Informed consent for the use of plasma samples for research purposes was obtained from all participants. Clinical samples were obtained from subjects who visited Severance Hospital (Seoul, South Korea), according to the guidelines of the independent Ethics Committee of Yonsei University College of Medicine (IRB approval no. 4-2020-1292; approval date January 4, 2021; Seoul, South Korea). A total of 35 patients with BC (20 with BC exhibiting no pCR after NAC and 15 patients showing pCR after NAC) who visited Severance Hospital between May 2015 and August 2020 were retrospectively registered in the present study. Preoperative plasma samples were collected from the participants before anesthesia. The criteria for inclusion in the analysis were as follows: i) NAC prior to plasma specimen acquisition; ii) confirmed pathological diagnosis of BC; and iii) hemolysis assessed before the isolation of exosomes to evaluate plasma sample quality. Moreover, the study excluded participants with breast cancer types other than invasive ductal carcinoma and those with a history of other malignancies or existing medical conditions. The clinical characteristics of participants enrolled in the present study are summarized in Table SI and the research design is illustrated in Fig. 1.

Doxorubicin hydrochloride (brand name Adriamycin; cat. no. D4000) and docetaxel (brand name Taxotere; cat. no. D1000) were purchased from LC Laboratories. Cyclophosphamide (brand name Cytoxan; cat. no. S1217) was purchased from Selleck Chemicals. Adriamycin (5 mg/ml) and Cytoxan (20 mg/ml) were dissolved in saline (JW Pharmaceutical Co., Ltd.). Taxotere (5 mg/ml) was dissolved in dimethyl sulfoxide (Duchefa Biochemie; cat. no. D1370). Aliquots were stored at −20°C for a maximum of 6 months and thawed immediately before use. Human BC cell lines MDA-MB-231 (MM231; HTB-26), MDA-MB-468 (MM468; HTB-132), and HCC1395 (CRL-2324) were purchased from the American Type Culture Collection. All cell lines were grown in Roswell Park Memorial Institute-1640 medium (RPMI-1640; cat. no. 22400-089) supplemented with 10% fetal bovine serum (FBS; cat. no. 12483-020) and 1% penicillin-streptomycin (cat. no. 15140-122) (all from Gibco; Thermo Fisher Scientific, Inc.). All cells were grown as monolayer cultures and maintained in a humidified atmosphere of 5% CO₂ at 37°C.

Drug-resistant BC cells were established by consistently increasing the concentrations of clinically used regimens, including Adriamycin, Cytoxan, and Taxotere, in the culture medium (19). Adriamycin and Cytoxan (1:10 weight ratio) followed by Taxotere were alternately added to the cells at specific concentrations [1/120, 1/90, 1/60, 1/30, 1/10, and 1 time of the half-maximal inhibitory concentration (IC₅₀) values of each drug] when they reached 60–70% confluency as per the protocol reported in the previous study by the authors (20). After continuous exposure to chemo-drugs for 6 months, the parental BC cells (herein referred to as ‘MM231_wild-type’, ‘MM468_wild-type’ and ‘HCC1395_wild-type’) were allowed to grow in fresh medium for an additional month until the surviving cells recovered favorably. After 7 months, the stabilized drug-resistant BC cells (herein referred to as ‘MM231_{resistant-type}’, ‘MM468_{resistant-type}’ and ‘HCC1395_{resistant-type}’) were generated and stored in a deep freezer (−70 to −80°C) for further investigation. The drug resistance characteristics were confirmed by comparison of the IC₅₀ values between the cells (Fig. S1 and Table SII).

