Main

Immune-checkpoint blockade immunotherapy unleashes antitumor immune responses and has emerged as one of the most effective therapies in oncology. Nevertheless, most patients do not respond or acquire resistance to ICB, and this strategy is not active against all cancer types. The underlying tumor-intrinsic mechanisms of ICB resistance are thus still under intense investigation1,2. Among these, phenotypic heterogeneity might help to explain how patients with similar tumor types still show varying sensitivity to ICB3.

Interferon signaling plays a central role in tumor immunosurveillance and immunotherapy response, in part by modulation of APM in tumor cells3,4,5,6. APM includes major histocompatibility complex (MHC) class I genes (HLAs), aiding molecules (β2M), transporters (TAP complex and tapasin) and immunoproteasome genes (PSMBs) among others, which are the cellular effectors of antigen presentation allowing its recognition by the immune system7. Multiple studies and clinical reports have identified genetic alterations in antigen presentation components and the IFN pathway as major mechanisms of ICB resistance2,8,9,10. Epigenetic factors and transactivators, such as EZH2, NF-κB and NLRC5, can also modulate MHC-I and other APM genes affecting tumor immunogenicity11,12. Therefore, a better understanding of APM regulation in different tumor cell populations and phenotypes can help to elucidate the mechanisms underlying immunotherapy resistance.

Stem cell phenotypes, such as embryonic and slow-cycling adult stem cells, downregulate MHCs, favoring immune surveillance evasion13,14. CSCs resemble their normal counterparts in many aspects15, and may co-opt similar immune-evasive phenotypes relevant in ICB resistance. CSCs actively interplay with immune-secreted cytokines, including IFNs16,17,18, and some reports have shown low expression of MHCs and Transporter 1 (TAP1) in CSCs19,20, suggesting a role in immune surveillance evasion. CSCs also express CD274 (PD-L1), a mechanism that contributes to immunosuppression21 but not to anti-PD-L1 ICB escape. Epithelial-to-mesenchymal transition, often linked to stemness, also promotes immunosuppressive features22,23,24. In the context of immunotherapy resistance, skin squamous cell carcinoma (SCC) tumor-initiating cells (TICs) have been linked to adoptive cell transfer resistance by upregulation of CD80 (ref. 25); and head and neck SCC CSCs have ICB resistance by downregulating the secretion of T cell-recruiting chemokines26. However, the connection between stemness and APM pathway dysregulation leading to ICB escape, and the molecular mechanisms controlling it, are not established.

In the mammary gland, normal mammary stem cells (MaSCs) and CSCs are usually governed by differentiating cell fate determinants and stem cell transcription factors16,27,28,29,30,31. LCOR is a differentiation factor sensitizing these to IFN, which drives intrinsic tumor cell differentiation and reduced tumor growth16. However, how LCOR intersects the IFN response has never been explored and may be critical to understanding cellular immunity. Therefore, we were interested in exploring their mechanistic connections with immunity and ICB resistance in TNBC. This is clinically relevant, because there is an urgent need to improve the efficacy of ICB in breast cancer, approved in TNBC with only limited clinical benefit to date32,33. Here, we show that breast LCORlow CSCs shut down both antigen processing and presentation, contributing to immune-checkpoint therapy resistance in TNBC. LCOR activates transcription of the APM independently of IFN, rendering CSCs visible and vulnerable to immune attack during ICB. Our results demonstrate the relevance of phenotypic heterogeneity and LCOR biology in APM regulation as an excellent therapeutic partner of ICB therapy.

Results

ICB resistance emerges from MaSC-like states

To understand breast cancer ICB resistance, we generated a preclinical immunocompetent syngeneic model of anti-PD-L1 ICB resistance in vivo. Mouse breast cancer 4TO7 cells were orthotopically implanted into the mammary fat pad (MFP). Once tumors had reached a size of 0.5 × 0.5 cm2, mice were treated with anti-PD-L1. After an initial transient response, all tumors developed resistance to anti-PD-L1 treatment (Fig. 1a). Anti-PD-L1 immunotherapy-resistant tumors (IRTs) were harvested in cell culture and their resistance was confirmed again in vivo (Extended Data Fig. 1a).

Fig. 1: CSCs lead immunoediting escape and ICB resistance.
figure 1

a, Growth curves for 4TO7 tumors in Balb/c mice treated with anti-PD-L1 ICB. Dose and regime are indicated. n = 3 tumors per condition. b, GO analysis of upregulated (red) and downregulated (blue) pathways by RNA-seq in 4TO7 anti-PD-L1 IRT and control cells (I-Ctrl). P value computed by Fisher exact test. c, GSEA of APM (GO: 0048002) in IRT versus control cells. Family-wise error rate (FWER) P value and FDR q value are provided. d, GSVA of previously described stemness signatures: embryonic stem cells (ES_1), NANOG-OCT4-SOX2 targets (NOS_Targets) and breast_CSCs (LIU breast ER CSCs). Circle size represents significance and colors the mean z-row score. bd, Two independent biological replicates. e, RT–qPCR of stem cell and differentiation genes in IRT versus I-Ctrl tumor cells. Log2 fold change of IRT versus control cells. Data represent mean ± s.e.m. f, Flow cytometry analysis of CD24hi/CD29hi population in IRT and control cells. Data represent mean ± s.e.m. e,f, n = 3 individual biological replicates. g, MFP injection and LDA of IRT 4TO7 cells in NSG mice. Table represents serial dilution injections with the take rate; n indicates the number of MFP injections. TIC frequency calculated by ELDA software is shown in red. P value obtained by Pearson’s chi-squared two-tailed test. h, CTL assay representation and CD24lo/CD44hi CSC flow cytometry analysis. Data represent percentage of viable cells in a representative experiment. Anti-PD-L1 treatment at 1 μg ml–1. i, OT-I/JEDI CTL assay with CSCs and non-CSCs isolated from AT3-OVA and 4TO7-EGFP cells. E:T denotes effector to target ratio. h,i, Data represent mean ± s.e.m. n = 3 independent biological replicates. j,k, GSEA of transcriptomic ranked list for responder versus nonresponder patients from the TONIC trial40, with our IRS (j) and ES_1 (k) and LIU breast ER CSC signatures. e,i, Exact P value by two-tailed Student’s t-test; d, Wilcoxon signed-rank test. f,h,i, Cells were gated from P3 (Extended Data Fig. 9a).

Source data

Transcriptomic analysis of IRT cells revealed loss of APM among the most downregulated Gene Ontology (GO) pathways, followed by an expected loss of IFN signaling (Fig. 1b,c). Interestingly, among upregulated pathways we found enrichment of stem cell signatures (Fig. 1b) and, using gene set variation analysis (GSVA), we verified an important enrichment of CSC-like signatures associated with breast cancer aggressiveness, such as ES_1 (ref. 29), NOS_Targets29 and breast_CSCs34 (Fig. 1d). Moreover, chromatin immunoprecipitation (ChIP)–enrichment analysis (CHEA) ranked stem cell transcription factors (SC-TFs) most enriched in IRT (Extended Data Fig. 1b). We validated these transcriptomic findings by reverse transcription quantitative PCR with reverse transcription (RT–qPCR), showing that IRT cells are enriched for SC-TF genes Pou5f1 (Oct4), Sox2, Sox9 and Nanog and depleted in mammary differentiation factors—in particular, Lcor (Fig. 1e). Importantly, IRT cells also showed increase in the MaSC markers CD24+/CD29hi (ref. 35), a population that has also been reported in 4TO7 cells and others36 (Fig. 1f). This cellular selection by ICB treatment was not observed when we disrupted antigen presentation using β2m knockout (KO) cells, consistent with absence of selective immune-mediated killing (Extended Data Fig. 1c). Of note, IRT cells maintained high Pd-l1 expression (Extended Data Fig. 1d), suggesting that they resisted anti-PD-L1 despite expressing the target. We propose that this resistance has emerged through downregulation of APM in CSCs, as suggested by the absence of CSC-like cell enrichment in β2m KO cells after treatment (Extended Data Fig. 1c). Further characterization showed that IRT cells possess increased tumor sphere formation ability across sphere generations in vitro (Extended Data Fig. 1e), despite no differences in cell cycle and with proliferation rates similar to the control condition, in both systems IRT and CD24hi/CD29hi CSCs (Extended Data Fig. 1f,g). To evaluate their TIC frequency in vivo, a hallmark of CSCs, we performed orthotopic MFP injections in limiting dilution assays (LDAs) in immunodeficient NOD-SCID Gamma (NSG) mice. IRT cells were enriched >tenfold for TIC frequency (Fig. 1g), confirming that CSC properties are selected by ICB therapy.

Next, we performed in vitro cytotoxic T lymphocyte (CTL) assays by coculture of AT3-OVA cells and OT-I CD8 T cells, which specifically recognize the OVA peptide 257-264 (SIINFEKL). After 3 days of coculture, surviving tumor cells that had evaded immune killing were enriched in CD24lo/CD44hi AT3 CSCs37,38, particularly after anti-PD-L1 treatment (Fig. 1h). Accordingly, sorted AT3 and 4TO7 CSCs were resistant to immune-mediated cell killing compared to non-CSCs when cocultured with OT-I CD8 and JEDI CD8 T cells, respectively (Fig. 1i and Extended Data Fig. 1h). AT3 CSCs induced less antitumor T cell activity than non-CSCs (Extended Data Fig. 1i). We utilized the SORE6 CSC reporter system, which reports SOX2 and OCT4 activity31, in AT3-OVA and Py8119-OVA cells. We first validated SORE6 fidelity in our models, showing that SORE6+ expressed SC-TFs and had >sixfold higher TIC frequency in vivo when injected into the MFP in LDAs (Extended Data Fig. 1j,k). Similar to previous results, the residual surviving cell population in CTL assays was highly enriched in SORE6+ CSCs due to selective elimination of SORE6 cells, an effect further enhanced by anti-PD-L1 treatment (Extended Data Fig. 1l). Overall, these results demonstrate the anti-PD-L1 ICB escape ability of CSC populations.

