Characterization of Drug-Tolerant Persister Cells in Triple-Negative Breast Cancer Identifies a Shared Persistence Program across Treatments and Patients

Abstract

Acquisition of resistance to anti-cancer therapies is a multistep process initiated by the survival of drug-tolerant persister cells. Accessibility of drug-tolerant persister cells in patients is limited, which has hindered understanding the mechanisms driving their emergence. Here, using multiple patient-derived models to isolate persister cells, we showed that these cells are transcriptionally plastic in vivo and return to a common treatment naïve-like state upon relapse, regardless of treatment. Hallmarks of the persister state in TNBC across treatment modalities included high expression of basal keratins together with activation of stress response and inflammation pathways. These hallmarks were also activated in HER2⁺ breast and lung cancer cells in response to targeted therapies. Analysis of gene regulatory networks identified AP-1, NF-κB and IRF/STAT as the key drivers of this hallmark persister state. Functionally, FOSL1, an AP-1 member, drove cells to the persister state by binding enhancers and reprogramming the transcriptome of cancer cells. On the contrary, cancer cells without FOSL1 had a decreased ability to reach the persister state. By defining hallmarks of TNBC persistence on multiple therapies, this study provides a resource to design effective combination therapeutic strategies that limit resistance.

Introduction

Acquiring a treatment-resistant cancer phenotype is a multi-step process (1). Following exposure to an anti-cancer therapy, cancer cells can initially tolerate the treatment but then will undergo massive cell death, with only a few so-called persister cancer cells surviving. Initially non-cycling, a fraction of these persister cells will ultimately start cycling again (2), thereby constituting a pool from which resistant cancer cells can emerge and actively grow under treatment (3,4). Persistence to cancer therapy is driven by non-genetic mechanisms, such as transcriptomic reprogramming (5), epigenomic remodeling (6,7), or translation deficiencies (8); correspondingly, persistence was shown to be reversible in vitro (2,6,9,10) and in vivo (11). In particular, in TNBC, persistence to therapy can be driven by transcriptional plasticity (11), metabolism reprogramming (12), or alterations in DNA methylation or histone modifications (7,13,14).

Numerous studies suggest that the switch from a treatment-naïve state to a persister state is the result of active reprogramming of a fraction of primed cancer cells, rather than the selection of a single clone. Lineage tracing studies show that the persister state is multi-clonal, and that a variety of treatment-naïve cancer cells can reach the persister state (7,11,15–17). Following treatment exposure, a fraction of predisposed cells can reach the persister state (7,18). However, as this pool of cells is constantly evolving, a potentially wider range of cancer cells can reach the persister state over a longer time (15,18). While preventing this process holds promise as a therapeutic strategy for preventing the subsequent acquisition of drug resistance, we first require a detailed understanding of the mechanisms that drive the reprogramming of treatment-naïve cancer cells to persister cancer cells.

Transition to the persister state is known to be regulated by several chromatin modifiers, and in particular histone demethylase enzymes KDM5A/B (9,19–21) and KDM6A/B (6,7). However, these enzymes are hard to target with high specificity, limiting their use as a therapeutic option in humans. Other drivers of the switch from treatment-naïve to persister include receptors (5,22), metabolic enzymes (23), and transcription factors (5,24). Targeting these drivers is currently under intense investigation as an option to enhance response to initial treatment and limit recurrence. Nonetheless, the range of patients who could benefit from these therapeutic options remains unknown. Indeed, most studies have focused on in-depth understanding of mechanisms of drug persistence in in vitro model systems or in a few patient-derived models.

Characterizing the persister state in individual patients remains challenging. In the standard of care, biopsies are often performed pre-treatment (at diagnosis) or post–neoadjuvant treatment (after surgery). On-treatment biopsies, which could catch persister cells in the making, are rarely performed. Patient-derived xenograft (PDX) models represent a powerful strategy for deriving persister cells from whole tumors: tumor cells can be isolated during treatment as required, and their response to multiple treatments can be determined in parallel. To date, however, the lengthy process of generating PDXs has only been used to profile persister cells in a few different patients tumors in parallel in solid cancers: melanoma in response to BRAF-MEK (5), TNBC in response to anthracyclin (11) or capecitabine (7), and colon cancer in response to Irinotecan (17).

Here, we analyze the drug persister state across patient-derived models and therapies, based on persister cells isolated from multiple patient models with distinct treatment regimens. We find that no matter the cancer treatment, cancer cells in individual patient models reach a common persister state that is reversible and can be reset to the initial untreated state. We further identify a recurrent persister expression program across patients, potentially driven by a set of transcription factors, namely members of AP-1, IRF/STAT, and NF-kB complexes. We showcase the key role of these transcription factors in driving persistence with a detailed study of AP-1. We show that the presence of the AP-1–activating transcription factor FOSL1 is sufficient for the emergence of persister cells in vitro and that it drives the reprogramming of the expression program of cancer cells.

Materials and Methods

Experimental approaches

PDX models

Eight xenograft models were used that had been previously generated from eight different patients with TNBC, whereby HBCx-10, HBCx-14, and HBCx-33 were generated from primitive tumors, and HBCx-218, HBCx-39, HBCx-95, HBCx-172, and HBCx-221 were generated from residual tumors post–neoadjuvant chemotherapy. All models were established at the Curie Institute with written informed consent from the patient, studies were conducted in accordance with recognized ethical guidelines (Declaration of Helsinki, CIOMS), and approved by Institut Curie’s review board. Female Swiss nude mice (RRID: MGI:5649767) were purchased from Charles River Laboratories and maintained under specific pathogen-free conditions. Mouse care and housing were in accordance with institutional guidelines and the rules of the French Ethics Committee (project authorization no. 02163.02).

As indicated, mice were treated with various methods (Table S3): capecitabine (Xeloda; Roche Laboratories) at a dose of 540 mg/kg, orally 5 d/week, for 25 to 51 days (i.e. pCR); adriamycine (Accord) at 2 mg/kg, intraperitoneal (i.p.) 1×/21 days; cyclophosphamide (Baxter) at 100 mg/kg, i.p. 1×/21 days; cisplatin (Mylan) at 6 mg/kg, i.p. 1×/21 days; or carboplatine (Accord) at 90 mg/kg, i.p. 1x/21 days. Relative tumor volumes (mm³) were measured as described previously. Before downstream analysis, cell suspensions were prepared as previously described (7).

Cell lines, culture conditions, and drug treatments

All cell lines were obtained from ATCC. MDA-MB-468 cells (ATCC, ref. HTB-132, RRID: CVCL_0419) were cultured in DMEM (Gibco-BRL, ref. 11966025), supplemented with 10% heat-inactivated fetal calf serum (Gibco-BRL, ref. 10270-106). The HCC38 (ATCC, ref. CRL-2314, RRID: CVCL_1267) and BT-20 (ATCC, ref. HTB-19, RRID: CVCL_0178) cell lines were cultured in RPMI 1640 (Gibco-BRL, ref. 11875085), supplemented with 10% heat-inactivated fetal calf serum. All cell lines were cultured in a humidified 5% CO₂ atmosphere at 37°C and were tested for mycoplasma infection monthly (Lonza, ref. 4043248). Cells were treated with 5 μM of 5-FU (Sigma, ref. F6627) for 7 to 21 days, as indicated or with 5μM of Irinotecan (Sigma, ref. I1406) for 7 days.

Viability assays

1x10^4 cells were plated in 96-well plates in triplicate and 5-FU or DMSO were added. Cell viability was measured after 7 days (GOF FOSL1) or 21 days (KO FOSL1) using the Cell proliferation Kit II–XTT (Sigma, Ref: 11465015001) and the TESCAN Multi Detection System according to the manufacturer’s instructions.

