Genomic characteristics of triple negative apocrine carcinoma: a comparison to triple negative breast cancer

Apocrine carcinoma is a rare breast cancer subtype. As such, the genomic characteristics of apocrine carcinoma with triple negative immunohistochemical results (TNAC), which has been treated as triple negative breast cancer (TNBC), have not been revealed. In this study, we evaluated the genomic characteristics of TNAC compared to TNBC with low Ki-67 (LK-TNBC). In the genetic analysis of 73 TNACs and 32 LK-TNBCs, the most frequently mutated driver gene in TNAC was TP53 (16/56, 28.6%), followed by PIK3CA (9/56, 16.1%), ZNF717 (8/56, 14.3%), and PIK3R1 (6/56, 10.71%). Mutational signature analysis showed enrichment of defective DNA mismatch repair (MMR)-related signatures (SBS6 and SBS21) and the SBS5 signature in TNAC, whereas an APOBEC activity-associated mutational signature (SBS13) was more prominent in LK-TNBC (Student’s t test, p < 0.05). In intrinsic subtyping, 38.4% of TNACs were classified as luminal A, 27.4% as luminal B, 26.0% as HER2-enriched (HER2-E), 2.7% as basal, and 5.5% as normal-like. The basal subtype was the most dominant subtype (43.8%) in LK-TNBC (p < 0.001), followed by luminal B (21.9%), HER2-E (21.9%), and luminal A (12.5%). In the survival analysis, TNAC had a five-year disease-free survival (DFS) rate of 92.2% compared to 59.1% for LK-TNBC (P = 0.001) and a five-year overall survival (OS) rate of 95.3% compared to 74.6% for LK-TNBC (P = 0.0099). TNAC has different genetic characteristics and better survival outcomes than LK-TNBC. In particular, normal-like and luminal A subtypes in TNAC have much better DFS and OS than other intrinsic subtypes. Our findings are expected to impact medical practice for patients diagnosed with TNAC.


INTRODUCTION
Apocrine carcinoma is a rare breast cancer (BC) subtype accounting for 1-4% of all BCs 1 . The 2019 WHO classification of breast tumors defined apocrine carcinoma as a distinct, special type of BC characterized by large round nuclei with plump, eosinophilic, granular, and sharp-bordered cytoplasm 2 . Apocrine carcinomas frequently express androgen receptor (AR) but are estrogen receptor (ER) and progesterone receptor (PgR) negative 3 . Therefore, apocrine carcinomas may be divided into two BC subtypes based on human epidermal growth factor receptor-2 (HER2) status: triple negative breast cancer (TNBC) and HER2-positive BC 1,3 .
TNBC is ER-and PgR-negative, does not overexpress HER2, and is usually included in the "basal-like BC" group according to the intrinsic subtype [4][5][6][7] . TNBCs account for 15% of all BCs, and they generally have poor clinical outcomes 4,7 . While this aggressive phenotype is common, there is a small group of TNBCs with favorable clinical outcomes showing a low risk of recurrence and death 8 . Low-risk TNBCs comprise salivary gland-type BC, TNBC with high tumor-infiltrating lymphocytes (TILs), and carcinoma with apocrine differentiation 9 . Apocrine carcinoma with triple negative immunohistochemical results (TNAC) has a low Ki-67 labeling index and is very different from most aggressive TNBCs, although the underlying molecular characteristics of apocrine breast cancer have been poorly studied 10 .
Previous study has demonstrated that TNAC has a poorer response to (neo)adjuvant chemotherapy than other non-apocrine TNBC 11 . However, survival analyses have shown discordant results. One previous study showed similar survival outcomes in TNAC and TNBC, whereas another reported poor prognosis in TNAC, and yet another reported better survival outcome in TNAC than in TNBC 9,12,13 . Therefore, comprehensive genomic profiling of TNAC could provide a detailed understanding of the molecular biology and prognosis and suggest potential therapeutic targets for individualized treatment. In this study, we evaluated the genomic and clinical characteristics of TNAC and compared them to those of TNBC with low Ki-67 expression.

METHODS Patients
Tumors were selected from 73 TNAC patients who underwent curative surgery at Samsung Medical Center, Seoul, Korea. In addition, we selected 32 TNBCs that had Ki-67 levels similar to those of TNACs from patients who underwent curative surgery. Independent pathological review was performed to determine the apocrine type and triple negativity of BC. The Institutional Review Board of Samsung Medical Center approved the study protocol (IRB No: 2020-05-159).

