Therapeutic synergies that overcome carboplatin resistance in triple-negative breast cancer.

Background
Breast cancer is the most frequently diagnosed cancer and the second leading cause of cancer-related death in women in the United States [1–3] These malignancies have rich molecular heterogeneity [4–6] and diverse subtypes with distinct pathological responses [4, 7]. Among breast cancer diagnoses, the Triple Negative Breast Cancer (TNBC) subtype, which lacks expression of human epidermal growth factor receptor 2 (HER2), progesterone receptor (PR), or estrogen receptor (ER), are associated with poor patient outcomes [8–10]. TNBC accounts for a significant proportion of breast cancer cases, disproportionately impacting younger women and those with African and Hispanic ancestry [10, 11]. Clinically, TNBCs are considered biologically aggressive and are characterized by relatively short disease-free survival compared to other breast cancer subtypes [12].
The lack of targetable receptors in TNBC limits treatment options, and cytotoxic chemotherapy remains the mainstay of care [13–15]. Carboplatin, a platinum-based chemotherapy agent, has emerged as a frontline treatment option for TNBC due to its ability to induce DNA damage and apoptosis in cancer cells [16, 17]. However, despite its initial effectiveness, the development of acquired resistance to carboplatin remains a significant clinical challenge [18–21]. Resistance mechanisms in TNBC are complex and multifaceted, often involving alterations in a number of key regulatory pathways, making overcoming acquired resistance a difficult task [20, 22]. Further complicating this, TNBCs display rich histological and molecular variations [23, 24]. Recent transcriptomic and proteomic studies highlight both intra- and inter-tumoral diversity within TNBC, underscoring the need to better define its molecular landscape to inform therapeutic development [24–26].
Recent advances in this field underscore the importance of understanding TNBC’s molecular diversity through transcriptomic and proteomic analyses [27, 28]. RNA sequencing and other transcriptomic investigations can provide informative insights into alterations in gene expression patterns and various molecular subtypes found within TNBC models [27, 29]. Meanwhile, proteomic analysis, which makes use of methods such as mass spectrometry, reveals information at the protein level, facilitating the discovery of functional networks and post-transcriptional alterations [30, 31]. Taken together, these omics data can aid in evaluating the information transfer between omics levels and help close the gap between genotype and phenotype [32]. These multi-omic approaches provide a holistic view of TNBC heterogeneity, aiding in the identification of novel therapeutic targets as well as alterations following treatment or acquired resistance.
Although many TNBC patients initially respond to platinum-based chemotherapy, resistance commonly develops [33]. And those patients with residual disease after treatment have poorer prognosis and low survival rates. Elucidating the molecular basis of acquired resistance is essential for identifying strategies to improve patient outcomes by preventing or improving acquired resistance. Prior studies have demonstrated that transcriptional reprogramming accompanies resistance development, making gene expression profiling a powerful tool for uncovering actionable pathways [34, 35].
Patient-derived xenograft (PDX) models serve as a clinically relevant platform to interrogate these mechanisms, as they recapitulate TNBC’s heterogeneity more faithfully than traditional cell lines [36]. In this study, we employ isogenic PDX models of carboplatin resistance generated through successive drug exposure to investigate transcriptional changes at single-cell resolution. By defining the pathways and molecular programs underlying acquired platinum resistance, we aim to identify novel therapeutic targets and inform the development of more effective treatment strategies for TNBC. Given the lack of durable responses to single-agent therapy in this setting, we also sought to evaluate whether other standard-of-care treatments, such as sacituzumab govitecan (SG), retain efficacy in carboplatin-resistant tumors and whether their activity can be enhanced through rational combination with synergistic agents. Integrating these studies with our transcriptional profiling offers the opportunity to identify both mechanisms of cross-resistance and potential vulnerabilities that can be exploited to overcome therapeutic failure in TNBC.

Materials and methods

TNBC cell lines & culture conditions
The human TNBC cell lines HCC1143 and MDA-MB-468 were obtained from the American Type Culture Collection (ATCC, Manassas, VA, USA). Cells were routinely cultured in RPMI-1640 medium (Gibco, Thermo Fisher Scientific) supplemented with 10% fetal bovine serum and 1% penicillin-streptomycin (100 U/mL penicillin, 100 µg/mL streptomycin). Cells were maintained at 37 °C in a humidified incubator with 5% CO₂ and routinely monitored for contamination as previously described [10].

PDX culture
The BCM-2147, BCM-3887, and BCM-7482 PDX models were obtained from Baylor College of Medicine, and WHIM2 and WHIM30 PDX models were sourced from Washington University in St. Louis. All animal experiments were performed under protocols approved by the Institutional Animal Care and Use Committee (IACUC) at Virginia Commonwealth University (Protocol# AD10001247) in accordance with institutional guidelines. Non-obese diabetic severe combined immunodeficient gamma (NSG) mice, bred in-house, were used for all in vivo studies. Tumor cells were suspended in Matrigel (Corning) and injected into the fourth mammary fat pad. Once tumors reached approximately 10 × 10 mm, they were harvested and digested in a solution of DMEM/F12 supplemented with 5% fetal bovine serum (FBS), 300 µL collagenase (Sigma), and 100 µL hyaluronidase (Sigma) [11]. Following digestion, tumors were trypsinized and single-cell suspensions were prepared.

