MIF-Induced CD74⁺ Microglia and Macrophages Promote Progression of Brain Metastasis and Are Clinically Relevant across Central Nervous System Disorders

A reprogrammable subset of CD74⁺ microglia/macrophages is a shared population with translational relevance across neurologic diseases that drives pathology in brain metastases.

Abstract

The upregulation of CD74, a chaperone involved in MHC-II antigen processing, has been mostly interpreted as indicative of antigen presentation in multiple brain disorders. However, CD74 expression has also been described in cancer cells across multiple tumor types and in the tumor microenvironment, notably in glioma. In this study, we found that the presence of CD74⁺ microglia/macrophages, which was induced by increased levels of interferon γ in brains affected by metastases, did not relate to its canonical pathway. Instead, the alternative function of CD74 as a cytokine receptor was pivotal. Proliferating cancer cells produced high levels of the ligand migration inhibitory factor (MIF) that bound the CD74 receptor and induced its translocation to the nucleus where it activated an NF-κB–dependent program that promoted metastatic progression. In patients, a CD74 signature was associated with more aggressive progression of brain metastatic disease, although it had no clinical correlation with the matched primary tumor. Interestingly, a pan-disease noncanonical and clinically relevant signature derived from the CD74⁺ myeloid population was identified that occurred in additional brain disorders, including Alzheimer's disease and multiple sclerosis. The brain-penetrant drug ibudilast, which prevents the binding of MIF to CD74, decreased brain metastasis in experimental models in vivo and in patient-derived organotypic cultures ex vivo in a primary tumor–agnostic manner. These findings suggest that MIF/CD74-induced reprogramming of myeloid cells in brain disorders is a vulnerability that could be exploited therapeutically against brain metastases and possibly other brain disorders.

Significance:

A reprogrammable subset of CD74⁺ microglia/macrophages is a shared population with translational relevance across neurologic diseases that drives pathology in brain metastases.

See related commentary by Lee and Kang, p. 3103

Graphical Abstract

Introduction

Secondary brain tumors are a frequent progression of cancer that involves a decreased survival and quality of life despite multimodal treatment approaches, which, however, remain palliative in most patients (1). Whether brain metastases should be treated like their primary tumors is still debated, as brain relapses are increasingly common even in patients who respond well to systemic therapies that work in both sites for certain cancer types (2–4). Emerging strategies target brain-specific microenvironment vulnerabilities (5), including the blood–brain barrier (BBB), whose structure forces metastatic cells to activate specific extravasation programs. For instance, targeting cathepsin S, which disrupts BBB junctions, prevents brain metastases (6). After completing extravasation, metastatic cells must prevent brain-specific responses reacting to the presence of micrometastases (7). With time, the increasing metastatic load influences the brain, inducing altered molecular patterns in subpopulations of reactive astrocytes that promote brain metastasis progression (8). This occurs through their role in blocking the CD8⁺ T cells infiltrating from the periphery (9) as well as by preventing the expression of MHC-I in cancer cells (10). Drugs targeting some of these mechanisms are in clinical trials (NCT05689619). One of the potential benefits of exploiting brain-specific vulnerabilities is the possibility that they could be expanded to other disorders affecting this organ. Given the high demand of effective therapeutic strategies for brain disorders in general, identifying microenvironment-based pan-disease targets remains largely unexplored.

In this study, we show that the expected antitumor nature of microglia/macrophages positive for CD74 (also known as the invariant chain), due to its canonical involvement in antigen processing through the MHC-II complex (11), is rewired into a noncanonical, prometastatic program in the brain metastatic niche. CD74, beyond its classic antigen-presenting role, is also a receptor for the cytokine macrophage migration inhibitory factor (MIF), a proinflammatory molecule involved in various cell processes, including cell survival, proliferation, and migration (12). MIF binding triggers CD74 endocytosis and cleavage, releasing an intracellular domain (ICD) that enters the nucleus and activates NF-κB (13, 14). This signaling cascade reprograms microglia/macrophages into a metastasis-supporting state.

MIF levels are elevated in many tumors, including glioblastoma (GBM), where it can act as a dual driver of tumor progression: acting within tumor cells to promote growth and survival, while simultaneously remodeling the microenvironment to favor immune evasion, myeloid recruitment, and angiogenesis (15). The MIF–CD74 axis promotes progression in multiple cancers, including melanoma (16), non–small cell lung cancer (NSCLC; ref. 17), or breast cancer (18). However, prior studies mostly examined CD74 in cancer cells (16–18) rather than immune cells within the tumor microenvironment. In several tumor types, including breast and melanoma, CD74 expression is enriched in stromal compartments such as macrophages, dendritic cells, and endothelial cells, often correlating with immune suppression and tumor progression (19, 20). In glioma, CD74 is expressed primarily by glioma-associated microglia/macrophages (GAM), with studies supporting both antitumoral effects (21) and immune evasive roles (22). However, secondary brain tumors represent a distinct entity with a more complex and heterogeneous immune microenvironment. Our previous work identified a protumoral role for MIF–CD74 signaling in microglia/macrophages associated with brain metastases (8). Recent work shows that MIF–CD74 inhibition promotes M1 polarization and improves radiotherapy response in brain metastases from NSCLC (23), although the role of CD74⁺ cells in disease progression was not addressed.

We characterize a CD74⁺ microglia/macrophages’ subpopulation in brain metastases with distinct transcriptional features, prometastatic activity, and clinical correlations. This population emerges during metastatic colonization, driven by tumor-derived signals and sustained by mitochondrial plasticity, a known mechanism of macrophage adaptation (24). The resulting CD74⁺ transcriptomic signature also appears in other neurodegenerative and neuroinflammatory disorders.

Ibudilast, a BBB-penetrant MIF–CD74 inhibitor (25) with a good safety profile (26), also affects phosphodiesterases (PDE) in inflammation (27), making on-target validation important. Currently, ibudilast is under investigation in combination with temozolomide for newly diagnosed and patients with recurrent GBM (NCT03782415; ref. 28) and for the treatment of multiple sclerosis (NCT01982942; ref. 26), highlighting its potential role in targeting CD74-mediated pathways in the brain. We show that ibudilast blocks the MIF–CD74 program in CD74⁺ microglia/macrophages, reversing their prometastatic phenotype and restoring immune surveillance. This effect was validated both in preclinical mouse models and in patient-derived organotypic cultures (PDOC) from a range of primary tumor types, including lung, breast, and melanoma.

Together, our findings reveal a brain-specific immune vulnerability in the MIF–CD74 axis in microglia/macrophages, offering a tractable therapeutic strategy for secondary brain tumors and other brain disorders. The accompanying liquid biopsy-compatible biomarker and transcriptomic predictive signature support clinical translation.

Materials and Methods

Animal studies

All animal experiments were performed in accordance with protocols approved by the Spanish National Cancer Research Centre (CNIO), Instituto de Salud Carlos III, and Comunidad de Madrid Institutional Animal Care and Use Committee (PROEX299/15 and PROEX135/19).

CX3CR1-EGFP mice were obtained from The Jackson Laboratory (B6.129P-Cx3cr1tm1Litt/J; 005582). cKO-Oma1 was generated by breeding Ccr2-Cre/ERT2-GFP [C57BL/6-Ccr2 < em1 (cre/ERT2)Peng >/J Stock No. 035229, The Jackson Laboratory] with Oma1^loxP/loxP, kindly shared to Dr. José Antonio Enriquez (CNIC) by Dr. Thomas Langer (MPIBA). Alzheimer's disease mice (P301S Tauopathy, C57BL/6 background) were provided by Dr. Jesús Avila’s laboratory (CBMSO). These transgenic mice develop neurologic symptoms at 3 months, with a median lifespan of 9 to 12 months. The cuprizone demyelinating mice as multiple sclerosis model were provided by F. de Castro’s laboratory (Instituto Cajal-CSIC), as previously described (29) in agreement with the protocols approved by the Instituto Cajal-CSIC, Consejo Superior de Investigaciones Científicas-CSIC, and Comunidad de Madrid Institutional Animal Care and Use Committee (PROEX044_19 and PROEX030.1-24), and demyelination was induced in 6-week-old male C57BL/6 mice by feeding them 0.2% cuprizone ad libitum for 5 weeks.

Brain colonization assays were performed in C57BL/6 mice and athymic nu/nu (ENVIGO #069) for 6 to 10 weeks as previously described (8) by injecting 100 μL of phosphate-buffered saline (PBS) into the left ventricle containing 100,000 cancer cells or 1 μL of PBS intracranially (right frontal cortex, approximately 1.5-mm lateral and 1-mm caudal from the bregma, and to a depth of 2 mm) containing 40,000 cancer cells using a gas-tight Hamilton syringe and a stereotactic apparatus.

Metastatic colonization was analyzed in vivo weekly and ex vivo at the endpoint by bioluminescence imaging (BLI). Anesthetized mice with isoflurane were injected with D-luciferin (150 mg/kg) retro-orbitally and imaged in an IVIS machine. The average radiance (photons/second/cm²/steradian) from each region of interest (ROI) was measured using Living Image software, version 4.5 (SCR_014247). ROIs were created in the 2D bioluminescent image in the head or body of each mouse for in vivo analysis or around the corresponding organ for ex vivo analysis. For intracranial injections, ex vivo values at the end point were normalized to the BLI values of the head in vivo 3 days after injection of the cancer cells.

Cre expression was induced by tamoxifen (TMX) 3 days after intracardiac (iC) inoculation of cancer cells, when it is considered that the first brain metastatic cells have crossed the BBB. TMX (T5648, Sigma-Aldrich) was prepared as a 20-mg/mL stock solution in corn oil. Mice were administered TMX via intraperitoneal injection at a dose of 50 mg/kg body weight. The administration schedule consisted of five consecutive days of TMX injection, followed by a 2-day rest period, and then an additional 5 days of TMX injection until the experimental endpoint. Ibudilast (S4837, Selleck Chemicals and CAS 50847-11-5, ChemSpace) was administered by oral gavage daily (20 mg/kg for immunocompromised models or 30 mg/kg for immunocompetent models), beginning 7 days (immunocompromised model) or 3 days (immunocompetent model) after iC injection, and treatment continued until mice reached the end point of the experiment.

For bone marrow (BM) transplant experiments, host mice were sublethally irradiated with two 4.5-Gy doses within a 4-hour period. BM cells were harvested from donor mice by flushing the tibias and femurs with PBS. The collected BM cells were then filtered through a 70-μm cell strainer. The required number of BM cells were resuspended in PBS for injection. For transplantation, the irradiated host mice were anesthetized, and BM cells were injected retro-orbitally in a volume of 100 μL of PBS containing 2 × 10⁶ cells. Three weeks later, B16/F10-BrM cells were intracardially injected to the mice.

Mouse organotypic cultures

Organotypic cultures from adult mouse brains were prepared as previously described (30). In brief, brains were dissected in HBSS supplemented with HEPES (2.5 mmol/L, pH 7.4), CaCl₂ (1 mmol/L), NaHCO₃ (4 mmol/L), MgCl₂ (1 mmol/L), and D-glucose (30 mmol/L) embedded in 4% low‐melting agarose (Lonza) preheated at 42°C. The embedded brains were cut into 250-mm slices using a vibratome (Leica), and each slice was divided into two pieces at the hemisphere. The slices were placed with flat spatulas on top of 0.8‐μm pore membranes (Sigma-Aldrich) floating on slice culture media [DMEM, supplemented with HBSS 27%, FBS 5%, D-glucose (27 mmol/L), L‐glutamine (1 mmol/L), and 100 IU/mL of penicillin/streptomycin] and subjected to the corresponding procedure.

For the interferon-gamma (IFNγ) experiments, slices were derived from wild-type (WT) brains from C57BL/6 mice 6 to 8 weeks of age. Slices were cultured in the presence or absence of recombinant murine IFNγ 250 ng/mL (315-05-100UG, PeproTech) and recombinant mouse MIF protein 100 ng/mL (1978-MF-0257CF, R&D Systems) for 12 hours. For some experimental conditions, BM-derived cells were isolated from tibias and femur from WT CX3CR1-EGFP mice and placed on top of the brain slice (5,000 cells suspended in 2 μL of RPMI medium).

When the experiments were finished, brain slices were fixed with 4% paraformaldehyde (PFA) overnight at 4°C, and immunofluorescence was performed.