Whole-exome sequencing
Extraction of DNA. DNA was extracted from formalin-fixed paraffinembedded (FFPE) blocks or fresh frozen (FF) tissues. A skilled pathologist reviewed hematoxylin and eosin (H&E)-stained sections from the FFPE blocks and outlined areas containing representative invasive breast carcinoma on the slide. FFPE slides were manually microdissected to maximize tumor purity. Genomic DNA was extracted using the ReliaPrep FFPE gDNA Miniprep System (Promega), the QIAamp DNA Mini Kit (Qiagen), and the QIAamp DNA Blood Mini Kit (Qiagen) for FFPE, FF tissues, and buffy coats, respectively.
Sequencing: Illumina NovaSeq6000 platform. For generation of standard exome capture libraries, we used 1 µg of input gDNA and the Agilent SureSelect Target Enrichment protocol for generating an Illumina pairedend sequencing library. In all cases, the SureSelect Human All Exon V6 probe set was used. DNA concentration and DNA quality were measured by PicoGreen and agarose gel electrophoresis. We used 1 μg of each cell line's genomic DNA diluted in EB buffer and sheared to a target peak size of 150-200 bp using the Covaris LE220 focused-ultrasonicator (Covaris, Woburn, MA) according to the manufacturer's recommendations. An 8 microTUBE Strip was loaded into the tube holder of the ultrasonicator, and DNA was sheared using the following settings: mode, frequency sweeping; duty cycle, 10%; intensity, 5; cycles per burst, 200; duration, 60 s × 6 cycles; temperature, 4-7°C. The fragmented DNA was repaired, an 'A' was ligated to the 3′ end, and Agilent adapters were ligated to the fragments.
Once ligation had been assessed, the adapter ligated product was PCR amplified. The final purified product was then quantified using the TapeStation DNA screentape D1000 (Agilent). For exome capture, 250 ng of DNA library was mixed with hybridization buffer, blocking mixture, RNase block and 5 µl of SureSelect all exon capture library, according to the standard Agilent SureSelect Target Enrichment protocol. Hybridization to the capture baits was conducted at 65°C using a heated thermal cycler lid option at 105°C for 24 h on a PCR machine. The captured DNA was then washed and amplified. The final purified product was then quantified using qPCR according to the qPCR Quantification Protocol Guide (KAPA Library Quantification kit for Illumina Sequencing platforms) and qualified using the TapeStation DNA screentape D1000 (Agilent). Indexed libraries were then sequenced using the NovaSeq6000 platform (Illumina, San Diego, USA) by Macrogen Incorporated.
Sequence alignment. Paired-end reads were aligned to the GRCh38 human reference genome using the Burrows Wheeler aligner 14 (BWA v0.7.17), and BAM (binary alignment/map) files were produced for each sample. Local realignment and quality score recalibration processes were performed using Genome Analysis Toolkit 15 (GATK v4.2). Sequencing quality was evaluated using fastQC and the Picard CollectMultipleMetrics tool (https://broadinstitute.github.io/picard/). Mutations were called for each duplicate tumor sample with matched control. For nonduplicated tumor samples (C009, C012), fastq files were split randomly into two using the SeqKit tool. Somatic point mutations and short indels were called using MuTect2 16 from GATK v4.2. The filtering process was performed to reduce sequencing or data processing errors from FFPE breast samples 17 . We selected only variants detected in both duplicate bam files for one sample. To detect significantly mutated driver genes, the dNdScv algorithm 18 was applied. We selected variants with p values less than 0.001 in TNAC and LK-TNBC.
TCGA luminal A and TCGA TNBC somatic mutations were downloaded from the GDC data portal 5 .
Mutational signature analysis. Mutational signature analysis was performed with the deconstructSigs 19 R package. It calculates the proportion of 30 COSMIC signatures for each sample. We selected signatures that showed a significant difference in proportion between TNAC and LK-TNBC, and between hypermutated samples and nonhypermutated samples in TNAC. Signatures that showed no significant difference when outliers were removed were excluded. Significance was calculated by Student's t test.
To examine DNA mismatch repair (MMR)-related signatures in more detail, we assessed the frequency of MMR-related gene mutations . A total  of 13 MMR-related genes, including MSH2, MSH3, MSH4, MSH5, MSH6,  MLH1, MLH3, PMS1, PMS2, POLD1, POLB, POLE, and POLG, were used for  analysis. Somatic copy number alteration analysis. CNVkit v0.9.9 was used to call somatic copy number alterations (SCNAs) on whole-exome sequencing data. In the results, log 2 depth ratios greater than 0.2 were considered to indicate copy number gain, and ratios less than −0.2 were considered to indicate copy number loss.
The significance of somatic copy number alterations was evaluated using GISTIC2.0 20 (Version 2.0.23), which deconstructs SCNAs into broad and focal events. We added the ranges of 1,000,000 paddings considering information on centromere and telomere positions from UCSC hg38. Cancer-related gene sets used for matching the genes covering the resulting peaks were downloaded from the COSMIC Cancer Gene Census (CGC) database.