Cloning

FOSL1 gRNAs were selected from the Weissman library (25). Forward and reverse oligonucleotides were synthesized by Integrated DNA Technologies - IDT and cloned in pJR85-gRNA-BFP (Addgene, ref. 140095) using the BstXI and BlpI restriction enzymes.

FOSL1 cDNA was mutated by adding a HA tag after the ATG site and cloned by Genscript in pHIV-SFFV-MCS-IRES-GFP using BamHI and XhoI restriction enzymes.

Lentivirus packaging and cell transduction

Cas9WT-P2A-mCherry (Addgene, ref. 99154) and dCas9-VPR-P2A-mCherry (Addgene, ref. 154193) were used. gRNA lentivirus particles were produced by transfecting the plasmids pJR85-gRNA-BFP (Addgene, ref. 140095), psPAX2 (Addgene, ref. 12260), and pMD2.G (Addgene, ref. 12259) into HEK293T cells (ATCC, CRL-321, RRID: CVCL_0063) with PEIpro (Polyplus, ref. 101000017) and NaCl, following manufacturer’s instructions. Lentiviruses were titrated after infecting HEK293T cells with cascade dilutions and assessing the percentage of BFP fluorescence on a cytometer. MDA-MB-468 cells from ATCC at passage 11 were infected with lentivirus produced at a low multiplicity of infection (MOI 0.1). Ten days after transduction, cells were sorted for mCherry expression to select cells with Cas9WT or dCas9-VPR - mCherry expression. These cell lines were then transduced with FOSL1 gRNA lentivirus at MOI 0.3, and transduced BFP+ cells were isolated three weeks after transduction.

For tagged cDNA overexpression, lentivirus particles were produced by transiently transfecting the pHIV-SFFV-MCS-IRES-GFP or the HA-FOSL1-pHIV-SFFV-MCS-IRES-GFP into HEK293T cells using the calcium-phosphate transfection method. Lentiviral constructs were co-transfected with pMD2.G (Addgene, ref. 12259) and psPAX2 (Addgene, ref. 12260) (both kindly provided by Didier Trono). The culture media was collected 72 h after transfection of HEK293T cells. Lentiviral particles were concentrated by ultracentrifugation. Lentiviruses were titrated after infecting HEK293T cells with cascade dilutions and assessing the percentage of GFP fluorescence on a cytometer. MDA-MB-468 cells from ATCC at passage 2 were infected with lentivirus produced at a low multiplicity of infection (MOI 0.5). Ten days after transduction, cells were sorted for GFP expression to select cells with GFP expression or HA-FOSL1-GFP expression.

In all cell lines, expression of FOSL1 was validated at both the protein and mRNA level.

Western blot

Cell lysates were prepared as previously described (7). We used the following antibodies: , FOSL1 (Dilution: 1:500, Santa Cruz, ref. Sc-28310, RRID: AB_627632), JUN (dilution 1:100; Abcam, ref. ab31419, RRID: AB_731605 ), FOS (Dilution: 1:500, Cell Signaling, ref. 2250, RRID: AB_2247211), FOSB (Dilution: 1:500, Cell Signaling, ref. 2251, RRID: AB_2106903), JUNB (Dilution : 1:500, Santa Cruz, ref sc-8051, RRID: AB_2130023), JUND (Dilution: 1:500, Santa Cruz, ref. Sc-271938, RRID: AB_10650101), or Histone H3 (Dilution: 1:1000, Cell Signaling, ref. 4620, RRID: AB_10949503).

RNA extraction & RNA-seq

After mechanical dissociation with Mixer Mill for 2 min at 30 Hz in 500 µL of Qiazol, RNA was extracted following Qiazol/chloroform extraction protocol (Qiagen, ref. 139046, MaXtract High density) and RNeasy Mini column (Qiagen, ref. 74106), following manufacturer’s instructions. RNA quality was checked on the Agilent TapeStation using RNA reagents. RNA extracts were sequenced on a NovaSeq 6000 (Illumina) in PE100 mode.

CITE-seq

Cells (1 × 10⁶) were blocked for 10 min at 4°C with 5 µL of human Trustain FcX (Biolegend, ref. 422301) and stained for 30 min at 4°C with 1 µg of TotalSeq A antibodies (Biolegend, ref. 394601/RRID: AB_2750015, ref. 394603/RRID: AB_2750016, ref. 394605/RRID: AB_2750017, ref. 394607/RRID: AB_2750018, ref. 394609/RRID: AB_2750019), using 1 barcoded antibody per sample for the MDA-MB-468 GOF VPR control, VPR FOSL1 gRNA2, VPR FOSL1 gRNA3, the HA control, and HA FOSL1. After washing, cells were filtered and counted. A pool of 5 stained samples was prepared by mixing 25,000 cells per sample (finale concentration of the pool 2,000 cells/µL). Approximately 16,000 cells were then loaded on a Chromium Single Cell Controller Instrument (Chromium Single Cell 3ʹv3, 10X Genomics, ref. PN-1000075) according to the manufacturer’s instructions. Samples and libraries were prepared according to the manufacturer's instructions. Libraries were sequenced on a NovaSeq 6000 (Illumina) in PE 28-8-91 with a coverage of 50,000 reads/cell.

scRNA-seq

For each single-cell suspension or PDX dissociated cell sample (7), approximately 3,000 cells were loaded on a Chromium Single Cell Controller Instrument according to the manufacturer’s instructions. Samples and libraries were prepared according to the manufacturer's instructions. Libraries were sequenced on a NovaSeq 6000 (Illumina) in PE 28-8-91 with a coverage of 50,000 reads/cell.

CUT&Tag

The CUT&Tag-IT™ Assay Kit (Active Motif, ref. 53160) was used with 1:50 Anti-Rabbit antibodies for FOSL1 (Cell Signaling, ref. #5281, RRID: AB_10557418), HA (Epicypher, ref. 13-2010, RRID: AB_3094663) and H3K27ac (Cell Signaling, ref. #8173, RRID: AB_3094663) following manufacturer instructions. The starting cell number was 500,000 cells for control cells and GOF cells, and down to 100,000 for persister cells after 35 days of 5 uM 5-FU treatment. The same digitonin concentration (0.05%) was maintained in all buffers, completing the kit with additional digitonin (Cell Signaling, ref. 16359). Sequencing libraries were prepared using Nextera™- Compatible Multiplex Primers (Active Motif, ref. 53155) and profiles were checked on the Agilent TapeStation (Laptop (RRID:SCR_019547) using High-sensitivity D1000 reagents. CUT&Tag libraries were sequenced on a NovaSeq 6000 (Illumina) in PE100 mode.

Immunofluorescence

After media aspiration, cells in 24-well plates were washed with PBS and fixed with 3.7% formaldehyde in PBS for 15 min at room temperature. Cells were then washed with PBS three times, for 3 min each, and then permeabilized for 5 min at room temperature with 0.2% Triton X-100 in PBS. Samples were then incubated for 1 h at room temperature with a blocking buffer (1% heat-inactivated fetal calf serum and 1% BSA in PBS) and then for 2 h at room temperature with 50 µL primary antibody anti-FOSL1 (dilution 1:100; Santa Cruz, ref. #sc-376148, RRID: AB_11012022) in blocking buffer. After incubation, cells were washed three times with PBS with 0.2% Tween and then incubated for 1 h at room temperature in 50 µL secondary antibody solution consisting of the appropriate species-specific Alexa Fluor-conjugated antibodies (ThermoFisher, refs. A-21127, A-11008), diluted 1:1000 in blocking buffer. Coverslips were then washed three times with PBS with 0.2% Tween and mounted on slides with 10 µL ProLong Gold Antifade Reagent with DAPI (ThermoFisher, ref. P36971) at room temperature for 24 h. Confocal laser scanning microscopy (Carl Zeiss, LSM780) was used to scan GOF samples. Image processing was performed using Fiji Software, version 1.0 (RRID: SCR_002285).