Small RNA was isolated from cultured cells using a mirVana™ miRNA Isolation kit (Thermo Fisher Scientific, Inc.; cat. no. AM1560) according to the manufacturer's instructions, using TRIzol reagent (Thermo Fisher Scientific, Inc.; cat. no. 15596026). The extracted small RNA was quantified using UV absorption at a wavelength of 260 and 280 nm with a spectrophotometer (NanoDrop 3000; Thermo Fisher Scientific, Inc.) and then stored at −80°C until further analysis. Affymetrix GeneChip microarray (Thermo Fisher Scientific, Inc.) runs were performed on the RNA eluates. Intensity values of the CEL files were normalized to remove bias between the arrays using the Robust Multiarray Average and Detection Above BackGround algorithms implemented using the Affymetrix Expression Console software (version 1.4.1; Thermo Fisher Scientific, Inc.). Overall signal distributions of each array were plotted and compared using tools available from the Bioconductor Project (https://www.bioconductor.org/) to verify for normalization. After confirming that the data were properly normalized, differentially expressed miRNAs (DEMs) that exhibited >2-fold difference between the average signal values of the control and treatment groups were selected manually. In addition, the normalized data of selected DEMs were imported into R software (version 4.1.2; R Core Team) for t-test. Genes with a P<0.05 were extracted as significant DEMs for further study.

Microbeads attached to TDEs were fixed for 24 h in Karnovsky's fixative consisting of 2% glutaraldehyde (Merck KGaA; cat. no. 354400) and 2% paraformaldehyde (Merck KGaA; cat. no. 818715) dissolved in 0.1 M phosphate buffer (pH 7.4). Next, the fixed samples were washed twice for 30 min with 0.1 M phosphate buffer (Merck KGaA; cat. nos. S7907 and S9638). The beads were post-fixed with 1% OsO4 for 2 h and dehydrated using a gradually ascending ethanol series (50–100%) with a Critical Point Dryer (Leica Microsystems; cat. no. CPD300). They were then coated with platinum using an ion sputter (Leica Microsystems; cat. no. ACE600) and observed under a field-emission scanning electron microscope at ×20,000 magnification (SEM; MERLIN; Carl Zeiss AG). Confocal microscopy measurements were obtained to confirm the presence of TDEs using 3-µm microbeads. The captured exosomes were detected using a primary PE-Cy7-labeled antibody (BD Biosciences; cat. no. 561982) against the general exosome marker CD63. Fluorescence images were obtained using a Zeiss LSM 780 confocal microscope (Carl Zeiss AG). All data were obtained from at least three independent experimental replicates. The concentrations and size distributions of the TDEs resuspended in phosphate-buffered saline (PBS) were measured using a Nanoparticle Tracking Analyzer (NTA; NanoSight NS300 system; Malvern Panalytical Ltd.) Analysis was performed using NTA 3.1 software (Malvern Panalytical Ltd.) with default settings according to the manufacturer's software manual. The camera focus was adjusted to distinctly visualize TDEs that did not exceed a particle signal.

To validate the miRNA profile of exosomes, TDEs were isolated following a previously reported method (18,21). miRNAs were extracted from three pairs of TDEs originating from wild-type and resistant-type BC cell lines using a mirVana™ miRNA Isolation kit (Thermo Fisher Scientific, Inc.). RNA concentration was measured using a Qubit™ microRNA Assay kit (Thermo Fisher Scientific, Inc.; cat. no. Q32880) with a Qubit^® 2.0 Fluorometer (Thermo Fisher Scientific, Inc.; cat. no. Q32866). The extracted RNA was reverse transcribed using the TaqMan microRNA Reverse Transcription Kit (Thermo Fisher Scientific, Inc.; cat. no. 4366597). Candidate miRNAs were selected from the cellular microarray data (Table SIII), public datasets (GSE71142; Fig. S2) (22), and other reference studies on drug-resistant exosomal miRNAs (23–27). Differential expression levels of 12 miRNAs (miR-21-5p, miR-122-5p, miR-125b, miR-146a-5p, miR-148-5p, miR-155, miR-210-3p, miR-222-5p, miR-484, miR-501-5p, miR-1246-5p and miR-1260b) were measured via cDNA amplification reactions with a TaqMan Universal PCR Master Mix, No AmpErase UNG (Thermo Fisher Scientific, Inc.; cat. no. 4324018) and TaqMan microRNA Assay kit (Thermo Fisher Scientific, Inc.; cat. no. 4440887) in a CFX96 Real-time PCR system (Bio-Rad Laboratories, Inc.; cat. no. 3600037). Individual miRNAs were reverse-transcribed with the following conditions: 30 min at 16°C to anneal primers, 30 min at 42°C for the extension, and 5 min at 85°C to stop the reaction. qPCR was run using cDNA with the following conditions: 10 min at 95°C for enzyme activation, followed by 40 cycles consisting of denaturing at 95°C for 15 sec, and annealing and elongation at 60°C for 10 min. miRNA expression levels were normalized using miR-16 as a normalization control for exosomal miRNAs. All experiments were performed according to the manufacturer's protocol and repeated in duplicate. The 2^−ΔΔCq method was used to determine the relative expression of exosomal miRNAs (28).