Cell culture

Brain metastatic cell lines (BrM) from human and mouse origins have been previously described. H2030-BrM3 (abbreviated as H2030-BrM; obtained from the Joan Massagué Laboratory, MSKCC; ref. 31). Cells were cultured in RPMI-1640 media supplemented with 10% FBS, 100 of IU/mL penicillin/streptomycin, 1 mg/mL of amphotericin B, and 2 mmol/L of L-glutamine. B16/F10-BrM7 (abbreviated as B16/F10-BrM; generated at the Brain Metastasis Group, CNIO; ref. 8) cells were cultured in DMEM media supplemented with 10% FBS, 100 of IU/mL penicillin/streptomycin, 1 mg/mL of amphotericin B, and 2 mmol/L of L-glutamine. HEK293T cells (CVCL_0063) were cultured in DMEM media supplemented with 10% FBS, 100 IU/mL of penicillin/streptomycin, 1 mg/mL of amphotericin B, and 2 mmol/L of L-glutamine. All the cell lines harbored a plasmid-expressing luciferase under the PGK promoter.

For the in vitro proliferation assay, 1,000 B16/F10-BrM cells per well harboring either shControl (shCtrl) or shMif#1/#2 were seeded in black 96-well plates (Fisher Scientific, ref. 08-772-225). Plates were scanned by the Opera Phenix Plus High-Content Screening System (Perking Elmer) 12 hours, 5 days, and 7 days after seeding. Cells were automatically counted using Harmony v.5.2 software.

B16/F10-BrM and H2030-BrM cells harboring MID KD were generated by lentiviral transduction. HEK293T cells at 70% confluence were transfected in Opti-MEM with Lipofectamine 2000 (Invitrogen) and incubated at 37°C overnight with 8.75 μg of the following plasmids: pMDLg/pRRE (#12251, Addgene), pRSV-Rev (#12253, Addgene), VSV.G (#14888, Addgene), and lentiviral vectors carrying the corresponding shRNA against human MIF (TRC21254, Horizon Discovery), mouse Mif (TRC142338, Horizon Discovery), or the corresponding nontargeting control. The following day, the medium was replaced with DMEM supplemented with 10% FBS and 2 mmol/L of L-glutamine, and virus production was maintained for 36 to 48 hours. The viral supernatant was collected, passed through a 0.45-μm syringe filter, and added to H2030-BrM (human MIF) at 50% confluence in RPMI-1640 supplemented with 10% FBS, 2 mmol/L of L-glutamine, and polybrene (5 μg/mL; Sigma-Aldrich) or B16/F10-BrM (mouse Mif) at 50% confluence in RPMI-1640 supplemented with 10% FBS, 2 mmol/L of L-glutamine, and polybrene (5 μg/mL; Sigma-Aldrich). The following day, the medium was replaced with complete culture media. Selection with puromycin (2 μg/mL; Sigma-Aldrich) was started 24 to 48 hours after and maintained until complete cell death was observed in the noninfected cancer cells. Mycoplasma testing is routinely performed in all cell lines at least every 6 months. None of the cell lines used were Mycoplasma-positive. Cell line authentication is done by in vivo phenotyping based on their organotropism. Experiments have been performed on the lowest passage number possible using frozen stocks from liquid nitrogen. No more than 10 passages have been done with any of the cell lines used.

Immunofluorescence and immunohistochemistry

The tissue for free-floating immunofluorescence was processed after overnight fixation at 4°C with 4% PFA. Slices for PDOCs were obtained using a vibratome (250 mm). Mice brains were processed, and mouse slices were performed using a sliding microtome (80 mm, Thermo Fisher Scientific) or vibratome (250 mm). Slices were blocked using a blocking solution composed by 10% NGS, 2% BSA, and a 0.25% Triton X-100 in PBS for 2 hours at room temperature. Primary antibodies (Supplementary Table S1) were incubated at 4°C overnight in the blocking solution and the following day 30 minutes at room temperature. The appropriate fluorochrome-conjugated secondary antibody (Supplementary Table S2) was added in blocking solution and incubated for 2 hours at room temperature after extensive washing with PBS–Triton 0.25%. After extensive washing with PBS–Triton 0.25%, using DAPI (1 mg/mL, Sigma-Aldrich) for 7 minutes at room temperature, nuclei were stained. After extensive washing with PBS, slices were mounted on glass slides using Fluoromount-G mounting medium (SouthernBiotech).

For immunofluorescence in paraffin sections, WT C57BL/6 mouse brain, liver, lymph node, pancreas, kidney, lung, intestine, and spleen were formalin-fixed and paraffin-embedded. In both cases, sections (3 μm) were obtained. After deparaffinization and rehydration, antigen retrieval was performed by boiling in citrate buffer pH 6 for 10 minutes. Sections were incubated with 3% BSA/0.1% Triton in PBS for 45 minutes at room temperature and incubated with primary antibodies overnight at 4°C. After washing with 0.1% Triton/PBS, the appropriate fluorochrome-conjugated secondary antibodies were added for 45 minutes, sections were washed, and nuclei were counterstained with DAPI. After washing with PBS, sections were mounted with Prolong Gold Antifade Reagent (Life Technologies).

Immunohistochemistry (IHC) of paraffin-embedded tissues was performed at the CNIO Histopathology Core Facility. For the different staining methods, the slides were deparaffinized in xylene and rehydrated by a graded ethanol series to water. Several immunohistochemical reactions were performed on an automated immunostaining platform (Autostainer Link 48, Agilent; Discovery XT-ULTRA, Roche/Ventana). First, antigen retrieval was performed with the appropriate pH buffer, and endogenous peroxidase was blocked (3% hydrogen peroxide). The slides were then incubated with the appropriate primary antibody for single or double staining. Following the primary antibody, the slides were incubated with appropriate secondary antibodies and with horseradish peroxidase (HRP)–conjugated visualization systems when needed.

The immunohistochemical reaction was revealed using ChromoMap DAB, DISCOVERY Purple, or Teal Kit (Roche/Ventana). Nuclei were counterstained with hematoxylin. Finally, slides were dehydrated, rinsed, and mounted for microscopic evaluation. Positive controls for primary antibodies were included in each staining series.

Lysozyme IHC and RNAscope staining methods were performed in an automated immunostaining platform (Ventana DISCOVERY ULTRA, Roche), including deparaffinization and rehydration as a part of the platform protocol with the appropriate probe: probe against Mif/MIF mRNA (R&D Systems, 513809 Q-284534 for mouse and 451049 Q-284534 for human). After the probe, slides were incubated using the corresponding probe amplification kit (RNAscope VS Universal HRP Reagent Kit, ACD, cat. #323210), conjugated with HRP, and the reaction was developed using DAB (DAB Detection Kit, Roche/Ventana, cat. #760-224).

Image acquisition and analysis

Immunofluorescence images were acquired using a Thunder Imaging System (Leica-Microsystems) equipped with AFC, LED8 excitation light source, and a DFC9000GTC camera using 5×/NA 0.12 and 10×/NA 0.45 dry objective for whole-slide imaging under the Navigator software integrated in the LAS X v 3.8.1. For z-stack confocal imaging, a TCS SP8 STED 3X (Leica-Microsystems) equipped with AFC, a white light laser, and PMT and HyD SMD detectors under the Navigator software and a TCS SP5 (Leica-Microsystems) equipped AOBS under the LAS AF v2.6 software were used with 20×/NA 0.7 dry, 20×/NA 0.75 multiple, and 63×/NA 1.4 oil immersion objectives.

For analyzing the CD74 nuclei translocation (Supplementary Fig. 2C), confocal microscopy images were analyzed in 2D using ImageJ (SCR_003070) and custom Java scripts. Nuclei and cytoplasm were segmented using Otsu thresholding and median filtering, followed by binary morphologic operations to refine the segmented areas using MorphoLibJ library (32). Fluorescence intensity was quantified for each z-slice, with mean and sum values calculated for both nuclei and cytoplasm areas. GitHub repository at https://github.com/cnio-cmu-BioimageAnalysis/fluorescenceQuantification_code.

For detecting the different microglia positive populations, confocal microscopy images were analyzed using a combination of Groovy and Python scripts. Initial segmentation of cells was performed in 3D using the deep learning–based Cellpose (33) library with the cyto3 model to accurately identify microglia based on the Iba1 marker. The segmented labels were then processed and analyzed using the mcib3d (34) library. For each marker described, the mean and SD of intensity values were calculated across the entire dataset of microglia masks in 3D. Cells were identified as positive for a given marker if their intensity exceeded the threshold of mean plus one SD (approximately the 84th percentile). GitHub repository: https://github.com/cnio-cmu-BioimageAnalysis/3DmicrogliaAnalysis_code.

IHC images were captured using Zen Blue Software v3.1 (Zeiss). DAB IHC intensity was measured using QuPath v.0.4.2 (SCR_018257).

Processing of mouse brains for flow cytometry analyses

Mouse brains were extracted in precooled D-PBS 1× and were processed using the Adult Brain Dissociation Kit (Miltenyi, ref. 130-107-677) using gentleMACS C Tubes (Miltenyi, ref. 130-093-237) and the gentleMACS Octo Dissociator (Miltenyi, ref. 130-096-427). The resulting homogenates were processed as previously described (35). Homogenates were filtered with a 70-μm strainer and were centrifuged at 300 g for 10 minutes at 4°C. For myelin removal, the pelleted homogenate was then resuspended in 6 mL of 37% isotonic Percoll and then underlayed with 5 mL of 70% isotonic Percoll in a 15-mL centrifuge tube. Tubes were then centrifuged at 600 × g for 40 minutes at 16°C to 18°C, with no acceleration or deceleration. The top layer of myelin was removed using a 10-mL pipette, and cells at the 37% to 70% Percoll interphase were then recovered to 15-mL tubes and washed once in staining buffer. Cells were then incubated in fluorescence-activated cell sorting (FACS) buffer on ice with CD16/CD32 blocking antibody for 15 minutes and incubated for 30 minutes with the corresponding primary antibody (Supplementary Table S3). Fluorescence Minus One, single-color and only cells’ controls were used in all the experiments for fluorescence compensation and gating strategy. After washing, cells were resuspended in staining buffer and acquired using LSRFortessa or FACSCanto II cytometers (BD Biosciences). To collect samples for subsequent bulk RNA sequencing (RNA-seq) or single-cell RNA-seq (scRNA-seq), samples were purified by FACS using a BD FACSAria IIu cell sorter (BD Biosciences). Pulse processing and cell viability dyes were used to exclude cell aggregates and dead cells. Data were analyzed using FlowJo v10.0 (SCR_008520, Treestar, OR).

For bulk RNA-seq analysis, cells were recovered in TRIzol LS reagent (10296010, Thermo Fisher), and RNA was extracted using the PicoPure RNA isolation Kit (Thermo Fisher) according to the manufacturer’s instruction. For scRNA-seq analysis, cells were recovered in 1× PBS containing 0.04% BSA and diluted to a final concentration of 7 × 10⁵ cells/mL.

Multiplex immunoassay for detection of cytokines

The FirePlex-96 key cytokines (mouse) immunoassay panel (ab235656, Abcam) was used for unbiased detection of 17 different murine cytokines in brain tissue lysates from tumor-naïve immunocompromised mice and H2030-BrM intracardially injected mice vehicle- or ibudilast-treated. For generation of brain tissue lysates from tumor-naïve and metastatic mouse brains, brain metastasis was microdissected according to BLI signal and immediately snap-frozen in liquid nitrogen. Frozen tissue was subsequently homogenized and reconstituted in 1× Cell Lysis Buffer (9803, Cell Signaling Technology) plus protease inhibitors. Lysates were centrifuged for 20 minutes at 14,000 × g, and supernatants were used for the multiplex immunoassay at a concentration of 250 μg/mL.

Detection of IFNγ by ELISA

Lysis buffer (Cell Signaling, ref. 9803S), with the following protease inhibitors: 200 mmol/L of Na3VO4, 500 mmol/L of NaF, and 100 mmol/L of PMSF, was used to extract total protein. Protein lysates were obtained from WT C57BL/6 or intracranially injected mice with B16/F10-BrM cells and brain slices from WT athymic nude mice or intracardially injected with H2030-BrM cells. Tumors were obtained by dissecting Luciferase tissue immediately adjacent to Luciferase+ cancer cells. Tissue was mechanically desegregated with the FastPrep-24 5G lysis system (MP Biomedical) using zirconium beads at 6.0 m/s for 15 seconds followed by 10 minutes of incubation on ice before lysis. For protein quantification, the BCA Protein Color Kit was used (Fisher Scientific, ref. 23227). The Invitrogen novex IFNγ mouse ELISA Kit (Life Technologies, KMC4021) was used according to the manufacturer’s instruction.