Sample preparation for proteomics and mass spectrometry analysis

Lysis and wash buffers were prepared based on Hanson et al. 2016 protocol (26). FOSL1 (Santa Cruz, ref. #sc28310 (C12) - RRID: AB_627632) and IgG (Sigma, ref. 12-371, RRID: AB_145840) immunoprecipitations on 1mg of lysates using protein A/G magnetic beads (Pierce, ref. 88802) were performed once in each cell condition after 7 days of DMSO or 5-FU. Beads were washed twice with 100 μL of 100 mM NH4HCO3 and resuspended in 100 μL of 100 mM NH4HCO3 before digestion by adding 0.2 μg of trypsin-LysC (Promega, ref. V5071) for 1 h at 37 °C. Samples were then loaded into custom-made C18 StageTips packed by stacking three AttractSPE disk (Affinisep, ref. SPE-Disks-Bio-C18-100.47.20) into a 200 µL micropipette tip for desalting. Peptides were eluted using a ratio of 40:60 CH3CN:H2O + 0.1% formic acid and vacuum concentrated to dryness with a SpeedVac device. Peptides were reconstituted in 10µl of injection buffer in 0.3% trifluoroacetic acid (TFA) before liquid chromatography-tandem mass spectrometry (LC-MS/MS) analysis.

LC-MS/MS analysis

Liquid chromatography (LC) was performed with a Vanquish Neo LC system (Thermo Scientific) coupled to an Orbitrap Astral mass spectrometer (MS), interfaced by a Nanospray Flex ion source (Thermo Scientific). Peptides were injected onto a C18 column (Thermo Scientific, ref. 75 µm inner diameter x 50 cm double nanoViper PepMap Neo, 2 μm, 100 Å) regulated at a temperature of 50°C, and separated with a linear gradient from 100% buffer A (100% H2O + 0,1% formic acid) to 28% buffer B (100% CH3CN + 0,1% formic acid) at a flow rate of 300 nL/min over 104 min. Peptides were analyzed in the MS applying a 2200V spray voltage, funnel RF level at 40% and a heated capillary temperature set to 285°C. MS full scans were recorded on the Orbitrap mass analyzer in centroid mode for ranges 380-980 m/z with a resolution of 240,000 at m/z 200, a normalized AGC target set at 500% and a maximum injection time of 5 ms. The MS2 acquisitions in the Astral analyzer were performed on a 150-2000 m/z range after fragmentation using HCD with 25% normalized collision energy and a normalized AGC target of 500%. A window width of 2 Da without overlap, 299 scan events and a maximum injection time of 3 ms were chosen for DIA. The loop control was set to 2.1 sec.

Colony forming assay

TNBC cells (MDA-MB-468 and BT-20) were plated in 6 multi-well plates at a density of 300,000 cells per well and treated with the indicated drugs (5-FU or Irinotecan) for 7 or 21 days. Cultures were incubated in humidified 37 °C incubators with an atmosphere of 5% CO2 in air, cell density was evaluated after a staining with 0.5% Crystal Violet (Sigma, ref: C3886). Quantifications were performed with the Fiji software on 8 bit inverted images using « Measure » function.

Computational approaches

All statistical analyses were performed in R (v4.4) using custom R scripts.

Proteomics data processing

For identification, the data were searched against the Homo sapiens (UP000005640) Uniprot database (RRID: SCR_002380) using Pulsar search engine through Spectronaut v19 (Biognosys) by directDIA+ analysis using default search settings. Enzyme specificity was set to trypsin and a maximum of two missed cleavage sites was allowed. Carbamidomethyl cysteine, N-terminal acetylation and oxidation of methionine were set as variable modifications. The resulting files were further processed using myProMS v3.10. https://github.com/bioinfo-pf-curie/myproms (27). For protein quantification, ion XICs from proteotypic peptides shared between compared conditions (TopN matching) were used, with missed cleavages and carbamidomethylation allowed. Median and scale normalization at peptide level was applied on the total signal to correct the XICs for each biological replicate (N = 5). To evaluate the statistical significance of the change in protein abundance, a linear model (adjusted on peptides and biological replicates) was performed, and a two sided T-test was apply on the fold change estimated by the model. The p-values were then adjusted using the Benjamini–Hochberg FDR procedure. In order to eliminate non-specific isolated proteins, only proteins were selected with at least three distinct peptides in three replicates out of five replicates, two-fold enrichment and an adjusted p-value ≤ 0.05 and for the unique proteins three distinct peptides. Proteins selected with these criteria were further used for quantitative data analysis.

Single-cell (sc) RNA-seq analysis

The scRNA-seq sequencing files were preprocessed using the cellRanger pipeline (RRID: SCR_017344). For PDX samples, files were aligned to the hg38 and mm10 genomes, and only cells with a majority of human reads (> 75%) were retained for the analysis. For the MDA-MB-468, HCC38, and BT-20 human cell lines, sequences were aligned to the hg38 genome. For all datasets, cells were kept within a 100,00 read coverage limit. For PDX samples, cells had a coverage between 2,000 and 9,000 genes, with mitochondrial reads below 25%. For MDA-MB-468, HCC38, and BT-20 cells, cells had a coverage between 3,000 genes and 8,000 genes, with mitochondrial reads below 15%. Normalization (SCTransform method), dimensionality reduction, and Louvain clustering were performed using Seurat (v5.1.0, RRID: SCR_016341), keeping the first 30 principal components (PC). The cell cycle was scored for each cell using Seurat. For the PDXs, BT-20, HCC38, and MDA-MB-468 models, persister cells were compared to cells from the untreated model. For each model, differentially expressed genes were computed using the “FindMarkers” Seurat function with default parameters. UMAP plots of gene expression were obtained using Nebulosa (v1.14.0) to recover the signal from dropped-out features using kernel density estimation.

Bulk RNA-seq analysis and microarray analysis

The bulk RNA-seq samples were preprocessed using the Nextflow RNA-seq analysis pipeline from Institut Curie (10.5281/zenodo.7443721). PDX samples were aligned on hg38 and mm10 genomes and only reads aligning to the hg38 genome were retained for the analysis. Normalization was made using the TMM method from the edgeR package (v4.2.0, RRID: SCR_012802) (28). U133 Affymetrix microarrays were preprocessed with the GCRMA package (29).

Integrated analysis of microarray, bulk RNA-seq, and scRNA-seq samples

Raw count matrices of scRNA-seq samples were first merged, keeping all features (and putting 0 for missing values). A pseudo-bulk count matrix was then created by aggregating all cell counts coming from the same tumor sample with Presto (v1.0.0). After this step, all raw count matrices coming from microarray, bulk RNA-seq, and scRNA-seq samples were merged together, whereby a set of features common to all samples was maintained. LogNormalisation was performed using the TMM method from the edgeR package (v4.2.0). To mitigate the impact of technical and sample-specific confounding effects, we implemented a batch correction strategy using the removeBatchEffect function from the limma package (30) (v3.60.3, RRID: SCR_010943), following normalization with the TMM (Trimmed Mean of M-values) method from edgeR (v4.2.0). The correction was performed on log-transformed counts per million (logCPM) obtained from the normalized data. We specified the batch variable as the sample origin (i.e., the specific PDX model), which was sufficient to also include the technological effect, as no PDX model was sequenced on multiple modalities. The design matrix included the biological group of each sample (Untreated, Persister, or Relapse). Internally, removeBatchEffect performs a linear modeling step where the expression of each gene is regressed against the batch variable, and the residuals—adjusted for the design matrix—are retained. This effectively removes systematic differences due to batch while preserving structured biological signal. This strategy allowed us to correct for strong cross-platform differences while maintaining the biological variation of interest.