Generation of carboplatin-resistant isogenic PDX models
Carboplatin-resistant (CR) TNBC PDX models were generated from four parental lines: WHIM30, BCM-2147, BCM-3887, and BCM-7482. For each, matched carboplatin-sensitive (CS) founder tumors were maintained to create isogenic CR/CS pairs. Resistance was induced through serial in vivo passaging under carboplatin treatment (40 mg/kg). Standard dosing consisted of three treatments per passage administered at 3–5 day intervals; however, in some early passages, only one to two doses were given to ensure sufficient tissue for continued passaging. Tumors were designated CR when no significant reduction in tumor volume was observed following treatment as measured with digital calipers. Time to acquire resistance varied between models from 2 to 5 sequential passages. Continued selective pressure by carboplatin dosing was continued for passages after resistance was acquired. Endpoint tumor burden was defined as a maximum dimension exceeding 10 mm in any direction, consistent across all models.

Single-cell RNA-seq library preparation, quality control and preprocessing
Prior to library preparation, PDX cells were digested into single cell suspensions as previously described [5, 36]. For single-cell collection, cells were dissociated with TrypLE and subsequently resuspended in 0.04% BSA. Single-cell RNA-seq libraries were prepared using the 10x Genomics Chromium Single Cell 3′ v3.1 kit following the manufacturer’s protocol. Target capture was ~ 8,000–10,000 cells per sample. Libraries were quality-checked on an Agilent Bioanalyzer and sequenced at the VCU Genomics Core and samples were sequenced to a minimum read depth of 20,000 reads/cell. Raw sequencing reads were first evaluated for quality using FastQC v0.11.9 and MultiQC v1.11 [37, 38]. Alignment was performed using the 10X Genomics CellRanger v9.0.1 “count” algorithm. For PDX samples, initial alignment was carried out against the 10X Genomics GRCh38_GRCm39-2024-A multi-species genome using the --expect-cells parameter to enable separation of human and mouse cells. Species classification was determined using the CellRanger-generated “gem_classification.csv” file [39]. Barcodes corresponding to human cells were extracted and re-aligned to the 10X GRCh38-2024-A human genome using the --force-cells parameter to retain all identified human cells. Cell-level quality control was conducted using a multi-step adaptive filtering strategy for each sample in R v4.4.1 using Seurat v5 [40]. Briefly, low-quality cells were excluded based on unique molecular identifier (UMI) counts, number of detected genes, and mitochondrial gene expression. Initially, a coarse filter was applied to remove cellular debris and empty droplets by excluding barcodes with fewer than 500 UMIs or 200 detected genes. Subsequently, sample-specific upper thresholds, defined as three median absolute deviations (MADs) above the median for both UMI and gene counts, were used to remove potential doublets calculated on the pre-filtered cell population. The mitochondrial content threshold, set to three MADs above the median and constrained to between 10% and 25%, was calculated using all cells with ≤ 50% mitochondrial content to prevent skew from dying cells. The final set of high-quality cells consisted of those passing all UMI, gene count, and mitochondrial percentage filters (Table 1). Filtered samples were independently normalized using Seurat’s log2 normalization before being scaled with UMI regressed out, and then merged using Seurat’s merge() function. Dimensionality reduction (UMAP and t-SNE), graph-based clustering, and cell cycle analyses were performed as previously described [36]. Loupe-compatible output files were generated using the 10X Genomics LoupeR package (https://github.com/10xGenomics/loupeR).

Bulk RNA isolation, quality control and preprocessing
Bulk RNA was extracted from PDX tumor tissue using the RNeasy Mini Kit (Qiagen, Cat. No. 74104) following the manufacturer’s protocol with mechanical disruption and column-based purification. Briefly, approximately 700 µL of RLT buffer was added to freshly dissected tumor fragments, which were mechanically dissociated using autoclaved scissors and homogenized through QIAshredder spin columns. All RNA samples were quantified and assessed for integrity prior to downstream library preparation and sequencing. Library preparation and sequencing were performed by Novogene (Novogene Co., Ltd.). Raw RNA-Seq fastq files were processed by the VCU Massey Comprehensive Cancer Center Bioinformatics Shared Resource (BISR) using an in-house pipeline described previously [5, 8, 41]. Briefly, sequencing quality was assessed using FastQC v0.11.9 [38]. CutAdapt v4.1 [42] removed adapters and low-quality base pairs prior to alignment using STAR v2.7.11a to a merged human (GRCh38)/mouse (GRCm38) genome (for details see Alzubi et al.) using the command line options: “--outSAMtype BAM Unsorted --outSAMorder Paired --outReadsUnmapped Fastx --quantMode TranscriptomeSAM --outFilterMultimapNmax 1. Read counts and log2 TPM values were obtained with Salmon v0.8.2 “quant” algorithm using the “IU” library type [43]. PAM50 subtyping was carried out using the genefu v2.11.2 R package [44].

Single-cell RNA differential gene expression analysis
Differentially expressed gene (DEG) analysis was performed using the 10X Genomics Loupe Browser (v8.0.0). The “Run Differential Expression” tool was applied with the “Between selected cluster(s) themselves” setting to compare CR and CS tumors. For BCM‑3887, WHIM30, and BCM‑2147 models, time-matched CR and CS tumors were collected and processed together, with library preparation and sequencing performed within the same batch to minimize technical variation. For BCM‑7482, single-cell libraries from CR tumors were generated and sequenced separately from the previously collected CS pair; normalization steps were applied to mitigate batch effects prior to comparison. Results were exported as .csv files containing adjusted p-values, log2 fold changes, and median expression values for all genes excluding those annotated as “low average count” (< 1 count per cell across the dataset).