To compare miRNA signature profiles of wild-type, educated-type (BC cells treated with drug-resistant exosomes), and resistant-type BC cells using quantitative PCR in BC cells, 2×10⁵ wild-type BC cells were seeded in 6-well plates and incubated at 37°C for 24 h, followed by treatment with 1×10¹⁰ of drug-resistant exosomes re-suspended in a complete culture medium and incubated at 37°C for 24 h. Subsequently, the exosome-treated cells were harvested for real-time PCR to evaluate miRNA expression levels. To verify the drug tolerance of wild-type, educated-type, and resistant-type BC cells, 5×10³ of both wild-type and resistant-type BC cells were seeded in 96-well plates. After a 24-h exosome education using 1×10⁹ of drug-resistant exosomes, the cells were incubated with chemo-drugs for 48 h and their cell viabilities were measured using MTT assay. The purple formazan was then dissolved in methanol at room temperature for 30 min on an orbital shaker and the absorbance was recorded at 570 nm with a correction wavelength of 690 nm with a spectrophotometer (NanoDrop 3000; Thermo Fisher Scientific, Inc.).

ROC analysis of drug-resistant exosomal miRNA markers was performed on data from plasma samples of 35 patients with BC using MedCalc (version 20.014; MedCalc Software Ltd.). Univariate ROC analysis was utilized for each miRNA target to obtain the ROC curve, area under the curve (AUC), AUC standard error (SE), and 95% confidence interval (CI) for evaluating the diagnostic power of the drug-resistant miRNA marker combinations. After performing a univariate ROC analysis on each combination of drug-resistant miRNA targets, the ‘best’ combination with the highest AUC was selected and also the lowest SE of AUC.

The present study included the gene expression profiles corresponding to the GSE25066 (29) and GSE41998 (30) datasets from the Gene Expression Omnibus (GEO) database. GSE25066 (508 samples) included 409 and 99 patients with BC exhibiting no response and complete response after NAC, respectively. GSE41998 (253 samples) included 184 and 69 patients with BC showing no response and complete response after NAC, respectively. DEGs were identified using DEGSeq (version 1.48.0) (31) with P<0.05, whereas log2FC ≥1 and log2FC ≤-1 cutoffs were used to denote upregulated and downregulated DEGs, respectively. Volcano plots were generated using ggplot2 (32) (version 3.3.5). Gene ontology (GO) was analyzed using ClusterProfiler (33). Representations of GO were generated by the DAVID tool (http://www.geneontology.org/) for functionally annotated molecular functions, biological processes, and biological pathways. To predict the miRNA-mRNA targets, TargetScan software (version 6.0) (https://www.targetscan.org/vert_60/) was used, an open platform for the prediction of miRNA targets. Gene function was annotated based on the biological process of GO gene set (34) (C5; MSigDB collections, BROAD Institute) and gene reference into function (GeneRIF; http://www.ncbi.nlm.nih.gov/sites/entrez?db=gene) database.

The cBioPortal for Cancer genomics is an open-access resource (http://www.cbioportal.org/), providing survival analysis for >10,000 tumor samples from 23 breast cancer studies in TCGA pipeline. This database was applied to predict the patient's survival using miRNA target gene sets (BAK1, NOVA1, PTGER4, RTKN2, AGO1, CAP1 and ETS1) that are highly related to drug resistance (miRNA^{high R}) and miRNA target gene sets (E2F2, ITGA3, SKP2, RIPK2 and STAT3) that are moderately related to drug resistance (miRNA^{moderate R}) as queries. The generated results were displayed as Kaplan-Meier curves with P-values from the log-rank test. The search parameters included alterations (missense mutations, splice mutations, truncating mutation, structural variant, deep deletion and copy number alterations) from whole genome/exome sequencing and targeted sequencing data with the default setting. OS and DFS were calculated on the basis of cBioPortal's online instruction.