Principal component analysis (PCA) was applied on the 500 most variable features within this matrix. For the differential analysis, the voom method from limma was used on the raw matrix count, taking into account in the fitted linear model the group of the sample (persister, chemonaïve, or relapse) and the model (PDX of origin). The persister and chemonaïve groups were compared using the makeContrasts function.

Regulatory network inference

Regulon activity was inferred using DecoupleR (v2.10.0) (31). The network of transcription factors (TFs) and their transcriptional targets were retrieved from the CollecTRI database using OmnipathR (v3.11.10), with the option “split_complexes = False”. With this network, a consensus score of TF enrichment scores was obtained by combining 3 methods (ulm, wsum and mlm) on the T-value from the differential expression analysis obtained with limma. T-values were either directly retrieved from the differential expression analysis when limma was used, or calculated by -log10(p.adj)*log2FC when Seurat was used. The TFs with the top 10 scores were selected. Regulon activity per sample per TF/complex was computed using the normalized weighted sum method on the corrected logcount matrix obtained with limma.

CUT&Tag analysis

Raw sequencing files were processed using an in-house pipeline (https://github.com/vallotlab/bulk_Epigenomics). Raw reads were mapped in paired end mode using BowTie2 with options "-k 1 -m 1" on the human genome (hg38, RRID: SCR_016368). Peak calling by condition and antibody used (histone modification or transcription factor binding) was performed on bam files using macs3 (32), with the following parameters: "--max-gap 500 -f BAM -q 0.000001 -- nolambda". PCA and differential analysis were performed from count matrices on a merged peak set from the different conditions (untreated, 5-FU, or FOSL1 overexpression experiments) and antibodies (anti-H3K27ac, anti-FOSL1, or anti-HA). Peaks covered with less than 5 reads across all samples were filtered out. Normalization and differential analyses were performed with DESeq2 (v1.44.0, RRID: SCR_015687) (33). PCA were calculated on the log-normalized count matrices selected for the 500 most variable peaks. To compare H3K27ac and FOSL1 enrichments, a common set of consensus peaks was assembled by combining peaks from all conditions (chemonaïve, 5-FU, or GOF cells) for both antibodies. Profile plots for CUT&Tag were made using Plotgardener (v1.10.2). For each snapshot, consensus peaks, associated log2 fold-change (log2FC) and adj. P computed with DESeq2s are indicated on the plots.

Motif enrichment analysis

Motif enrichment analysis was performed on differentially enriched peaks for H3K27ac, comparing 5-FU–treated and wild-type MDA-MB-468 cells. The Hocomoco database (v12) and Memes (v1.12.0) were used to perform motif enrichment on the top 1000 differentially enriched peaks (ordered by decreasing T-values). A set of 5000 random peaks with no enrichment upon 5-FU treatment was used as background.

Results

Tumor cells from individual patients consistently reach a common drug tolerant persister state that resets upon relapse

To derive persister cells from tumors, we first engrafted patient tumors on nude mice and subsequently treated them with standard of care anti-cancer therapies (Table S2 and S3 for model and treatment regimens). We retrieved persister cells by pooling the fat pads from mice once they showed pathological responses, as previously described (Fig. 1A)(7). We leveraged this method to derive persister cells from eight TNBC PDX tumors (Fig. S1A–C), established either from treatment-naïve or residual tumors after patients had been exposed to several cycles of chemotherapy. We then treated mice with anthracyclines (AC) or platinum drugs, which are typically used in the neoadjuvant setting, or with capecitabine, the oral prodrug of 5-fluorouracil (5-FU), used in the adjuvant setting (Fig. S1A). For two PDXs (HBCx-33 and HBCx-14), we treated the mice independently with up to three chemotherapies (Fig. 1B and Fig. S1B). From a total of 127 mice, we performed transcriptomic analyses of n = 23 untreated tumors, n = 42 persister populations, and n = 22 relapsed tumors, using a combination of microarray, bulk RNA-seq, and single-cell (sc) RNA-seq to characterize the transcriptional program activated upon therapeutic stress in vivo (Table S1).

Fig. 1. Identification of a recurrent and reversible persister state across TNBC patient models.

A. Scheme of the PDX generation and persister isolation.

B. Table summarizing the PDX models and chemotherapy treatments. Total number of mice analyzed per model is indicated.

C. PCA of transcriptomic data for HBCx-10 (left) and HBCx-14 (right), with samples colored by collection timepoint or cluster ID.

D. PCA of transcriptomic data across all PDX models, with samples colored by collection time point, PDX model/Patient ID, treatment, or cluster ID.

E. Volcano plot showing differentially expressed genes between persister and chemonaïve samples in all PDX models.

F. Hallmark pathways upregulated (red) or downregulated (blue) in persister samples compared to chemonaïve samples in all PDX models.

G. Boxplot showing the expression of selected genes in chemonaïve, persister and relapsed samples in all PDX models. Log2FC between persister and chemonaïve samples and adjusted p-values are indicated.

H. Barplot showing the proportion of cells in each cell cycle phase in scRNA-seq data from chemonaïve and persister samples. P-values from Fisher’s exact test are indicated.

I. UMAP visualization of scRNA-seq data from HBCx-218 (left) and HBCx-39 (right) PDX models. Cells are colored by collection time point or persister signature score.

We first analyzed the variability of the persister state in individual patient models, to determine whether a patient-specific tumor always gives rise to the same pool of persister cells. For this, we tested multiple mice (n = 16) engrafted with the same tumor of origin (HBCx-10) (Fig. 1C) after treatment with an AC chemotherapy, which is a combination of doxorubicin (a topoisomerase II inhibitor) and cyclophosphamide (an alkylating agent). After performing an unsupervised analysis of transcriptomes from all mice, principal component analysis (PCA), and hierarchical clustering, we observed a striking similarity between the eight persister populations derived from different mice: they all shared common transcriptional features that account for most of the variations observed in the collected transcriptomes (PC1 corresponds to 79.8% variance within all mice), and they all clustered together away from untreated tumors (Fig. 1C). Critically, the transcriptome of these persister populations always reset back to that of initial treatment-naïve cells upon relapse, as all relapsed tumors clustered with untreated tumors (Fig. 1C).

We further tested the observation that the persister populations are similar and reset upon relapse for another PDX (HBCx-33). Here, we observed that 11/14 persister populations clustered together and away from the group of untreated tumors (PC1 corresponds to 78.5% variance). Likewise, they reset to initial state upon relapse (n = 14 relapsed) (Fig. S1D). Mice were treated with two different chemotherapies: the alkylating agent cisplatin, and the anti-metabolite, capecitabine. Notably, transcriptomes from persister populations generated with either treatment were hardly distinguishable.

To further investigate whether the persister state is treatment independent, we used a third model treated with three different chemotherapies: cisplatin, carboplatin (also an alkylating agent), and an AC (HBCx-14) (Fig. 1C). We found that all persister cell populations clustered together irrespective of the treatment. Altogether, we demonstrated in vivo that tumor cells recurrently reach a persister state that is common across therapies and resets to the same initial state upon relapse.

