{\mathrm{A}\mathrm{X}\mathrm{L}}(x,y)}{{\sum }_{(x,y)}{m^{\prime }}_{\text{tumor–associated IBA1}}(x,y)}.$$

4) To quantify the abundance of TNC deposits in the cellular niches surrounding individual tumor-associated microglia and macrophage cells, the cellular mask ${m^{\prime }}_{\text{tumor–associated IBA1}}(x,y)$ was expanded by 14.63 μm, approximately the diameter of a human IBA1⁺ cell, in all directions to obtain the corresponding neighborhood mask $n_{\text{tumor–associated IBA1}}(x,y)$ of a cell. The average of TNC intensities $I_{\mathrm{T}\mathrm{N}\mathrm{C}}(x,y)$ inside this mask was determined as:

$$s_{\text{niche TNC}}=\frac{{\sum }_{(x,y)}{n}_{\text{tumor–associated IBA1}}(x,y)\odot I_{\mathrm{T}\mathrm{N}\mathrm{C}}(x,y)}{{\sum }_{(x,y)}{n}_{\text{tumor–associated IBA1}}(x,y)}.$$

5) $s_{\mathrm{A}\mathrm{X}\mathrm{L}}$ and $s_{\text{niche TNC}}$ represented the AXL level in tumor-associated microglia/macrophage cells and the microenvironmental TNC deposit abundance of the cells, respectively. The distribution of ($s_{\text{niche TNC}},s_{\text{AXL}}$) data points from each patient were visualized using the Seaborn scatter plot and kernel density estimate (KDE) plot (with 10 contour levels) in Python (version 3.8.5). Correlation between $s_{\text{niche TNC}}$ and $s_{\text{AXL}}$ was evaluated by Spearman’s correlation analysis on data points of four patients compiled together. To test the dependence of s_AXL or s_{niche TNC} levels on tumor subtype, data were grouped per subtype (RA-1 and RA-2 for TNBC, RA-3 and RA-4 for HER2BC) and compared between subtypes using a Wilcoxon signed-rank test.

RNA isolation and bulk gene expression assays

RNA extraction

Cells grown to approximately 80% confluence in a 10 cm petri dish were collected as a cell pellet for total RNA extraction. Cells were lysed through QIAshredder columns (Qiagen, Cat: 79656) in Buffer RLT from QIAgen RNeasy Mini Kit (QIAgen, Cat: 74106) supplemented with 1:100 β-Mercaptoethanol (Sigma Aldrich, Cat: M3148100ML). After cell lysis, RNA isolation was performed with the RNeasy Mini Kit. Total extracted RNA was quantified using a NanoDrop (Fisher Scientific, Cat: 2741002PM22). Reverse transcription was performed using the Transcriptor First Strand cDNA Synthesis Kit (Roche, Cat: 04897030001).

qRT-PCR

For quantitative real-time polymerase chain reaction (qRT-PCR), 1 μg of total RNA was used to synthesize cDNA. Amplification of targets from cDNA was performed with the TaqMan Universal PCR Master Mix (ThermoFisher, Cat: 4304437) and TaqMan probes in a ViiA 7 Real-Time PCR System (ThermoFisher) via 40 cycles of PCR, each consisting of 10 seconds at 95 °C (denaturation) and 30 seconds at 60 °C (annealing/polymerization). Relative gene expression was calculated by the canonical 2^−ΔΔC_T method¹¹¹, using housekeeping genes ACTB or GAPDH as control for analysis of cancer cells. To assess the microglial responses to recombinant TNC or IFN-β in vitro, the housekeeping gene Hprt1 was employed as the reference control. This choice was based on the observation that Actb and Gapdh exhibited heightened expression in metastasis-associated DAM in both repeated scRNA-seq experiments characterizing the mCherry-labeled microglia harvested from in vivo (see Table S3), while the Hprt1 expression remained unchanged compared to unlabeled microglia distal from brain metastases. This expression pattern raised the possibility that the Actb and Gapdh could be upregulated by the TNC enriched in HER2BC brain metastases and did not support assuming them as housekeeping gene controls with stable expression for the in vitro assay of recombinant TNC treatment.

Flura-seq

As summarized in Figure S8A, Flura-seq was conducted as previously described^76,112 with minor modifications. Briefly, MDA231-BrM cells and HCC1954-BrM cells were transduced with lentivirus carrying PGK-UPRT-T2A-CD. For each BrM model, 1.0 × 10⁵ cells resuspended in 100 μL PBS supplemented with 2% FBS were inoculated into the mice by intracardiac injection as described above (see Animal Studies). 30 days post-inoculation, mice were intraperitoneally injected with 250 mg/kg 5-fluorocytosine (5-FC) (Sigma Aldrich, Cat: F71291G), and after 12 hours, mouse brains were harvested and homogenized. mRNA was extracted from homogenized cell lysate by 4 × 250 μL/brain Oligo (dT)25 magnetic beads (New England Biolabs, Cat: S1419S), and immunoprecipitated (IP) with 2.5 μg/brain anti-BrdU antibody (Abcam, Cat: ab6326). cDNA libraries were constructed using the immunoprecipitated mRNA and SMARTer PCR cDNA Synthesis Kit (Clontech, Cat: 634926), and sequenced on a HiSeq 2500 system (Illumina, Integrated Genomics Core, MSKCC).

RNA-seq

RNA-seq samples were processed and sequenced in the Integrated Genomics Core, MSKCC. The quality and quantity of RNA samples were determined using BioAnalyzer 2100 (Agilent). cDNA libraries were constructed using either TruSeq RNA Sample Prep Kit v2 (Illumina, Cat: RS-122-2001) or NEBNext Ultra RNA Library Prep Kit (New England Biolabs, Cat: E7530S) and sequenced on a HiSeq 2500 system (Illumina) at a depth of 25-50 million reads per sample.