Whole-transcriptome sequencing
Extraction of RNA. RNA was extracted from formalin-fixed paraffinembedded (FFPE) blocks or fresh frozen (FF) tissues. A skilled pathologist reviewed hematoxylin and eosin (H&E)-stained sections from the FFPE blocks and outlined areas containing representative invasive breast carcinoma on the slide. FFPE slides were manually microdissected to maximize tumor purity.
Total RNA was extracted using the ReliaPrep FFPE Total RNA Miniprep System (Promega) and RNeasy Mini Kit (Qiagen) for FFPE and FF specimens, respectively, according to the manufacturer's protocol. Nucleic acid yield and purity were assessed using a NanoDrop ND-1000 Spectrophotometer (NanoDrop Technologies, Thermo-Fisher Scientific, MA, USA).
Sequencing: Illumina NovaSeq6000 platform. The total RNA concentration was calculated by Quant-IT RiboGreen (Invitrogen). To determine the DV200 (% of RNA fragments >200 bp) value, samples were run on the TapeStation RNA screentape (Agilent). Then, 100 ng of total RNA was subjected to sequencing library construction using a TruSeq RNA Access library prep kit (Illumina, San Diego, CA, USA) according to the manufacturer's protocols. Briefly, the total RNA was fragmented into small pieces using divalent cations under elevated temperature. The cleaved RNA fragments were copied into first-strand cDNA using SuperScript II reverse transcriptase (Invitrogen, #18064014) and random primers. This was followed by second-strand cDNA synthesis using DNA polymerase I, RNase H, and dUTP. These cDNA fragments then went through an end-repair process; single 'A' bases were added, and the adapters were ligated. The products were purified and enriched via PCR to create cDNA libraries. All libraries were normalized, and six libraries were pooled into a single hybridization/capture reaction. Pooled libraries were incubated with a cocktail of biotinylated oligos corresponding to coding regions of the genome. Targeted library molecules were captured via hybridized biotinylated oligo probes using streptavidinconjugated beads. After two rounds of hybridization/capture reactions, the enriched library molecules were subjected to a second round of PCR amplification. The captured libraries were quantified using KAPA Library Quantification kits for Illumina Sequencing platforms according to the qPCR Quantification Protocol Guide (KAPA BIOSYSTEMS, #KK4854) and qualified using the TapeStation D1000 ScreenTape (Agilent Technologies, # 5067-5582). Indexed libraries were submitted to an Illumina NovaSeq6000 (Illumina, Inc.), and paired-end (2×100 bp) sequencing was performed by Macrogen Incorporated.
Sequence alignment and quantification. Whole transcriptome sequencing reads were mapped to the GRCh38 human reference genome from GENCODE v38 using STAR v2.7.9a. Fragments per kilobase of exon model per million mapped fragments (FPKM) values were obtained using RSEM v1.3.2. Sequencing quality was evaluated using fastQC and FastQ Screen 0.14.0.
Batch correction. Batch effects between FF and FFPE samples were corrected using ComBat-seq 21 , and then FPKM normalization was performed using the convertCounts function in DGEobj.utils of the R package. To conduct batch correction between the TCGA cohort and our batch-corrected dataset, the ComBat 22 function in the sva package was used. The expression profiles were transformed as follows: log 2 (FPKM + 1).
Intrinsic subtype prediction. Intrinsic subtypes of each sample were determined by the 'molecular.subtyping' function in the genefu 23 bioconductor package (v2.26.0). The subtyping classification model was 'pam50', which identifies breast cancer molecular subtypes as luminal A, luminal B, HER2-enriched, basal, and normal-like subtypes based on the PAM50 genes. Hierarchical clustering based on the Ward D2 method with a Euclidean distance measure was used to co-cluster our cohort and the TCGA BRCA cohort based on the PAM50 genes. We transformed the expression matrix by log 2 (FPKM + 1) and then used pheatmap v1.0.12 in the R package to scale the results.
Burstein and FUSCC subtyping. In Burstein subtyping 26 , hierarchical clustering based on the Ward D2 method with a Euclidean distance measure was used to cluster our cohort into Burstein subtypes of luminal androgen receptor (LAR), mesenchymal (MES), basal-like immunosuppressed (BLIS), and basal-like immune-activated (BLIA). Fifty-five genes significantly overexpressed in each subtype defined by Burstein et al. 26 were selected for clustering to determine the characteristics of TNAC and LK-TNBC. We transformed the expression matrix by log 2 (FPKM + 1) and then used pheatmap v1.0.12 in the R package to scale the results.
In FUSCC subtyping 27 , hierarchical clustering based on the Ward D2 method with a Euclidean distance measure was used to co-cluster our cohort and the TCGA TNBC cohort to identify the subtype mainly grouped with TNAC and LK-TNBC. The TCGA TNBC cohort annotated with both Lehmann subtype and FUSCC classification was provided by Jiang, Yi-Zhou, et al. 28 . The top 2000 most variable genes in the TCGA cohort were selected for clustering. We transformed the expression matrix by calculating log 2 (FPKM + 1) and then used pheatmap v1.0.12 in the R package to scale the results.
Differential gene expression (DGE) analysis and enriched pathway analysis. DGE analysis between TNAC and LK-TNBC was performed using DESeq2. Genes with |log 2 -fold change| > 2, p value < 0.01, and adjusted p value < 0.01 were considered differentially expressed genes (DEGs). We performed gene set enrichment analysis 29 (GSEA v3.0) to identify pathways enriched in TNAC or LK-TNBC. Pathways with FDR q-values less than 0.05 were selected.
DGE analysis was also performed between intrinsic subtypes of TNAC using NOISeqBIO 30 to compare one subtype to the others. Genes with | log 2 -fold change| > 1 and probability > 0.95 were considered DEGs. Gene set enrichment analysis was performed using Enrichr 31 . Significantly enriched pathways from MSigDB Hallmark 2020, KEGG 2021 Human, and GO Biological Process 2021 were selected (adjusted p value < 0.05).
Cell type estimation analysis. The abundance of 29 cell types as defined in Wu et al. 32 was used for deconvolution analysis of our bulk WTS data. CIBERSORTx 33 was mainly performed to estimate cell types for each sample, and MuSiC 34 v0.2.0 and BisqueRNA 35 v1.0.5 were used to verify the results of CIBERSORTx. Cell types that showed significant differences with at least one of these two additional tools were selected. Significance was calculated by the Wilcoxon signed-rank test.
Gene fusion analysis. Arriba 36 v2.1.0 was used to detect gene fusions and selected those with confidence of 'high' or 'medium'. Fusion genes detected in healthy individuals (n = 87) 37 with a frequency of less than 0.05 were filtered. The difference in the frequency of gene fusion events was compared between TNAC and LK-TNBC and between intrinsic subtypes. Significance was calculated by the Wilcoxon signed-rank test and Kruskal-Wallis test.