Identification of a recurrent drug tolerant persister state across TNBC patient models and treatments

Next, we aimed to identify potential recurrent features of the persister state across TNBC tumors and treatment regimens in TNBC. When combining datasets from all persister and control tumors (n = 87 datasets), we observed that persister states were first and foremost patient-specific (Fig. S1E). Thus, to search for recurrent features of the persister state across patient models, we next integrated datasets taking into account the patient and technology of origin(30) (Methods). We showed that the main remaining source of variation among the 87 transcriptomes is the treatment status of the tumor, whether samples were taken from persister, untreated, or relapsed cells (PC1 corresponds to 61.4% variance; Fig. 1D). Persister cell populations clustered together (39/42) away from untreated or relapsed cells, revealing the existence of a common and transient persister state across the different tumors and chemotherapy treatments (Fig. 1D and Fig. S1F). This recurrent consensus persister state is characterized by the activation of 836 genes (Fig. 1E-G and Table S4) with a log₂ expression fold-change (herein, log2FC) over 0.5 and adjusted P-value (herein, adj. P) < 1e-4, which we defined as the consensus persister signature. Compared to untreated cells, persister cells displayed higher expression levels of basal genes (e.g KRT14, KRT6A) and an enrichment for basal gene signatures (Fig. S1G). N-cadherin (CDH2), a mesenchymal gene, was also overexpressed suggesting the potential co-acquisition of mesenchymal features (Fig. 1G and Table S4). These observations confirmed our initial findings on n = 3 patient models (7), and that basal keratins and CDH2 are hallmark markers of the persister state in TNBC independently of the chemotherapy regimen.

Persister cell populations all activated the interferon and inflammatory responses (IRF/STAT), and stress response (JUNB), while blocking the cell cycle (Fig. 1E–H). In addition, across patient models, persister cells showed a common drastic switch in cholesterol metabolism, with a combined shutdown of lipogenesis (IDI1) and cholesterol import (LDLR) and a high expression of cholesterol efflux pump (ABCA1) (Fig. 1G). A major fraction of target genes of the SREBP2 transcription factor, which is responsible for the intracellular production of cholesterol, were repressed in persister cell populations across patient models; these repressed genes included IDI1, LDLR, SQLE, FDPS, and HMGCR (Fig. 1G and Table S4). Altogether, we show that in TNBC, across treatment regimens, persister tumors share a common basal and mesenchymal identity, with stress-associated features and an alternative cholesterol metabolism.

Focusing on two patient-derived models (PDX HBCx-218 and -39), we next quantified the intra-tumor heterogeneity of the identified consensus persister state using scRNA-seq. For both PDX models, we detected activation of our newly defined consensus persister gene signature in persister cells, whereby it was high in most persister cells for HBCx-39 and high in only a fraction of the persister cells in HBCx-218 (Fig. 1I). This suggests that persister cells are heterogeneous, and only a fraction of them could be activating the consensus persister gene signature that we identified. Further, we detected low activation of the persister gene signature in untreated tumors, suggesting that at least part of the genes are already expressed prior to treatment.

Hallmark persister signature can be activated across cancer types and is detected in on-treatment patient biopsies

We next investigated whether our hallmark TNBC drug-tolerant persister state shared features with persister states of other cancer types and treatments. We investigated another breast cancer subtype as well as lung persister cancer cells, that are each treated with a targeted therapy - rather than cytotoxic chemotherapies as performed in the current study. We found specific activation of our hallmark persister signatures in HER2+ breast persister cells, in multiple HER2+ cell line models (34,35) in response to anti-HER2 therapy (Fig. 2A-B). In BT474 cells (Fig. 2B), the persister expression program is gradually activated upon treatment, with maximum expression in fully resistant cells. Likewise, in lung cancer (2) (Fig. 2C), persister cells gradually activate the persister signature, with a maximum at day 14 upon anti-EGFR therapy. These results show that the persister hallmarks that we have identified across TNBC patient models upon cytotoxic treatments are also found in cells responding to targeted anti-cancer therapies, further stressing the hallmark nature of the persister signature identified in TNBC.

Fig. 2. Activation and regulation of hallmark persister signature in cancer cells.

A. Boxplot of Persister signature score on transcriptomic data from HER2+ breast cancer cell lines, for chemonaïve and persister samples. P-value from one-sided Wilcoxon rank-sum test is indicated.

B. Boxplot of Persister signature score on transcriptomic data from HER2+ breast cancer cell line (BT474), for chemonaïve, persister, cycling persister and resistant samples. P-values from one-sided Wilcoxon rank-sum tests are indicated.

C. UMAP of scRNA-seq data from PC9 lung cancer cells. Cells are colored by treatment time point and persister signature score.

D. Boxplot of persister signature scores on transcriptomic data from TNBC patient samples collected at three time points (T1: pre-treatment, T2: 3 weeks after the first cycle of NAC, T3: at surgery for patients who did not achieve pCR). P-values from one-sided Wilcoxon rank-sum tests are indicated.

E. Boxplot of the log10 normalized expression level of KRT14 in the PanCancer Breast Cancer Atlas, stratified by tumor subtype. Comparisons were made according to the basal-like subtype using the Wilcoxon rank-sum test. Red horizontal line represents the median expression value of the normal-like breast cancer subtype.

F. Kaplan-Meier disease-free survival curve for basal-like tumors only from the PanCancer Breast Cancer Atlas, according to KRT14 expression binarized values, using the median value of the normal-like subtype highlighted in (E). A log-rank test was used to compare the two groups; the P-value is indicated on the plot.

G. Inference of regulon activity on the common persister gene program. The top 10 scoring transcription factors are highlighted in red.

H. Heatmap of regulon activity inference for the indicated predicted transcription factors across all PDX models.

I. Boxplots showing the gene expression of the indicated transcription factors in chemonaïve, persister, and relapsed samples in all PDX models. Log2FC between persister and chemonaïve samples and adjusted p-values are indicated.

(ns: not significant, *: p-value<0.05, **: p-value<0.01, ****:p-value<0.0001)

We next investigated the potential activation of our hallmark persister signature in patients with TNBC. The study by Park et al. (36) (GSE123845) monitored the transcriptome of patients during treatment - in chemonaïve biopsies, on-treatment biopsies and in residual tumors at surgery post treatment. The hallmark persister signature is transiently activated in tumors on-treatment returning to initial level at surgery (Fig. 2D), regardless of the response of the patient to the treatment (Fig. S2A). This suggests that persister-like cells can be found in patients during treatment exposure, and that at least part of the persister genes are expressed in response to therapeutic stress, not necessarily driving a drug tolerance process. When inspecting the expression levels of individual persister genes, we found that 7 genes (Fig. S2A, Table S5) were significantly more expressed in non-responders vs responders prior to treatment or on-treatment, among which the basal marker gene KRT14. Two persister genes coding for calcium binding proteins (CALM2 & S100P) were upregulated at both timepoints in non-responders. Altogether these results suggest that patients activate a persister program upon several weeks of treatment (~3 weeks), and that a few key persister genes could be candidate biomarkers of response to therapy.

KRT14 is one of the persister genes that is associated with response to treatment prior to treatment. Of note, it is used clinically as a prognostic marker, which opens the perspective of using it in routine to predict response to therapy. KRT14 is used as an IHC marker to characterize breast cancers (in conjunction with KRT5/6, estrogen (ER), progesterone (PR), human epidermal growth factor receptor-2 (HER2) and Ki-67). KRT14 and/or KRT5/6 define the basal-like phenotype within TNBC subtypes that do not express ER, PR, or HER2 and are associated with a poor outcome (37). Interrogating the PanCancer Breast cancer atlas comprising chemonaïve tumors (38), we observed that basal-like tumors express KRT14 levels close to that of the normal adjacent tissue, which contains basal cells. In addition, about a fourth of these tumors (24%) express KRT14 at levels that are higher than that of the adjacent normal tissue (Fig. 2E). These tumors are associated with lower disease-free survival rates than other basal-like tumors (P = 3.10e-2) (Fig. 2F). These KRT14-high tumors could contain a subset of KRT14-positive persister-like or pre-persister cells prior to treatment.