Bulk gene expression data analysis

Expression signatures of brain tropism

RNA-seq profiles of HCC1954 and MMTV-ErbB2, including parental cells and BrM derivatives, were generated in this paper. Microarray gene expression data of MDA231 parental cells and BrM derivatives were obtained from previous breast cancer dataset³. Raw sequencing reads in FASTQ format were mapped by STAR RNA-seq aligner^113,114 (version 2.5.3a) , using human genome reference GRCh38 and mouse genome reference GRCm38 from GENCODE (www.gencodegenes.org/) for RNA-seq samples of human cancer cells and mouse cancer cells, respectively. Uniquely mapped reads were assigned to annotated genes by HTSeq (version 0.6.1p1) with default settings¹¹⁵ to measure read counts. The counts were normalized by library size to be subject to differential gene expression analysis, both performed with the DESeq2 package¹¹⁶ (Bioconductor, version 3.4) in RStudio (version 1.0.153) implementing R (3.4.1). A gene was defined as differentially expressed if its 1) absolute value of log2 fold change in normalized counts was greater or equal to 2, 2) adjusted P value was below 0.05, and 3) mean read counts was larger than 10. Flura-seq data was analyzed following the same procedure with one additional step preceding alignment. As the data were derived from xenograft mouse model samples containing a mixture of reads from human cancer cells and mouse stromal cells, Xenome¹¹⁷ (version 1.0.1) was run as the first step to classify the reads as belonging to either human genome or mouse genome using the default value (25) of parameter k-mer size. The web-based tool DAVID¹¹⁸ (Database for Annotation, Visualization and Integrated Discovery, version 6.8) was employed to identify over-represented gene ontology (GO) terms in differentially expressed genes (DEGs).

Analysis of MetMap dataset

Bulk RNA-seq profiles of brain metastases with multiplexed breast cancer cell line composition were obtained from the MetMap dataset, and the differential expression analysis that properly accounted for cancer cell composition variability was conducted as previously described⁷. The percentage of cells belonging to HER2BC versus TNBC cell lines in a sample were computed. To assess the statistical significance of the in vivo upregulation of genes encoding ECM components in brain metastases in relation to the HER2BC subtype, linear regression analysis was performed on the log2 fold changes in gene expression with respect to the fraction of HER2BC cells. Wilcoxon rank sum test was performed on the log2 fold changes in gene expression with respect to the dominance of HER2BC cells. P values were adjusted by BH multiple hypothesis correction. Data were plotted using ggplot2 in RStudio (version 1.2.5019).

Query of GAS6 and AXL levels in breast cancer cell lines in DepMap database

For all human breast cancer cell lines available in the DepMap database, consistent subtype annotation, including HER2BC, TNBC, luminal A, and luminal B, were confirmed by cross validating Ref.²⁰, DepMap, and breast cancer cell line classification resources²¹. DepMap transcriptomics data were originally processed through STAR alignment, RSEM quantification¹¹⁹, and TPM (transcripts per million) normalization. The normalized transcript levels (log2(TPM + 1) values) were obtained from the DepMap online portal (depmap.org/portal/). Proteomics data were measured by mass spectrometry¹²⁰, and the relative normalized protein levels published by Ref.¹²⁰ were obtained from the DepMap online portal as well. Correlation between AXL and GAS6 was assessed by subtype-specific Spearman’s correlation analysis. Transcript and protein levels were compared across subtypes using a Wilcoxon signed-rank test. All analysis was performed using R (version 4.2.0, www.r-project.org) and Bioconductor (version 3.15).

scRNA-seq data collection

HCC1954-BrM cells and MDA231-BrM cells were transduced with lentivirus expressing sLP-mCherry-P2A-eGFP under the control of EF1α promoter (plasmid constructed by Tammela Lab, MSKCC). As illustrated in Figure 2C, athymic mice were intracardially inoculated with the engineered MDA231-BrM cells or HCC1954-BrM cells (1.0 × 10⁵ cells/mouse). A mouse brain harboring MDA231-BrM metastases and one harboring HCC1954-BrM metastases were harvested 4 weeks post-inoculation as described above (see Animal studies), and both processed through 1) enzymatic tissue dissociation using the Adult Brain Dissociation Kit, mouse and rat, 2) removal of myelin using Myelin Removal Beads II, human, mouse, rat (Miltenyi Biotec, Cat: 130-096-733), and 3) removal of debris and dead cells using the Dead Cell Removal Kit (Miltenyi Biotec, Cat: 130-090-101). The brain of an age-matched uninoculated mouse was harvested and processed in parallel to serve as the negative control for downstream FACS sorting. The first enzymatic tissue dissociation step occurred at 37 °C, and all following steps were performed at 4 °C.

The resulting single-cell suspension of each brain metastasis sample was incubated with an anti-human Hashtag antibody (binding to human cancer cells) and an anti-mouse Hashtag antibody (binding to mouse cells), both conjugated to unique DNA barcodes as listed below, 1:100 human TruStain FcX (Fc Receptor Blocking Solution) (BioLegend, Cat: 422301) and 1:100 TruStain FcX (anti-mouse CD16/32) antibody (BioLegend, Cat: 101319) to prevent unspecific binding of Hashtag antibodies, and 40 ng/mL CellTrace Calcein Violet, AM, for 405 nm (Thermo Fisher Scientific, Cat: C34858) and 1:20 APC Annexin V (BioLegend, Cat: 649020) to allow filtering residual debris and apoptotic cells during FACS sorting. The Calcein Violet⁺ Annexin V⁻ live cells of each sample were FACS-sorted into three groups, i.e., the labeling GFP⁺ mCherry⁺ cancer cells, labeled GFP⁻ mCherry⁺ TME cells, and rest unlabeled GFP⁻ mCherry⁻ cells. The single-cell suspension of uninoculated mouse brain, which only contained GFP⁻ mCherry⁻ cells, was used to set the gates for GFP⁺ or mCherry⁺ cells.