Bulk RNA variant analyses

Variant calling
Aligned BAM files were processed on the Virginia Commonwealth University high-performance computing cluster. The human reference genome GRCh38 (primary assembly) was used, with “chr” prefixes removed for compatibility with downstream tools. BAM files were sorted and indexed using samtools v1.21 prior to variant calling. Variants (SNPs and indels) were called using bcftools mpileup and bcftools call with multiallelic calling enabled [45]. Raw VCF files were compressed and indexed with bcftools index. To reduce false positives, variants were filtered with thresholds QUAL > 20, DP > 10, and MQ > 40, and filtered VCFs were indexed for downstream analyses. RNA-seq variant calls may include RNA-editing events, typically A to G or T to C substitutions. These artifacts are minimized in our pipeline by: read-quality and mapping-quality requirements in bcftools mpileup/call, high-confidence filtering (QUAL ≥ 20, MQ ≥ 40) which removes the majority of low-allele-fraction RNA editing signals, depth requirement (DP ≥ 10) to exclude low-level editing events, and sensitive/resistant comparisons, which treat variants conservatively and do not interpret low-level, non-reproducible mismatches as biological differences.
Commands used were:

Variant annotation and comparison
Filtered VCFs were imported into R using the VariantAnnotation package v.1.54.1 [46]. Gene coordinates were obtained from Gencode v43 annotations via a TxDb object, and gene symbols were mapped to Ensembl IDs using org.Hs.eg.db. Variants were annotated for clinical significance using a ClinVar VCF (hg38, bgzipped and indexed). For each gene of interest (BRCA1, BRCA2, TP53), variants within the gene region were extracted from each sample VCF. Variants were classified as shared or unique between paired samples, and genotypes were compared side-by-side in a wide-format table. ClinVar pathogenicity annotations were added for each variant to identify previously reported disease-associated alleles. The final output was saved as CSV tables for downstream analyses.

Differential expression analysis of bulk RNA-seq data
Raw count data from bulk RNA-seq were imported into R (v4.3.2) for preprocessing and differential expression analysis. Count matrices were read in as .txt files and curated to standardize column names, parse sample identifiers, and remove non-coding transcripts. Genes with zero counts across all samples were excluded from downstream analyses. Differential expression analysis was performed using the edgeR v4.0.16 package in R [47]. Count data were filtered to retain only genes with counts per million > 1 in at least two samples. Library sizes were recalculated, and normalization was performed using the trimmed mean of M-values method. Multi-dimensional scaling plots were generated to visualize sample clustering based on biological coefficient of variation. For group comparisons, a design matrix was constructed based on PDX model identity. Common, trended, and tagwise dispersions were estimated using edgeR’s negative binomial model framework. Differential expression was assessed using the exactTest function in edgeR, comparing CR versus carboplatin-sensitive CS tumors within each PDX model. Resulting p-values were adjusted for multiple testing using the Benjamini–Hochberg method to control the false discovery rate (FDR). Genes with FDR < 0.05 were considered differentially expressed. Significant genes were visualized using smear plots. Gene-level results for each PDX pairwise comparison were exported as CSV files.

Nuclear area comparisons in BCM-7482

Nuclear area quantification
Formalin-fixed paraffin-embedded (FFPE) tumor sections from BCM‑7482 CS and CR PDX tumors were imaged using the Agilent Cytation 7 imaging system at 20× magnification. Automated nuclear segmentation was performed using the Gen5 Image + software (Agilent), with a DAPI channel mask applied to define nuclear regions of interest. Identical thresholding parameters and segmentation settings were applied to all images to ensure consistency between CS and CR samples. For each image, the software exported per-object nuclear measurements including area, integrated intensity, and shape descriptors. Data from multiple images across biological replicates were combined for downstream analysis.

Data processing and statistical analysis
Raw nuclear object measurements were exported into .xlsx format. Nuclear area was extracted and averaged per image, and distributions of nuclear sizes were compared between CS and CR tumors. Statistical significance was determined using Welch’s t-test to account for unequal variance between groups. All analyses and visualization of bar plots were performed in GraphPad Prism v10.6.0.

Immunohistochemistry (IHC) staining
FFPE sections were baked at 60 °C for 30 min and sequentially deparaffinized in xylene and graded ethanols prior to rehydration in deionized water. Antigen retrieval was performed in 10 mM Tris-EDTA buffer (pH 9.0) using a pressure cooker, followed by gradual cooling to room temperature. Endogenous peroxidase activity was quenched using a commercial peroxidase block, and sections were incubated with rabbit primary antibodies overnight at 4 °C in a humidified chamber. Following washes in TBST, sections were incubated with HRP-conjugated polymer detection reagent for 20 min, developed using DAB substrate, and counterstained with hematoxylin. Slides were dehydrated through graded ethanols, cleared in xylene, and mounted with Permount mounting medium (Table 2).