Each experiment was repeated at least three times independently. Data are presented as the mean ± standard deviation. All statistical analyses were performed with either unpaired Student's t-test, or multiple comparison tests following one-way analysis of variance (ANOVA), using R and R studio (version 4.1.2; R Core Team) and GraphPad Prism software (version 9.0.0; Dotmatics). P<0.05 was considered to indicate a statistically significant difference.

Microarrays were used to identify DEMs with drug resistance-associated properties. MiRNA expression levels were compared with BC cells that were untreated and continuously treated with drugs. After evaluating miRNA expression using 2,578 human mature miRNA probe sets, 144 DEMs were identified in resistant-type BC (compared with wild-type controls) using the following cut-off criteria: P<0.05; fold change ≥2 (Fig. 2A). To visualize the expression patterns of the identified DEMs, hierarchical clustering was performed, and heatmaps were generated for each cell line. The results showed that MM231, MM468, and HCC1395 cells contained 42, 31 and 88 DEMs, respectively (Fig. 2B). Among the 144 DEMs identified, only 64 miRNAs were significantly upregulated, indicating that these were high-confidence candidates involved in drug resistance. When the common DEMs were assessed, 16 were common between at least two cell lines, and 1 was found to be common across all three cell lines (Fig. 2C). These observations highlighted the heterogeneity of miRNA expression patterns across different BC cell lines. The upregulated miRNAs could be further investigated as potential targets or biomarkers for drug resistance, considering both the common and unique regulatory mechanisms in BC.

A panel of 12 candidate exosomal miRNAs was selected for analysis. These miRNAs (miR-21-5p, miR-122-5p, miR-125b, miR-146a-5p, miR-148a-5p, miR-155, miR-210-3p, miR-222-5p, miR-484, miR-501-5p, miR-1246-5p, and miR-1260b) were selected based on a comprehensive assessment that incorporated microarray data, public datasets, and relevant studies on drug-resistant exosomal miRNAs (Fig. 3A). This crosschecking approach provided a basis for selecting specific miRNAs as potential candidates for drug-resistant exosomal miRNAs and strengthened the reliability of the selection process. These miRNAs were further analyzed in exosomes isolated from the three BC cell lines (MM231, MM468, and HCC1395) and their expression patterns between the wild-type and resistant-type (Fig. 3B-D) were compared. Although the overall miRNA expression patterns varied among exosomes derived from BC cells, a significant finding emerged. Notably, miR-21-5p, miR-122-5p, miR-125b, miR-146a-5p, miR-148a-5p, miR-155, miR-484, miR-1246-5p and miR-1260b exhibited a marked increase in drug-resistant exosomes compared with the wild-type counterparts. This intriguing result suggested that these confirmed miRNAs exhibited consistent expression patterns in both exosomes and parental cells and were potentially involved in the drug resistance mechanisms of BC through exosomal communication.

To investigate the impact of drug-resistant exosomal miRNAs on drug tolerance, wild-type BC cells were exposed to resistant-type exosomes and the relationship between miRNA expression levels and cell viability was analyzed. The radar plots revealed significant differences in miRNA expression patterns between wild-type and resistant-type BC cells (Fig. 4A). Notably, the miRNA expression patterns of wild-type BC cells became nearly identical to those of resistant-type BC cells after exposure to resistant-type exosomes. Furthermore, examination of the cell viability of wild-type BC cells following treatment with exosomes derived from resistant-type BC cells revealed a substantial increase in chemoresistance to both Adriamycin and Taxotere (Fig. 4B). Specifically, after Adriamycin treatment at a drug concentration equivalent to the IC₅₀ value of wild-type BC cells, a survival increase of 36.9, 10.1 and 9.6% was noted for MM231_wild-type, MM468_wild-type, and HCC1395_wild-type cells, respectively. Under the same conditions, treatment with Taxotere resulted in a survival increase of 27.9, 28.3 and 21.8% for BC cells. These findings indicated that exosomal miRNAs from resistant-type BC cells have the capacity to reshape miRNA expression patterns in wild-type BC cells and can confer the wild-type BC cells with an increased resistance to chemotherapy drugs, causing their behavior to more closely resemble that of drug-resistant BC cells.