Despite the limitations in studying immune-specific interactions in human models, CD104hi/CD44hi CSCs isolated from MDA-MB-231 cells39 were more resistant to anti-PD-L1 when cocultured with human peripheral blood mononuclear cells (PBMCs) in an allogenic in vitro setting (Extended Data Fig. 1m). Moreover, when immune-humanized mice with PBMCs were orthotopically injected with MDA-MB-231 cells, after 5 weeks tumors treated with anti-PD-L1 were enriched for the CD104hi/CD44hi CSC population (Extended Data Fig. 1n).

To assess the clinical significance of our preclinical IRT resistance model, we generated a breast cancer ‘ICB-resistance signature’ (IRS) using the top 300 IRT upregulated genes (Supplementary Table 1). We applied IRS to transcriptomic RNA sequencing (RNA-seq) data from the TONIC trial, which collected 53 metastatic TNBC cases treated with anti-PD1 after preinduction therapy40. Gene set enrichment analysis (GSEA) showed that IRS was highly enriched in those patients not benefitting from ICB therapy (nonresponders) (Fig. 1j), validating its clinical value. Moreover, the nonresponding group was also enriched with stem cell-like signatures ES_1 (ref. 29) and breast_CSCs34 (Fig. 1k), consistent with our experimental observations. We also reanalyzed a single-cell RNA-seq dataset of 29 patients with breast cancer before and during anti-PD-1 therapy41. On-treatment tumors reduced LCOR and gained breast_CSC34 expression (Extended Data Fig. 1o–q). Overall, these results support the existence of CSC immunoediting driving ICB resistance.

LCORlow breast CSCs have reduced antigen-presenting ability

Based on the finding that ICB-resistant cells have a stem cell-like phenotype and low APM expression (Fig. 1b,c), we analyzed available MaSC transcriptomic profiles42. As expected, fetal mammary stem cells (fMaSCs) showed simultaneous downregulation of APM pathway genes, including immunoproteasome factors (PSMB8 and PSMB9), transporters (TAP1 and TAP2), β2M and MHC-I genes (Extended Data Fig. 2a).

To study the APM pathway in human CSCs, we generated patient-derived organoids (PDOs) from clinical samples and isolated CD24lo/CD44hi CSCs37 from four different patients with TNBC (Fig. 2a). Our analysis confirmed that APM genes are downregulated in clinical CD24lo/CD44hi low-LCOR CSCs, by RT–qPCR analysis (Fig. 2a) and by flow cytometry of pan-HLAs-ABC and β2M (Fig. 2b). Additionally, three-dimensional (3D) PDO imaging with advanced light-sheet fluorescence microscopy (LSFM) also demonstrated the absence of pan-HLA staining in SORE6+ CSCs (Fig. 2c). In human mammary cell lines, HMLE CD24lo/CD44hi stem cells43 and MDA-MB-231 CD104hi/CD44hi CSC populations39 also showed downregulation of APM genes and LCOR compared to non-CSCs (Extended Data Fig. 2b–d), as well as in ALDH+ populations (Extended Data Fig. 2c,d). Accordingly, mouse 4TO7 CSCs and IRT cells also showed low Lcor and low APM (Extended Data Figs. 2c,e and 1e). Next, we generated a fused LCOR-green fluorescent protein (GFP) knock-in in MDA-MB-231 cells reporting the endogenous protein levels and localization of LCOR (Fig. 2d and Extended Data Fig. 2f,g). As expected, LCORlow MDA-MB-231 cells were enriched in OCT4+/SOX2+ CSCs (SORE6+) (Extended Data Fig. 2h). We then used fluorescence-activated cell sorting (FACS) to isolate cells based on GFP (LCOR-GFP) levels and measured the expression of APM genes in LCOR−/low and LCORhigh cells by RT–qPCR. APM gene expression progressively increased with increasing levels of LCOR (Fig. 2e). This was further validated by immunofluorescence (IF) of pan-HLA-ABC and LCOR-GFP (Fig. 2f). These results indicate reduced APM activity in human LCORlow CSCs.

Fig. 2: LCORlow breast CSCs have low APM activity.
figure 2

a, PDOs from four patients with TNBC and isolation of CD24lo/CD44hi (CSC) and CD24hi/CD44lo (non-CSC). Right: RT–qPCR analysis of LCOR and APM genes. n = 4 independent biological replicates; data represent mean ± s.e.m. b, Flow cytometry analysis of pan-HLAs and β2M in CSC and non-CSC cells from PDOs. n = 3 independent biological replicates; data represent relative mean fluorescent intensity (MFI; versus isotype control) ± s.e.m. c, Representative LSFM imaging of a PDO with the CSC SORE6-GFP reporter (green), pan-HLA-ABC (purple) and DAPI (blue). Panels on the right show single channels. n = 3 independent organoids. d, LCOR-GFP knock-in flow cytometry analysis in MDA-MB-231 cells. e, RT–qPCR analysis for APM genes of LCOR knock-in MDA-MB-231 cells isolated by LCOR, LCORlow/med and LCORhigh. d,e, n = 3 individual biological replicates; data represent mean ± s.e.m. f, IF of pan-HLA-ABC in LCOR knock-in MDA-MB-231 cells: LCOR-GFP (green), HLA-ABC (purple) and DAPI (blue). n = 3 independent biological replicates; image of one representative experiment. g,h, Flow cytometry isolation (g) and RT–qPCR analysis (h) of OVA−/low and OVAhigh (SIINFEKL OVA peptide) from AT3 cells with ectopic OVA overexpression (AT3-OVA). g,h, n = 3 individual biological replicates for; data represent mean ± s.e.m. i, Flow cytometry analysis of OVA−/low and OVAhigh cell distribution within CD24lo/CD44hi CSC markers of AT3 cells. j, Orthotopic MFP injection and LDA of OVA−/low and OVAhigh AT3-OVA cells in immunodeficient NSG mice. Table represents serial dilution injections with the corresponding take rate, and n indicates the number of MFP injections for each dilution. TIC frequency calculated by ELDA software is shown in red. P value was obtained by Pearson’s chi-squared two-tailed test. c,f, Scale bars, 100 and 40 μm, respectively. a,b,h, Exact P values determined by two-tailed Student’s t-test; e, *P < 0.05, **P < 0.01, ***P < 0.005 by one-way ANOVA with Bonferroni post hoc test. a,b,e,g, Cells were gated from P3 (Extended Data Fig. 9a).

Source data

To demonstrate impaired APM in CSCs, we isolated breast cancer cells based on APM activity using the OVA antigen-peptide presentation as readout. We ectopically expressed the full-length native chicken egg ovalbumin (OVA) in AT3 and Py8119 cells, which are H2-K1b haplotype cells established from mouse breast cancer C57BL/6J PyMT tumors44. Ectopic OVA is processed by the immunoproteasome generating the OVA257–264 (SIINFEKL) peptide, transported and presented in a MHC-I H2-K1b context45. As we hypothesized, isolated OVA−/low cells (Fig. 2g and Extended Data Fig. 2i) expressed higher levels of CSC genes (Oct4, Sox2, Sox9 and Nanog) compared to OVAhigh cells as measured by RT–qPCR, and substantially lower levels of Lcor (Fig. 2h and Extended Data Fig. 2j). Accordingly, the AT3 CSC population CD24lo/CD44hi37,38 is largely segregated in OVA−/low AT3 cells by flow cytometry (Fig. 2i) whereas CSCs are only partially segregated by Pd-l1 (Extended Data Fig. 2k). Functional assays in both cell lines demonstrated increased tumor sphere formation (Extended Data Fig. 2l) and increased TIC capacity of OVA−/low cells in orthotopic MFP LDAs in immunodeficient NSG mice in vivo (Fig. 2j and Extended Data Fig. 2m), reflecting an inherent tumor-initiation stem cell potential of OVA−/low cells. Therefore, within tumor cell heterogeneity, LCORlow CSCs have a defective APM system as an important immune-evasive property for ICB escape.

LCOR regulates APM independently of interferon signaling

LCOR is among the top downregulated mammary differentiation factors in our IRT model, and LCORlow CSCs are associated with low APM. Here, we show expression correlation of LCOR and APM components in estrogen receptor (ER)-negative breast cancer cell lines of the Cancer Cell Line Encyclopedia (CCLE) (Fig. 3a). Moreover, patient stratification of the TNBC METABRIC dataset based on LCOR levels demonstrates a strong correlation between LCOR expression and both the APM pathway (Kyoto Encyclopedia of Genes and Genomes (KEGG): M16004) (Fig. 3b) and IRS (Fig. 3c). These results suggest that LCOR may play a key role in APM regulation.

Fig. 3: LCOR regulates APM genes independently of IFN stimulation.
figure 3

a, Correlation of LCOR mRNA with APM (KEGG: M16004) signature in 39 ER-negative breast cancer cell lines from CCLE. P value calculated by Rho correlation. b, GSEA of the METABRIC dataset TNBC stratified as LCOR-high versus -low patients based on the median and interrogated with the APM (KEGG: M16004) signature. FWER P value and FDR q value are provided. c, Left: GSEA of our ICB resistance signature in the TNBC METABRIC dataset stratified by LCOR median. Right: correlation of GSVA score of ICB resistance signature with LCOR in the METABRIC TNBC dataset. Blue line represents lineal regression analysis by Rho correlation and P value. d, Heat map of RT–qPCR analysis of APM genes in LCOR-OE and LCOR-KD transduced MDA-MB-231 cell lines (see OE and KD validations in Extended Data Fig. 8b) relative to their respective controls. n = 3 individual biological replicates; data represent the mean. eg, Flow cytometry analysis of TAP1 (e), β2M (f) and pan-HLA (g) levels under basal conditions, IFN-γ treatment treatment at 10 ng ml–1 and ruxolitinib (1 μM) in MDA-MB-231 LCOR-OE and LCOR-KD cells with respect to their controls and isotype negative control. n = 3 independent biological replicates; data represent MFI relative to the isotype ± s.e.m. h, Flow cytometry analysis of SIINFEKL OVA peptide presented by H2-K1b in AT3-OVA cells at the indicated conditions (IFN-γ treatment at 10 ng ml–1 and ruxolitinib at 1 μM). n = 3 independent biological replicates; data represent MFI relative to the isotype ± s.e.m. dh, Exact P values calculated by two-tailed Student’s t-test. eh, Cells were gated from P3 (Extended Data Fig. 9a).