Transient activation of AP-1, NF-kB, and STAT/IRF transcription factors in persister state

We next inferred regulon activity to nominate transcription factors that could drive the persister expression signature (Fig. 2G). To do so, we used the decoupleR framework (31) that extracts regulon activities from transcriptomic datasets. We show that the top 10 regulons explaining the consensus persister expression signature are associated with NF-kB (NFKB1, RELA), IRF/STAT (STAT1/IRF1), and AP-1 (JUN) (Fig. 2G). In addition, a series of individual transcription factors (ETS1, CNNTB1, and SPI1) were also predicted as potential regulators of the consensus expression signature. The predicted regulons are transiently activated in persister cells and are then shut down in cells at relapse (Fig. 2H). These predictions match the pathways that we have found to be characteristic of the consensus persister state. Transcription factors of the IRF and STAT families are indeed orchestrators of the interferon response (39), while the AP-1 and NF-kB complexes are major actors of stress-response (40,41). Differences in regulon activity could result from changes in the expression or activity of transcription factors. Analyzing our transcriptome datasets, we found that only a fraction of these transcription factors are overexpressed in persister cells (Fig. 2I, Fig. S2B-D, and Table S4). Among the AP-1 family, JUNB is the most recurrently overexpressed member across patient models and treatments (log2FC = 0.80, adj. P = 1.41e-14), while RELB is the most overexpressed genes within the NF-kB family (log2FC = 0.50, adj. P = 2.54e-5), and IRF7 (log2FC =1.49; adj. P = 2.12e-9) and STAT1 (log2 FC = 0.84; adj. P = 2.10e-6), within the IRF/STAT families. These changes in expression could explain in part the regulon activities observed in the persister state.

Using the TCGA PanCancer atlas, we analyzed the patterns of expression of these transcription factors prior to treatment. As a follow-up of our previous observation on KRT14-high chemonaïve tumors, we observed that FOSL1 and (to a lesser extent) JUNB are the two transcription factors that most significantly correlated with KRT14 prior to treatment (correlation score= 0.51 and 0.39; P = 3.20e-12 and 1.70e-7, respectively; Fig. S2E). This observation suggests that KRT14/AP-1–high tumors exist prior to treatment.

Remodeling of the H3K27ac epigenome upon drug insult

To study the mode of action of the potential TF drivers we identified, we first tested in vitro, whether TNBC cell lines could recapitulate at least part of the consensus persister state and the transcriptional regulons we observed in PDX samples (Fig. 3A). We performed scRNA-seq on three different TNBC cell lines (MDA-MB-468, HCC38, and BT-20) treated with chemotherapy (5-FU) and identified a set of genes common to all persister cell populations (Fig. S3A and B). To characterize the regulatory landscape of the persister state, we used inference of regulon activity and observed that part of the top 10 regulons potentially driving a persister program in vitro belong to NF-kB (RELA) and AP-1 (JUN) regulons (Fig. 3B). This regulatory network recapitulates in part what we observed in persister cells obtained from our TNBC PDX models.

Fig. 3. Epigenomic remodeling upon treatment is dedicated to AP-1 sites.

A. Schematic representation of TNBC cell line evolution under 5-FU treatment. Time points and associated phenotypes observed by microscopy are represented.

B. Inference of regulon activity for the persister program in 3 TNBC cell lines, MDA-MB-468, BT-20 and HCC38. Top 10 scoring transcription factors are highlighted in red.

C. Volcano plot of differentially enriched H3K27ac peaks in persister samples (35 days of 5-FU treatment) compared to chemonaïve samples in MDA-MB-468 cell line. The red horizontal line marks adj. P < 1e-10.

D. Enrichment of transcription factor binding motifs within H3K27ac peaks gained in persister samples (35 days of 5-FU treatment) versus chemonaïve samples in MDA-MB-468 cell line. The identified consensus motif in gained H3K27ac peaks and the genomic binding motif of AP-1 are shown.

E. Representative images of FOSL1 and JUN protein expression (canonical AP-1 members) in the indicated cell lines treated with 5-FU for 21 days. Histone H3 serves as a loading control (n = 3 biological replicates).

F. Volcano plot of differentially enriched proteins from FOSL1 immunoprecipitation after 7 days of 5FU versus DMSO treatment (MDA-MB-468 cells, n=5 replicates, n=84 proteins). The vertical dashed line marks the log2FC for FOSL1 protein.

G. PCA of H3K27ac distribution and FOSL1 genomic binding profiles in MDA-MB-468 chemonaïve and persister samples (35 days of 5-FU treatment).

H. Scatterplot comparing Log2FC values for FOSL1 binding vs H3K27ac enrichment upon 5-FU treatment in MDA-MB-468 cells treated for 35 days. Pearson’s correlation scores and associated P-value are indicated.

I. Genomic tracks of H3K27ac and FOSL1 profiles over the LAMB3 and CAV1 loci in MDA-MB-468 chemonaïve and persister cells (35 days of 5-FU treatment). Log2FC enrichment values and adjusted p-values are indicated.

To further understand the epigenomic mechanisms driving transcriptional changes, we used H3K27ac profiling to map active enhancers in chemonaïve or persister cells from the MDA-MB-468 cell line (Fig. 3C and Fig. S3C-E). The persister state was characterized by a rewiring of the H3K27ac landscape, with the most significant changes corresponding to localized gains of H3K27ac (n = 106 enriched peaks vs n = 4 depleted peaks, with adj. P < 1e-10). These changes in chromatin are associated to the activation of expression of the nearest genes (Fig. S3D) in persister cells. To understand which transcription factors could be associated with the chromatin remodeling at these peaks, we performed motif enrichment on peaks enriched for H3K27ac after treatment (Fig. 3D). We found that the AP-1 related motifs were by far the most significantly enriched in these persister cells. Thus, we focused on the AP-1 regulon for the rest of the study.

The FOSL1 protein is overexpressed upon treatment, switches partners and accumulates at enhancers together with H3K27ac

We next studied the expression level of the AP-1 members in vitro upon chemotherapy treatment. At the protein level, we detected significant changes in the expression of the core AP-1 members JUN and FOSL1 (Fig. 3E and Fig. S3F). Specifically, both are upregulated in all three cell lines upon 21 days of treatment, while JUNB, JUND, and FOSB are expressed at similar levels in chemonaïve cells and persister cells (note that no suitable antibodies were available for the other AP-1 members). As FOSL1 displayed the highest fold-change in expression (average FC = 6.7 in protein expression in all three cell lines), we focused our efforts on this AP-1 member. We showed that FOSL1 is highly expressed starting at day 7, with a slight activation detected already at day 3 (Fig. S3G). Notably, after chemotherapy exposure, day 7 corresponds to the time point at which a large fraction of the cell population has died, with only persisters left (Fig. S3G). Altogether, these results suggest that FOSL1 is highly expressed in persister cells at their emergence.

In addition to the activation of FOSL1 expression we investigated the evolution of the repertoire of FOSL1 partners upon 5-FU with mass-spectrometry (Fig. 3F, Fig. S3H). Consistent with FOSL1 overexpression upon treatment, we observed a significant increase of FOSL1 peptides after 5-FU. Ratio between FOSL1 peptides and several partners were increased upon treatment: we detected higher levels of YBX1-2 and JUNB and JUN peptides compared to FOSL1, indicating that interaction with these proteins is significantly enriched independently of the gain in FOSL1 expression. On the contrary, interaction of FOSL1 with several members of the SWI/SNF family - expected to form a canonical BAF complex - are lost upon treatment. Altogether, these results indicate that FOSL1 could be switching partners upon treatment.