RNA isolation and qRT-PCR

Whole RNA was isolated using the RNAeasy Mini Kit (QIAGEN). RNA from brain metastatic cell lines (H2030-BrM and B16/F10-BrM cells containing shCTRL or shMIF/shMif) was obtained from a confluent well from a six-well plate. RNA (1,000 ng) was used to generate complementary DNA (cDNA) using the iScript cDNA Synthesis Kit (1708890, Bio-Rad). Gene expression was analyzed using SYBR green gene expression assays (GoTaq qPCR Master Mix, A6002, Promega). The relative gene expression was normalized to a “housekeeping” gene. The following primers were used (5′→3′, forward; reverse):

• B2M (AGA​TGA​GTA​TGC​CTG​CCG​TG; TCA​TCC​AAT​CCA​AAT​GCG​GC).

• B2m (GAC​CGG​CCT​GTA​TGC​TAT​CC; CAG​TAG​ACG​GTC​TTG​GGC​TC).

• ACT (CAA​GGC​CAA​CCG​CGA​GAA​GAT; CCA​GAG​GCG​TAC​AGG​GAT​AGC​AC).

• Act (GGC​ACC​ACA​CCT​TCT​ACA​ATG; GTG​GTG​GTG​AAG​CTG​TAG​CC).

• MIF (TCT​GCC​ATC​ATG​CCG​ATG​TT; TTG​CTG​TAG​GAG​CGG​TTC​TG).

• Mif (CTT​TGT​ACC​GTC​CTC​CGG​TC; CGT​TCG​TGC​CGC​TAA​AAG​TC).

Quantitative PCR was performed on QuantStudio 6 Flex Real-Time PCR System (Applied Biosystems) and analyzed using the software QuantStudio 6 and 7 Flex software.

Bulk RNA-seq

Total RNA samples were processed using the “NEBNext Single Cell/Low Input RNA Library Prep” kit (NEB #E6420) by following the manufacturer’s instructions. Briefly, an oligo(dT) primed reverse transcription with a template switching reaction was followed by double-stranded cDNA production by limited-cycle PCR. (Nondirectional) sequencing libraries were completed using the “NEBNext Ultra II FS DNA Library Prep Kit for Illumina” (NEB #E7805) and subsequently analyzed on Illumina NextSeq 550 (using v2.5 reagent kits) in single-read fashion (85 bases) by following the manufacturer’s protocols. Image analysis, per-cycle basecalling, and quality score assignment were performed with Illumina Real Time Analysis software. Conversion of BCL files into FASTQ format was performed with Local Run Manager Generate FASTQ Analysis Module (Illumina).

Mouse reads were analyzed with the Nextpresso (36) pipeline as follows: Sequencing quality was checked with FastQC v0.11.0. Reads were aligned to the mouse genome (GRCm39) with TopHat-2.0.10 (37) using Bowtie 1.0.0 (SCR_005476; ref. 38) and Samtools 0.1.19 (SCR_002105; ref. 39), allowing three mismatches and 20 multihits. The Gencode (SCR_014966) vM26 gene annotation for GRCm39 was used. Read counts were obtained with HTSeq (SCR_005514; ref. 40). Differential expression and normalization were performed with DESeq2 (41), filtering out those genes in which the normalized count value was less than 2 in more than 50% of the samples. From the remaining genes, those that had an adjusted P value below 0.05 FDR were selected. GSEAPreranked (42) was used to perform gene set enrichment analysis for several gene signatures on a preranked gene list, setting 1000 gene set permutations. Only those gene sets with significant enrichment levels (FDR q value < 0.25) were considered.

The pipeline used to preprocess RNA-seq data is available at GitHub: https://github.com/osvaldogc/nextpresso1.9.2. This section does not report original code.

Bulk RNA-seq data have been deposited, and they are available at GSE293921.

Survival signature

From the deregulated genes identified by the bulk RNA-seq, 151 had orthologues in the human genome, corresponding to 153 unique Ensembl values. Of these, 128 IDs had mRNA expression estimates in a 45-breast cancer brain metastasis (BCBM) patients’ cohort previously described (43) and were further used to build a predictive gene signature model. The “glmnet” R package was used to construct a prognostic model by least absolute shrinkage and selection operator (LASSO) Cox regression. The penalty regularization parameter lambda was selected through 10 times cross-validation using the Harell C index [function cv.glmnet, family = “cox,” type.measure = “C,” set.seed (123) for reproducibility]. Using the minimum deviance parameter lambda.min, 37 genes with coefficients higher or lower than 0 were selected for GS1, and 23 genes were selected for GS2. A risk score model was constructed based on the mRNA expression of each selected gene (log₂CPMTMM normalized values) together with the coefficient generated by LASSO Cox regression.

The risk score was calculated for each patient, and the cohort was divided into high-risk and low-risk groups based on the median value of the risk score. The Kaplan–Meier survival curves were plotted to visualize the association between the risk model and survival post brain metastasis (SPBM).

The association between the risk score model constructed using the cohort of 45 patients and SPBM was validated further by extending the initial cohort with 17 additional patients with BCBM (newly sequenced, not previously published; n = 62).

scRNA-seq library preparation and data analysis

FACS-purified CD45⁺ cells were processed for library preparation using the Chromium Next GEM Single Cell 3′ GEM, Library & Gel Bead Kit v3.1 (10x Genomics, PN-1000121), following the manufacturer’s instructions. The protocol included GEM generation, barcoding, GEM-RT clean-up, cDNA amplification, and library construction. The sequencing was conducted on an Illumina NextSeq 550 instrument equipped with v2.5 reagent kits, using a PairEnd run type of 28-bp R1 and 56-bp R2. Primary data processing, including image analysis, per-cycle basecalling, and quality score assignment, was carried out using Real Time Analysis software (RTA v2, Illumina). Additionally, the conversion of BCL files into FASTQ format was executed with bcl2fastq2 within the Local Run Manager RNAfusion Analysis Module v2 (Illumina).

Seven 10× barcoded sequencing libraries were generated using the 10× Single Cell 3′ v3 kits (10x Genomics, Inc.). These libraries were sequenced in multiple rounds until each sample achieved a mean depth of 25,000 reads per cell. The mean number of reads per cell was calculated by processing the accumulated raw FASTQ files from each sequencing round with CellRanger Count (10X Genomics Cell Ranger v7.1.0) using default parameters. The sequencing data were aligned to the prebuilt 10x mouse GRCh38 (GENCODE v32/Ensembl 98) scRNA-seq reference genome (mm10-2020-A). This analysis yielded a median of 264 million sequenced reads per sample, with individual sample reads as follows: CTR_BrM: 264,577,836; BrM_VEH: 236,382,953; BrM_IBU: 295,303,588; CTR_MS: 96,922,258; MS: 222,018,909; CTR_AD: 335,324,101; and AD: 275,467,604.

To remove potential contamination with human environmental RNA, the raw FASTQ data from the gene expression libraries of all samples were aligned to the human GRCh38 (GENCODE v32/Ensembl 98) and mouse GRCh38 (GENCODE v32/Ensembl 98) 10x prebuilt scRNA-seq reference genome (GRCh38_and_mm10-2020-A). This alignment was processed using CellRanger Count (10X Genomics Cell Ranger v7.1.0) with default parameters.

In total, 12,219 cells were retrieved from CTR_BrM, 9,445 from BrM_VEH, 11,850 from BrM_IBU, 29,818 from CTR_AD, 11,388 from AD, 13,189 from CTR_MS, and 18,849 from MS. Background ambient RNA and barcode-swapped reads were subsequently removed from the count matrix using the Cellbender (SCR_025990) v0.3.0 Python (SCR_024202) package (Python v3.7.12) with default parameters. This processing yielded a refined count matrix, identifying barcodes corresponding to genuine cells.

After applying Cellbender, the cell counts were adjusted to 10,237 in CTR_BrM, 9,452 in BrM_VEH, 11,383 in BrM_IBU, 12,143 in CTR_AD, 10,262 in AD, 3,665 in CTR_MS, and 8,530 in multiple sclerosis. The removal of doublet artifacts was performed using the Scrublet v0.2.3 Python package (Python v3.7.10; ref. 44) with default parameters. Following doublet removal, the final cell counts were 10,159 in CTR_BrM, 9,379 in BrM_VEH, 11,370 in BrM_IBU, 12,143 in CTR_AD, 9,283 in AD, 3,659 in CTR_MS, and 8,528 in multiple sclerosis.

All subsequent single-cell transcriptomic analyses were conducted using R Statistical Software (SCR_001905; v4.2.0) with the Seurat (SCR_007322) v5 (v4.9.9.9059) package (R: The R Project for Statistical Computing; ref. 45). For quality control, human genes and those expressed in fewer than three cells were excluded. After filtering out human genes, cells with more than 5% mitochondrial content or fewer than 1,500 detected mouse genes were also excluded. This filtering resulted in a total of 43,245 high-quality cells, distributed as follows: 7,360 in CTR_BrM, 5,202 in BrM_VEH, 5,870 in BrM_IBU, 9,108 in CTR_AD, 6,363 in AD, 2,885 in CTR_MS, and 6,457 in multiple sclerosis.

For each sample, unique molecular identifier (UMI) count data were normalized using a regularized negative binomial regression approach via SCTransform (SCR_022146; sctransform v0.3.5). SCTransform was executed with the parameter vst.flavor set to “v2” to enhance regularization, and the method was set to “glmGamPoi” to fit a generalized linear Gamma–Poisson model, thereby improving the speed of the learning procedure and enhancing computational efficiency. In the subsequent step, mitochondrial content was regressed out in a second nonregularized linear regression.

To ensure that SCTransform residuals were computed for the top variable genes across all samples, the SelectIntegrationFeatures function was utilized to identify the top 3,000 highly variable genes (HVG). These features were then used in the PrepSCTIntegration function to recompute residuals for any missing values across the selected HVGs, utilizing stored model parameters. Upon obtaining the residuals for the common HVGs, principal component analysis (PCA) was conducted on the SCT-normalized data of each sample, focusing on the top 3,000 HVGs.

Integration anchors were identified using the FindIntegrationAnchors function, specifying SCT as the normalization method, the top 3,000 HVGs as anchor features, and using reciprocal PCA (RPCA) for reduction over the top 30 PCA components, with k set to 20 for anchor selection.

The seven datasets were integrated into a single dataset using the IntegrateData function, applying the previously identified anchors, SCT normalization, and the top 30 dimensions for the anchor weighting procedure. To facilitate dimensionality reduction, a new PCA was performed on the integrated dataset, followed by a nonlinear dimensionality reduction using Uniform Manifold Approximation and Projection (UMAP) applied to the 30 most significant PCA dimensions.

A shared nearest-neighbor graph was constructed utilizing the FindNeighbors function, based on the 30 most significant principal components. Cell clusters were subsequently identified through the implementation of the FindClusters function with default parameters.

For differential expression analysis within the integrated object comprising multiple SCT models, the dataset was prepared using the PrepSCTFindMarkers function. Differentially expressed genes (DEG) between identified clusters were determined via a Wilcoxon rank-sum test, implemented through the Seurat package’s FindAllMarkers function, with default parameters. This analytic approach culminated in the identification of 16 distinct immune subpopulations. Gene expression plots were generated using the FeaturePlot function.

Enrichment scores for the identified gene signatures were computed using the AddModuleScore function with default settings, facilitating a systematic assessment of gene expression patterns across clusters and conditions. The visualizations were enhanced with the “RdBu” gradient color scheme from the RColorBrewer package. To enhance the interpretability of the expression data, the maximum and minimum cutoff values for visualization were adjusted, emphasizing biologically relevant expression ranges.

scRNA-seq data have been deposited and are available at GSE293921.