Source data

To investigate the mechanistic link of LCOR with the APM pathway, we used LCOR gain-and-loss modifications in MDA-MB-231 and HMLE cells. Remarkably, ectopic LCOR expression induced APM pathway genes while LCOR-KD reduced their expression in both cell types (Fig. 3d and Extended Data Fig. 3a). Moreover, we measured the expression of different APM components via flow cytometry in these models. LCOR-KD cells phenocopy CSCs and display low APM as measured by the reduced 26S proteasome activity reporter pQCXIN/ZsGreen46, and reduced expression of transporters (TAP1) and presenting molecules (β2M and HLA-ABC) (Fig. 3e–g and Extended Data Fig. 3b–d). As expected, the inverse results were obtained with LCOR-OE boosting antigen processing and presentation (Fig. 3d–g and Extended Data Fig. 3a–d,f).

Surprisingly, these effects were not dependent on IFN stimulation. Despite IFN-γ treatment enhancing the effects of LCOR on APM components, ruxolitinib—an inhibitor of JAK1/JAK2-mediated activation of STATs/IRFs—did not affect the ability of LCOR to induce different APM components in vitro (Fig. 3e–g and Extended Data Fig. 3b–d). Importantly, the effects of IFN-γ treatment were abrogated in LCOR-KD, suggesting that LCOR levels modulate sensitivity to IFN and its impact on the APM pathway (Fig. 3e–g and Extended Data Fig. 3b–d). To validate global APM activity, murine AT3-OVA and Py8119-OVA Lcor-KD cells showed reduced OVA presentation while Lcor-OE cells had high presentation in all three conditions, thus demonstrating the essential role of LCOR in modulating and priming APM activity (Fig. 3h and Extended Data Fig. 3e,f). Again, IFN-γ treatment was unable to increase OVA presentation in Lcor-KD cells. Overall, our data demonstrate that LCOR levels determine APM activity in tumor cells with and without IFN signals, highlighting a dominant role of LCOR in antigen presentation.

LCOR directly regulates APM factors through ISRE binding

To understand how LCOR regulates the APM pathway, we transduced MDA-MB-231 cells with ectopic expression of LCOR and LCOR mutant forms16, including a double-point mutation in the nuclear receptor (NR) binding domain (LSKLL to LSKAA) preventing binding to NRs and deletion of the HTH DNA-binding domain (ΔHTH), abolishing its putative binding to DNA. We verified the correct overexpression and nuclear localization of each variant (Extended Data Fig. 4a–c). Next, we performed chromatin immunoprecipitation sequencing (ChIP–seq) analysis. The highest cluster of peaks genome wide was found on the short arm of chromosome 6 (chr6) for the wild-type and LSKAA forms of LCOR, but not for the ΔHTH mutant (Fig. 4a). Importantly, this genomic region of approximately 8 Mb (chr6: 27,810,120–35,980,577) contains the MHC-I cluster47 and all APM genes (Fig. 4a,b), except β2M located on chr15. Consistently, gene expression analysis in MDA-MB-231 cells showed that the wild-type and LSKAA forms of LCOR, but not the ΔHTH mutant, induce APM genes (Fig. 4a,b). Of note, LCOR peaks were located in gene regulatory elements (GREs) of these genes (Fig. 4c), including β2M, suggesting direct master regulation of the APM cluster, also validated by endogenous LCOR ChIP–qPCR using an LCOR-HA knock-in system (Extended Data Fig. 4d). Accordingly, ChIP–seq peak enrichment analysis ranked the APM pathway as the most highly enriched pathway among BioCarta biological processes (Fig. 4d). Interestingly, by analysis of available assay for transposase-accessible chromatin using sequencing (ATAC–seq) data, we observed that this region is shut down in fMaSCs42 (Extended Data Figs. 4e and 2a), reflecting a conserved gene regulatory mechanism of mammary cell immunogenicity. Therefore, we investigated the evolutionary conservation of LCOR in other species and found that it is highly conserved in all vertebrates, especially the NR and HTH domains, the latter being the only domain preserved beyond vertebrates (Extended Data Fig. 4f,g). Therefore, it is an ancient DNA-binding domain originating in prokaryotes as a transcription factor domain. These findings support a conserved role of LCOR in coordinating APM transcriptional regulation through its HTH domain.

Fig. 4: LCOR transcriptional regulation of the APM genomic region through ISRE binding.
figure 4

a, ChIP–seq analysis of pLEX-HA, pLEX-LCOR-HA, pLEX-LSKAA-HA and pLEX-ΔHTH-HA in MDA-MB-231 cells. Genome-wide distribution of LCOR-HA peaks and in chr6 of indicated conditions. Two independent biological replicates. b, Heat map RT–qPCR analysis of APM genes located in the MHC cluster of chr6 in MDA-MB-231 cells, represented by log2 fold change (log2(FC). n = 3 individual biological replicates; data represent the mean. c, ChIP–seq LCOR peak occupancy in APM genes in this specific region of chr6 and β2M on chr15. Blue horizontal shading shows either GREs defined by H3K27ac or H3K4me3 defined by UCSC (https://genome.ucsc.edu/). Occupancy and scale are indicated. Data show a replicate from a. d, Pathway analysis of genes with LCOR ChIP–seq peaks using BioCarta-2016. Data represent –log10(P); P value computed from Fisher’s exact test. e, ChIP–seq HOMER analysis and top five enriched binding motifs of LCOR peaks in the MHC cluster region. P value of enrichment and percentage of motif sequence labeled and classified as positive among peaks (%TP). P value computed from Fisher’s exact test. f, Scheme of the interferon-stimulated regulatory element (ISRE) red fluorescent protein (RFP)-reporter containing six tandem ISRE motifs and minimal CMV promoter (mCMVp). g, Flow cytometry analysis normalized to control condition using ISRE and ISRE-mutated reporters at the indicated conditions in MDA-MB-231 cells. n = 3 individual biological replicates; data represent MFI relative to isotype control mean ± s.e.m. h, Flow cytometry analysis of ISRE reporter in MDA-MB-231 LCOR-GFP knock-in under control (ctrl) conditions and after 24 h treatment with IFN-γ (10 ng ml–1), ruxolitinib (0.5 μM) or BX-795 (TBK1i, 0.5 nM). Left: density plots and MFI of mCherry in LCOR, LCORmed/low and LCORhigh. Right: MFI quantification. n = 3 individual biological replicates; data represent relative MFI mean ± s.e.m. g,h, Exact P values calculated by one-way ANOVA and Bonferroni post hoc analysis; cells were gated from P3 (Extended Data Fig. 9a).

Source data

Our transcriptomic and conservation analyses suggest that LCOR can act as a transcriptional activator. However, LCOR has previously been described as a transcriptional corepressor48 and its function as a possible transcriptional activator had not yet been proven. Performing motif discovery analysis of LCOR ChIP–seq peaks using HOMER software revealed ISREs among the top-ranked prediction motifs—in particular, the interferon-regulatory factor 1 (IRF1) binding site (Fig. 4e), typical of interferon-stimulated genes and APM genes in antiviral responses49. To demonstrate that LCOR can activate APM genes, we generated a promoter-reporter with six tandem repeats of the ISRE motif (6 × CAGTTTCACTTTCCC) upstream of a minimal CMV promoter, and also a mutated version (6 × CAGTGGCACGGTCCC) to discern specific LCOR binding (Fig. 4f). As expected, only LCOR and LSKAA, but not ΔHTH, increased reporter activity as quantified by red fluorescent protein (RFP) flow cytometry in MDA-MB-231 cells (Fig. 4g). Remarkably, no induction was detected when the ISRE sequence was mutated (Fig. 4g), showing that LCOR binds to ISREs and activates transcription. Next, we combined the ISRE reporter with the LCOR-GFP knock-in, generating a unique flow cytometry system to study ISRE activity dependent on endogenous LCOR regulation. We explored the effects of different conditions, including IFN-γ treatment and different IFN inhibitors—ruxolitinib (inhibitor of JAK1/JAK2) and BX-795 (TBK1 inhibitor blocking STING signaling)—to further extricate the interference by IRFs and LCOR on reporter activity. We observed that IFN-γ treatment increased ISRE in LCOR+ cells (medium and high), but not in LCOR cells. Moreover, ruxolitinib inhibited ISRE activity only in LCOR cells but not in LCOR+ cells, demonstrating that LCOR induction of ISRE is independent of IFN and essential for the transcriptional activation of ISRE-controlled genes (Fig. 4h). Accordingly, these treatments did not suppress ISRE induction in LCOR-OE cells (Extended Data Fig. 4h) and IFN-γ failed to induce ISRE in LCOR-KD cells (Extended Data Fig. 4i). This is consistent with the independent induction of APM activity (Fig. 3e–h).

We further confirmed the transcriptional activity of LCOR by performing RNA polymerase II subunit B (POLR2B) ChIP. POLR2B bound MHC-I genes only in the presence of LCOR (Extended Data Fig. 4j) but not in LCOR-KD cells (Extended Data Fig. 4k). Moreover, POLR2B is enriched with either LCOR-OE or LSKKA-OE but not with the ΔHTH mutant (Extended Data Fig. 4j), indicating the requirement of LCOR binding at MHC-I gene promoters. Overall, these results reveal that LCOR is a transcriptional activator of the APM pathway.