Next, to determine whether AP-1 factors, particularly FOSL1, are involved in the regulation of the persister expression program, we first used in situ tagmentation (CUT&Tag) to profile FOSL1 binding sites in persister cells and compared FOSL1 and H3K27ac profiles. Applying PCA on a common consensus peak annotation, we demonstrated that the H3K27ac and FOSL1 genomic binding patterns are extremely similar, grouping together on the first principal component of maximal variation (Fig. 3G). In particular, changes at genomic peaks in H3K27ac and FOSL1 enrichments upon treatment were significantly correlated (Fig. 3H). Chromatin remodeling upon treatment was characterized by the simultaneous enrichment of FOSL1 and H3K27ac on persister genes in persister cells as compared to those in chemonaïve cells (Fig. 3I and Fig. S3I). Indeed, we found that 26% of all persister genes in MDA-MB-468 cells gained both FOSL1 and H3K27ac upon 5-FU treatment, suggesting that this AP-1 member is a key regulator of the persister expression program.

FOSL1 is sufficient to reach the persister state and FOSL1 KO impairs treatment tolerance

To test whether FOSL1 is sufficient to reach the persister state in vitro, we generated gain-of-function (GOF) models that constitutively overexpress FOSL1. Specifically, MDA-MB-468 cells were infected with a tagged FOSL1 ORF (‘HA-FOSL1’) or with a CRISPR-Cas9 VPR system and two different gRNAs targeting FOSL1 (‘gRNA1’ and ‘gRNA2’ cell populations; Fig. 4A-B), or respective control constructs (Methods). We selected (via the fluorescent reporter genes) for transduced cell populations, rather than individual clones, to retain maximum heterogeneity of the population. Indeed, depending on the site of integration of the cassette (tagged FOSL1 ORF or gRNAs targeting FOSL1), cells can express variable levels of FOSL1 protein (Fig. 4C). We next measured global expression levels of the FOSL1 protein in the various cell populations. HA-FOSL1 cells express high levels of FOSL1 as compared to gRNA1 and gRNA2 cell populations, while cells with the CRISPR-Cas9 activating system, which activates the endogenous FOSL1 promoter, have expression levels closer to that of 5-FU–exposed cells (Fig. 4B). We then evaluated the phenotypic consequences of FOSL1 overexpression by comparing the ability of GOF and control cells to survive a 7-day 5-FU treatment. Constitutive FOSL1 upregulation leads to the emergence of more persister cells under 5-FU at day 7, and with higher metabolic activities (Fig. 4D-F). These results suggest that even a low expression of the FOSL1 protein is sufficient to confer the cells with the ability to survive chemotherapy. In parallel, we tested the effect of FOSL1 overexpression on the emergence of drug tolerant persister cells in another TNBC model system (BT-20 cell line). FOSL1 overexpression in BT-20 cells increased the capacity of cancer cells to tolerate 5-FU treatment compared to control cells (Fig. S4A and B). In addition, we observed that MDA-MB-468 FOSL1 overexpressing cells are also more tolerant to Irinotecan treatment than control cells, a topoisomerase I inhibitor proposed to treat advanced breast cancer including TNBC (42) (Fig. 4F). FOSL1 appears crucial in the transition from chemonaïve to persister phenotype in TNBC cells upon several cytotoxic treatments.

Fig. 4. FOSL1 is sufficient to reach the persister state.

A. Scheme representation of the genomic location of CRISPR VPR gRNAs at the FOSL1 locus.

B. Protein expression of FOSL1 in the indicated MDA-MB-468 FOSL1 overexpressing cell lines treated 7 days with 5-FU. Histone H3 is used as a loading control. Representative images are shown (n = 3 biological replicates).

C. FOSL1 protein localization in the indicated MDA-MB-468 FOSL1 overexpressing cell lines treated or not with 5-FU for 7 days. Representative images are shown.

D. Histogram representing the number of MDA-MB-468 cells after 7 days of treatment with 5-FU relative to DMSO in the indicated cell lines for n = 3 replicates; mean sd, Anova Bonferroni's multiple comparison test.

E. Histogram representing the mitochondrial activities of MDA-MB-468 cells after 7 days of treatment with 5-FU relative to DMSO using a XTT assay in Control and FOSL1 overexpressing cell lines. n = 3; mean sd, Anova Bonferroni's multiple comparison test.

F. Colony forming assays for 5-FU–treated MDA-MB-468 Control and FOSL1 overexpressing cells after 7 days of treatment with 5-FU. Representative images are shown (left) and quantification (right) for n=3 replicated, mean sd, Anova Bonferroni’s multiple comparison test.

G. Scheme representation of the genomic location of the CRISPR KO gRNA at the FOSL1 locus.

H. Protein expression of FOSL1 in the indicated MDA-MB-468 KO cell lines treated 21 days with 5-FU. Histone H3 is used as control. Representative images are shown (n = 3 biological replicates).

I. Histogram representing the mitochondrial activities of MDA-MB-468 cells after 21 days of treatment with 5-FU relative to DMSO using a XTT assay in Control and FOSL1 KO cell lines. n = 3; mean sd, paired t-test.

J. Colony forming assay for 5-FU–treated FOSL1 MDA-MB-468 KO cells after 21 days of treatment with 5-FU. Representative images (left) and quantification (right) are shown. n = 3; mean sd, paired t-test. (*: p-value<0.05, **: p-value<0.01, ***:p-value<0.001, ****:p-value<0.0001)

We generated KO FOSL1 using CRISPR cells to understand whether the TF was necessary for cells to reach a persister phenotype (Fig. 4G). Such KO partially prevented activation of FOSL1 expression upon treatment at day 21 (Fig. 4H). In these conditions, less tumor cells tolerate treatment, and the few obtained persisters showed decreased metabolic activity (Fig. 4I-J), further proposing FOSL1 as a facilitator of drug tolerance.

FOSL1 can drive the activation of the persister expression program in chemonaïve cells independently of treatment

To investigate how FOSL1 induces the persister state, we analyzed the transcriptome and epigenome of cells with an exogenous activation of FOSL1 expression in the absence of chemotherapy. We first analyzed the effect of different levels of FOSL1 expression on the expression programs of cancer cells using multiplex scRNA-seq analysis. As expected, VPR-FOSL1 (gRNA1 & gRNA2) and HA-FOSL1 cell populations showed heterogeneous FOSL1 gene expression levels (Fig. 5A). HA-FOSL1 cells (with highest protein expression of FOSL1) fully resembled 5-FU–treated persister cells (Fig. S5A), with the maximal expression of the KRT14 gene (Fig. S5B), which is characteristic of the persister state in vivo and in vitro (7). Increasing FOSL1 expression levels led to a gradual shift of the expression program toward the persister state (Fig. 5B-C and Fig. S5C): FOSL1-low cells activated low levels of the persister expression program as compared to mid- or high-level expressing cells. FOSL1-low cells correspond to control cell populations, in which FOSL1 is lowly expressed in absence of chemotherapy, or to a fraction of VPR-FOSL1 or HA-FOSL1 cells in which FOSL1 mRNA is not detected. Altogether, expression changes induced by FOSL1 activation are significantly similar to changes induced by 5-FU chemotherapy (Fig. 5D), suggesting that FOSL1 activation alone can recapitulate the persister state without chemotherapy exposure.

Fig. 5. FOSL1 drives the activation the persister expression program in chemonaïve cells independently of treatment.