Sampling of human tissues

Human brain metastasis tissue and CSF were collected by CNIO Biobank as the backbone of a collaborative nationwide multicenter cohort, RENACER, integrated by 19 different hospitals and coordinated from the CNIO Biobank. Written informed consent for each donor was collected from each patient included in this study, and surplus diagnostic samples were shipped to CNIO in less than 24 hours from surgery, under controlled temperature and other preanalytical variables, to warranty homogeneity and quality of the cohort. All the studies were conducted in accordance with recognized ethical guidelines (Declaration of Helsinki) and were approved by our institutional review board (IRB; CEI PI 25_2020-3). Comprehensive clinical information was also collected by CNIO Biobank associated with the samples.

Postmortem cortical brain tissue blocks from patients with multiple sclerosis and controls, as well as the associated clinical and neuropathologic data, were supplied by the Multiple Sclerosis & Parkinson’s Tissue Bank at Imperial College (London, United Kingdom), funded by the Multiple Sclerosis Society of Great Britain and Northern Ireland, registered charity 207495 (F. de Castro’s request approved by communication dated on January 19, 2023). Alternate coronal slices or brain hemispheres were fixed in 4% PFA for ∼2 weeks and were then cryoprotected in 30% sucrose for ∼1 week and frozen in isopentane precooled on a bed of dry ice. Microtome sections (50 μm) were obtained from all the cortical blocks, containing gray and white matters for immunohistochemical analysis as previously described (46). Blocks from patients with multiple sclerosis with white matter lesions and controls were studied here, from individuals with no history of neuropsychiatric disease in either case. The use of the UK Multiple Sclerosis Tissue Bank by F. de Castro’s group at Instituto Cajal-CSIC has been approved (code 073/2021) by the Ethics Committee of Consejo Superior de Investigaciones Científicas (CSIC, Madrid, Spain). Control samples were chosen among those available to best match age, sex, and postmortem interval with multiple sclerosis cases.

PDOCs

A total of 26 brain metastases from patients with lung cancer (11 cases), breast cancer (seven cases), colorectal cancer (three cases), uterine cancer (two cases), melanoma (two cases), and prostate cancer (one case) were obtained from RENACER collaborating hospitals: Hospital 12 de Octubre (Madrid), Hospital La Princesa (Madrid), Complejo Hospitalario Universitario de Albacete, Hospital Universitario de Burgos, Hospital Rio-Hortega (Valladolid), and Hospital Universitario Bellvitge (Barcelona). PDOCs were generated as described previously (30). Briefly, after neurosurgical resection, brain metastasis samples were directly collected in neurobasal media A supplemented with 1× B27 (17504-044, Gibco), 1× N-2 (17502-048, Gibco), 25 ng/mL of EGF (E9644, Sigma-Aldrich), 25 ng/mL of bFGF (13256029, Gibco), 10 ng/mL of NRG1-b1/HRG-b1 (396-HB, R&D Systems), 100 ng/mL of IGF1 (291-G1, R&D Systems), 100 IU/mL of penicillin/streptomycin, and 1 μg/mL of amphotericin B. Tissue was embedded in 4% low-melting agarose and cut in 250-μm slices using a vibratome.

PDOCs were treated with either DMSO or 50 μmol/L of ibudilast for 3 days. The human brain slices were fixed with 4% PFA overnight at 4°C, and a free-floating immunofluorescence was performed afterward. Proliferation was evaluated by manually counting Ki67+ cells in three brain slices per condition, and three fields of view were captured per slice.

Bulk RNA-seq data processing

Raw RNA-seq reads were processed as follows: We used FASTQC (SCR_014583) v0.12.1 to check the quality of the sequencing reads and BBDuk [BBMap (SCR_016965) v38.93] to remove the adapter sequences and read ends with Phred (SCR_001017) quality scores lower than 10 and discard reads shorter than 20 bp. Trimmed reads were aligned to the human genome (GRCh38) with STAR (SCR_004463) v0.74.0 (47) and Samtools v1.14 (39). Finally, mapped reads were counted and aggregated to a matrix of gene-level counts with HTSeq v0.13.5 (40). Genes with less than 10 counts across samples were filtered out. RNA-seq–based counts were normalized using a variance-stabilizing transformation with DESeq2 (SCR_000154) v1.34.0 (41) and then corrected for batch variation while preserving variation associated with the primary tumor with limma (SCR_010943) v3.50.1 (48). Differential expression was performed using DESeq2 (41). We only considered 25 patients from the RENACER cohort from six different primary sites (lung, n = 11; breast, n = 7; melanoma, n = 3; colorectal, n = 3; prostate, n = 1; and uterus, n = 1), who had PDOCs treated with ibudilast and available RNA-seq data.

Definition of an 11-gene signature associated with response to ibudilast

To select genes that were differentially expressed between the responder (n = 16) and nonresponder groups (n = 7), we applied an adjusted P value below 0.05 FDR and identified 51 genes. The responder group comprised nine intermediate responders and nine superresponders. We quantified a 51-gene signature score using the weighted mean method from decoupleR (49) v2.8.0. We assigned a positive unit weight to genes associated with favorable response and a negative unit weight to genes associated with nonresponse.

Considering the original 51-gene signature of ibudilast response, we used a drop out inference approach to optimize our signature genes while retaining predictive value. First, we computed the scores 1,000 times, by dropping a variable percentage of the signature genes each time. This percentage ranges between 5% and 95%. Second, by checking all the occurrences in which a gene was present, we calculated the average mean square error (MSE) to compare the actual predictions with the predictions from the original signature (full score with all genes). Third, we built our ranking on the basis that a gene removal has a bigger effect on the predictions (higher MSE is associated with a gene being most useful/important). To assess the stability of this ranking, we applied bootstrap on the average MSE of each gene creating 100 samples of size equal to the minimum number of occurrences of all genes. The different gene rankings derived from the bootstrap samples showed a nice overlap of the gene ranks. Finally, we reduced the original 51-gene signature to a final signature comprising the 11 top-ranked genes (i.e., dropping out 78% of the original signature genes), as it showed a reasonable compromise between the loss in performance (MSE = 1.83) and the ability to predict ibudilast response [area under the curve (AUC) = 0.89]. Patients were stratified into responders and nonresponders based on the median value of the final 11-gene signature score.

Therapeutic predictions using experimental data from Genomics of Drug Sensitivity in Cancer

A total of 102 cell lines present in both the dataset generated by Kinker and colleagues (50) and the Genomics of Drug Sensitivity in Cancer (GDSC) database were selected for analysis. Enrichment scores for the 11-gene nonresponder biomarker signature were computed at the single-cell level using the UCell package (SCR_027109), and the average enrichment values were obtained for each cell line. Pearson correlations were then calculated between these average enrichment scores and drug AUC values from GDSC. Correlations were performed within each primary cancer type represented in the study (breast cancer, colon adenocarcinoma, GBM, lung adenocarcinoma, lung squamous lung adenocarcinoma, small cell lung cancer, large cell lung cancer, and skin cutaneous melanoma) by grouping the corresponding cell lines. Potential drugs for nonresponders were defined as those showing a significant positive correlation between the markers signature and AUC (meaning less expression of the signature will correlate with lower AUC, aka higher sensitivity). A total of 13 drugs were selected.

Estimation of differential pathway activity

We used the univariate linear model method from decoupleR (49) v2.8.0 to infer pathway activities, based on the t-values of DEGs between nonresponders and responders and the PROGENy (SCR_006647) pathway weighted interaction network (14 pathways; ref. 51).

MIF quantification in clinical samples

Brain metastases were obtained from the CNIO Biobank that previously received them from Hospital Universitario 12 de Octubre and Hospital La Princesa. All samples followed protocols approved by our IRB (CEI PI 25_2020-3) and the IRB of the Department of Neuroscience, University of Turin. Written informed consent was signed by each patient included in this study. IHC was performed at the CNIO Histopathology Core Facility using standardized automated protocols.

MIF detection in CSF

CSF samples from eight noncancer patients were obtained from the Biobank of Hospital Universitario Virgen de la Macarena, and CSF samples from patients with lung cancer brain metastasis (six cases), BCBM (two cases), melanoma brain metastasis (one case), and brain metastasis with other primary tumors (two cases) were obtained from the CNIO Biobank that previously received them from Hospital Universitario 12 De Octubre, Complejo Hospitalario Universitario de Vigo, Hospital Virgen De La Salud de Toledo, Hospital Universitario Vall d’Hebron, and Complejo Hospitalario Universitario Santiago De Compostela. All samples were in compliance with protocols approved by their respective IRB (B.0001601, CEI PI 25_2020-v2, and CEI PI 25_2020-3). Written informed consent was signed by each patient included in this study. MIF levels in patients’ CSF were measured by ELISA following the manufacturer’s instructions (Merck, RAB0360-1KT).

Meta-analysis of our gene signature across publicly accessible human scRNA-seq datasets

The analysis of scRNA-seq data from murine models of BrM, Alzheimer disease, and multiple sclerosis identified a specific gene signature for CD74⁺ myeloid cells, comprising a total of 672 genes. To evaluate the relevance of this signature within human patient datasets for multiple sclerosis and Alzheimer disease, publicly available single-cell expression datasets corresponding to the studies by Jäkel and colleagues (52) for multiple sclerosis and Gabitto and colleagues (53) for Alzheimer disease were accessed, downloaded, and reprocessed as follows.

The multiple sclerosis scRNA-seq expression matrix and the corresponding annotation table were imported into R using the read.delim function. Row names in the annotation table were standardized by replacing colons and slashes with periods to ensure uniformity. Subsequently, a Seurat object was created to encapsulate both the expression data and its associated metadata. This Seurat object was then segmented into a list of individual datasets based on unique identifiers found in the “orig.ident” metadata column. Each dataset underwent normalization and variance stabilization via the SCTransform function, utilizing the “glmGamPoi” method to enhance computational efficiency.

Following normalization, SCTransform residuals were computed for the top variable genes across all datasets. The top 3,000 HVGs were selected as integration features using the SelectIntegrationFeatures function and were subsequently utilized in the PrepSCTIntegration function to recompute residuals for any missing values across these 3,000 HVGs, leveraging the stored model parameters. Once the residuals for the common HVGs were obtained, PCA was performed on the SCT-normalized data for each dataset, focusing on the top 3,000 HVGs.

Integration anchors were then identified using the FindIntegrationAnchors function, specifying SCT as the normalization method, the top 3,000 HVGs as anchor features, and using RPCA as the reduction method over the top 30 PCA components, with 20 neighbors (k) selected for anchor identification. The individual datasets were subsequently integrated into a single dataset using the IntegrateData function, with the previously established anchors, SCT normalization, and the top 30 dimensions utilized in the anchor weighting procedure. To facilitate dimensionality reduction, a new PCA was executed on the integrated dataset, followed by a nonlinear dimensionality reduction using UMAP applied to the 30 most significant PCA dimensions.

For the analysis of two single-nucleus RNA-seq datasets from Alzheimer disease, specifically derived from the dorsolateral prefrontal cortex and the middle temporal gyrus, data were imported from RDS files into Seurat objects using the readRDS function, ensuring that all relevant information was retained. The disease status for each sample was categorized as either “normal” or “dementia.” Both Alzheimer disease datasets underwent normalization using the LogNormalize method, applying a scaling factor of 10,000 to standardize gene expression levels across the datasets.

Finally, the enrichment score for the Cd74⁺ cell-specific gene signature was computed across the three human datasets using the AddModuleScore function with default parameters. Feature plots were generated using the FeaturePlot function to visualize the spatial distribution of the calculated module scores across the UMAP embeddings for the respective datasets. The visualizations were enhanced with the “RdBu” gradient color scheme from the RColorBrewer (SCR_016697) package, and cutoff values were judiciously adjusted to emphasize the relevant expression ranges, thereby improving the interpretability of the results of the meta-analysis.

Quantification and statistics

Data were analyzed using GraphPad Prism 8 software [GraphPad Software (SCR_002798)]. For comparisons between two experimental groups in datasets that followed a normal distribution, an unpaired or paired (depending on the data), two-tailed Student t test was used. For comparisons that did not follow a normal distribution, a Mann–Whitney U test was performed. For multiple comparisons, ANOVA test was performed. For survival curves, P values were obtained with log-rank (Mantel–Cox) two-sided tests. The χ² test was performed for the comparison of group proportions.