Experiment	Model	Hashtag antibody	Barcode sequence	Dilution
1	MDA231-BrM	TotalSeq-A0255 anti-human Hashtag 5 Antibody	AAGTATCGTTTCGCA	1:100
		TotalSeq-A0307 anti-mouse Hashtag 7 Antibody	GAGTCTGCCAGTATC	1:100
	HCC1954-BrM	TotalSeq-A0253 anti-human Hashtag 3	TTCCGCCTCTCTTTG	1:100
		TotalSeq-A0306 anti-mouse Hashtag 6 Antibody	TATGCTGCCACGGTA	1:100
2	MDA231-BrM	TotalSeq-A0258 anti-human Hashtag 8 Antibody	CTCCTCTGCAATTAC	1:100
		TotalSeq-A0308 anti-mouse Hashtag 8 Antibody	TATAGAACGCCAGGC	1:100
	HCC1954-BrM	TotalSeq-A0256 anti-human Hashtag 6 Antibody	GGTTGCCAGATGTCA	1:100
		TotalSeq-A0306 anti-mouse Hashtag 6 Antibody	TATGCTGCCACGGTA	1:100

The sorted cells from MDA231-BrM and HCC1954-BrM samples were then pooled per FACS-sorted group for single-cell droplet encapsulation and sequencing, and subsequently assigned to their original sample source according to the sequencing reads of different Hashtag sequences used (see Pre-processing of scRNA-seq data in scRNA-seq data analysis). Two independent scRNA-seq experiments were conducted following similar steps. In experiment 2, 5 μg/mL Actinomycin D (Sigma, Cat: A1410), 10 μM Triptolide (Sigma, Cat: T3652), and 27.1 μg/mL Anisomycin (Sigma, Cat: A9789) were added to the PBS used for perfusing mice and soaking the harvested brains and to the enzymatic dissociation buffer as well, in order to inhibit ex vivo transcription and translation as previously reported⁴³.

scRNA-seq data analysis

Data from both experiments (1, 2) were combined to be pre-processed (i.e., filtering ambient RNA and non-cell droplets, normalizing and log-transforming single-cell UMI counts, calling cell types, computing signature scores) together before being separated for downstream analysis following the same procedure, presented in Figures 2, S2, and S4 (experiments 1, 2), 3 (experiment 1), and S4 (experiment 2). All our analysis results were consistent between two scRNA-seq experiments, except that in the second one, pro-inflammatory cytokine gene transcripts were diminished likely due to the transcription and translation inhibitors added⁴³.

Pre-processing of scRNA-seq data

FASTQ files of sequenced samples were individually processed using the SEQC pipeline⁵² (version 0.2.1) with default parameters for the 10x single-cell 3′ library. Samples from GFP⁺ cancer cells and from GFP⁻ non-cancer cells in the brain were mapped to the hg38 human and mm38 mouse genome references, respectively. The SEQC pipeline performed read alignment, multi-mapping read resolution, and cell barcode and UMI correction to generate a raw UMI-based (cell × gene) count matrix per sample. Following SEQC, we ran multiple complementary quality control (QC) algorithms in parallel, and used the Scanpy software¹²¹ (version 1.6.0) in Python (version 3.8.5) to incorporate their results to identify high-quality, background-removed single-cell transcriptome from the count matrices. The algorithms used were: 1) CellBender¹²² (version 0.2.0) removed counts due to ambient RNA molecules and random barcode swapping from the (raw) SEQC output count matrices, using the remove-background command with the parameters of expected-cells = 10000, total-droplets-included = 50000, fpr = 0.01 (default value), and epochs = 150 (default value). 2) DropletUtils¹²³ (version 1.10.3) distinguished the droplets containing cells from those containing ambient RNA by the function of emptyDrops on the (raw) SEQC output count matrices with the parameters of retain = Inf, test.ambient = True, lower = 50, niters = 50000. 3) An in-house method called SHARP (github.com/hisplan/sharp, version 0.2.1) assigned a cell as labeled by a particular Hashtag oligo and thereby belonging to the specific mouse (bearing either MDA231-BrM metastases or HCC1954-BrM metastases) barcoded with this oligo or as a doublet or low-quality droplet. We subsequently uploaded the CellBender-processed count matrices, subset to droplets that were called to be real cells by both emptyDrops (FDR ≤ 0.01) and SHARP, and computed additional metrics on the subset matrices to eliminate residual droplets that were likely to be empty droplets, doublets or apoptotic cells. We used Scanpy’s pp.normalize_total function to normalize the count matrices by library size (i.e., total number of UMIs after CellBender filtering of background counts). Using the library size-normalized count matrices (base = 2, pseudo count = 1), we performed principal component analysis (PCA) using the top 5000 highly variable genes (HVGs) that were robustly detected (in at least 10 cells), excluding mitochondrial or ribosomal genes, and ran PhenoGraph¹²⁴ (version 1.5.6, k = 30) to cluster the droplets. To exclude apoptotic cells, we filtered out droplets with a high fraction of mitochondrial gene transcripts (mtRNA% ≥ 0.2), indicative of cellular stress. To remove droplets of low library complexity (i.e., with few unique genes), we followed the SEQC⁵² approach of fitting a linear model to the relationship between the number of genes and the number of UMIs across droplets, and discarded droplets with lower than expected number of genes (residuals > 0.15, default value in SEQC). To clean out any remaining lower-quality droplets, we examined the gene-gene expression covariance of each PhenoGraph cluster computed on its top 500 highly expressed genes and removed cells constituting the PhenoGraph clusters that did not exhibit apparent covariance structures. After the above filtering steps, we ran DoubletDetection (version 2.5.2, github.com/dpeerlab/DoubletDetection) to infer and remove droplets that may contain more than one cell. The above pre-processing steps yielded:

Experiment	Cells	FACS group	Source	Number of cells	Number of genes	Median of UMIs/cell	Median of genes/cell
1	Cancer	GFP+ mCherry+	TNBC	5659	19284	15185	3866
			HER2BC	2025	18610	20773	4356
	Non-cancer	GFP− mCherry− (Unlabeled)	TNBC	2303	16992	3929	1704
			HER2BC	8570	18940	4232	1783
		GFP− mCherry+ (Labeled)	TNBC	5691	16820	4756	1749
			HER2BC	5187	17492	6816	2093
2	Cancer	GFP+ mCherry+	TNBC	2444	18045	11062	3499
			HER2BC	3054	19016	16792	4157
	Non-cancer	GFP− mCherry− (Unlabeled)	TNBC	1790	17963	6443	2440
			HER2BC	4644	18980	7335	2623
		GFP− mCherry+ (Labeled)	TNBC	1054	17425	13623	3490
			HER2BC	2418	18518	14167	3706

Basic analysis of single-cell transcriptome

After constructing a count matrix as described in Pre-processing of scRNA-seq data, we concatenated the original (un-normalized) CellBender-processed count matrices of both experiments (1, 2) for filtered GFP⁺ cancer cells (GFP+ mCherry⁺) and GFP⁻ non-cancer cells (GFP⁻ mCherry⁻/unlabeled and GFP⁻ mCherry⁺/labeled), respectively, and used scran¹²⁵ (version 1.14.6) to normalize the concatenated count matrices. We chose scran instead of the simpler library size normalization performed for pre-processing, because scran is designed to better recapitulate the endogenous variability in library size among a heterogeneous population of cells, and has been verified to be a top-performing normalization method by multiple independent comparisons¹²⁶. After scran normalization, we log-transformed the count matrices (base = 2, pseudo count = 1), and performed principal component analysis (PCA) using the top 5000 highly variable genes (HVGs) that were robustly detected (in at least 10 cells), excluding mitochondrial or ribosomal genes. The top principal components (PCs) selected according to the knee point of total variance explained (as listed below) were used to construct k-nearest neighbor (kNN) graph (n_neighbors = 15) to generate uniform manifold approximation and projection (UMAP) layouts, and to cluster cells by PhenoGraph¹²⁴ (version 1.5.6, k = 30). The PhenoGraph clusters of GFP⁻ non-cancer cells were proceeded with for cell-type annotation.

Cells	Number of PCs	Variance explained
Cancer	56	0.43
Non-cancer	42	0.56

Cell-type annotation of non-cancer cells

We annotated the cell types of GFP⁻ non-cancer cells using a reference single-cell transcriptome atlas of the mouse nervous system⁴⁴ (hereinafter referred to as “atlas”, mousebrain.org) and additional immune cell type markers that complement the atlas (see Figure S2; Table S2). The atlas determined a hierarchical classification of cells in the nervous system. At the most refined level of classification, it identified 265 cell types, and provided 1) cluster-averaged expression, 2) six marker genes, and 3) associated metadata including the upper-level taxonomy annotation of the clusters. The additional immune cell type markers were collected from commonly used scRNA-seq cell type markers and included genes for calling B cells (BC), neutrophils (NEUT), bone marrow-derived macrophages (BMDM), and NK cells (NK) – infiltrated immune cells absent in the atlas database of healthy brain tissue. See Table S2 for cell type marker genes and metadata.

We first mapped the query PhenoGraph¹²⁴ clusters (see Basic analysis of single-cell transcriptome in scRNA-seq data analysis) to their respective reference cell type by evaluating two metrics – 1) cellular detection score of marker genes and 2) correlation of global gene expression levels. 1) Cellular detection score was quantified as described previously⁵². For a given cell c and marker gene k:

$$d_{ck}=\frac{I_k}{N_c},$$

where a cell was scored 1 $\left(I_k=1\right)$ if it contained a non-zero UMI count of the marker gene and 0 otherwise $\left(I_k=0\right)$, and corrected by the detection rate of the cell, defined as the fraction $(\frac{1}{N_c})$ of total number of genes detected in it $\left(N_c\right)$. The corrected scores $\left(d_{ck}\right)$ were subsequently averaged across marker genes of a reference cluster and all cells from certain query PhenoGraph cluster to generate the query cluster-level cellular detection score for certain group of marker genes as:

$$d_{qr}=\frac{{\sum }_{c=1}^{N_q}{\sum }_{k=1}^{N_r}{d}_{ck}}{N_q{N}_r},$$

where $N_q$ is the total number of cells in the query cluster $q$, and $N_r$ total number of marker genes in the reference cluster $r$. 2) Correlation in cluster-level global expression profiles. Specifically, we summarized the gene expression of query clusters by averaging single-cell normalized UMI counts across cells per query cluster. We obtained the cluster-level normalized UMI counts of 265 reference clusters from the atlas database. We then computed the correlation in the log-transformed cluster-averaged count matrices between query PhenoGraph clusters and 265 reference atlas clusters on 5036 robustly detected genes. These genes were selected from genes with ≥ 1 average normalized UMI counts per query cluster, excluding abundant mitochondrial and ribosomal genes. The selection was performed to ensure that the correlation computed was representative of the expression patterns of all query clusters, and not overwhelmed by individual clusters with higher number of cells (e.g., macrophages) or genes ubiquitously expressed at high or low levels across clusters. Correlation with infiltrated immune cell types was not computed, as the atlas does not contain transcriptome data of these cell types. For each query cluster of non-cancer cells, we noted that both metrics had one or several high-scored reference clusters which were well separated from the rest, and that the top-scored reference cluster(s) were largely identical or overlapping between metrics, indicative of consistent cell type calling by either specific marker genes or global gene expression (if available). We manually examined the common (of both metrics) top-scored reference clusters for each query cluster, especially in the cases of more than one reference clusters and in most of such cases, the reference clusters shared the same upper-level taxonomy (e.g., MGL1, MGL2, and MGL3 clusters in the atlas corresponding to different states of the same taxonomy of microglia, MG), which we used to annotate the cell population. In the remaining rare cases where multi-mapping arose from certain query cluster containing cells of closely related but different cell types (e.g., MG, BMDM, and border associated macrophages, BAM, all belong to macrophages), we repeated the PCA and PhenoGraph clustering for this query cluster specifically to obtain more refined grouping and re-ran the above cell type annotation steps of re-grouped query clusters. After finalizing the annotation through an iterative optimization process, we re-computed and plotted the two metrics restricted to identified populations and relevant reference clusters to generate the plots in Figures 2E, 2F, and S2C.