LCOR facilitates tumor-immune infiltration and killing

Next, we tested the effects of LCOR-mediated APM induction on tumor immunity. In vitro CTL assays of AT3-OVA and Py8119-OVA cells cocultured with CD8+ OT-I T cells, and of 4TO7-EGFP cocultured with CD8+ JEDI T cells, showed increased immune killing of Lcor-OE cells in both systems (Fig. 5a,b and Extended Data Fig. 5a,c). Knockdown and knockout of β2m in Lcor-OE cells rescued survival and evasion of T cell killing, demonstrating that this effect is dependent on antigen presentation capacity (Fig. 5c and Extended Data Fig. 5d). In addition, Lcor-KD cells avoided immune-mediated cell killing (Fig. 5d and Extended Data Fig. 5b,e), consistent with their reduced OVA presentation ability (Fig. 3h). Lcor-OE cells increased T cell activation as measured by CD69, while Lcor-KD reduced it (Fig. 5e). In tumor growth experiments, 4TO7 Lcor-OE cells showed higher Lcor-mediated reduction in immunocompetent mice (Balb/c) than in immunodeficient NSG mice (Extended Data Fig. 5f,g), indicating an elicited immune reaction due to LCOR-mediated immunogenicity. The inverse effect was true for Lcor-KD tumors (Extended Data Fig. 5h,i). 4TO7 Lcor-OE tumors showed a substantial increase in CD4 and CD8 infiltrating lymphocytes compared to control tumors as measured by immunohistochemistry (IHC) and flow cytometry (Fig. 5f,g), and a nonsignificant increase in CD45+ leukocytes (Fig. 5f and Extended Data Fig. 5j). The CD45+/CD3 compartment showed no significant changes in dendritic cells (CD45+/CD3/CD11c+/F4/80) and macrophages (CD45+/CD3/CD11b+/F4/80+) (Extended Data Fig. 5k,l). These results support a lymphocytic immunoreactive response to Lcor-OE tumors due to their high immunogenicity.

Fig. 5: Tumor infiltration, immunogenicity and killing.
figure 5

ae, CTL assays of AT3-OVA cells after 72 h of coculture with OT-1 CD8+ T cells at the indicated conditions, and effector (E) and tumor (T) cell ratios. a, Lcor-OE versus control vector. b, Time-lapse confocal imaging of a, ratio 1:1. Tumor cells are labeled red, CD8+ T cells blue and cell dead tracker (SYTOX) green (white shading when mixed with tumor cells). Blue arrows indicate tumor-immune interactions, and white arrows killed tumor cells. c, Lcor-OE siRNA β2m versus Lcor-OE siRNA control. d, Lcor-KD versus control vector. e, Flow cytometry analysis of CD69 expression in CD8+ T cells after CTL. ae, n = 3 independent biological replicates; a,c,d, data represent mean ± s.e.m.; e, relative MFI of a representative replicate. f, IHC and quantification analysis of percentage of tumor-infiltrating immune cells (CD45+) and CD8+ T cells (CD8+) in control versus Lcor-OE 4TO7 tumors. n = 3 individual biological replicates; data represent mean ± s.e.m. g, Flow cytometry analysis of percentage of tumor-infiltrating lymphoid cells (CD45+CD3+), CD8+ and CD4+ T cells (Extended Data Fig. 9b) in control or Lcor-OE 4TO7 tumors. n = 10 individual biological replicates; data represent mean ± s.e.m. h,i, Computational immunophenotype clustering by deconvolution of TNBC METABRIC dataset using xCell50. h, Heat map representing clusters of patients based on four IPs as indicated. i, LCOR levels in each immunophenotype. b,f, Scale bars, 20 μm. a,c,d,g,h, Exact P values by two-tailed Student’s t-test; i, one-way ANOVA with Bonferroni post hoc analysis.

Source data

We performed deconvolution xCell50 analysis of the METABRIC dataset to estimate the immune content of 186 samples from patients with TNBC51. We generated immunophenotype clusters representative of the immune landscape: cluster 1, low-immune-infiltrated tumors; cluster 2, immunosuppressive populations; cluster 3, cytotoxic populations; and cluster 4, highly cytotoxic populations. Next, we analyzed LCOR levels in these different clusters and found higher expression in the cytotoxic clusters (Fig. 5h,i). LCOR-high tumors were also more enriched for CD4, CD8 and γδ T lymphocyte gene signatures (Extended Data Fig. 5m). Overall, these results support the premise that LCOR promotes immunogenicity and adaptive immune infiltration in TNBC, mediating antitumor immunity.

LCOR levels and ICB responsiveness in patients with TNBC

We performed LCOR IHC analysis on a small set of matched clinical TNBC samples before and after ICB treatment in the neoadjuvant setting. Two patients (P1 and P2) were treated with anti-PD-L1 atezolizumab plus polychemotherapy (abraxane and carboplatin) while two others (P3 and P4) were treated with anti-PD1 nivolumab plus SYK/FLT3 inhibitor. In all cases, LCOR expression was lower in residual disease (Fig. 6a). These data support the supposition that LCOR cells are eliminated by combinatorial neoadjuvant ICB therapy.

Fig. 6: Preclinical and clinical implications of LCOR in combination with ICB therapy.
figure 6

a, IHC analysis of LCOR in samples from patients pre and post treatment for TNBC: P1 and P2 were treated with anti-PD-L1 atezolizumab + abraxane and carboplatin while P3 and P4 were treated with anti-PD1 nivolumab + SYK/FTL3 inhibitor (TAK-659). Left: nIRS of LCOR in pre- versus post-treated samples. Right: representative images of pre- and post-treated samples with LCOR nIRS = 6 and nIRS = 2, respectively. P value by Wilcoxon signed-rank test. b,c, Analysis of LCOR mRNA levels in all preinduced responders (CR + PR) versus nonresponders (PD + SD) patients in the TONIC trial (b) and in the I-SPY2 trial (c), comparing responders (CR + PR) and patients with progression disease (SD + PD). n is indicated separately for each condition; data represent mean ± s.e.m. in both. d, Growth curves of orthotopic 4TO7 transplanted tumors and individualized treatment starting at the reach point of tumor size 0.5 × 0.5 cm2, with the indicated conditions. Treatments were administered intraperitoneally at the indicated dose regimes weekly. n = 10 mammary glands for each condition; data represent mean ± s.e.m. e, Waterfall plot representing change of percentage in tumor volume from the first day of treatment to the respective endpoint. n = 10 mammary glands for each condition. f, Take rate of complete response (CR) to treatment and tumor-free mouse survival 12 months after treatment withdrawal. Compilation of five independent experiments (n = 10 per experiment) treated with anti-PD-L1 (10 mg kg–1) every 3–7 days. g, Carmine staining of a representative cured mammary gland. n = 3 cured mammary glands. h, Growth curve of orthotopic 4TO7 control versus Lcor-KD tumors at the indicated dose regime every 3 days. n = 10 mammary gland tumors; data represent mean ± s.e.m. i, Preclinical lung metastasis assay by TV injection of 4TO7 cells. Once metastases were established, anti-PD-L1 treatment was commenced at the indicated dose regimen. n = 6 mice per condition; data represent mean ± s.e.m. in log10 scale. Veh., vehicle. b,c, Exact P values by one-tailed Student’s t-test; h,i, one-way ANOVA. N.T., nontreated d,h,i.

Source data

We also explored the value of LCOR in ICB response in larger clinical trials. Analysis of the TONIC trial, which included 53 metastatic TNBC patient samples pretreatment40, showed higher levels of LCOR in responders (Fig. 6b). Additionally, data from the phase-II I-SPY2 trial, using durvalumab, olaparib and neoadjuvant paclitaxel in patients with TNBC52, also showed higher levels of LCOR in responders (Fig. 6b). These findings demonstrate that LCOR levels are associated with response to ICB-containing combination therapy in TNBC. To assess whether these observations also apply in other cancer types, we examined a melanoma cohort treated with single-agent anti-PD-1, showing higher levels of LCOR in responders (Extended Data Fig. 6a). Moreover, our breast IRS overlaps with a melanoma ICB resistance signature53, and patients with LCORhigh TNBC inversely correlated with the latter signature (Extended Data Fig. 6b–d). Overall, these results support the association of LCOR with ICB clinical benefit.

LCOR overcomes resistance to ICB leading to tumor eradication

Antigen presentation and PD-L1 are hallmarks of anti-PD-1/PD-L1 therapy response3. Lcor slightly increased Pd-l1 expression in 4TO7 and AT3 cells (Extended Data Fig. 6e,f), and showed positive correlation in patients with TNBC (Extended Data Fig. 6g). This may be explained by LCOR priming IFN sensitivity, which can induce PD-L1 (refs. 3,5). Remarkably, the positive levels of PD-L1 in combination with the potent LCOR induction of APM sets an ideal tumor configuration for anti-PD-1/PD-L1 therapy. Therefore, we performed orthotopic MFP transplantation of syngeneic 4TO7 cells in immunocompetent mice, allowed tumors to reach a size of 0.5 × 0.5 cm2, then initiated ICB therapy with anti-PD-L1 once per week. Control-nontreated and control-treated tumors continued growing; however, Lcor-OE tumors treated with anti-PD-L1 ICB totally regressed, with complete response (CR) by 20 days in all mice (Fig. 6d,e). Importantly, depletion of the CD4/CD8 compartment led to no response to ICB, validating that the observed CR to anti-PD-L1 in Lcor-OE tumors was mediated by the adaptive immune system. We performed up to five independent experiments and observed CR in 49 out of 50 Lcor-OE tumors (Fig. 6f). Of note, the only Lcor-OE tumor that did not respond had lost the ectopic expression of Lcor (Extended Data Fig. 6h). All 49 mammary glands with CR were tumor free after 2 months of anti-PD-L1 discontinuation. We followed up 15 of these mice for 1 year, and in none had tumors recurred and all had tumor-free glands, suggesting that we had irreversibly eradicated the tumors and cured these mice (Fig. 6f,g). Accordingly, 4TO7 Lcor-KD tumors demonstrated higher resistance to anti-PD-L1 compared to control 4TO7 cells (Fig. 6h). The AT3 syngeneic model in C57BL/6J mice also confirmed the CR of Lcor-OE tumors to ICB therapy, both in vivo (Extended Data Fig. 6i) and in vitro (Extended Data Fig. 6j).