Exosomes were isolated from the plasma of patients with BC. These plasma exosomes exhibited a particle size distribution of 30–350 nm in diameter, with an average of 155.3 nm (Fig. 5A). TDEs were specifically targeted using the immunoaffinity method, and then a specific population of exosomes of interest released from the tumor cells was isolated. This specificity allowed for focused analysis of exosomes relevant to research on drug-resistant miRNAs. Furthermore, SEM was employed to confirm whether the isolated TDEs possessed the characteristic features of exosomes. Microscopy revealed that the isolated TDEs expressed CD63, a positive tetraspanin marker associated with exosomes (35), and possessed the expected morphological and structural characteristics of exosomes (Fig. 5B). The expression of selected 12 drug-resistant miRNAs in 35 patients with BC were analyzed, including 20 and 15 patients showing no response and complete response following NAC, respectively, using real-time PCR. The raw Ct values of the 12 miRNAs are listed in Table SIV. The results showed that nine miRNAs (miR-21-5p, miR-125b, miR-146a-5p, miR-155, miR-222-5p, miR-484, miR-501, miR-1246-5p, and miR-1260b) were significantly upregulated in patients with BC who showed no response following NAC (Fig. 5C). These findings suggested that the upregulated miRNAs may be strongly associated with drug resistance in BC. No significant differences were observed in the other three miRNAs (miR-122-5p, miR-148a-5p and miR-210) between the groups in the patient cohort.

Considering the elevated levels of several miRNAs in patients with BC who exhibited no response to NAC, their roles as prognostic markers were determined using ROC curve analysis. A total of five miRNAs (miR-125b-5p, miR-146a-5p, miR-484, miR-1246-5p and miR-1260b) demonstrated relatively high discriminatory abilities, as indicated by their AUC values ranging from 0.817 to 0.898 (Fig. 6A). These miRNAs had the potential to serve as prognostic markers for identifying patients with BC likely to develop drug resistance following NAC. Conversely, miR-21-5p, miR-155-5p, miR-222-5p, and miR-501-5p showed moderate outcomes, with AUC values <0.8 and miR-122, miR-148a-5p, and miR-210 exhibited poor outcomes, with AUC values <0.6 (Table SV). The integrated model consisting of the five miRNAs showed even greater discriminatory ability, with an AUC of 0.950 (95% CI: 0.819–0.995; P<0.001), 75% sensitivity and 95% specificity (Fig. 6B). These results suggested that the combined 5-miRNA signature, among various combinations, could provide enhanced predictive outcomes for identifying patients with BC showing drug resistance (Fig. S3; Tables SVI and SVII). Therefore, compared with standalone miRNA markers, more effective combinations are proprosed, to maximize AUC and achieve improved diagnostic/prognostic accuracy (Table SVIII).

To reinforce the limited scale of clinical validation, a meta-analysis of datasets containing gene expression profiles in patients who received sequential NAC, such as doxorubicin, cyclophosphamide and paclitaxel, after histologically confirming primary invasive breast adenocarcinoma was performed to identify DEGs between the response and no-response groups. Analysis of GSE25066 identified 2,266 downregulated and 1,843 upregulated DEGs. Analysis of GSE41988 identified 455 downregulated and 564 upregulated DEGs (Fig. 7A). Subsequently, cutoffs of adjusted P-values <0.05 and |log2FC|>1 were used for each analysis. Moreover, GO analysis was performed for each significantly dysregulated DEG to identify GO terms representing the biological function of the genes. Clustering of GO terms based on their semantic similarity suggested two and seven GO clusters for the upregulated and downregulated genes, respectively (Fig. 7B). Notably, the analysis demonstrated that the upregulated GO clusters were enriched in the ‘regulation of cellular response to growth factor stimulus’, ‘intracellular transport’ and ‘cytokine-mediated signaling pathway’, strongly suggesting their involvement in drug resistance mechanisms. Conversely, the downregulated GO clusters were associated with the regulation of ‘DNA recombination’, ‘DNA replication’, ‘RNA splicing’, ‘mRNA processing’, and ‘regulation of mitotic cell cycle’, suggesting a potential impairment of the G2/M checkpoint and DNA repair pathways. These findings aligned with the documented effects of drug resistance mechanisms in solid tumors, as elucidated by several pivotal studies (36–38).