Inferring tissue dissociation-associated ex vivo activation and cell cycling status in macrophages

As noted above (see Cell-type annotation of non-cancer cells), macrophages include microglia (MG), bone marrow-derived macrophages (BMDM), and border associated macrophages (BAM). To infer the cell cycling status of macrophages and evaluate the impact of adding transcription and translation inhibitors⁴³ during tissue dissociation in experiment 2, we utilized the score_genes function of Scanpy¹²¹ to compute on all macrophages the signature scores of 1) genes associated with S phase or G2/M phase¹²⁷ and 2) genes whose expression can be induced ex vivo by the enzymatic dissociation process⁴³, respectively (both listed in Table S2). The Scanpy¹²¹ gene signature score is calculated as the average expression of a set of genes, subtracted with the average expression of a reference set of genes that were randomly sampled to match the expression distribution of the given gene set. To ensure that the score values of the cells from two experiments were directly comparable to each other, signature scores were computed on both experiments together by running the score_genes function once on their concatenated matrices rather than twice on separate matrices. Two PhenoGraph clusters of microglia displayed markedly higher scores of S phase and G2/M phase genes, and were inferred to be cycling (Figure S2D) (and the majority of microglia as non-cycling, in G0 or G1 phase). The substantially lower scores of tissue dissociation-associated ex vivo activation (Figure 2G) in the macrophage cells from experiment 2 validated the reported effects of including transcription and translation inhibitors⁴³.

Differential sample abundance visualization and analysis

We adapted the Milo algorithm⁴⁵ (version 0.1.0) built upon the generalized linear model (GLM) to analyze the differential abundance in sub-populations between conditions. Milo expects replicates in its input but is exceedingly sensitive to batch effects and even minor imperfections in data integration. Thus, we chose to analyze experiment 1 and 2 separately, which required a change in Milo’s test statistic, and looked for correspondence in Milo’s results across the two experiments. We aimed to compare the subpopulation frequencies in MDA231-BrM TNBC- or HCC1954-BrM HER2BC-labeled (GFP⁻ mCherry⁺) cells in reference to unlabeled (GFP⁻ mCherry⁻) cells. We isolated the log-transformed, normalized UMI counts of the cells of interest (i.e., either all non-cancer cells or non-cycling microglia only) from each experiment to construct a kNN graph (k = 15) and generate UMAP embeddings, using the following indicated number of PCs computed as described above (see Basic analysis of single-cell transcriptome in scRNA-seq data analysis). As described in Identifying gene programs associated with homeostasis-to-DAM transition, the same number of PCs were also selected to compute diffusion components, using the Palantir algorithm⁵³ (version 1.1) that runs diffusion maps with an adaptive anisotropic kernel. To intuitively visualize the differential abundance in the phenotypic space, we used the Seaborn kernel density estimate (KDE) plot (with 7 contour levels) to show the distribution of each group of non-cycling microglia cells in the UMAP embedding computed on all four groups of cells (i.e., TNBC-labeled, TNBC-unlabeled, HER2BC-labeled, HER2BC-unlabeled) together.

Experiment	Cells	FACS group	Source	Number of cells	Number of PCs	Variance explained
1	All non-cancer cells	Unlabeled	TNBC	2303	43	0.51
			HER2BC	8570
		Labeled	TNBC	5691
			HER2BC	5187
	Non-cycling microglia (MG)	Unlabeled	TNBC	142	98	0.24
			HER2BC	2147
		Labeled	TNBC	5068
			HER2BC	231
2	All non-cancer cells	Unlabeled	TNBC	1790	40	0.61
			HER2BC	4644
		Labeled	TNBC	1054
			HER2BC	2418
	Non-cycling microglia (MG)	Unlabeled	TNBC	37	128	0.41
			HER2BC	180
		Labeled	TNBC	600
			HER2BC	920

We followed the steps of the Milo algorithm⁴⁵ to select index cells, using a graph sampling strategy previously devised by Palantir⁵³, and to construct partially overlapping neighborhoods defined on the index cells. To increase statistical power, since both TNBC-unlabeled and HER2BC-unlabeled groups comprised cells distal to tumor lesions in the brains of athymic mice, we assumed these two groups as the replicates of unlabeled cells. Let ${\mu }_{ns}$ denote the mean counts of cells of group s in neighborhood $n,M_s$ the total number of cells of group $s$, and $I_{S\text{TNBC–labeled}}$ and $I_{s\text{HER2BC–labeled}}$ the indicator variables of whether group $s$ is TNBC-labeled ($I_{s\text{TNBC–labeled}}=1$ for the TNBC-labeled group, and 0 otherwise) or HER2BC-labeled (similarly, $I_{s\mathrm{HER2BC–labeled}}=1$ for the HER2BC-labeled group, and 0 otherwise), following the Milo framework, we modeled the cell counts from certain neighborhood $n$ as

$$\mathrm{l}\mathrm{o}\mathrm{g}{\mu }_{ns}=I_{s\mathrm{TNBC–labeled}}{\beta }_{n\text{TNBC–labeled}}+I_{s\text{HER2BC–labeled}}{\beta }_{n\text{HER2BC–labeled}}+\mathrm{l}\mathrm{o}\mathrm{g}{M}_s,$$