Immunohistochemistry (IHC) image quantification
Quantification of IHC staining was performed using Fiji (ImageJ) across biological replicates, with a minimum of three representative images analyzed per tumor [48]. TIF or PNG images were imported into Fiji and processed using the Color Deconvolution plugin with the “H DAB” vector setting. Following deconvolution, the DAB signal (Color 2) was selected, and thresholding was applied under Image > Adjust > Threshold, ensuring the “Dark Background” option was unchecked. Thresholds were set with the minimum value fixed at 0, and the maximum adjusted manually to isolate DAB-positive staining. The mean intensity of the DAB signal was then recorded using the Analyze > Measure function.
Subsequently, the hematoxylin signal (Color 1) was selected to quantify nuclei. The image was thresholded using the same parameters (minimum = 0, maximum adjusted), followed by application of Process > Binary > Fill Holes and Process > Binary > Watershed to improve nuclear segmentation. Particle analysis was performed under Analyze > Analyze Particles, with the size parameter set to exclude small artifacts (minimum size: 51 pixels, approximating the average nuclear diameter). The total nuclei count per image was recorded. For each image, the mean DAB intensity was normalized to the number of nuclei by dividing the DAB signal by the nuclei count, generating a per-cell intensity estimate for each replicate. Final ratios were multiplied by 100 for axis readability. Protocol is an adapted version of prior published method [49].
BRCA1 IHC comparison significance was calculated using linear mixed model with Satterthwaite’s method in R using lmerTest package [50]. Of note, IHC images quantified for WHIM30 and BCM-2147 CS/CR models for proteomic validation were done on 4–5 representative images from a single biological sample, and significance was calculated using unpaired student t-test in Prism by GraphPad.

Drug combination and synergy analysis
High throughput drug screening was performed as previously described using a library of 555 compounds (NCI NExT Oncology Interrogation Tools Library) at a concentration of 1 µM [4, 5, 36, 51] in combination with the IC20 dose of sacituzumab govitecan (Supplemental File 5). Drug combination experiments were conducted to assess potential synergy between SG and secondary agents in CR PDX-derived tumor models. Sacituzumab govitecan was tested at four concentrations (0–40 nM) in combination with eight-point dose ranges (0–100 µM) of Everolimus, Talazoparib, Selinexor (KPT-330), or Ribociclib. Cell viability was measured after 72 h of treatment using the CellTiter-Glo luminescent assay [52]. Dose–response matrices were analyzed using the SynergyFinder platform to calculate synergy scores based on the BLISS reference model [53].

Protein extraction and quantification
Freshly flash-frozen tumor fragments were lysed in buffer containing 8 M urea, 1% SDS, and 200 mM EPPS (pH 8.0) supplemented with protease and phosphatase inhibitor cocktails. Lysates were sonicated on ice to ensure complete disruption and solubilization of proteins. Protein concentrations were determined using the Bradford method according to the manufacturer’s instructions. For each sample, 200 µg of total protein was aliquoted and submitted for total proteome analysis to the Thermo Fisher Scientific Center for Multiplexed Proteomics at Harvard Medical School (https://tcmp.hms.harvard.edu).

Proteomic data analysis
Differential proteins were identified using log2 protein abundances and the limma v.3.64.3 R package (Supplemental File 4) [54]. Proteomic pathway enrichment was performed using two complementary approaches: Enrich (over-representation analysis) and GSEA (rank-based enrichment using ordered fold changes). KEGG pathways and MSigDB gene-set collections (H and C1–C7) were used. Results are provided in Supplemental Files 2 & 3 with worksheets labeled by analysis and signature (e.g., Enrich.KEGG, GSEA.C7).

In vivo drug treatment studies in PDX models
Therapeutic studies were performed using CR PDX models WHIM30CR and BCM-2147CR, as well as the intrinsically carboplatin-resistant WHIM2 model. For in vivo expansion and propagation, tumor cells were prepared and implanted as previously described. Briefly, 5 × 10^5 tumor cells suspended 1:1 in Cultrex basement membrane extract (Bio-Techne) were injected bilaterally into the fourth mammary fat pads of female NSG mice (500,000 cells per injection) as previously described [5]. Tumors were allowed to establish for approximately three weeks (~ 21 days) before treatment initiation. To minimize baseline variability, mice were informally checked to ensure that tumors were of comparable size across groups prior to enrollment.

Vehicle formulation
For in vivo studies, drugs were either dissolved in 1% methylcellulose + 0.1% Tween-80 and administered via oral gavage or saline to be administered through intraperitoneal injection as indicated.

Group sizes and experimental cohorts
For WHIM30CR, the initial cohort consisted of 3 mice per group; an expanded cohort of 4 mice per group was subsequently performed to evaluate the Everolimus + Selinexor (KPT-330) combination. BCM-2147CR cohorts included 4 mice per group. For WHIM2 mammary gland tumor (MGT) studies, groups included vehicle (n = 4), Selinexor (KPT-330) (n = 4), Everolimus (n = 5), and combination (n = 5). For WHIM2 metastasis studies (intracardiac injection), groups consisted of vehicle (n = 4), Selinexor (KPT-330) (n = 3), Everolimus (n = 3), and combination (n = 4).

Drug administration
Sacituzumab govitecan (SG) was administered intraperitoneally once weekly at 5 mg/kg. Everolimus was given by oral gavage 10 mg/kg (in 100 µL) three times per week. Selinexor (KPT-330) was administered by oral gavage at 5 mg/kg (100 µL) three times per week. For combination groups, agents were given on the same schedules as their respective monotherapy arms. Vehicle groups received the corresponding vehicle formulation(s) on the same schedule. Carboplatin (40 mg/kg, three doses per passage at four-day intervals) was used for the generation of resistant PDX lines, but was not used in subsequent combination efficacy experiments.