Results

MIF reprograms brain metastasis–associated macrophages/microglia

The experimental and human brain metastases showed high cancer cell–derived MIF (Fig. 1A; Supplementary Fig. S1A–S1E; Supplementary Table S4) and its receptor colocalized with microglia/macrophage markers (Fig. 1B–D). Notably, CD74⁺ cells are absent from the healthy brain parenchyma, making it the only organ without basal CD74 expression under homeostatic conditions (Supplementary Fig. S1F). Thus, the emergence of CD74⁺ microglia/macrophages correlates with the local progression of brain metastases, as this cell subpopulation appeared only in established brain metastases, not in early micrometastatic stages (Supplementary Fig. S1G). In contrast, MIF is already highly expressed by cancer cells during the early stages of brain colonization (Supplementary Fig. S1H), whereas CD74⁺ microglia/macrophages remained absent at these early stages (Supplementary Fig. S1I). Additionally, analysis of other receptors involved in MIF signaling (11) showed that CD74⁺ microglia/macrophages also express CXCR4 and CD44 (Supplementary Fig. S1J–S1L), whereas CXCR2 was not detected in this population (Supplementary Fig. S1J). IFNγ, the main inducer of CD74 expression (54), was increased during metastatic colonization of the brain (Supplementary Fig. S2A and S2B). Recombinant IFNγ induced CD74 in microglia and BM-derived macrophages (BMDM) when evaluated in an ex vivo system (Supplementary Fig. S2C–S2H). Thus, tumor-derived MIF and CD74⁺ microglia/macrophages are consistent features of mouse and human brain metastases irrespectively of their primary source.

Figure 1.

MIF as a reprogramming factor for brain metastasis macrophages/microglia. A, Representative image of human brain metastasis with RNA in situ hybridization using RNAScope against MIF and GFAP IHC. Scale bar, 20 μm. B, IHC against CD74 and Iba1 in human brain metastasis from two different patients. The black dotted lines delineate the tumoral area. Scale bar, 10 μm. C, Immunofluorescence against CD74 and Iba1 in mouse-established brain metastasis from H2030-BrM (GFP, left; scale bar, 25 μm) and B16/F10-BrM (Luc, right; scale bar, 50 μm; magnification scale bar, 10 μm). D, Quantification by immunofluorescence of CD74⁺ Iba1+ cells, which are microglia or BMDMs, in H2030-BrM and B16/F10-BrM experimental brain metastasis models, from representative images in C. n = 6 brains per mouse model, with 25–30 fields of viewper mouse. Bar graph represents mean + SD. E, Schema of the experimental design. H2030-BrM cells harboring a knockdown for MIF using two different shRNAs were intracardially injected in immunocompromised WT mice. F, BLI in vivo of mice injected with either H2030-BrM WT cells or knockdown for MIF at endpoint (32–35 days after iC injection). Ex vivo BLI of brains (down) is also shown. G, Quantification of ex vivo BLI of brains from F. Values are shown in a box-and-whisker plot in which each dot represents a different animal and the line corresponds to the median. n = 17 (shCTRL), n = 18 (shMIF#1), and n = 19 (shMIF#2) mice from two independent experiments. P value was calculated using an unpaired two-tailed t test of each experimental condition against control. H, Immunofluorescence images in brain metastasis tumors derived from H2030-BrM cell line harboring or not a knockdown for MIF (shMIF). Top, the tumors are stained against CD74, Iba1, and GFP (labeling the H2030-BrM cancer cells). Bottom, the same tumoral area stained against CD74 and DAPI. White dotted lines outline DAPI⁺ nuclei. Scale bar, 10 μm. I, Quantification of CD74 found in the nucleus (DAPI) vs. the membrane/cytoplasm on images shown in H. Values are shown in a box-and-whisker plot in which each dot represents the mean percentage of CD74-ICD found in the nuclei per animal, and the line corresponds to the median. Quantification was done in Iba1⁺ cells. In all, 3–12 fields of view were captured per brain in n = 8 animals from each experimental condition from two independent experiments. P value was calculated using an unpaired two-tailed t test of each experimental condition against control. J, Immunofluorescence images in brain metastasis tumors derived from H2030-BrM cell line harboring or not a knockdown for MIF (shMIF). Top, the tumors are stained against CD74, MDK, Iba1, and GFP (labeling the H2030-BrM cancer cells). Bottom, the same tumoral area but only with the CD74 and MDK staining. White dotted lines outline the CD74⁺ MDK⁻ cells, whereas yellow dotted lines indicated with an arrow delineate the double-positive CD74⁺ MDK⁺ cells. Images with the red channel only (MDK) are shown for clarity. Evaluation of the expression was done in Iba1⁺ cells. Scale bar, 10 μm. K, Quantification of images shown in J, represented as the mean percentage of CD74⁺ microglia/macrophages (Iba1⁺) positive for MDK normalized to the total CD74⁺ microglia/macrophages (Iba1⁺). Values are shown in a box-and-whisker plot in which each dot represents the mean percentage of CD74⁺ MDK⁺ cells per animal. In all, 3–12 fields of view were captured per brain in n = 6 brains from each experimental condition from two independent experiments. P value was calculated using an unpaired two-tailed t test of each experimental condition against control.

Independently of its role as an MHC-II chaperone, CD74 can act as a cytokine receptor. Upon MIF binding, CD74 undergoes endocytosis and cleavage, generating an ICD that enters the nucleus and activates NF-κB–dependent gene expression (14). To test whether MIF influences CD74⁺ microglia/macrophages, we first treated ex vivo control metastasis-naïve brain slices with IFNγ alone or IFNγ and MIF (Supplementary Fig. S2I). According to existing bibliography (13, 14, 54), only when both cytokines were combined, nuclear translocation of CD74-ICD could be observed (Supplementary Fig. S2J and S2K).

To evaluate MIF’s contribution, we knocked down MIF in brain-tropic lung adenocarcinoma (31) and melanoma (8) cell lines (Fig. 1E; Supplementary Fig. S3A–S3C). MIF knockdown did not affect in vitro growth (Supplementary Fig. S3D). In vivo, MIF loss significantly reduced brain metastatic burden by BLI (Fig. 1F and G; Supplementary Fig. S3E–S3H). When systemic inoculation was used [intracardiac (iC)], MIF loss also modestly reduced lung metastases (Supplementary Fig. S3I–S3L), but the effect was far weaker than in brain (Fig. 1F and G; Supplementary Fig. S3G and S3H), reinforcing the organ-specific relevance of this phenotype. This reduction in brain metastasis corresponded with decreased nuclear accumulation of CD74-ICD in metastasis-associated microglia/macrophages (Fig. 1H and I). Consistent with MIF-induced CD74 signaling, NF-κB was active in CD74⁺ microglia/macrophages (Supplementary Fig. S3M and S3N). Among NF-κB–dependent genes involved in cell survival and inflammation (13, 14), midkine (MDK; ref. 55), a heparing-binding growth factor, plays a crucial role in cancer progression. MDK has been shown to contribute to metastatic progression by promoting cell survival, proliferation, and migration of cancer cells (56). We confirmed reduced expression of MDK in the CD74⁺ microglia/macrophages upon MIF knockdown in the cancer cells (Fig. 1J and K).

Thus, tumor-derived MIF seemed required to rewire microglia/macrophages into a prometastatic CD74⁺ subpopulation in secondary brain tumors.

A brain-specific poor prognosis signature derived from CD74⁺ macrophages/microglia

Our functional data indicate that CD74⁺ microglia/macrophages form a disease-associated population. To evaluate its contribution to the human disease, we scored whether a transcriptomic signature derived from experimental brain metastases–associated CD74⁺ microglia/macrophages could correlate with clinical parameters. We first compared bulk RNA-seq profiles of CD74⁺ and CD74⁻ Cx3cr1⁺ Cd11b⁺ microglia/macrophages from two brain metastasis models (Fig. 2A; Supplementary Fig. S4A and S4B). CD74⁺ cells showed upregulation of tumor-associated and protumoral macrophage markers (Supplementary Fig. S4C–S4G; Supplementary Table S5). We next scored these DEGs (Fig. 2B; Supplementary Table S6) in a breast cancer brain metastasis cohort with matched primary tumors (43). This analysis identified a 37-gene signature (Supplementary Table S7) that stratified survival after brain metastasis diagnosis (Fig. 2C). This stratification was absent in matched primary breast tumors (Fig. 2D), suggesting an organ-specific contribution of CD74⁺ microglia/macrophage subpopulations. We next sought to interrogate this signature in additional primary tumors using The Cancer Genome Atlas (TCGA) datasets. As a positive control to validate that the scoring approach captures a known CD74⁺ protumoral program, we first examined GBM (TCGA-GBM; Supplementary Fig. S5A) and lower grade glioma (TCGA-LGG; Supplementary Fig. S5B), in which CD74⁺ GAMs have been previously reported to promote tumor progression (22). As expected, high signature expression stratified patients with significantly poorer survival in these primary brain tumors. Applied to primary tumors that most frequently metastasize to the brain (lung, breast, melanoma, and colorectal carcinomas; Supplementary Fig. S5C–S5H), the signature did not stratify outcomes. These findings support that the CD74⁺ program’s prognostic value is brain-specific.

Figure 2.

The prometastatic signature of reprogrammed CD74⁺ macrophages/microglia has clinical relevance. A, Representative flow cytometry plot from healthy and H2030-BrM metastatic brains, coming from populations shown in Supplementary Fig. S4B. Each graph represents a pool of three brains. CD74⁻ microglia/macrophages are outlined in red and CD74⁺ in gray. B, Heatmap of the top 150 upregulated and 150 downregulated genes from bulk RNA-seq analysis of CD74⁺ microglia/macrophages. Genes listed on the right constitute the 37-gene human signature. C, Survival graph of 62 patients with brain metastasis from breast cancer. CD74 signature was scored in RNA-seq data from brain metastasis human samples. The orange line represents patients enriched in the 37-gene signature derived from the CD74⁺ bulk RNA-seq signature (B), and the black line represents patients with lower enrichment of this signature. D, Survival graph of 62 patients with brain metastasis from breast cancer. CD74 signature was scored in RNA-seq data from breast primary tumors (matched samples from C). The orange line represents patients enriched in the 37-gene signature derived from the CD74⁺ bulk RNA-seq signature (B), and the black line represents patients with lower enrichment of this signature.

We next analyzed CD74⁺ populations using existing scRNA-seq data from brain metastasis (9). Cd74 was expressed in both microglia and BMDMs (Supplementary Fig. S6A and S6B). Although minimal expression was also detected in astrocytes (Supplementary Fig. S6B) as it has been previously reported (57), we could not confirm this finding at the protein level (Supplementary Fig. S6C). We therefore focused subsequent analysis on these two myeloid populations. When microglia and BMDMs were segregated using established lineage signatures (Supplementary Fig. S6D and S6E; ref. 58), Cd74 expression was enriched in the BMDMs (Supplementary Fig. S6F). Immunofluorescence and flow cytometry confirmed the predominance of the BM origin within the brain metastasis–associated CD74⁺ population in experimental models in situ (Supplementary Fig. S6G–S6K).

To test whether CD74⁺ BMDM alone reproduced the clinical correlation (Fig. 2C), we use Ccr2 expression as an established marker of the BM origin of macrophages (Supplementary Fig. S6L; ref. 59) to identified DEGs between Ccr2⁺; Cd74⁺ and Ccr2⁻; and Cd74⁺ populations (Supplementary Table S8). A Ccr2⁺; Cd74⁺–derived signature best reproduced the clinical association in brain metastasis (Supplementary Fig. S6M) and, like the full signature, did not stratify matched primary tumors (Supplementary Fig. S6N).

Thus, our data confirmed the clinical value of a signature derived from CD74⁺ microglia/macrophage and pointed to CD74⁺ BMDMs as a contributor to the poor prognosis signature in brain metastasis.

The protumoral activity of CD74⁺ macrophages/microglia relies on mitochondria-mediated plasticity

Transcriptomic analysis of CD74⁺ cells showed strong oxidative phosphorylation (OXPHOS) upregulation (Supplementary Figs. S4A and S7A; Supplementary Table S9), consistent with protumoral macrophages’ metabolism (60). Because CD74 and Ccr2 expression overlapped (Supplementary Figs. S6L, S7B, and S7C) and Ccr2⁺; Cd74⁺ BMDM correlated with poor prognosis (Supplementary Fig. S6M), we tested whether targeting components of the mitochondria required for OXPHOS-mediated plasticity could disrupt the prometastatic function.