where ${\beta }_{n\text{TNBC–labeled}}$ and ${\beta }_{n\text{HER2BC–labeled}}$ corresponded to the regression coefficients by which the effects of originating from the TNBC-labeled or HER2BC-labeled group were mediated for neighborhood $n$. We deviated from Milo in the statistical test applied to identify significantly differential abundances from each experimental condition per neighborhood. Instead of the quasi-likelihood F-test selected by Milo, we utilized alternative log-likelihood ratio test functions in EdgeR package¹²⁸ (version 3.32.1) best fit for 1-2 replicates per condition, smaller cell counts, and larger dispersion in the cell counts. To evaluate the variability in cell counts in replicate samples for each neighborhood, we ran the EdgeR¹²⁸ function estimateDisp with default parameters, which maximized the negative binomial likelihood to estimate the dispersion in cell counts across phenotypic neighborhoods. To obtain ${\beta }_{n\text{TNBC–labeled}}$ and ${\beta }_{n\text{HER2BC–labeled}}$ values as the log fold change (logFC) in labeled cell abundance, we ran glmFit to fit a negative binomial generalized log-linear model to the cell counts of each neighborhood. To test whether TNBC-labeled or HER2BC-labeled sample abundance differed from unlabeled sample abundance, we ran glmLRT to conduct log-likelihood ratio test on whether ${\beta }_{n\text{TNBC–labeled}}$ or ${\beta }_{n\text{HER2BC–labeled}}$ differed from 0. Lastly, we applied the Milo function graphSpatialFDR using default parameter values to correct the glmLRT-derived P values for multiple testing while accounting for the non-independence of overlapping neighborhoods. The corrected P values and logFC values were used for visualization (Figures 3B, 3E, S2C, and S4D) and downstream analysis. It should be noted that, compared to the quasi-likelihood F-test which accounts for uncertainty in dispersion estimates to obtain more conservative and rigorous type I error rate control, the log-likelihood ratio test runs a higher risk of generating false positives¹²⁸. Aware of this potential caveat, we based the majority of our downstream analysis on tracking expression patterns along with continuous changes in logFC values without thresholding on corrected P values, and utilized the scRNA-seq analysis to generate hypotheses which we tested experimentally.

Differential gene expression in tumor-associated microglia

To detect genes characteristic of tumor-associated microglia in both subtypes of breast cancer in experiment 1, we first identified two groups of phenotypic neighborhoods in non-cycling microglia that were enriched in or depleted of tumor-associated microglia, respectively, and ran MAST¹²⁹ (version 1.16.0) to call genes differentially expressed between these two groups of microglia. As the first step, neighborhoods that were connected in kNN graph (i.e., sharing at least one common cell) and exhibited concordant significant differential abundance in TNBC-labeled and HER2BC-labeled cells (i.e., corresponding logFC values being both positive or both negative, BH-adjusted P values ≤ 0.1) were merged into two groups of neighborhoods – one enriched in (positive logFC values) and one depleted of (negative logFC values) labeled non-cycling microglia from both brain metastasis models. By comparing between the transcriptome of cells in these two groups using MAST, 838 DEGs with absolute log2 fold change ≥ 0.322 (= log2(1.25)) and Benjamini-Hochberg (BH)-adjusted P value ≤ 0.01 were defined to be differentially expressed in tumor-associated microglia in reference to unassociated homeostatic microglia. Given fewer non-cycling microglia captured in experiment 2 (1737 cells, compared to 11407 cells from experiment 1), to enhance robustness in calling DEGs, we compared cells from connected phenotypic neighborhoods that displayed concordant significant enrichment in TNBC-labeled and HER2BC-labeled cells (i.e., corresponding logFC values being both positive, BH-adjusted P values ≤ 0.1) to all remaining cells, which allowed pooling more cells to be used as control. Running MAST with the same thresholds identified 546 DEGs between these two groups of cells, 501 of which (92%) overlapped with the list of 838 DEGs from experiment 1. As shown in Figure S4C, the logFC values of experiment 2 spanned a narrower dynamical range compared to experiment 1, possibly due to the limited total number of cells, which affected statistical power, and the less restricted selection of unlabeled control microglia that was performed to counteract cell number limitation. As evaluated by linear regression performed with the scikit-learn library (version 0.24.1), despite differences in amplitude, the logFC values of all 11379 genes tested by MAST (which dropped inestimable genes) were well correlated between experiments 1 and 2 (R² = 0.96, see Table S3 for MAST results), indicating consistent metastasis-induced changes in the global transcriptome of microglia in both experiments.

Identifying gene programs associated with homeostasis-to-DAM transition

Considering potentially higher sensitivity in calling DEGs in experiment 1, we used its resulting larger set of 838 DEGs to identify gene programs associated with homeostasis-to-DAM transition by pre-ranked gene set enrichment analysis of these DEGs (Table S3). The analysis was run using the GSEA software^130,131 (version 4.0.3) with the default of 1000 permutations. DEGs were pre-ranked according to their Spearman correlation coefficients, calculated with the SciPy package (version 1.4.1) in python, between the diffusion component 1 (DC1) values and neighborhood-level gene expression of all index cells of phenotypic neighborhoods. As noted above, DC1 values were computed by Palantir⁵³ on the 98 PCs of non-cycling microglia from experiment 1 (see Differential sample abundance in scRNA-seq data analysis). Neighborhood-level expression of a given gene and index cell was calculated as the mean log-transformed, normalized UMI counts of the gene among cells in the corresponding neighborhood. GSEA results obtained using a comprehensive collection of reference gene sets, including curated gene programs (organized in Table S2) and mouse Hallmark (MH) and Biological Process (BP) subset of ontology gene sets (M5)^130,132,133 displayed multiple positively enriched oxidative phosphorylation-related BP gene sets that largely overlapped with each other and with the oxidative phosphorylation gene set of Hallmark as well. To reduce redundancy and provide more informative graphical presentation, Figure 3D summarized the results of GSEA run without BP gene sets (see Table S4 for both sets of GSEA results, run with or without BP gene sets). Among the gene sets shown, oxidative phosphorylation and glycolysis gene sets were obtained from the Hallmark database¹³², and actin remodeling gene set from the GO term of actin polymerization or depolymerization (GO:0008154)¹³². Following curated genes sets were provided in Table S2: DAM signature (88 genes)⁴⁶; axon tract-associated microglia (ATM) signature and CNS pathology-shared microglia responses (i.e., a core, common set of 12 genes upregulated in ATM, DAM, and LPC-induced injury responsive microglia or IRM in abbreviation)⁴⁹; TGF-β-dependent genes, determined specifically in microglia either in vitro or in vivo⁵⁴; genes dependent on LRRC33 (a membrane-anchored carrier and activator of latent TGF-β required for microglia homeostasis)⁵⁵; homeostasis signatures^47,134; type I IFN response genes^82,135.