Tumor monitoring and humane endpoints
Tumor dimensions were measured twice weekly using digital calipers, and tumor volume was calculated as (length × width^2) / 2, where width is the smaller dimension. Mice were euthanized when tumors exceeded 15 mm in any direction, when body weight loss exceeded 20%, or at planned experimental endpoints. In experiments where vehicle-treated tumors reached burden earlier than treated groups, those mice were euthanized and remaining treatment cohorts continued until endpoint criteria were met.

WHIM2 metastasis protocol and imaging
To assess metastatic burden, WHIM2 cells (500,000 cells in 100 µL HF buffer) were injected intracardially into five-week-old NSG mice. Successful injections were confirmed immediately using IVIS® Spectrum imaging (PerkinElmer) with Living Image® 4.7.4 software. After 35 days of treatment, mice were injected intraperitoneally with 200 µL luciferin (150 mg/kg), incubated for 5 min, and imaged for whole-body metastasis. Brains and ovaries were subsequently harvested and imaged ex vivo on the IVIS. Metastatic burden was quantified by organ radiance, and radiance values were normalized across groups to compare metastatic load between treatments.

Statistical analysis
For the initial WHIM30CR cohort (all mice harvested simultaneously), Welch’s two-sample t-test was used to compare tumor volumes. For the WHIM30CR expansion cohort and BCM-2147CR studies, where vehicles reached endpoint earlier than treatment groups, tumor growth trajectories were analyzed using linear mixed-effects models. The fixed-effects structure was specified as Measurement ~ Days * Condition, with random intercepts for Mouse and for Mouse: Side (nested), and significance of the Days: Condition interaction was used to test treatment effects. The standardized effect size was calculated as the interaction estimate divided by the model residual SD. For WHIM2 MGT and metastasis studies, tumor volume and organ radiance were compared between groups using one-way ANOVA, with p < 0.05 considered significant.

Results

Generation of isogenic platinum-resistant PDX models
The novelty of this study lies in the generation of isogenic paired TNBC PDX models of carboplatin resistance from four founder lines (WHIM30, BCM-2147, BCM-3887, BCM-7482). All four lines were initially characterized to be CS, showing notable reduction in primary tumor volume following treatment. Through serial in vivo passaging under carboplatin treatment, CR variants were established (Fig. 1A). Resistance was defined as the absence of a significant reduction in tumor mass after treatment. Across the four PDX lines, resistance was achieved after 2–5 passages (Fig. 1B). This approach yielded four matched CR/CS PDX pairs encompassing diverse TNBC backgrounds.

These paired models provide a controlled system for dissecting resistance mechanisms. Once resistance was confirmed, tumors from both CS and CR lines were collected for downstream omics analyses, including bulk and single-cell RNA sequencing and tandem mass tag mass spectrometry (TMT-MS) (Fig. 1A). These datasets also offer a clinically relevant platform for further testing rational drug combinations with potential synergy alongside standard therapies.

Histopathological and morphological differences between isogenic CR/CS PDX pairs
Following model generation, tumors were resected for CR/CS pairs for histology. Histological and immunohistochemical evaluation revealed both shared and model-specific phenotypic changes between CR and CS tumors. Across all four PDX pairs (BCM‑2147, BCM‑3887, BCM‑7482, and WHIM30), hematoxylin and eosin (H&E) staining demonstrated preserved overall tumor architecture with variable regional necrosis (Fig. 2A-D: a, e). Pan‑cytokeratin (PanCK) staining confirmed maintained epithelial identity in both CR and CS tumors for BCM-2147 (Fig. 2A: b, f) and WHIM30 (Fig. 2D: b, f) models, with no appreciable loss of cytokeratin expression following resistance. BCM-3887 (Fig. 2B: b, f) and BCM-7482 (Fig. 2C: b, f) models maintained low PanCK staining levels prior to and following resistance. Ki67 and phospho-histone H3 (pHH3) staining did not show consistent changes in proliferative indices between CR and CS models (Fig. 2A-D: c, d, g, h).

Upon initial inspection of BCM‑7482 tumor sections under 20× magnification, CR tumors appeared to contain noticeably larger cells/nuclei than CS tumors, prompting quantitative analysis. Since there were clear differences in the size of tumor cells within BCM-7482CR, automated nuclear segmentation of BCM‑7482 tumor sections was performed and a significant increase of nuclear size within CR tumors compared to their CS counterparts was observed (Supplemental Fig. S1; Supplemental File 1). Quantification of segmented nuclear objects demonstrated a significant shift in mean nuclear area in CR tumors, consistent across multiple images and biological replicates (p < 0.0001, Unpaired t-test). This nuclear enlargement aligns with the striking morphological differences observed in H&E sections of BCM‑7482CR tumors, supporting the phenotypic distinction between CR and CS pairs in this model. No notable differences in nuclear size or morphology were observed in the three other basal-like TNBC models.