To study the dominant role of LCOR over IFN signaling in vivo, we performed a MFP experiment comparing the effects of Lcor-OE and the IFN type-I inducer Poly (I:C)54. The combination of Poly (I:C) treatment and anti-PD-L1 did not reach the efficiency shown by the Lcor + anti-PD-L1 condition (Extended Data Fig. 6k), again demonstrating the dominant role of LCOR in ICB therapy beyond IFN-mediated effects.

Next, we conducted preclinical lung metastasis assays using 4TO7 tail vein (TV) administration and allowed for the establishment and growth of lung metastases before starting anti-PD-L1 therapy (Fig. 6i). To achieve synchronous metastases among the different conditions, we injected threefold more Lcor-OE cells due to their inherent reduced tumorigenicity. 4TO7 Lcor-OE metastasis were cured after 4 weeks of anti-PD-L1 treatment in five out of six mice, while all control metastatic tumors progressed despite anti-PD-L1 exposure (Fig. 6i and Extended Data Fig. 6l). Overall, these preclinical assays demonstrate a conclusive curative response of anti-PD-L1 treatment mediated by the effects of LCOR on tumor cell immunogenicity, which represents a promising therapeutic target for early and advanced TNBC.

mRNA-based LCOR therapy in combination with ICB in vivo

Therapeutic mRNA delivery using nanoparticles has enabled the rapid development of highly effective vaccines against COVID-19 and may similarly revolutionize cancer therapy55. Based on current knowledge on extracellular vesicles (EVs) and mRNA delivery56, we designed a proof-of-concept approach to restore Lcor expression through the introduction of Lcor mRNA into tumor cells, and to combine this treatment with ICB. We ectopically expressed Lcor tagged with HA in HEK293T cells, which produce large quantities of EVs containing ectopic Lcor-HA mRNA transcripts (Fig. 7a). In vitro treatment of 4TO7 cells with Lcor-HA mRNA EVs showed incorporation and translation of the Lcor protein as detected by immunoblot using anti-HA (Fig. 7b), and corrected nuclear localization (Fig. 7c). Lcor EVs reduced the CSC population and upregulated APM genes in 4TO7 cells (Extended Data Fig. 7a,b), increasing CD8+ T cell-mediated killing (Extended Data Fig. 7c). Next, we designed a preclinical lung metastasis assay to test Lcor mRNA therapy. After 5 days of EV administration, lung metastases were already showing incorporation and translation of Lcor-HA protein in most tumor cells (Fig. 7d) as proof of principle in vivo. Next, mice serially treated with the combination of EV-based Lcor mRNA therapy and anti-PD-L1 showed significantly longer survival and complete elimination of lung metastasis compared with the EV-control and anti-PD-L1 therapy (Fig. 7e–g). These results suggest that LCOR mRNA therapy is a potential therapeutic partner of ICB therapy.

Fig. 7: Lcor mRNA therapy with EVs in combination with ICB in breast cancer metastasis.
figure 7

a, Lcor mRNA levels in EVs from HEK293T cells with ectopic overexpression of Lcor-HA (pLEX-Lcor-HA) versus control (pLEX-HA). Data represent mean ± s.e.m. of three technical replicates. Representative experiment of five independent experiments listed in Source data. b, In vitro treatment with control and Lcor-HA-EVs over 3 days and anti-HA immunoblot in 4TO7 cells. Representative immunoblot of two biological replicates. c, Lcor-HA protein expression and localization in recipient cells. Representative image of three independent biological replicates. d, EV delivery, uptake and protein translation in vivo: retro-orbital administration of EVs and metastatic tumor tissue IHC with anti-HA and quantification of nuclear positivity. n = 3 independent biological replicates; data represent mean ± s.e.m. e, Preclinical lung metastasis assay by TV injection of 50,000 4TO7 cells. After 8 days, metastases were established; EV + anti-PD-L1 treatment was started at the indicated dose regime and repeated every 3 days. n = 6 mice per condition; data represent mean ± s.e.m. in log10 scale. Representative BLI images (left) and growth curves (right) of the indicated conditions. f, Mouse survival Kaplan–Meier plot of the lung metastasis assay in e. Tx, treatment. P value calculated by log-rank test. g, Representative images of established lung metastases treated with control EVs + anti-PD-L1 (left), Lcor EVs (middle) and Lcor EVs + anti-PD-L1 (right). h, Schematic model of the main thrust of this study: (1) LCORlow CSCs led to immunoediting escape and resistance to ICB therapy in TNBC; (2) LCOR increased the expression of APM genes through ISREs, independently of IFN; (3) LCOR-mediated immunogenicity determined response to ICB, CSC elimination and overcoming resistance in TNBC; and (4) LCOR mRNA therapy led to increased APM in CSC and avoided their escape to ICB, resulting in complete pathological response and cure. c,d, Scale bars, 60 µm; g, 300 µm. d, Exact P values by two-tailed Student’s t-test; e, one-way ANOVA; f, logrank test.

Source data

Discussion

Our study provides integrative knowledge of TNBC ICB resistance, with insights into intratumoral phenotypic heterogeneity and a key molecular mechanism that controls tumor cell immune detection independently of IFN. We have shown how, in tumor heterogeneity, LCORlow breast CSCs possess an abrogated APM gene program, representing a tumor-intrinsic mechanism of ICB resistance (Fig. 7h). Therefore, LCOR modulation represents a potential therapeutic strategy for improving the clinical benefit of immunotherapy in TNBC.

LCOR is known as a corepressor of agonist-activated NR signaling48, and has been associated with reduced proliferation in different cancer types57,58,59. In ER BC, although LCOR was reported as a tumor suppressor through mediation of IFN sensitivity16, it remains unknown what molecular mechanism explains the overlap of LCOR and genes downstream of IFN. Our findings mechanistically dissect their overlap and uncouple LCOR from IFN signaling. We demonstrate that LCOR is a transcription factor binding to ISREs—commonly found in APM genes and other genes downstream of IFN—activating their transcription with or without IFN inputs. Another conclusion from our study is that IFNs barely activate APM in the absence of LCOR, demonstrating that LCOR is essential for the transcriptional coordination of APM. Although the role of LCOR as a transcription factor has not been reported, previous reports are aligned with this function since it belongs to the transactivator factor homolog family Mblk1 (ref. 57) and its ablation reduced specific gene expression60. Our conservation analysis further supports the transcription factor function of LCOR, showing high conservation of the HTH DNA-binding domain in ancestral organisms57. Therefore, we reveal a relevant activity of LCOR as a conserved transcription factor controlling cellular immunity.

We report a remarkable genomic binding zonation of LCOR on the MHC genomic region and APM genes of human chromosome 6 (ref. 47). It is believed that these genes are evolutionarily clustered to favor highly coordinated regulation of cellular immunity47. Our results suggest that the LCOR-HTH domain might be evolutionarily favored to orchestrate these genes, and thus cellular immunogenicity. Therefore, in a malignancy context with high genomic instability and neoantigens, LCOR-mediated immunogenicity promotes selective antitumor immunity while leaving healthy cells exempt from immune attack due to low neoantigen load. LCOR modulation thus represents an excellent opportunity to be exploited in immune-based strategies specifically targeted at elimination of cancer cells.

Cancer stem cells with defective APM fit immunoediting principles, in which a poorly immunogenic subpopulation is not eliminated but persists, self-renews and finally escapes due to immune-evasive properties and tumor initiation capability. Importantly, we demonstrate that LCORlow CSCs exploit this mechanism by abrogating the entire APM system to escape ICB immunotherapy, suggesting that eradication of LCORlow CSCs will prevent immunoediting escape and relapse and thus improving long-term ICB clinical benefit. Recent studies also support the belief that CSCs and tumor dedifferentiation are implicated in immunotherapy resistance25,26,61, although in different cancer types and by mechanisms not related to the APM. In TNBC, recent data on ICB therapy show no efficiency in the mesenchymal subtype, which is enriched in stem cell-like phenotypes62. These evidences are aligned with our unique findings in regard to TNBC phenotypic heterogeneity and CSCs driving ICB resistance.

In breast cancer, we show that the LCOR–APM axis is highly relevant in triple-negative disease, is associated with ICB response and thus provides an opportunity to improve TNBC therapy. While most immunotherapies are designed to avoid immunosuppression or to potentiate immune system antitumoral activity, tumor antigen presentation is still critical for ICB response3,6. A key translational breakthrough of our study is a potential therapy designed to specifically modulate antigen presentation in tumor cells using LCOR mRNA delivery in concert with ICB. The current focus on mRNA therapies, and development of nanotechnology for its therapeutic delivery, makes this strategy a compelling case for potential future cancer therapies. In summary, LCOR represents an unprecedented opportunity to reconfigure tumor immunity, constrain CSCs and exploit alongside ICB therapy. Future clinical studies are required to apply mRNA LCOR therapy in combination with ICB in patients with TNBC.