Visualizing gene expression patterns

We computed the signature scores for homeostasis and DAM-related gene programs (all listed in Table S2) on individual non-cycling microglia cells from both experiments, using the score_genes function of Scanpy¹²¹ as described above (see Inferring tissue dissociation-associated ex vivo activation and cell cycling status in macrophages). We then calculated the neighborhood-level signature scores in particular experiment as the mean scores of cells in a neighborhood, similar to that of the neighborhood-level gene expression of individual DEGs (see Identifying gene programs associated with homeostasis-to-DAM transition in scRNA-seq data analysis). We evaluated the correlation in neighborhood-level Axl expression and actin remodeling signature score across neighborhoods by linear regression, using scikit-learn library (version 0.24.1). To fit the trends of neighborhood-level signature scores and gene expression along the DC1 values of phenotypic neighborhood index cells (see Differential sample abundance analysis for sampling index cells), we used Palantir⁵³, which employs the generalized additive model (GAM)¹³⁶ to derive robust estimate of the nonlinear expression trends and estimate the standard error of prediction¹³⁶. As a convenient way of implementing the trend fit, we ran the compute_gene_trends function in Palantir⁵³, using the neighborhood-level expression and DC1 values of index cells (instead of the default single-cell expression and pseudo-time) as the input of the function. As visualized by the plot_gene_trend_heatmaps function of Palantir⁵³, the GAM-predicted expression trends were presented as heatmaps, with the values z-normalized across the index cells per gene or gene set. To detect genes whose expression trends followed the differential abundance in TNBC-labeled microglia along DC1, we clustered the DEGs by the cluster_gene_trends function of Palantir⁵³, which z-normalizes the gene expression trends (as performed for visualization) to bring the DEGs to the same scale, and performs PhenoGraph¹²⁴ clustering on the z-normalized trends, using a high value of k = 150 to avoid over-fragmented clustering (see Table S3 for full clustering results). We identified the cluster four genes that contained Tnf, Ccl4, and Il1b, whose transcripts were enriched in TNBC-associated microglia and known to be potently upregulated in the stage 1 DAM of Alzheimer’s disease⁶¹. The gene trend clustering was performed only for experiment 1, as the aforementioned genes that encode proinflammatory cytokines were found to be depleted in experiment 2 possibly due to the transcription and translation inhibitors added⁴³. All other steps were conducted for both experiments 1 and 2, which replicated the transition from homeostasis to differential stages of DAM in metastasis-associated microglia.

Type I IFN response analysis and visualization

The in vitro TNC treatment time course revealed progressive upregulation of DAM marker genes that succeeded the type I IFN response following IFN-β production (Figure 7D). To explore potentially analogous gene expression dynamics associated with the homeostasis-to-DAM transition in vivo, we first ran Phenograph¹²⁴ (version 1.5.6, k = 30) to cluster the non-cycling microglial cells from each scRNA-seq experiment, using the top PCs selected as detailed in Differential sample abundance visualization and analysis. We then used MAST¹²⁹ (version 1.16.0) to call the top DEGs of each cluster, compared to the rest of the cells from the experiment. This analysis detected, in each experiment, a particular Phenograph cluster distinguished by potent induction of type I IFN response genes^82,135 as its top upregulated DEGs (Figures 7E and S9A; Table S6). Of note, due to the high homology and short length of IFN genes, the Ifnb1 transcripts captured by 3′ end scRNA-seq often failed to accurately map to the reference genome and were filtered out by SEQC⁵² in generating the cell × gene count matrix.

Experiment 1	Phenograph cluster	0	1	2	3	4	5	6	7	8	9
	Number of cells	1992	1813	1682	1257	1202	1198	859	857	363	184
Experiment 2	Phenograph cluster	0	1	2	3	4	5	6	7
	Number of cells	381	287	286	275	271	119	81	37

We visualized gene expression dynamics using the matrixplot function of Scanpy¹²¹, displaying cluster-averaged Axl expression (i.e., log-transformed, normalized UMI counts) and indicated gene signature scores. The Phonograph clusters were organized by increasing DAM scores, and each expression feature standardized to between 0 and 1 across the clusters.

Analysis of minority cycling microglia

As with non-cycling microglia, the minority cycling microglia (MG) from each experiment were individually extracted from the log-transformed, normalized count matrix encompassing all macrophages and microglia in both experiments to be subject to further analysis. To test whether these subsets of macrophages (MΦ) and microglia exhibit tumor subtype-dependent responses to metastases, for each subset of cells from an experiment, we performed PCA with their top 5000 highly variable genes (HVGs), constructed a kNN graph (k = 15) using the following indicated number of PCs, and ran the customized Milo analysis as detailed above for non-cycling microglia (see Differential sample abundance visualization and analysis).