Single-cell transcriptional divergence accompanies carboplatin resistance in TNBC PDXs
To define transcriptional changes associated with CR at the single-cell level, we performed scRNA-seq on matched CR and CS tumors (Fig. 3A). In t-SNE embeddings, CR cells within each model occupied distinct transcriptional space from their paired CS cells, yet CR cells were more similar to their CS pair than to CR cells from other models (Fig. 3A-B). These patterns indicate that transcriptionally distinct profiles emerged during resistance development, while overall, each PDX maintained its unique genomic identity.
To assess cellular heterogeneity within and between models, we performed Seurat clustering and visualized the resulting clusters using tSNE embeddings (Fig. 3C). CS and CR cells projected into multiple discrete transcriptional sub-clusters, indicating substantial intra-tumoral heterogeneity. Each PDX pair shared at least one major clusters, although the relative proportions of cells within these clusters differed between sensitive and resistant tumors. While some clusters within WHIM30 and BCM-7482 segregated almost exclusively CR or CS cells (clusters 3–4 & 5–6), BCM-2147 and BCM-3887 model pairs showed less strict cluster segregation based on resistant status alone. Together, these data indicate that carboplatin resistance can arise through different modes across models, ranging from shifts in the abundance of pre-existing transcriptional states to the apparent emergence of distinct, resistance-associated clusters.

Differentially expressed gene (DEG) analysis was performed within each CS/CR pair to identify model-specific transcriptional changes. Each PDX exhibited a unique set of up- and down-regulated genes, reflecting distinct molecular programs driving resistance. No individual genes were significantly altered across all models after multiple testing correction, highlighting the heterogeneity of CR mechanisms (Fig. 3D-O). Nonetheless, subsets of genes were shared among two or more models, suggesting partial convergence of resistance pathways.

Homologous repair related pathways in acquired carboplatin resistance
Bulk RNA sequencing of the paired CS and CR PDX models was performed following scRNAseq. From the bulk RNA profiles, DEG analysis was performed in R for each model set. When lists of differently expressed genes were compared across models, there was some overlap; however, after multiple testing correction there was only one slightly upregulated gene that overlapped in all 4 CR/CS model comparisons (Fig. 4A). The identified gene was SIX3, a member of the sine oculis homeobox transcription factor family that has been associated with cancer signaling [55, 56].
Since prior literature suggests that homologous recombination deficiency (HRD) can be linked to resistance mechanisms for DNA damaging agents and carboplatin is commonly prescribed to patients with BRCA1 mutations, we sought to stratify genomic changes that occurred during CR with particular emphasis on BRCA1- mediated repair and alternative HRD-related pathways [57, 58]. The DEG analyses were contrasted with known HRD genes [59]. Within the BCM‑2147 pair, the CR tumors exhibited a marked increase in BRCA1 expression relative to the CS line (Fig. 4B-C). This transcriptional upregulation suggests restoration or enhancement of BRCA1-dependent homologous recombination as a primary driver of acquired resistance. This same trend was observable under IHC probing for BRCA1 (Supplemental Fig. 2; p-value < 0.1). Consistent with this, variant interrogation of bulk RNA sequences did not detect BRCA1 mutations in this model, supporting a functional repair axis. In WHIM30, we identified heterogeneity between different CR passages (Fig. 4F-G). One set of CR derivatives demonstrated significant BRCA1 upregulation, implicating the same BRCA1-mediated homologous repair pathway in resistance. In contrast, an earlier CR derivative lacked this BRCA1 increase but had consistent downregulation of HORMAD1 and MAGE‑A4. Both transcripts are cancer-testis antigens associated with DNA damage tolerance and have been reported as elevated in HRD-high breast tumors. Their suppression in WHIM30CR suggests a possible shift away from this alternative HRD mechanism, indicating more than one route to resistance within this model. For BCM‑3887, which harbors a pathogenic splice-site BRCA1 mutation resulting in loss of functional protein [60], CR tumors did not show increased BRCA1 expression (Fig. 4E). Instead, we observed significant upregulation of mismatch repair genes MLH1, MLH2, and MLH3 in the resistant tumors. These findings suggest that, in the absence of a functional homologous recombination axis, enhanced mismatch repair may underlie carboplatin resistance in this model.

The BCM‑7482 model, carrying a pathogenic BRCA1 frameshift mutation [61], similarly lacks a functional BRCA1 repair pathway (Fig. 4D). Here, CR tumors showed strong upregulation of HORMAD1 compared to CS tumors, consistent with an adaptive increase in DNA damage tolerance and replication stress response. This claudin-low model displayed a broader DEG profile in HRD-related genes but without a uniform trend across the canonical homologous recombination pathways, highlighting a potentially distinct resistance biology.
A variant analysis was performed on the bulk RNA sequencing data for BRCA1, BRCA2, and TP53. While known mutations in TP53 were observable by this analysis in BCM-7482 and BCM-3887, no novel mutations related to these genes was observed in CR models when compared to their CS counterpart.