By integrating miRNAs, their respective target genes, and GO data, the authors successfully confirmed a drug resistance regulatory network. TargetScan predictions were utilized for identifying the target genes of the nine candidate miRNAs. To ensure reliability, the authors focused on high-confidence interactions with binding scores ≥90 (Table SIX). Subsequently, the miRNA-target gene interactions were subjected to GO analysis, with specific emphasis on the GO terms showing an association with drug-resistant genes. This narrowed the analysis to the molecular functions, biological processes, and biological pathways most relevant to drug resistance in BC (Fig. 8A). The significant drug-related categories of molecular function were ‘transcription regulator activity’, ‘transporter activity’ and ‘catalytic activity’. GO biological processes included ‘regulation of nucleic acid metabolism’, ‘signal transduction’, ‘apoptosis’, ‘cell communication’, and ‘cell growth and maintenance’. Moreover, the GO terms for target mRNAs of the nine candidate miRNAs were associated with multiple biological pathways participating in ‘mTOR signaling pathway’, ‘IFN-gamma pathway’, ‘IL3-mediated signaling events’, ‘ErbB receptor signaling network’, ‘VEGF and VEGFR signaling network’, ‘LKB1 signaling events’, ‘integrin family cell surface interactions’ and ‘TRAIL signaling pathway’. Notably, these results were similar to the findings in Fig. 7, presenting the analysis of the metadata for drug resistance genes, implying that the nine miRNAs are related to drug resistance and potentially regulate target genes to influence tumor response to chemotherapy. Furthermore, miRNA^{high R} target gene sets (BAK1, NOVA1, PTGER4, RTKN2, AGO1, CAP1 and ETS1) and miRNA^{moderate R} target gene sets (E2F2, ITGA3, SKP2, RIPK2 and STAT3) that are related to multiple biological activities related to drug resistance were revealed (Fig. 8B). Notably, NOVA1, ITGA3, SKP2 and RIPK2 were common after cross-validation of metadata analysis, DEMs and target DEGs, suggesting that these four genes are more importantly involved in drug resistance.

To evaluate the clinical significance of the identified drug-resistant miRNAs and their target genes, survival analyses was performed using data from a comprehensive collection of cancer studies, accessed through cBioPortal, encompassing thousands of patient samples filtered by BC classification. Kaplan-Meier curves and log-rank tests were performed for individual genes identified in the present study, followed by the analysis of combined gene-altered signatures involving miRNA^{high R} or miRNA^{moderate R}, which showed significant results (Fig. S4). The gene-altered group exhibited a decreased lifespan than the unaltered group, indicating a high correlation with patient prognosis. For example, the median overall survival in the miRNA^{high R} target gene signature-altered and unaltered groups was determined to be 110.77 and 154.50 months, respectively (Fig. 8C). Similarly, the median overall survival rates in the miRNA^{moderate R} target gene signature-altered and unaltered groups were 124.20 and 161.13 months, respectively (Fig. 8D). In addition, both gene signatures showed significant differences in progression-free survival and relapse-free survival (Fig. S5). These findings highlighted the prognostic significance of the drug resistance target genes, including BAK1, NOVA1, PTGER4, RTKN2, AGO1, CAP1, ETS1, E2F2, ITGA3, SKP2, RIPK2 and STAT3, in patients with BC. Taken together, these results provided compelling evidence for a strong association between the altered expression of drug-resistant genes induced by the regulation of drug-resistant miRNAs and poor prognosis among patients with BC.