Experiment	Minority MΦ/MG subsets	FACS group	Source	Number of cells	Number of PCs	Variance explained
1	Cycling microglia (MG)	Unlabeled	TNBC	1	111	0.62
			HER2BC	7
		Labeled	TNBC	107
			HER2BC	231
2	Cycling microglia (MG)	Unlabeled	TNBC	2	90	0.64
			HER2BC	4
		Labeled	TNBC	72
			HER2BC	254

As summarized in the table above, cycling microglia contained few unlabeled cells (altogether 8 and 14 cells in experiment 1 and 2, respectively). Such scarcity of unlabeled cycling microglia was consistent with the non-cycling microglia result, showing that the unlabeled cells distal to brain metastases predominately existed in the homeostatic state devoid of active proliferation¹³⁷. To map the responses of cycling microglia to TNBC versus HER2BC brain metastases, we tested whether the sampled phenotypic neighborhoods were more abundant in HER2BC-labeled cells than in TNBC-labeled cells by running the log-likelihood ratio test on whether β_{nHER2BC–labeled} – β_{nTNBC–labeled} differed from 0. The neighborhoods displaying positive logFC values and BH-adjusted P values ≤ 0.1 were classified as associated with HER2BC-labeled cycling microglia, while the others as associated with both TNBC- and HER2BC-labeled cycling microglia (Figure S4H, the following table). We subsequently calculated the neighborhood-level signature scores for all cycling microglia neighborhoods (as we did for non-cycling microglia, see Visualizing gene expression patterns). Of note, we did not perform cell cycle regression on cycling microglia to avoid potential arbitrary alternation to their global transcriptome. Also, as explained in Inferring tissue dissociation-associated ex vivo activation and cell cycling status in macrophage, the single-cell signature scores that were averaged per neighborhood to obtain neighborhood-level scores for either cycling or non-cycling microglia were originally computed on the unified count matrix that included all macrophages and microglia. These two details altogether guaranteed that the signature scores of interest were directly comparable between different microglia subsets and experiments.

To examine the phenotypic neighborhoods associated with HER2BC-labeled or with both TNBC- and HER2BC-labeled cycling microglia in reference to the full spectrum of homeostasis-to-DAM transition detected in non-cycling microglia, we inspected the neighborhood-level signature scores for the following categories of neighborhoods.

Phenotypic neighborhoods associated with various microglia sources		Criteria for categorizing the neighborhoods
		Threshold logFC and BH-adjusted P values		Biological interpretation
Non-cycling microglia	Unlabeled	Experiment 1	logFCnLabeled vs.Unlabeled < 0, adj. PnLabeled vs.Unlabeled ≤ 0.1	Statistically enriched in unlabeled cells
		Experiment 2*	logFCnTNBC–labeled vs.Unlabeled < 1, logFCnHER2BC–labeled vs.Unlabeled < 1	Not enriched in either TNBC- or HER2BC-labeled cells
	TNBC-labeled	logFCnTNBC–labeled vs.Unlabeled > 0, logFCnTNBC–labeled vs.HER2BC–labeled > 0, adj. PnLabeled vs.Unlabeled ≤ 0.1		Statistically enriched in labeled cells, with highest abundance in TNBC-labeled cells
	HER2BC-labeled	logFCnHER2BC–labeled vs.Unlabeled > 0, logFCnHER2BC–labeled vs.TNBC–labeled > 0, adj. PnLabeled vs.Unlabeled ≤ 0.1		Statistically enriched in labeled cells, with highest abundance in HER2BC-labeled cells
Cycling microglia	TNBC- & HER2BC-labeled	The rest		Undifferentiable abundance in TNBC-labeled or HER2BC-labeled cells
	HER2BC-labeled	logFCnHER2BC–labeled vs.TNBC–labeled > 0, adj. PnHER2BC–labeled vs.TNBC–labeled ≤ 0.1		Statistically enriched in HER2BC-labeled cells

^*Given the smaller number of unlabeled non-cycling cells collected in experiment 2 (217 cells, compared to 2289 cells in experiment 1), which limited the sensitivity of Milo analysis, the criteria for calling neighborhoods associated with unlabeled non-cycling microglia was relaxed to identify sufficient, representative ones for comparison with the other categories of neighborhoods.

As summarized in Figure S4H, both scRNA-seq experiments corroborated the conserved trends of transition from homeostasis to TNBC-enriched stage 1 DAM and HER2BC-enriched stage 2 DAM in the metastasis-associated microglia displaying either low or high scores of S and G2/M phases, corresponding to the majority non-cycling and minority cycling microglia, respectively.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical analysis for non-sequencing data

Sample sizes were indicated in figures or figure legends. Statistic tests were performed in the Prism software (GraphPad, version 8.4.3). P values were computed by two-tailed Mann-Whitney U test unless otherwise noted. Box-whisker plots were shown to summarize the values of minimum, lower quartile, median, upper quartile, and maximum. Web-based survival analysis tool km-plotter (www.kmplot.com) was used to assess the correlation between Affymetrix gene expression profiles and relapse-free survival in the clinical samples of HER2BC breast cancer patients (n = 695). For each queried gene, km-plotter classified all samples into two cohorts using three pre-defined quantiles (i.e., median, upper and lower quartiles) as the cutoff expression value, computed Cox (proportional hazards) regression for each cutoff value and selected the quantile that yielded the highest significance (lowest P values) as the best cutoff value to output the final analysis results displayed as forest plots. Given only six matrisome component-encoding genes were examined, multiple hypothesis test correction was unnecessary and thus not performed.

Supplementary Material

2

Table S2. Reference marker genes of cell type and state clusters and curated gene sets of microglia phenotypes, related to Figures 2 and 3.

3

Table S3. DEGs of breast cancer-labeled microglia, related to Figure 3.

4

Table S4. Microglia GSEA results with and without BP gene sets, related to Figure 3.

5

Table S6. DEGs of type I-responsive microglial clusters, related to Figure 7.

6

Video S1. 3D animation of brain metastases imaged by iDISCO, related to Figures 1 and S1.

Highlights.

• Brain metastases initiate with distinct tumor architectures and stromal interfaces

• Perivascular and spheroidal micrometastases elicit different microglial responses

• Breast cancer cell-derived tenascin C (TNC) drives spheroidal colony growth

• TNC triggers a disease-associated microglia state via type I interferon response