Candidate resistance-associated pathways identified by proteomic profiling
Having identified some potential routes of resistance in these models, we sought to determine if any common targetable pathways existed. We performed Tandem Mass Tagged Mass Spectrometry performed on the BCM-2147 and WHIM30 CS/CR sets. Of note, several proteins were seen to be significantly different between CR and CS pairs in these models (Supplemental File 4). We further validated some of those targets with IHC characterization in relevant tumor sections (Supplemental Fig. 3).
To further identify coordinated biological processes associated with carboplatin resistance, we performed pathway enrichment analysis on the differentially expressed proteins. Both over-representation analysis (Enrich) and rank-based gene set enrichment analysis (GSEA) were applied using KEGG pathways and MSigDB collections (H, C1–C7). Across models, carboplatin-resistant tumors showed differential enrichment of several metabolic and stress-response pathways—including glycolysis/gluconeogenesis, amino sugar and nucleotide sugar metabolism, and glutathione metabolism—as well as pathways related to extracellular matrix organization such as focal adhesion and ECM–receptor interaction, providing pathway-level context for the proteomic differences between CR and CS tumors (Supplemental Files 2 & 3).
We identified several genes which we thought might be promising in our models, including UCHL1, an emerging target in breast cancer [62]. Despite this, preliminary testing of inhibitory agents targeting these proteins in vivo largely failed to produce meaningful effects on reducing overall tumor burden, indicating that while these proteins may mark resistant tumors, more testing is needed to fully determine their utility as direct therapeutic targets in these models.

Synergistic therapeutic strategies in CR TNBC models
Beyond characterizing pathways associated with carboplatin resistance, these models also serve as a clinically relevant platform for testing rational drug combinations, including the evaluation of synergistic agents alongside current standard of care. For this reason, we pursued identifying agents which would work synergistically with the standard care agent SG, a TROP2-targeted antibody-drug conjugate currently in clinical use [63, 64].
To identify actionable vulnerabilities in CR TNBC, we performed high-throughput drug screening with an emphasis on synergistic interactions with SG. Across models, several strong synergistic interactions emerged, with several mTOR pathway inhibitiors consistently ranking among the top candidates when quantified by coefficient of drug interaction (CDI; values < 1 indicate synergistic effects) (Fig. 5A & Supplemental File 5; Vistusertib (AZD2014) & AZD8055 shown).
Drug combination assays were carried out in vitro in WHIM30CR model and MDA-468 cell line using a four-point SG dose range (0–40 nM) in combination with eight-point dose ranges (0–100 µM) of Everolimus, Talazoparib, Selinexor (KPT-330), or Ribociclib (Fig. 5B-I). Cell viability was assessed after 72 h using the CellTiter-Glo assay.

Everolimus, an FDA-approved mTORC1 inhibitor with established clinical utility, demonstrated robust and reproducible synergy with SG in multiple CR lines (Fig. 5E, H). Notably, the non-novel combination of SG and Talazoparib, a PARP inhibitor which is currently being tested with SG in clinical trials, was seen here as well (Fig. 5D, I) [65]. Based on these results and its translational potential, the SG + Everolimus combination was prioritized for further evaluation. Prior work from our group has demonstrated synergy between XPO1 inhibitors and mTOR blockade in TNBC models, including the carboplatin-insensitive PDX model WHIM2 [10]. For this reason combination of Selinexor (KPT-330), an XPO-1 inhibitor, and Everolimus were tested in parallel.

In vivo validation of synergistic combinations in CR TNBC
Given the strong synergy observed in vitro, we next evaluated SG-based drug combinations in vivo using NSG mice bearing WHIM30CR and BCM2147CR tumors. In WHIM30CR SG + Everolimus showed significantly reduced tumor volume when compared to either drug alone in an initial cohort of three mice per group (Fig. 6A). Noticing a trend in the Everolimus + Selinexor (KPT-330) group, we sought to determine if cohort expansion would yield significant results. Expanding the cohort to 4 mice per group did result in a more robust findings with a significant difference noted in the combination group when compared to either mono-agent (Fig. 6B).
In BCM-2147CR we again tested Everolimus + Selinexor (KPT-330) and Everolimus + SG combinations (Fig. 6C). While Everolimus + Selinexor (KPT-330) showed a synergistic effect in this model, the Everolimus + SG combination did not show significant improvement over Everolimus alone, suggesting that this effect is being driven primarily by the Everolimus in this model. Across both models, Everolimus + SG consistently yielded the most robust effect size, supporting this combination as a leading candidate for overcoming carboplatin resistance in TNBC.
Additionally, we assessed blood taken from treatment mice without tumors and noted some shifts with treatment similar to those seen in patients, but did not note significant weight loss in the mice following treatment (Supplemental Fig. 4).

Dual mTOR and XPO1 inhibition reduces primary and metastatic burden in WHIM2
To further evaluate the therapeutic potential of combined mTOR and XPO1 inhibition, WHIM2 PDX tumors were treated with Selinexor (KPT-330) and Everolimus. WHIM2 was selected due to its intrinsic carboplatin resistance and high XPO1 expression. In the mammary gland tumor (MGT) model, combination therapy resulted in a marked reduction in tumor volume compared with single-agent Selinexor, single-agent Everolimus, and vehicle controls (Fig. 7A–B). Tumors in the dual-treatment group demonstrated both delayed growth onset and sustained regression over the treatment course.
In mice bearing metastatic WHIM2 disease, the combination also significantly reduced metastatic burden. Quantitative imaging showed decreased whole-body signal relative to controls (Supplemental Fig. 5A; Supplemental Fig. 6). Organ-specific analyses confirmed reduced metastatic lesions in the brain (Fig. 7C–D; Supplemental Fig. 5B) and ovary (Fig. 7E–F; Supplemental Fig. 5C). Notably, in several mice, metastases in these organs were undetectable following dual treatment, whereas they remained prominent in vehicle and single-agent treated cohorts.
Together, these data demonstrate that combined inhibition of XPO1 and mTOR suppresses both primary tumor growth and metastatic dissemination in the WHIM2 model.