Beyond the enrichment of OXPHOS in the CD74⁺ microglia/macrophages’ bulk RNA-seq (Supplementary Figs. S4A and S7A; Supplementary Table S9), we also identified the Oma1-related pathway (Fig. 3A). Oma1 is a major determinant of mitochondria quality and critical to ensure its ability to sustain OXPHOS-mediated cellular plasticity (61). To study the potential dependency of the protumoral role of CD74⁺ macrophages on mitochondria-mediated plasticity, we performed brain metastasis in vivo assays using the Ccr2-Cre^ERT2;Oma1^loxP/loxP genetically engineered mouse model (GEMM; Fig. 3B). The GEMM showed a decreased viability of cancer cells in the brain as evaluated with both bioluminescence and histology (Fig. 3C and D; Supplementary Fig. S7D–S7F), however, without affecting extracranial metastases (Supplementary Fig. S7G and S7H). Although brain metastasis–associated CD74⁺ microglia/macrophages numbers were unchanged (Fig. 3E and F), MDK⁺CD74⁺ cells were reduced in the GEMM, indicating impaired NF-κB activation (Fig. 3G and H).

Figure 3.

Brain metastasis–promoting activity depends on CD74⁺ macrophages/microglia mitochondria-mediated plasticity. A, Heatmap of Oma1 pathway–related genes that are upregulated in the CD74⁺ vs. CD74⁻ microglia/macrophages. B, Schema of the mouse model. Mice carry a loxP site flanking the Oma1 sequence and Cre under the Ccr2 promoter. Mice were intracardially injected with B16/F10-BrM cells, and TMX was administered to induce Cre expression 3 days after iC injection. C, Representative BLI in vivo images of WT and cKO-Oma1 mice 14 days after iC injection. Ex vivo BLI of brains (down) is also shown. D, Quantification of ex vivo BLI of brains from C. Values are shown in a box-and-whisker plot in which each dot represents a different animal, and the line corresponds to the median. n = 33 (WT) and 35 (cKO) mice from three independent experiments. P value was calculated using an unpaired two-tailed t test. E, Immunofluorescence against CD74, GFP (labeling Cre expression under the Ccr2 promoter), and Iba1 in B16/F10-BrM tumors from WT and cKO mice. F, Quantification of images shown in E, representing the total numbers of CD74⁺ microglia/macrophages (MG/BMDM). Values are shown in a box-and-whisker plot in which each dot represents the mean number of cells per brain, and the line corresponds to the median. n = 10 (WT) and 9 (cKO) mice, and 3–6 fields of view per mouse, from two independent experiments. P value was calculated using an unpaired two-tailed t test. G, Immunofluorescence in B16/F10-BrM tumors from WT and cKO mice stained against Iba1, CD74, midkine (MDK), and Luciferase (Luc; labeling the B16/F10-BrM cancer cells). Bottom, the staining for CD74 and MDK only. Dotted lines surround the CD74⁺ cells that are MDK⁺ (yellow) or MDK⁻(white). Double-positive CD74⁺ MDK+ cells are also indicated with a yellow arrow. Images with the red channel only (MDK) are shown for clarity. Evaluation of the expression was done in Iba1⁺ cells. Scale bar, 25 μm. H, Quantification of images shown in G representing the percentage of CD74⁺ MDK⁺ microglia/macrophages (MG/BMDM) out of the total population of CD74⁺ Iba1⁺ cells. Values are shown in a box-and-whisker plot in which each dot represents the mean percentage of double-positive CD74⁺ MDK⁺ cells per brain, and the line corresponds to the median. n = 10 (WT) and 9 (cKO) brains and 4–12 fields of view per mouse, from two independent experiments. P value was calculated using an unpaired two-tailed t test.

Thus, targeting mitochondria-mediated plasticity in Ccr2⁺CD74⁺ limits brain metastasis.

Ibudilast challenges experimental brain metastases blocking CD74⁺ macrophages/microglia reprogramming

Given the absence of CD74⁺ microglia/macrophages in healthy brain (Supplementary Fig. S1F), and the availability of the brain-penetrant MIF–CD74 inhibitor ibudilast (25, 26) tested in other brain disorders (NCT03782415 and NCT01982942) with a high safety profile, we evaluated its therapeutic potential. First, we confirmed that ibudilast did not affect cancer cells’ viability in vitro (Supplementary Fig. S8A–S8C). In contrast, in two brain metastasis models (H2030-BrM and B16/F10-BrM), ibudilast reduced intracranial tumor burden (Fig. 4A and B; Supplementary Fig. S8D–S8K), largely recapitulating MIF knockdown phenotypes (Fig. 1E–G; Supplementary Fig. S3E–S3H). Similarly to the genetic approach, extracranial metastasis showed only modest reduction (Supplementary Fig. S8L–S8O).

Figure 4.

Ibudilast challenges brain metastases by targeting reprogrammed CD74⁺ macrophages/microglia. A, Representative BLI in vivo images of vehicle- or ibudilast-treated mice at endpoint (35 days after iC injection). Ex vivo BLI of brains is also shown. B, Quantification of ex vivo BLI of brains from A. Values are shown in a box-and-whisker plot in which each dot represents a different animal, and the line corresponds to the median. n = 28 (vehicle) and 27 (ibudilast) mice from three independent experiments. P value was calculated using an unpaired two-tailed t test. C, Immunofluorescence against CD74 and Iba1 in H2030-BrM tumors from vehicle- and ibudilast-treated mice. Scale bar, 50 μm. D, Quantification of images shown in C representing the percentage of CD74⁺ area normalized to the Iba1⁺ area. Values are shown in a box-and-whisker plot in which each dot represents the mean percentage of CD74⁺ normalized to the Iba1+ area per brain, and the line corresponds to the median. n = 7 mice and 4–8 fields of view per mouse. P value was calculated using an unpaired two-tailed t test. E, Immunofluorescence images in H2030-BrM tumors from vehicle- and ibudilast-treated mice. Top, brain tumor sections are stained against GFP (H2030-BrM cancer cells), Iba1, CD74, and DAPI. Bottom, CD74 and DAPI shown for clarity. White dotted lines outline the DAPI⁺ nuclei. Scale bar, 10 μm. F, Quantification of CD74-ICD found in the nucleus (DAPI) vs. the membrane/cytoplasm on images shown in E. Measures were performed in Iba1⁺ cells by an automatic software. Values are shown in a box-and-whisker plot in which each dot represents the mean percentage of CD74-ICD found in the nuclei per animal, and the line corresponds to the median. Ten fields of view were captured per brain in n = 6 (vehicle) and n = 8 (ibudilast) mice. P value was calculated using an unpaired two-tailed t test of each experimental condition against control. G, Immunofluorescence in H2030-BrM tumors from vehicle- and ibudilast-treated mice. Top, brain tumor sections were stained against CD74, MDK, Iba1, and GFP (labeling H2030-BrM cancer cells). The panel below only includes CD74 and MDK staining. White dotted lines outline CD74⁺ MDK⁻ cells, whereas yellow dotted lines indicated with an arrow delineate the double-positive CD74⁺ MDK⁺ cells. Images with the red channel only (MDK) are shown for clarity. Evaluation of the expression was done in Iba1⁺ cells. Scale bar, 10 μm. H, Quantification of images shown in G; values are shown in a box-and-whisker plot in which each dot represents the mean percentage of Iba1⁺ CD74⁺ cells positive for MDK normalized to the total CD74⁺ Iba1⁺ per brain. Twelve fields of view were captured per brain, in n = 6 (vehicle) and n = 7 (ibudilast). P value was calculated using an unpaired two-tailed t test of each experimental condition against control. I, UMAP showing the expression of CD74 signature obtained by bulk RNA-seq (Fig. 2A and B; Supplementary Fig. S3A) in CD45⁺ cells found in the healthy and brain metastasis microenvironment of H2030-BrM model, categorized by condition—left, healthy brain; middle, H2030-BrM metastatic brain; and right, H2030-BrM metastatic brain treated with ibudilast. J, Violin plot presenting the distribution of Cd74 signature score in the myeloid CD74⁺ populations (disease-associated microglia, reactive microglia, and reactive macrophages) in H2030-BrM vehicle- and ibudilast-treated brains.

Consistently with the MIF knockdown phenotype, ibudilast did not alter the amount of CD74⁺ microglia/macrophages in vivo (Fig. 4C and D; Supplementary Fig. S8P and S8Q) but reduced both nuclear CD74-ICD levels in microglia/macrophages (Fig. 4E and F) and MDK expression (Fig. 4G and H; Supplementary Fig. S8R and S8S).

To confirm that the reduction in brain metastasis burden observed in vivo was not mediated by ibudilast’s additional inhibition of PDEs, we tested AV1013, a PDE-inactive analog (25). AV1013 replicated ibudilast’s antimetastatic effect in organotypic brain slice cultures (Supplementary Fig. S9A–S9C), indicating MIF blockade as the primary mechanism.

To further dissect the mechanism of action of ibudilast, we performed scRNA-seq on CD45⁺ sorted cells from the brain metastasis microenvironment (Supplementary Figs. S9D–S9G and S10; Supplementary Table S10). Ibudilast-treated brains showed a clear shift in the immune landscape, with a nearly twofold increase in homeostatic microglia and a marked decrease in reactive macrophages and infiltrative populations, showing a regression to control levels (Supplementary Fig. S9E–S9G). Notably, several subpopulations of microglia and macrophages in vehicle-treated mice expressed the protumorigenic signature identified in CD74⁺ microglia/macrophages (Figs. 2B and 4I), mimicking the emergence of a disease-associated cell type. The effect of ibudilast reduced the expression of the CD74 prometastatic signature (Fig. 4I and J). In addition, pathway analysis of CD74⁺ myeloid population reveals a significant downregulation of NF-κB signaling (Supplementary Fig. S11A; Supplementary Table S11), which suggests its ability to impair MIF-dependent reprogramming of CD74⁺ macrophages/microglia (Fig. 1E–K; Supplementary Fig. S2A–S2C).

Furthermore, analysis of reactive macrophages, as the CD74⁺ cell subpopulation with the highest expression of the signature (Fig. 4I; Supplementary Fig. S11B), identified additional potential biomarkers of response (Supplementary Fig. S11C; Supplementary Table S12). Among them, we validated Cd274 (PD-L1) in CD74⁺ microglia/macrophages (Supplementary Fig. S11D and S11E) given its implications on local immunosuppression and known dependency on NF-κB (62). Therefore, the reduction of PD-L1 expression observed upon ibudilast treatment aligns with its activity blocking the MIF–CD74–NF-κB axis. Furthermore, it is tempting to speculate that ibudilast would facilitate the antimetastatic action of immunotherapies locally, at least, in part, by limiting the expression of immune checkpoint blockade molecules.

To further characterize the immunosuppressive compartment of metastasis-induced CD74⁺ microglia/macrophages, we further investigated the cytokine profiling associated with brain metastases in the presence of ibudilast. Despite overall similar cytokine profiles (Supplementary Fig. S12A; Supplementary Table S13), ibudilast restored IL2 levels, which were reduced in metastasis-bearing mice (Supplementary Fig. S12A). IL2 plays a major role as an activator of natural killer (NK) cells (63). Given our interest to understand the prometastatic roles of ibudilast both in immunocompetent and in immunosuppressed models of brain metastases (Fig. 4A and B; Supplementary Fig. S8D and S8E), we explored the potential effect on NK cells. Both mouse and human NK cells’ activation and expansion are IL2-dependent (63) and have been shown to influence brain metastasis progression (64). We observed a significantly higher infiltration of NK cells in the ibudilast-treated mice compared with the untreated controls (Supplementary Fig. S12B and S12C).

These results suggest that ibudilast inhibition of the MIF–CD74 signaling axis in microglia/macrophages impairs acquired prometastatic functions (MDK, PD-L1) but also suggests a broader influence in the brain metastatic niche toward a less supportive state for tumor growth (increased NK cells). This is consistent with reports showing that MDK can limit NK cytotoxicity (65) and that PD-L1 restrains NK cell persistence and effector function in vivo (66).

Use of ibudilast on human brain metastases depicts a tumor-agnostic strategy guided by biomarkers of therapeutic response

To evaluate the effect of ibudilast on human brain metastases, we used PDOCs from freshly resected brain metastases obtained in RENACER (67).