Methods

Ethical regulations

This study complies with all ethical regulations. Clinical patient samples have approval from the Ethical Committee of Clinical Investigation—Mar Park of Health, the institutional review boards at Vall d’Hebron Hospital and from the INCLIVA ethical committee. All individuals gave their informed consent before inclusion. All animal procedures presented in this study were approved by the Ethical Committee for Animal Research of the Barcelona Biomedical Research Park and by regulation from the Departament de Medi Ambient i Habitatge de la Generalitat de Catalunya (Catalonia Government). For all experimental procedures, euthanasia was applied once tumors reached a volume of 1,500 mm3 or when animal health was compromised.

Animal studies

For this study, mouse strains Balb/c, C57BL/6J, C57BL/6-Tg (TcraTcrb)1100Mjb/J (OT-1), Ptprca (TcrbTcra)Ln1Bdb H2d/J Just Enhanced GFP (JEDI) and NOD.Cg-Prkdcscid Il2rgtm1Wjl/SzJ (NSG) were used. Tumor cells were orthotopically injected into the MFP using 1:1 PBS:Matrigel (Corning). For TIC evaluation, severely immunocompromised NSG female mice were orthotopically transplanted with series of LDA. After 2 weeks, tumor incidence was evaluated by palpation and TIC frequency calculated using extreme limiting dilution analysis (ELDA) software16. For tumor comparison between immunocompetent and immunodeficient strains, 5,000 AT3 or 4TO7 cells were transplanted into the MFP of ten glands of NSG and Balb/c or C57BL/6J mice. Tumor volume was measured twice weekly with digital calipers, and calculations were applied (π × length × width2/6). In ICB assays, mice were treated with either anti-human/mouse anti-PD-L1 (atezolizumab, clone SP142, Tecentriq) or anti-mouse anti-PD-L1 (BioXCell, clone 10F.9G2, catalog no. BE0101) versus vehicle. Treatments were initiated when tumor size reached 0.5 × 0.5 cm2. Dose regimes were applied at 10 mg kg–1 every 3 or 7 days (as indicated in the different experiments). Depletion of CD4 and CD8 was achieved using anti-CD4 and anti-CD8 neutralizing antibodies, respectively, at 400 μg per mouse once tumors reached 0.5 × 0.5 cm2, followed by a weekly dose of 250 μg. For lung metastasis experiments, 50,000 transduced control and 150,000 Lcor-overexpressing 4TO7 cells were TV injected in Balb/c mice to obtain similar metastatic growth among conditions. After 1 week, metastases were treated with 10 mg kg–1 anti-PD-L1 once weekly. Metastatic lesions were monitored by photon flux bioluminescence (BLI) imaging once per week, and data were collected using Live Image v.4.3.1 in a Perkin Elmer Living Image system. For the humanized xenograft model, NSG mice were orthotopically transplanted with 15,000 MDA-MB-231 cells. Once tumors reached 0.3 × 0.3 cm2, animals were intraperitoneally injected with 10,000,000 peripheral blood mononuclear cells (PBMCs) from a healthy donor63. When tumor size reached 0.5 × 0.5 cm2, the animals were treated with either vehicle or human/mouse crossreactive anti-PD-L1 at 10 mg kg–1 every 3 days.

Paraffin-embedded tissue samples and PDOs

Paraffin-embedded tissue samples were obtained from a total of four patients with early-stage TNBC pre and post treatment. Two patients were treated with polychemotherapy (abraxane + carboplatin) in combination with atezolizumab (NeoTRIP study) while the other two were treated with TAK649 in combination with anti-PD1 nivolumab (NCT02834247 study). Samples were obtained from the Hospital Clínic of Valencia. Hormonal receptor status was evaluated by IHC (ER- and PR- were defined as <1% positive stained nuclei), and HER2 was assessed by IHC and fluorescence in situ hybridization.

Small patient-derived tumor pieces (~1–5 cm3) were obtained from surgical resection of patients with TNBC at Hospital del Mar (Barcelona), with previous informed consent. Small tumor pieces were mechanically and enzymatically digested at 37 °C for 2 h in mammary epithelial cell growth medium16 supplemented with enzymes (Supplementary Table 2), followed by a short incubation with 5 mg ml–1 dispase (Merck, catalog no. D4693), 0.1 mg ml–1 DNase I (Merck, catalog no. D5025-150) and 0.25% trypsin (Cultek, catalog no. 25300-062). The digested tissue suspension was strained through a 100-μM filter, and single cells were cultured and expanded in vitro as previously described64.

Cell lines, culture conditions and treatments

All cell lines used in this study—breast cancer cell lines (human MDA-MB-231 and mouse Py8119, 4TO7 and AT3)—were obtained from Y. Kang at Princeton University and cultured according to the American Type Culture Collection (ATCC). HMLE cells were obtained from R. Weinberg at MIT. HEK293T cells were obtained from ATCC. Cells were routinely checked for Mycoplasma and all were negative. For this study, anti-human/mouse anti-PD-L1 (atezolizumab, clone SP142, Tecentriq) was used in vitro at 10 μg ml–1 for 72 h; recombinant mouse or human IFN-γ (R&D systems, human catalog no. 285-IF-100; mouse: catalog no. 485-MI-100) at 10 ng ml–1 for 24 h; the JAK1-JAK2 inhibitor ruxolitinib (LC laboratories, catalog no. R-6688) at 1 μM for 24 h; and BX-795 (TBK1i) (Merck, catalog no. SML0694) at 0.5 nM for 24 h.

3D tumor sphere assay

Single cells were plated in Ultra-Low-attachment plates (Cultek, catalog no. 3473) in standard mammosphere medium (Supplementary Table 2)16. For multiple-generation sphere formation assay, tumor spheres were collected by centrifugation at 200g for 2 min and incubated with trypsin 0.25% for 5 min at 37 °C. Single cells were strained through a 40-μM filter, centrifuged at 300g for 2 min and reseeded at an equal number of cells for the next round of tumor sphere formation.

CTL assay

For mouse cell coculture assays, splenocytes from OT-I and JEDI mouse models were obtained. Spleens were mechanically minced, and single cells were incubated with ACK buffer (Fisher, catalog no. A1049201). The single-cell suspension was strained over a 70-μM filter and activated overnight with 2 μg ml–1 SIINFEKL (Merck, catalog no. S7951) or HYLSTQSAL (Proimmune, catalog no. F198-2A-E) peptide for OT-1 and JEDI, respectively, in mouse splenocyte medium (Supplementary Table 2). CD8+ T cells were purified using the CD8a+ T cell Isolation kit (Milteny, catalog no. 130-104-075) for negative selection, then cocultured with tumor cells for 72 h with or without anti-human/mouse PD-L1 (ref. 65) (atezolizumab, clone SP142, Tecentriq) at 10 μg ml–1. For the T cell-killing assay with β2m knockdown, cell lines were transfected using lipofectamine 3000 (Life Technologies, catalog no. L3000015) with either 100 ρM siRNA negative control (Life Technologies, catalog no. AM4613) or 100 ρM Silencer Pre-designed siRNA against murine β2m (Life Technologies, catalog no. 160820). For human coculture assays, PBMCs were isolated by gradual centrifugation63 of blood samples from healthy adult donors from the Tissue Bank of Catalonia, and cocultured with tumor cells in coculture medium (Supplementary Table 2). Cell viability or enrichment was determined after 72 h by the crystal violet method or flow cytometry, respectively.

EV isolation and in vitro/in vivo delivery assays

HEK293T cells were cultured in DMEM with 10% extracelular vesicles (EVs)-reduced fetal bovine serum (FBS) (Hyclone) and transfected with either control pLEX-HA vector (control condition) or mouse pLEX-Lcor-HA overexpression construct using Lipofectamine 3000. Medium was changed 24 h after transfection, and supernatants were collected at 48 and 72 h for EV isolation. Cells, debris and large vesicles were first removed by serial spinning at 500g for 10 min, 2,000g for 15 min and 10,000g for 30 min, followed by ultracentrifugation in a Beckman Coulter L-90K ultracentrifuge at 70,000g for 60 min. The resulting EV pellet was resuspended in PBS. EV RNA content was estimated by RNA isolation (Qiagen) and quantification with Nanodrop; protein content was measured by bicinchoninic acid assay. EVs were treated with 10 µg ml–1 RNAse A (Fisher Scientific, catalog no. 12091021) and 5 µl of proteinase K at 20 mg ml–1 (Fisher Scientific, catalog no. EO0491) for 30 min at 37 °C before assay. 4TO7 cells in culture were treated with 4 ng µl–1 EV protein for 72 h at 37 °C. For experimental metastasis in vivo, 50,000 4TO7 cells were TV injected. Mice were treated with 10 mg kg–1 anti-PD-L1 (atezolizumab, Tecentriq) and 8 µg of EV protein (in a total volume of 100 µl of PBS) from control or Lcor-HA EVs via retro-orbital venous sinus injection every 3 days, when BLI imaging (>2 × 106 photon flux) showed metastatic colonization. Metastatic lesions were monitored by photon flux BLI, then lungs were harvested for anti-HA IHC and hematoxylin and eosin staining.

IF, LSFM and IHC analysis

For IF analysis of LCOR-OE and LCOR mutants, cells were seeded in coverslips, fixed for 1 h with methanol at –20 °C and washed with acetone. Samples were blocked for 30 min using blocking buffer (Supplementary Table 3), incubated for 2 h at room temperature with anti-HA, PBS washed and incubated for 1 h at room temperature with a secondary antibody Alexa Fluor 488 anti-rabbit. For IF analysis of LCOR-GFP knock-in reporter system and pan-HLA expression, cells were seeded, fixed for 15 min at room temperature with 4% paraformaldehyde (PFA), washed and blocked for 1 h at room temperature with blocking buffer. Cells were incubated with anti-pan-HLA-ABC for 2 h at room temperature, followed by incubation with an Alexa Fluor 647 anti-mouse. Images were taken using an upright Nikon Eclipse Ni-E fluorescence microscope (Nikon), and data were collected using Ni Setup Tool v.1.2.2. Fiji software was used for further analysis.