Discussion
In this study, we established and comprehensively profiled isogenic CR TNBC PDX models to investigate mechanisms of resistance and identify potential therapeutic strategies. By generating matched CR/CS pairs from four genetically and phenotypically distinct TNBC PDX lines, we created a clinically relevant system to dissect tumor-intrinsic adaptations that emerge during carboplatin treatment. This model set represents a uniquely diverse and well-controlled resource for investigating platinum resistance in TNBC.

Heterogeneity of resistance mechanisms across TNBC models
Our findings highlight the heterogeneity of molecular programs underlying carboplatin resistance. Single-cell and bulk transcriptomic analyses revealed distinct transcriptional changes in each CR/CS pair, with no single gene consistently altered across all models after multiple testing correction. Instead, subsets of resistance-associated genes were shared among two or more models, suggesting partial convergence of pathways rather than a universal mechanism. This aligns with prior observations that TNBC is a highly heterogeneous disease, both at baseline and under therapeutic pressure [23, 24, 66–68].
Despite this heterogeneity, several models demonstrated resistance programs linked to DNA repair capacity. In BCM-2147CR, we observed marked upregulation of BRCA1 at both RNA and protein levels, consistent with restoration of homologous recombination repair as a driver of acquired resistance [69, 70]. WHIM30 also demonstrated passage-dependent changes, with later CR passages showing increased BRCA1 expression, while some earlier passages displayed suppression of cancer-testis antigens HORMAD1 and MAGE-A4, both previously associated with HRD-high states [59, 71, 72]. In contrast, BCM-3887 and BCM-7482, which carry pathogenic BRCA1 mutations, exhibited distinct adaptive pathways. BCM-3887CR tumors upregulated mismatch repair genes (MLH1, MLH2, MLH3), suggesting compensation via alternative repair programs [59], while BCM-7482CR tumors showed pronounced induction of HORMAD1, consistent with increased tolerance of replication stress [72, 73]. Together, these results support the concept that platinum resistance in TNBC can arise through restoration of homologous recombination in BRCA1-wildtype tumors or through activation of alternative repair/tolerance pathways in BRCA1-deficient backgrounds.

Proteomic profiling identifies markers but limited direct targets
Proteomic analysis of CR versus CS tumors in BCM-2147 and WHIM30 revealed several proteins differentially expressed in resistant tumors, including UCHL1. Although UCHL1 has been implicated as an oncogenic driver in breast cancer [62], functional validation in our models did not demonstrate therapeutic vulnerability worthy of further investigation. This suggests that while such proteins may serve as markers of resistance, they may not represent effective direct targets in this setting. These findings underscore the need for multi-level validation and highlight the importance of distinguishing between biomarkers of resistance and actionable dependencies.

Therapeutic implications: exploiting vulnerabilities with combination strategies
Beyond mechanistic insights, these isogenic CR PDX models provide a platform for preclinical evaluation of rational therapeutic strategies. High-throughput drug screening identified consistent synergy between the standard-of-care ADC SG and inhibitors of the mTOR pathway. Notably, the non-novel combination of SG and PARP inhibitor also emerged as synergistic, supporting previous findings [65]. SG + Everolimus demonstrated robust synergy in vitro across multiple CR models and was further validated in vivo, where the combination significantly reduced tumor growth in WHIM30CR and BCM-2147CR models. While efficacy varied between models, the reproducibility of synergy with mTOR inhibition suggests this pathway may represent a tractable vulnerability in carboplatin-resistant TNBC.
We also evaluated the combination of Selinexor (KPT-330) and Everolimus. In WHIM2, an intrinsically carboplatin-resistant model with high XPO1 expression, dual inhibition markedly suppressed both primary tumor growth and metastatic dissemination. This is particularly notable given the limited number of therapeutic strategies available for metastatic TNBC and highlights the potential of XPO1/mTOR co-targeting to address both local and disseminated disease. This finding is supported in part by other studies which have noted synergistic effects of XPO1 inhibition with mTOR pathway [74].

Limitations and future directions
Several limitations should be acknowledged. First, while the use of multiple genetically distinct TNBC PDX lines increases generalizability, resistance mechanisms observed here may still not capture the full spectrum present in patients [23, 24]. Second, although our omics profiling identified model-specific pathways, functional validation was limited for many candidates, and additional mechanistic studies will be required to establish causality. Third, in vivo combination studies were conducted with modest cohort sizes, which, while sufficient to identify significant effects, may underestimate variability. Finally, while Everolimus and Selinexor are clinically available, their tolerability in combination with ADCs such as SG in patients remains to be established.

Conclusions
Together, these studies establish a robust platform of isogenic carboplatin-resistant TNBC PDX models and demonstrate their utility for dissecting mechanisms of resistance and testing rational drug combinations. Our findings emphasize the heterogeneity of resistance programs, reveal both BRCA1-dependent and -independent adaptations, and identify SG + mTOR inhibition as a promising therapeutic avenue. Moreover, dual mTOR and XPO1 inhibition demonstrates potent activity against both primary and metastatic disease in intrinsically resistant TNBC. These results provide both mechanistic insights and translational strategies to inform therapeutic development for patients with platinum-refractory TNBC.

Supplementary Information