Brain metastasis PDOCs were established from 26 patients, capturing the clinical and biological heterogeneity of the disease across primary tumor types and clinical management (Supplementary Table S14). The rationale for such tumor-agnostic approach derives from the broad and abundant presence of CD74⁺ microglia/macrophages among human brain metastases (Fig. 5A; Supplementary Table S15) and our preclinical functional findings in both lung and melanoma brain metastases (Fig. 4A and B; Supplementary Fig. S8D and S8E). The 26 PDOCs were tested for sensitivity to ibudilast using a well-established pipeline (30) that consists on the incubation of the fresh tissue slice with the drug over a 3-day period (Fig. 5B). Remarkably, 19 of 26 (73.08%) of the PDOCs responded to treatment, defined as a reduction in cancer cell viability greater than 30%, following clinical benchmarks (Fig. 5C and D; ref. 68). Importantly, ibudilast reduced PD-L1 expression in CD74⁺ microglia/macrophages in PDOC (Supplementary Fig. S13A and S13B), consistent with the phenotypic shift observed in vivo (Supplementary Fig. S11D and S11E), supporting inhibition of the MIF–CD74 axis as a mechanism to modulate local immunosuppression.

Figure 5.

Ex vivo clinical use of ibudilast on PDOCs suggests a tumor-agnostic strategy against brain metastases. A, Quantification of CD74 in microglia/macrophages in human brain metastasis. In total, 17 of 17 (100%) showed positive staining for CD74, from which 10/17 (58.82%) scored with 1, 6/17 (35.29%) with 2, and 1 of 17 (5.88%) with 3 according to the abundance and signal intensity of CD74 in the microenvironment. B, Schema of the experimental design. Fresh neurosurgeries from patients with brain metastasis cut with a vibratome were cultured to generate PDOCs. The slices were cultured with 50 μmol/L of ibudilast for 3 days. C, Representative immunofluorescence against Ki67 labeling the proliferating cancer cells in PDOCs derived from a lung cancer brain metastasis (top) and a breast cancer brain metastasis (bottom) after 3 days of ex vivo treatment with ibudilast. Scale bar, 75 μm. On the right down corner of each image, IHC staining against CD74 in the corresponding PDOC is shown. Scale bar, 50 μm. D, Quantification of the relative number of Ki67⁺ cancer cells in PDOCs after treatment with ibudilast with respect to the corresponding vehicle-treated PDOC with DMSO (from pictures shown in C). Dots are colored according to the primary source of the metastasis—blue, lung cancer 42.31% (11/26); pink, breast cancer 26.92% (7/26); dark red, colorectal cancer 11.54% (3/26); purple, uterine cancer 7.69% (2/26); black, melanoma 7.69% (2/26); and yellow, prostate cancer 3.85% (1/26). Values are shown in a box-and-whisker plot in which each dot represents a different patient (mean value obtained from all PDOCs from the same condition and patient), and the line in the box corresponds to the median. n = 26 patients, three slices per condition (DMSO or ibudilast), and three fields of view per slice. P value was calculated using a paired two-tailed t test. E, Heatmap of DEGs in nonresponders (NR) vs. ibudilast responders (R) from samples in D, as determined by bulk RNA-seq performed on fresh, untreated neurosurgical samples. The 11 genes composing the ibudilast-response prediction signature are listed on the right. F, Pathway analysis of differential induced pathways in ibudilast responders vs. nonresponders. G, AUC–ROC curve for the 11-gene signature, obtained by comparing nonresponders vs. ibudilast responders, demonstrating its ability to predict treatment response. H, Quantification of CD74⁺ cells in the microenvironment in paraffin sections from ibudilast nonresponders (n = 7) and responders (n = 17) patients. Values are shown in a violin plot in which each dot represents one human sample, and the line corresponds to the median. P value was calculated using an unpaired two-tailed t test. I, Quantification of MIF⁺ cancer cells in paraffin sections from ibudilast nonresponders (n = 7) and responders (n = 17) patients. Values are shown in a violin plot in which each dot represents one human sample, and the line corresponds to the median. P value was calculated using an unpaired two-tailed t test. J, Measure of MIF levels (pg/mL) determined by ELISA in CSF from healthy donors and patients with brain metastasis. Values are shown in a box-and-whisker plot in which each dot represents a patient, and the line corresponds to the median. P value was calculated using a Mann–Whitney U test.

Importantly, 53.85% of the organotypic cultures showed a reduction of 50% or more, with responses observed across multiple primary tumor origins (Fig. 5D). We then examined the 34.62% brain metastases that do not respond to ibudilast, classified as nonresponders, and identified a transcriptomic signature that separates them from responders (Fig. 5E and F). Interestingly, a reduced panel of 11 genes was sufficient to predict response, offering a clinically actionable biomarker set (Fig. 5E and G). In addition, downregulation of this signature in nonresponders enabled in silico identification of alternative candidate therapeutics (Supplementary Fig. S13C; Supplementary Table S16).

The potential to develop a clinical trial using ibudilast in brain metastasis based on its brain penetrance, safety profile (26), and its potential to facilitate immunotherapies (Supplementary Fig. S11C–S11E) encouraged us to identify additional biomarkers to select patients. We evaluated whether protein levels of CD74 or MIF correlated with ibudilast response. Although CD74 expression levels did not significantly differ between responders and nonresponders (Fig. 5H), MIF levels were higher in responders (Fig. 5I), suggesting that MIF abundance might be a predictive marker of treatment efficacy. However, relying on tumor tissue biopsy for biomarker detection requires invasive procedures. Alternatively, the secreted nature of MIF (Supplementary Fig. S1D) and its high levels in brain metastatic cells (Fig. 1A; Supplementary Fig. S1A–S1E) suggested its potential as a biomarker compatible with liquid biopsy. CSF from nontumor hosting patients (intracerebral pressure diagnosis, Supplementary Table S17) and from RENACER patients (Supplementary Table S17) were evaluated with an MIF-specific ELISA. Remarkably, MIF levels were increased only in patients with brain metastases (Fig. 5J), confirming its potential clinical value as a noninvasive liquid biopsy biomarker.

CD74⁺ in macrophages/microglia is a biomarker of several brain disorders

Given that CD74⁺ microglia/macrophages have been broadly reported in other brain disorders beyond oncology (69–71), and generally considered a marker of the MHC-II complex in spite of the limited functional evaluation for its contribution to the disease, we asked whether our findings on brain metastasis–associated CD74⁺ microglia/macrophages could be applied to this cell subpopulation present in Alzheimer's disease and multiple sclerosis.

First, we confirmed that CD74⁺ microglia/macrophages were present in pathology-associated regions in Alzheimer's disease and multiple sclerosis models (i.e., phospho-Tau and myelin debris, respectively; Fig. 6A–C), along with MIF-positive areas (Fig. 6D–F). Furthermore, our comparative scRNA-seq approach on CD45⁺ cells (Supplementary Fig. S14A–S14F) confirmed increased Cd74 expression across disease models compared with their age-matched controls (Fig. 6G–L).

Figure 6.

CD74⁺ in macrophages/microglia is a biomarker of brain disorders. A, Immunofluorescence against CD74, GFP (H2030-BrM cells), and DAPI in H2030-BrM metastatic brain. Scale bar, 75 μmol/L. Magnification scale bar, 25 μmol/L. B, Immunofluorescence against CD74, pTau, and DAPI in brain from an Alzheimer's disease mouse model. Scale bar, 75 μmol/L. Magnification scale bar, 25 μmol/L. C, Immunofluorescence against CD74, myelin basic protein (MBP), and DAPI in a brain from a cuprizone murine model of multiple sclerosis. Scale bar, 75 μmol/L. Magnification scale bar, 25 μmol/L. D, Immunofluorescence against MIF and DAPI in H2030-BrM metastatic brain. White dotted lines outline the tumoral area. Scale bar, 25 μmol/L. E, Immunofluorescence against MIF and thioflavin S (ThioS) in the brain (somatosensory cortex) from an Alzheimer's disease mouse model. Scale bar, 25 μmol/L. F, Immunofluorescence against MIF and myelin (MBP) in a demyelinating plaque from a patient with secondary progressive multiple sclerosis. Scale bar, 25 μmol/L. G–l,Cd74 expression in UMAPs of CD45⁺ populations found in healthy brain from immunocompromised mice (G), in H2030-BrM metastatic brains (H), in healthy aged brain from immunocompetent mice (I), in the brains of mouse models of Alzheimer's disease (J), in healthy brain from immunocompetent mice (K), and in the brains of mouse models of multiple sclerosis (L). DAM1, disease-associated microglia; Interferon Macro, IFN-responsive macrophages. M, Heatmap depicting DEGs with significant upregulation (adjusted P value < 0.05, log₂ fold change >0.5) and significant downregulation (adjusted P < 0.05, log₂ fold change < −0.5) shared by microglia and macrophages across each disease context. AD, Alzheimer's disease; BrM, brain metastasis; MS, multiple sclerosis. The common signature of 15 protein-coding genes exhibiting significant differential enrichment (adjusted P < 0.05, log₂ fold change >0.5) across all contrasts is labeled as the pan-disease CD74⁺ gene signature (C). N–Q, UMAPs showing the cell scores for the experimental pan-disease CD74 gene signature (PD-CD74) from M in human brain scRNA-seq data. N–Q, From the dorsolateral prefrontal cortex (DPC) of healthy individuals (N), from the DPC of patients with Alzheimer's disease (O), from healthy individuals (P), and from patients with multiple sclerosis (Q).

The analysis of upregulated pathways in the CD74⁺ versus CD74⁻ myeloid compartments revealed signatures that are also part of protumorigenic macrophages in tumors such as NF-κB (72), mTOR (73), or OXPHOS (Supplementary Table S18; ref. 60).

Importantly, gene expression analysis in CD74⁺ versus CD74⁻ microglia/macrophages suggested the existence of a conserved 67-gene signature across disorders, including 15 protein-coding genes (Fig. 6M; Supplementary Table S19). We refer to this shared program composed of 15 genes as the pan-disease CD74⁺ signature (PD-CD74). The experimental PD-CD74 was evaluated in publicly available datasets of patients affected with Alzheimer disease (53) and multiple sclerosis (52). Remarkably, the PD-CD74 was enriched in human Alzheimer disease and multiple sclerosis datasets (Fig. 6N–Q; Supplementary Fig. 14G–S14H).

Together, these findings position CD74⁺ microglia/macrophages as a shared, disease-associated population with translational relevance across brain metastasis, neuroinflammation, and neurodegeneration.

Discussion

Our findings demonstrate the prometastatic nature of CD74⁺ microglia/macrophages, challenging the prevailing assumption of being a surrogate of MHC-II activity (69–71). Although CD74 is a known MHC-II chaperone, it can also act as a cytokine receptor for MIF. Upon ligand binding, CD74 is cleaved, and its ICD translocates to the nucleus, triggering a noncanonical transcriptional program via NF-κB (13, 55). We confirmed this mechanism in CD74⁺ microglia/macrophages in brain metastases, using both nuclear staining and MDK expression as functional readouts of NF-κB activation.

Although MIF has previously been suggested to originate from the brain metastatic niche (8), we provide functional evidence that metastatic cancer cells are the primary source of MIF in both experimental and patient-derived models. Importantly, this tumor-derived MIF rewires microglia/macrophages into a disease-promoting state. We also show that this switch depends on mitochondrial plasticity, as we can suppress nuclear translocation of the CD74-ICD–dependent NF-κB program upon targeting components that regulate mitochondrial quality.

Based on our single-cell data and depletion experiments, this phenotype seems to be predominantly mediated by BMDMs; however, we cannot exclude a contribution from microglia. This is particularly relevant given recent findings in melanoma brain metastases showing that activation of the Rela/NF-κB pathway in microglia drives a protumoral program, whereas its inhibition reprograms them toward a proinflammatory, antitumor state, enhancing responses to immune checkpoint blockade (74).