A. UMAP of the MDA-MB-468 FOSL1 overexpressing cells scRNA-seq dataset, colored by sample of origin or FOSL1 expression.

B. UMAP of the MDA-MB-468 FOSL1 overexpressing cells scRNA-seq dataset, colored by Persister signature score.

C. Boxplot showing persister program activation in MDA-MB-468 control and FOSL1 overexpressing populations, stratified by FOSL1 expression status. The distribution of cells per group is shown above the plot. p-value from t-test comparing FOSL1 expression categories is indicated.

D. Dot plot comparing Log2FC in gene expression induced by 5-FU treatment or FOSL1 overexpression in MDA-MB-468 cells. Pearson correlation coefficient and associated p-value are indicated.

E. PCA of FOSL1 genomic binding profiles in MDA-MB-468 chemonaïve, 5-FU–treated persister (35 days), control GOF and FOSL1 overexpressing cells.

F. Genomic tracks of FOSL1 binding over LAMB3 and CAV1 loci in MDA-MB-468 chemonaïve, persister (35 days) and HA-FOSL1 cells.

G. Heatmap showing signal in FOSL1 peaks in MDA-MB-468 chemonaïve, 5-FU–treated persister cells (35 days), and FOSL1 overexpressing cells. FOSL1 peaks in persister cells were used as reference and ordered by decreasing signal. The average FOSL1 signal per condition is indicated above the heatmap.

H. Scheme, created with BioRender.com, representing FOSL1 and H3K27ac role in regulating the persister expression program.

(****:p-value<0.0001)

We also investigated the genomic binding profile of FOSL1 in cells with forced FOSL1 activation in the absence of chemotherapy. We showed that FOSL1 activation recapitulated the FOSL1 binding pattern induced by chemotherapy exposure (Fig. 5E-G), showing that FOSL1 is able to reach its target genes in absence of therapeutic stress. Increasing levels of the FOSL1 protein led to a wider range of target sites (Fig. 5G and Fig. S5D). Notably, the H3K27ac landscapes of cells with activated FOSL1 expression were similar to that of chemonaïve cells (Fig. S5E), with no H3K27ac changes even at genes with induced expression (Fig. S5F-G). These results suggest that pre-existing levels of H3K27ac could be sufficient for FOSL1 to activate gene expression (Fig. 5H). Other histone modifications or chromatin modifiers could also facilitate gene activation by FOSL1.

Discussion

By using eight PDXs from different patients with TNBC, which we treated with up to three different chemotherapies in parallel, we identify recurrent features of the persister state in TNBC. We observe that persistence to chemotherapy is a transient and reversible state, as has been previously described in other cancers (2,5,6,9,11). We further show that persister cells derived from an individual patient model share common features across different treatments and revert back to a common initial state at relapse. We identify hallmarks of persister cells in TNBC shared across patient models irrespective of the treatment. These hallmarks are a high level of expression of basal keratins, a switch in cholesterol metabolism, and the activation of the interferon as well as of the stress responses, whereby the latter being in agreement with previous studies investigating early response to chemotherapy in breast and ovarian cancers in vivo (43,44). Additionally, we demonstrate that these hallmarks of persistence are found in HER2+ breast tumor cells in response to targeted therapies, and beyond breast cancer in lung cancer models.

We identify basal keratins, and KRT14 in particular, as a recurrent marker of persister cells in vitro and in vivo in TNBC. Whether KRT14 is only indicative of the basal-like identity or has a functional role in the persister phenotype remains to be investigated. So far, studies have demonstrated that KRT14 can mediate tumorigenesis and invasion in breast cancer (45), ovarian cancer (46), and bladder cancer (47). Furthermore, a high KRT14 expression is closely associated with advanced tumor stage and poor prognosis in melanoma (48) and ovarian cancer (46) and may regulate the treatment response to cancer therapeutics in bladder cancer (49). Here, we show that prior to treatment, some tumors expressed extremely high levels of KRT14, associated with the expression of the AP-1 members FOSL1 and JUNB. The corresponding patients showed a lower disease-free survival, suggesting that pre-existing persister-like features could be associated with poor response to neo-adjuvant chemotherapy and high recurrence rates.

Through epigenomic and transcriptomic analyses of persister cells, we identify the AP-1 complex as a key regulator of the persister state. We show that FOSL1 is sufficient to launch the persister expression program, and that it appears necessary, in the model we tested, for cells to reach the persister state, indicative of a regulatory role in the switch to persister state. FOSL1 was previously shown to regulate the epithelial-to-mesenchymal (EMT) transition in aggressive breast cancer cells (50) and to lead to shorter metastasis-free survival in patients with basal-like breast cancer (51). More generally, several studies have highlighted the role of the AP-1 complexes in response to chemotherapy or to immuno- or targeted-therapies in cases of melanoma (18,52,53), lung or ovarian cancers. Here, we show that AP-1 might be a general regulator of the persister state in TNBC by characterizing the persister cells in response to multiple treatments.

We had previously shown that the activation of persister genes is accompanied by the loss of the Polycomb associated mark H3K27me3 at promoters (7). Here we further show that H3K27ac is actually enriched at a set of these loci upon treatment, suggesting an enhancer rewiring upon the switch to the persister state. Transcription factors, including AP-1, might facilitate the chromatin remodeling observed at persister genes. Consistently with this hypothesis, AP-1 transcription factors have been recently described as pioneer factors capable of rewiring 3D enhancer landscapes together with the multimeric chromatin remodeler SWI/SNF complex (54) or by cooperating with chromatin remodelers, such as CBP/p300 (55). Investigating additional histone marks and DNA methylation distribution dynamics in response to treatment will be instrumental to better understand the intimate link between chromatin remodeling and identified TFs activity in response to anticancer therapies.

We predict that other families of transcription factors could also regulate—together or not with AP-1—the recurrent persister program we identified, such as members of the IRF/STAT and NF-kB families, which are known to contribute to the therapeutic stress response in some cancers (44). Further mechanistic investigations are needed to demonstrate whether transcription factors actually drive persistence in TNBC alone or in combination. Such cooperation of AP-1 with other transcription factors has interestingly been described for therapy resistance in prostate cancer and melanoma, with AP-1 factors driving phenotype switching together with the YAP/TAZ-TEAD complex (56). The co-occupancy between AP-1 factors and these partner transcription factors on chromatin has also been experimentally demonstrated in breast cancer cells (57), and loss-of-function of the partner transcription factor is sufficient to prevent AP-1-driven tumorigenesis in vivo. Therefore, inhibiting the interaction between transcription factors might offer additional opportunities to prevent its activity. Specific pharmacological compounds against transcription factors are not yet available given the absence of catalytic activities or binding pockets for small-molecule inhibitors (58). New strategies like degraders-based technology are currently being developed to eliminate TFs in cancer cells (59). They could represent an opportunity, in combination with standards of care in TNBC, to prevent the emergence of persister cells and delay recurrence. Neo-adjuvant treatments in TNBC have recently evolved with the inclusion of immunotherapies. We will need to understand, using immuno-competent models, which are the most relevant vulnerabilities of persister cells in this treatment scheme.

Statement of Significance.

Elucidation of the features of the drug-tolerant persister state in triple-negative breast cancer reveals shared programs across patients and treatments that offer opportunities to prevent persistence and delay tumor recurrence.

Data availability

Genomic datasets are listed in Table S1 and deposited in GEO repository GSE280454, GSE280455 and GSE303201. The mass spectrometry proteomics raw data are deposited on the ProteomeXchange Consortium via the PRIDE partner repository (accession number: PXD066019). Public data analyzed in this study were obtained from GEO: GSE231525, GSE155341 and GSE123845. All other raw data are available upon request from the corresponding author. Code related to the manuscript are deposited on Github (https://github.com/vallotlab/AP1_Multi_Chemopersistence_TNBC).