For imaging of living organoids transduced with the SORE6+ CSC reporter, we used the LS1 live light-sheet microscope system (Viventis). Individual organoids were incubated with anti-pan-HLA-ABC for 2 h at room temperature, washed five times with PBS and incubated with Alexa Fluor 647 anti-mouse for 1 h at room temperature. Hoechst 33342 (Fisher, catalog no. H3570) was used. 3D rendering of organoids was performed using the Clear Volume plugin in FIJI software.

For IHC analysis of infiltration by CD45+ and CD8+ T cells, paraffin-embedded mouse tumor samples were stained with an Auto stainer Plus (Dako) using the 3,3’-diaminobenzidine (DAB) staining method. Tissue sections (3 µm) were stained for CD8 and CD45 antibodies; necrotic areas were excluded from quantification. For LCOR evaluation in clinical samples, paraffin-embedded sections (3 µm) from tumor tissue blocks were stained with anti-LCOR at room temperature for 1 h, followed by incubation with an anti-rabbit Ig dextran polymer (Flex, Agilent). Sections were visualized with DAB and counterstained with hematoxylin. All incubations were performed on the Agilent Link platform. Nuclear immunoreactive score (nIRS) was used to evaluate LCOR expression in samples. nIRS provides a range between 0 and 12 as the product of positive cell proportion score (0–4) (0, 0%; 1, 1–30%; 2, 31–60%; and 3, >60%) and staining intensity score (0–3) (0, no reaction; 1, weak signal; 2, mild signal; and 3, strong signal). For confirmation of Lcor-HA delivery in metastatic lesions, metastatic lungs treated with control or Lcor-HA EVs were collected and fixed in 4% PFA overnight. Paraffin-embedded sections (3 µm) from tissue blocks were stained with anti-HA at room temperature for 2 h. For all IHC analyses performed in this study, slides were visualized and analyzed using either QuPath-0.2.0 (ref. 66) or CellSens software.

Time-lapse confocal microscopy imaging

Stained tumor cells (CellTracker Deep Red; Fisher, catalog no. C34565) were cocultured with stained CD8+ T cells (CellTracker CM-Dil Dye; Fisher, catalog no. C7000) at a 1:1 ratio in coverslips µ-Slide 8 Well (Ibidi, catalog no. 80826) for 16 h in RPMI medium containing SYTOX green dye (Invitrogen, no. S7020). Forty random zones were selected and analyzed using Zeiss Cell Observer HS (Zeiss), taking images every 5 min in the green, red, deep red and phase contrast channels. Images were processed and visualized using FIJI software.

Evolutionary conservative analysis

LCOR homologous sequences from 12 representative vertebrate species were obtained using Ensembl (https://2020.ensembl.org/). We aligned the sequences with MAFFT67 and displayed the alignments with the NCBI Multiple Sequence Alignment Viewer (https://www.ncbi.nlm.nih.gov/projects/msaviewer/, National Center for Biotechnology Information, 2020). We searched for conserved protein domains in LCOR using PFAM68. Information on other proteins containing HTH-psq domains, and their domain organization, was also obtained from PFAM.

Viral production and transduction of cell lines

HEK293T cells were transfected with lentiviral plasmids jointly with pocket plasmid (VSVG) and gag-pol plasmid (pCMV-R8.91), following the standard lentiviral packaging protocol. Cell lines were transduced in six-well plates, with concentrated viruses in the appropriate medium for each cell line containing 8 μg ml–1 Polybrene and selected with the corresponding antibiotic resistance.

ChIP and ChIP–seq library preparation

For ChIP–qPCR and sequencing, cells were grown on 150-mm2 plates, fixed, lysed and sonicated for seven cycles of 30 min on/30 min off using Bioruptor Pico Tubes (Diagenode) with Sonication Beads (Diagenode) in a Bioruptor Sonicator (Diagenode). Samples were incubated overnight at 4 °C with rotation, with either 5 μg of anti-HA antibody or 5 μg of goat anti-rabbit IgG (R&D, no. SC-2025). Dynabeads Protein A for Immunoprecipitation (Fisher Scientific) was used for chromatin isolation. Immunoprecipitated chromatin was washed, eluted and purified using the Qiagen DNA purification kit (Qiagen). All primers for ChIP–qPCR used in this study are listed in Supplementary Table 4. For ChIP–seq, chromatin quality and quantity were checked using the Agilent Bioanalyzer. Libraries were sequenced (50-base pair (bp) single-end) on a HiSeq 2500 platform (Illumina). The quality of fastq files was checked with FastQC software (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Reads were aligned with the bowtie2 (ref. 69) mapper to release 27 of the Homo sapiens Gencode version of the genome (GRCh38/hg38 assembly) (https://www.gencodegenes.org/human/release_27.html). Quality of the mapped files (BAM format) was checked with QualiMap. Peaks70 on individual samples were identified using the MACS2 (ref. 71) program (narrow peaks with q < 0.1 were initially selected). Overlap of peaks across biological replicates was retrieved using the Bedtools72 suite of tools.

Knock-in generation

We generated two different LCOR knock-in types in MDA-MB-231 in the endogenous LCOR of human chromosome 10: LCOR tagged with GFP (LCOR-GFP) and LCOR tagged with HA (×3) (IRES-GFP). We cloned single-guide RNAs targeting the 3’ of exon 8 into the pX330-U6 plasmid73 (sgRNAs shown in Supplementary Table 4). Insertion of each guide was checked by PCR and Sanger sequencing. To construct donor vectors for LCOR-GFP and LCOR-3 × HA-IRES-GFP knock-in, homology arms of 1 kb targeting the cutting site were amplified and inserted in a homology direct repair (HDR)-plasmid donor backbone. Plasmid sequences were verified by Sanger sequencing. MDA-MB-231 was cotransfected with lipofectamine 3000 with plasmids HDR-GFP-HDR or HDR-3 × HA-IRES-GFP-HDR and pX330-U6-sgRNA at a 10:1 ratio, respectively. Cells were harvested and single-cell sorted for GFP+. Single clones were tested by flow cytometry and Sanger sequencing.

β2m KO generation

For 4TO7-EGFP β2m KO generation, three independent guides (Supplementary Table 4) were annealed and cloned into digested pSpCas9(BB)-2A-GFP (px458) with FastDigest Bpil (Fisher Scientific, catalog no. FD1014). 4TO7 cells were transfected with lipofectamine 3000, stained for β2m using a mouse PE anti-β2m and sorted in single-cell, 96-well plates. β2m KO clones were validated by flow cytometry (Extended Data Fig. 8a), PCR and Sanger sequencing.

Molecular cloning and plasmids

LCOR-overexpression plasmids (controls pLEX-MCS and pLEX-HA; pLEX- mouse Lcor and human LCOR-HA, LCOR-LSKAA-HA and LCOR-ΔHTHHA) and an OVA-overexpressing plasmid (pLEX-OVA-IRES-mCherry) were obtained from the laboratory of Y. Kang. pLEX-Lcor-HA was generated by PCR amplification of mouse Lcor complementary DNA and the addition of HA at the 3’ end, inserted together into pLEX-MCS after digestion with SpeI and AgeI (NEB). For gene knockdown assays, shRNA were purchased from Sigma-Aldrich for targeting of mouse gene Lcor (no. TRCN0000085107) and human gene LCOR (nos. TRCN0000016306 and TRCN0000436034). Other plasmids used were the SORE6 series31 including minCMV-GFP, minCMV-RFP, SORE6-minCMV-GFP and SORE6-minCMV-RFP, plus the proteasomal-activity pQCXIN/ZsGreen reporter46. The ISRE reporter was generated by the introduction of ISRE sequences (six tandem repeats of the interferon-stimulated response element (ISRE) 5’-CAGTTTCACTTTCCC-’3) in front of the minimal CMV promoter (mCMVp)-RFP after removal of binding sites of the SORE6-minCMV-RFP construct31 using BstZ17I-HF and ClaI (NEB) restriction enzymes. A mutated version of the ISRE reporter was also designed by mutation of two thymines to guanines on two highly conserved thymine triplets (Supplementary Table 5). Sequence insertion was assessed by Sanger sequencing. Details of all constructs and recombinant DNA used can be found in Supplementary Table 6.

Flow cytometry analysis and cell sorting

Flow cytometry analysis data were collected on either a Fortessa Flow Cytometer (BD) or with LSRII Flow cytometry (BD) using FACSDiva v.9.0 software, and analyzed by FlowJo software v.10.8.1 (FlowJo). All gating strategies for Figs. 1–5 and Extended Data Figs. 17 are summarized in Extended Data Fig. 9. Cells were strained through a 70-μM filter, counted and diluted in PBS + 10% FBS buffer before proceeding with a general staining protocol using anti-SIINFEKL (OVA), anti-mouse or human PD-L1, intracellular staining anti-TAP1, proteasomal activity using the pQCXIN reporter system46, pan-HLA-ABC, anti-mouse or human β2M, pan-H2-Kd/Dd, CSC SORE6 reporter31 and ISRE reporter in cell lines or organoids; and anti-CD69 in OT-1 CD8+ T cells. For analysis of tumor-infiltrating lymphocytes, tumors were disaggregated as described above. Cells were strained through a 70-μM filter, counted and diluted in PBS + 10% FBS buffer before proceeding with the staining protocol using anti-CD45, anti-CD3, anti-CD8, anti-CD4, anti-CD11c, anti-CD11b and anti-F4/80. For cell cycle analysis, cells were fixed with ice-cold methanol for 2 h at 4 °C, stained with propidium iodide buffer (Supplementary Table 3) and analyzed for DNA content by flow cytometry.