Although MIF–CD74 signaling has been implicated in tumorigenesis across multiple cancers, most studies outside of primary brain tumors have focused on CD74 expression in tumor cells themselves (16–18), rather than in the immune microenvironment. In contrast, gliomas have revealed roles for CD74 in GAMs, with evidence for both immune suppression and tumor-promoting activity (22). Our work extends this concept beyond gliomas to brain metastases, where we identify a distinct, MIF-responsive CD74⁺ subpopulation of myeloid cells that drive disease progression. Moreover, although CD74–NF-κB activation may represent a general mechanism, our findings suggest that its pathologic relevance is context-dependent. Specifically, we observed only modest effects in lung metastases, in strong contrast to the brain-specific phenotype using the different approaches. This suggests that the MIF–CD74 axis plays a particularly prominent role in the brain metastatic niche. Nevertheless, we cannot exclude its relevance in other metastatic sites or primary tumors without further investigation. Comparative studies across tumor types and compartments will be essential to delineate the spatial and functional specificity of this pathway.

In addition to the benefit provided by genetic targeting of MIF in cancer cells, we demonstrate the potential of a pharmacologic intervention strategy using ibudilast, a brain-penetrant MIF–CD74 inhibitor already in clinical trials for GBM. Although ibudilast is known to sensitize GBM cells to temozolomide and improve survival in preclinical models (28), the mechanism has remained unclear. In GBM, ibudilast has been shown to reduce the immunosuppressive function of M-MDSCs and enhance CD8⁺ T-cell activity in the tumor microenvironment (75). Moreover, additional studies using CD74-neutralizing antibodies or MIF siRNA in glioma have linked the MIF–CD74 pathway to tumor immune evasion and myeloid reprogramming (22), supporting the broader therapeutic relevance of MIF–CD74 signaling. Our findings point out to a broader influence of ibudilast on the immune landscape as depicted by its modulation on IL2 levels, which correlated with increased NK cells’ infiltration.

To assess clinical relevance, we tested ibudilast in 26 PDOCs from a range of primary tumors. Although most patients responded to ibudilast, the profiling of the metastases suggests a predictive molecular classifier based on 11 genes. Given the safety, brain penetrance, and the molecular classifier of response together with the liquid biopsy biomarker that could facilitate patient selection and response follow-up, we envision the potential of developing a clinical trial with ibudilast in patients with brain metastases.

Interestingly, ibudilast has also shown clinical benefit in progressive multiple sclerosis, slowing brain atrophy in phase II trials (26). Although the exact mechanism that explains the clinical benefit remains to be clarified, it is often assumed to be primarily mediated through ibudilast’s inhibition of PDE. However, our findings raise the possibility that ibudilast’s benefit in multiple sclerosis may also involve modulation of disease-associated CD74⁺ microglia/macrophages. This same subpopulation, which we identify in both multiple sclerosis and Alzheimer's disease models, shares a conserved NF-κB-driven signature that parallels the prometastatic program observed in brain metastases.

Taken together, our work identifies a pan-disease, reprogrammable subpopulation of CD74⁺ myeloid cells that drive pathology in brain metastases and potentially other neurologic diseases. Therapeutically targeting this subpopulation via the MIF–CD74 axis could open new avenues in neuro-oncology, neurodegenerative, and neuroinflammatory disorders. Comparative multidisease analyses will be critical to validate these findings and broaden their clinical effect.

Supplementary Material

Supplementary Table 1

List of primary antibodies used for immunofluorescence and immunohistochemistry analysis in fresh tissue and paraffin-embedded sections

Supplementary Table 2

List of secondary antibodies used for immunofluorescence

Supplementary Table 3

List of flow cytometry antibodies

Supplementary Table 4

Brain metastasis samples analysed for MIF in either the tumor or the microenvironment

Supplementary Table 5

Differentially expressed genes in CD74+ vs CD74-microglia/macrophages obtained by bulk RNA-seq

Supplementary Table 6

List of the top 150 up and 150 down differentially expressed genes in CD74+ vs CD74- microglia/macrophages obtained by bulk RNA-seq

Supplementary Table 7

List of the human 37- gene signature derived from the bulk RNA-seq of CD74+ microglia/macrophages

Supplementary Table 8

Differentially expressed genes in CD74+ CCR2+ cells vs CD74+ CCR2- cells

Supplementary Table 9

Gene set enrichment analysis of CD74+ microglia/macrophages

Supplementary Table 10

List of marker genes defining each CD45⁺ immune cell population identified by single-cell RNA-seq

Supplementary Table 11

Pathway analysis of the top significantly downregulated pathways in CD74 positive myeloid cells following ibudilast treatment

Supplementary Table 12

Differentially expressed genes for the effect of ibudilast in reactive macrophages from mice with brain metastasis (ibudilast vs. vehicle treatment)

Supplementary Table 13

Cytokine and chemokine profiling by FirePlex assay in healthy brain, vehicle-treated, and ibudilast-treated brain metastasis-bearing mice

Supplementary Table 14

Patient-derived organotypic cultures (PDOC) treated with Ibudilast

Supplementary Table 15

Brain metastasis samples analyzed for CD74 in the microenvironment

Supplementary Table 16

GDSC-derived therapeutic candidates for non-responder patients based on the 11-gene responder signature

Supplementary Table 17

MIF levels from ELISA applied to CSF from patient samples

Supplementary Table 18

Upregulated pathways in the CD74+ myeloid clusters vs CD74- myeloid clusters

Supplementary Table 19

List of deregulated genes in the CD74+ microglia (disease associated microglia, reactive microglia)/macrophages (reactive macrophages) vs CD74- microglia (homeostatic microglia)/macrophages (homeostatic macrophages) in brain metastasis, Alzheimer disease and multiple sclerosis

Supplementary Figure 1

Cancer cells are the main source of MIF and CD74 expression is induced in the brain tumor microenvironment

Supplementary Figure 2

CD74 expression is induced by IFN-γ in the brain tumor microenvironment

Supplementary Figure 3

MIF induces CD74-ICD translocation to the nucleus and its silencing reduces B16/F10-BrM brain metastasis in vivo

Supplementary Figure 4

CD74+ microglia/macrophages are enriched in protumoral markers

Supplementary Figure 5

Survival analysis of the 37-gene signature derived from CD74+ microglia/macrophages across TCGA primary tumors cohorts

Supplementary Figure 6

CD74+ are mainly bone marrow-derived macrophages with high expression of CCR2 and the Cd74+ Ccr2+ signature has clinical relevance

Supplementary Figure 7

CD74+ cells are mainly BMDM with transcriptional enrichment in OXPHOS pathway

Supplementary Figure 8

Ibudilast reduces brain metastasis in vivo without comprising the viability of cancer cells in vitro

Supplementary Figure 9

Ibudilast acts mainly through MIF inhibition

Supplementary Figure 10

Heatmap of differentially expressed genes defining CD45⁺ cell populations in healthy and metastatic brains

Supplementary Figure 11

Ibudilast reverts the protumorigenic signature of CD74+ microglia/macrophages in vivo

Supplementary Figure 12

Ibudilast-treated brains show a recovery in IL-2 levels, which correlates with higher infiltration of natural killer cells

Supplementary Figure 13

Ibudilast reduces PD-L1 expression in CD74+ from patient-derived organotypic cultures

Supplementary Figure 14

A CD74 pan-disease signature is present in Alzheimer’s disease and multiple sclerosis patients

Data Availability

Bulk RNA-seq data from CD11b⁺; CD74⁺ versus CD11b⁺; and CD74⁻ have been deposited to Gene Expression Omnibus [GEO (SCR_005012)] with the dataset identifier GSE 294461. scRNA-seq data from CD45⁺ cells obtained from brain metastasis, multiple sclerosis, and Alzheimer's disease experimental models have been deposited to GEO with the dataset identifier GSE293921. scRNA-seq data from experimental brain metastasis (9) analyzed in this study were obtained from GEO at GSE228368. TCGA datasets were analyzed at the cohort level via Survival Genie 2.0 web-based platform, which integrates TCGA data. All other raw data generated in this study are available upon request to the corresponding author.

Authors’ Disclosures

J. Bernhagen reports grants from Deutsche Forschungsgemeinschaft during the conduct of the study, as well as a patent for PCT/EP2021/066015 pending. F. de Castro reports grants from Ministerio de Ciencia, Innovación y Universidades and Fundación Ramón Areces during the conduct of the study. M. Valiente reports grants from AstraZeneca outside the submitted work. M.-J. Artiga reports grants from Asociación Española Contra el Cáncer during the conduct of the study. Á. Pérez-Nuñez reports grants from Asociacion Española Contra el Cancer and Instituto de Salud Carlos III (MInisterio de Ciencia, Innovación y Universidades) during the conduct of the study. E. Martínez-Sáez reports personal fees from Servier outside the submitted work. No disclosures were reported by the other authors.

Authors’ Contributions

L. Alvaro-Espinosa: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–review and editing. A. Marquez-Galera: Data curation, formal analysis, investigation, visualization, methodology, writing–review and editing. N. Priego: Formal analysis, validation, investigation, visualization, methodology. V. García-Calvo: Investigation, visualization, methodology. M. Perea-García: Formal analysis, validation, investigation, visualization, methodology. C. Hernandez-Oliver: Data curation, formal analysis, validation, investigation, visualization, methodology. D. Retana: Investigation, visualization, methodology. O. Sanchez: Validation, investigation, visualization, methodology. A. de Pablos-Aragoneses: Formal analysis, validation, investigation, visualization, methodology. P. García-Gómez: Formal analysis, validation, investigation, visualization, methodology. O. Graña-Castro: Data curation, formal analysis, validation, investigation, visualization, methodology. Ó Lapuente-Santana: Data curation, formal analysis, validation, investigation, visualization, methodology. L. Serrano-Ron: Data curation, formal analysis, investigation, visualization, methodology. F. Al-Shahrour: Supervision. A. Cayuela López: Software, formal analysis, validation, investigation, visualization, methodology. I. Peset: Supervision, investigation, methodology. D. Megías: Supervision, visualization, methodology. M. Ola: Data curation, formal analysis, methodology. D. Varešlija: Resources, supervision. L.S. Young: Resources, supervision. Y. Martí-Mateos: Resources. J.A. Enríquez: Resources, supervision. E. Hernández-Encinas: Methodology. C. Blanco-Aparicio: Supervision, methodology. M.S. Soengas: Resources. J. Bernhagen: Resources, supervision, investigation, methodology. A. Antón-Fernández: Resources, formal analysis, methodology. J. Ávila: Resources, supervision. M.A. Marchena: Resources, formal analysis, methodology. M. Torres: Resources, methodology. F. de Castro: Resources, supervision. M. Márquez-Ropero: Resources, methodology. A. Sierra: Resources, supervision. J.P. Lopez-Atalaya: Resources, data curation, formal analysis, supervision, investigation, visualization, methodology, writing–review and editing. RENACER Group: Resources. M. Valiente: Conceptualization, resources, formal analysis, supervision, funding acquisition, writing–original draft, project administration, writing–review and editing. P. Baena Galán: Resources. C. Sobrino: Resources. I. Almenara: Resources. D. Alba-Olano: Resources. C. Ortega-Sabater: Resources. M.-J. Artiga: Resources. E. Ortega-Paino: Resources. A. González Piñeiro: Resources. C. Fiaño Valverde: Resources. A. de la Lama Zaragoza: Resources. A. Londoño Quiroz: Resources. P.D. Delgado López: Resources. M. Pascual-Llorente: Resources. Á. Díaz-Piqueras: Resources. Á.A. Arriaga Aragón: Resources. S. Nam-Cha: Resources. C. Barrena-López: Resources. G. Plans Ahicart: Resources. B. Escolano Otín: Resources. I. Gil Aldea: Resources. J. Delgado-Fernández: Resources. J.M. Sepúlveda-Sánchez: Resources. Á. Pérez-Nuñez: Resources. A. Hernández-Laín: Resources. Ó. Toldos González: Resources. R. Gargini: Resources. D. Alcivar: Resources. J.A. Fernández Alén: Resources. G. Blasco García de Andoain: Resources. S. Cepeda Chafla: Resources. E. Martínez Zamorano: Resources. M. Mollejo Villanueva: Resources. M.-S. Opazo Rodríguez: Resources. Á. Rodríguez de Lope Llorca: Resources. M. Arbaiza Martínez: Resources. G.M. Múzquiz Rueda: Resources. S. Benavente: Resources. F. Martínez-Ricarte: Resources. S. Ramón y Cajal: Resources. M. Sesé Faustino: Resources. L. Fernández Cabré: Resources. J. Hernández-Losa: Resources. E. Martínez-Sáez: Resources. L. Calero Félix: Resources. K. Vargas-Osorio: Resources.