Either FacsAria (BD) or Influx (BD) cell sorter equipment was used to isolate CD24lo/CD44hi CSC populations from PDOs, which were then incubated with anti-CD24 and anti-CD44. For cell sorting from cell lines, cells were washed with PBS, trypsinized, counted and diluted. Poorly and highly immunogenic cells, and stable OVA-expressing AT3 and Py8119 cells, were incubated with anti-OVA and isolated gating for OVA−/low and OVAhigh populations. For SORE6 system validation in our mouse models, Py8119 and AT3 cells SORE6-pCMV-GFP were separated into GFP+ and GFP cells31. For CSC isolation from MDA-MB-231, cells were stained using anti-CD44 and anti-CD104 and sorted by CD44hi/CD104hi-enriched CSC and CD44hi/CD104lo-nonenriched CSC populations39; ALDH was compared with ALDH+ using the Aldefluor Kit (Stemcell technologies) and gated using DEAB control to delimit the ALDH+ population, following the manufacturer’s instructions. HMLE cells were sorted by CD44hi/CD24lo-enriched CSC and CD44lo/CD24hi-nonenriched CSC populations43. For CSC isolation in cell lines 4TO7 and AT3, CD24hi/CD29hi CSC and CD24lo/CD29lo non-CSC populations and CD44hi/CD24lo CSC and CD44lo/CD24hi non-CSC populations were isolated, respectively. Cells were incubated with anti-CD29, anti-CD24 and anti-CD44.

RT–qPCR analysis

Total mRNA was purified using the RNeasy Mini Kit (Qiagen) and reverse transcribed into cDNA using the High-Capacity cDNA Reverse Transcription Kit (Life Technologies). RT–qPCR was performed using LightCycler 480 SYBR Green I Master (Merck). Data were collected using QuantStudio 12 K Flex software. mRNA expression was normalized by the expression of GAPDH (RT–qPCR primers are listed in Supplementary Table 4).

Immunoblot analysis

Cells were lysed with RIPA buffer (Supplementary Table 3) and SDS Laemmli buffer (Supplementary Table 3). Primary antibodies were rabbit anti-HA and mouse monoclonal anti-β-actin; HRP-secondary antibodies against rabbit IgG and mouse IgG were used. Data were collected using Nine Alliance Q9 software in an UVITEC Cambridge system.

RNA-seq analysis

RNA was isolated from tumors using the Qiagen RNA extraction Kit. Poly-A sequencing was selected for library preparation, and samples were sequenced using the Illumina Hi-Seq 2500 platform with 1 × 50-bp settings at the Centre of Genomic Regulation (CRG). Quality checking of raw data (fastq files) was done with FastQC software (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Estimation of ribosomal RNA in the raw data was obtained with riboPicker74. Raw reads were aligned with the STAR mapper75 to the Mus musculus genome (Gencode release M24 of the GRMm38/mm10 assembly: https://www.gencodegenes.org/mouse/release_M24.html). The raw count of reads per gene and per sample was obtained with STAR (quantMode TranscriptomeSAM GeneCounts option). The R/Bioconductor package DESeq2 v.1.30.1 was used to assess differential expression between experimental groups (Wald statistical test plus false discovery rate (FDR) correction). Genes for which the sum of raw counts across all samples was <1 were discarded. Genes are considered differentially expressed if their absolute log2 fold change was >1 and their adjusted P < 0.05. With these parameters, we generated an ICB-resistant signature taking the top 300 genes upregulated in IRT versus control cells.

RNA-seq reanalysis of fMaSCs, basal (Ba), luminal progenitor (LP) and luminal (L) cells was carried out based on a published dataset (GSE116386)42. This dataset was reanalyzed for mRNA expression of genes of interest using z-row score (z = (x – μ)/σ) where x is the raw score, μ is the population mean and σ is the population standard deviation.

Bioinformatic ATAC–seq analysis

Previously published ATAC–seq of fMaSCs, Ba, LP) and L cells (GSE116386)42 was obtained and analyzed using IGV_2.8.0 software. The various conditions were aligned, autoscaled and analyzed for the mouse MHC cluster (Chr17:33,681,276–38,548,659) region.

Clinical dataset analysis

Clinical assessment of LCOR expression was performed using RNA-seq transcriptomic data from the TONIC trial (no. EGAS0001003535)40, which included 53 metastatic TNBC preinduced patient samples with nine responders (one CR and eight partial responses) and 44 nonresponders (one stable disease and 43 progression disease); RNA-seq transcriptomic data of stage I–III TNBC pretreated samples (durvalumab in combination with olaparib) from the I-SPY2 trial (no. GSE173839)52 with nine responders and 12 nonresponders; and RNA-seq transcriptomic data from 30 on-treatment metastatic melanoma biopsies with either pembrolizumab or nivolumab (no. phs001919)76 including 13 patients with progression disease, three with stable disease, ten with partial response and two with CR. The METABRIC dataset was obtained from the open access cBioportal website (http://www.cbioportal.org/index.do), and mRNA expression data from 186 patients with TNBC were analyzed. METABRIC TNBC and TONIC trial data were also used in GSEA and GSVA.

GSEA, GO and ChEA

Transcriptomic and ChIP–seq data generated in this study were submitted and analyzed for GO and ChEA. For GO analysis, up- and downregulated genes in IRT cells and promoter/enhancer regions from LCOR ChIP–seq were interrogated for enriched pathways using BioCarta signatures. For ChEA77, upregulated genes in IRT were interrogated for enriched transcription factors.

GSEA was used for correlation of signatures in stratified datasets, performing 1,000 random permutations of their labeled phenotypes and obtaining P and q values and normalized enrichment score (NES). Genes were ranked using the formula r(g) = sign (FCg) (1 – Pg), where r(g) is the enrichment score of each gene, FCg reflects median fold change (FC) between two conditions and Pg represents the P value of the Wilcoxon rank-sum test40. The following datasets were used: METABRIC, in which samples were stratified as higher or lower than LCOR expression levels based on median expression after normality was confirmed; and the TONIC dataset40, stratified by responders and nonresponders. IRT transcriptomic data were also compared to IRT control cells. Datasets were interrogated with different signatures: the public signature of embryonic stem cell-like (BENPORATH_ES_1)29 with 379 genes; the LIU breast ER CSC signature (generated by median expression of ER breast CSCs isolated by CD24/CD44+ from HCC1954, MC1, SUM149 and SUM159 cells)34 with 164 UP genes qualifying as >2FC; the human Kyoto Encyclopedia of Genes and Genomes (KEGG: M16004) signature of APM with 88 APM-related genes; the mouse GO 0048002 signature of APM) with 76 APM-related genes; our generated ICB-resistant signature of 300 UP genes qualifying as >1FC and adjusted P > 0.05; and the previously described immunoresistant signature53 in melanoma patients.

GSVA

The GSVA package was used to establish an enrichment score representing the enrichment of signatures across patients78. GSVA was applied using the following datasets: IRT transcriptomic RNA-seq data, and METABRIC TNBC clinical data and transcriptomic data from 39 ER breast cancer cell lines from the CCLE. Datasets were interrogated with the following gene signatures: BENPORATH_ES_1 (ref. 29); NOS (NANOG, OCT4, SOX2) targets29; LIU breast ER CSC34; APM signature (KEGG: M16004); our IRS; and the melanoma immunotherapy-resistant signature (Jerby-Arnon)53.

Deconvolution and immunophenotype analysis

For estimation of immune infiltrating cells, the xCell algorithm (https://xcell.ucsf.edu/)50 was used on METABRIC TNBC data. A nonhierarchical k-means cluster analysis of four clusters based on Euclidean distance was applied, and LCOR expression was assessed in the various clusters or immune phenotypes.

Single-cell RNA-seq analysis

Available count data from single-cell RNA-seq41 were downloaded in the form of an R-readable RDS file, along with corresponding metadata. Only cohort 1 (treatment-naive patients receiving anti-PD1 treatment, n = 29) was considered. Count data were preprocessed and analyzed with the Seurat R package79. Low-quality cells were first filtered in order to have a number of features <6,000 and >200, and there should be <15% mitochondrially derived genes. Counts were then normalized with the global-scaling normalization method ‘LogNormalize’. The remaining analysis was performed on cancer cells only. Highly variable features were calculated and subsequently used to scale the data. The ‘addModuleScore’ Seurat function was used to calculate a score for each cell for LCOR expression and LIU breast ER CSC signature34. The uniform manifold approximation and projection dimensional reduction technique was run, and cells were color coded according to that score. We then calculated the percentage of cells for which the given signature was expressed more highly than expected, in the pretreatment group on one side and in the on-treatment group on the other. We next used Wilcoxon’s paired test to check whether the fraction of this signature had significantly changed following treatment.

Statistical analysis and reproducibility

No statistical method was used to predetermine sample size in animal studies; rather, this was guided by pilot and previous studies. Randomization among litters was performed before injection time, with animals of a similar age and female sex. Researchers were not blinded to allocation during experimentation and outcome since it was necessary to know the treatment groups. Tumor initiation assays in vivo were monitored by a trained technician in a blinded fashion. For in vitro experiments no statistical method was used to predetermine sample size, all samples being analyzed equally, and thus no randomization was required. Tissue staining scores were determined by three independent researchers blinded to sample information. For all in vivo and in vitro experiments, independent biological replicates are indicated in the figure legends. Each result is represented by mean ± s.e.m. For all experiments, normality and variance equality were checked by Kolmogorov–Smirnov and Bartlett’s test, respectively. Statistical significance was determined by applying either a one- or two-tailed unpaired t-test or paired Wilcoxon signed-rank test for parametric, and Welsh’s t-test for nonparametric. For multiple independent groups, either two-/one-way analysis of variance (ANOVA) or the Mann–Whitney test was applied. Tests used to check for statistical differences are specified in the figure legends. For TIC, ELDA software was used with Pearson’s χ2-test. P value and Rho index were used to assess significance for correlations. For free-survival analysis, P log-rank tests were applied using GraphPad Prism. All other statistics were calculated with the R and R-studio interface (https://www.r-project.org/). All experiments were reproduced in independent biological experiments at least three times, unless indicated.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.