Statistical analysis
We evaluated the differences between TNAC and Ki-67-matched TNBC (low Ki-67 TNBC; LK-TNBC) using Fisher's exact test. Disease-free survival (DFS) was defined as the duration of survival without any signs or symptoms of disease after primary treatment. Overall survival (OS) was defined as the duration between curative surgery and death. DFS and OS were analyzed using the Kaplan-Meier method. Cox proportional hazard regression was used to estimate hazard ratios and 95% confidence intervals (CIs).

Baseline characteristics
We described the baseline clinical and pathological characteristics of TNAC and LK-TNBC in Supplementary Table 1. Only stage at diagnosis was different between TNAC and LK-TNBC (P = 0.03), while no significant differences were observed in other characteristics, including nuclear grade, histologic grade, Ki-67, and status of (neo)adjuvant treatment.
We found eight TNAC and seven LK-TNBC samples with a high tumor mutational burden (TMB) (> 10 mutations per Mb) ( Fig. 2A). High-TMB TNAC samples had more than 20 times higher TMB values (mean value: 368.72) than high-TMB LK-TNBC samples (mean value: 17.85). Previously, APOBEC activity (SBS2 and SBS13) and MMR-related mutational signatures (SBS3, SBS6, SBS10, SBS15, and SBS20) were reported to be the most common in high-TMB BC samples 41 . In our TNAC cohort, SBS20 was significantly enriched in hypermutated samples compared to nonhypermutated samples (Student's t test, p < 0.05) (Fig. 2B)  Somatic copy number alterations Overall, somatic copy number alterations (SCNAs) in the entire genome were less prevalent in TNAC than in LK-TNBC (Wilcoxon signed-rank test, p < 0.05) (Supplementary Fig. 1). However, TNAC had significantly more frequent amplification of 7p and 8q, and deletion of 6q, 9p, and 18 than LK-TNBC, with q-values less than 0.05 based on a GISTIC 20 analysis (Supplementary Table 2). All the TNAC-specific broad events of significant gains and losses were frequently observed in the luminal A subtypes of the TCGA BRCA dataset 5 . In addition, 6q was reported to be frequently deleted in the luminal androgen receptor (LAR) subtype among Burstein TNBC subtypes 26 .
Burstein 26 and Fudan University Shanghai Cancer Center (FUSCC) 27 TNBC subtypes were also evaluated ( Supplementary  Figs. 2 and 3). The expression levels of a total of 55 genes representing each Burstein subtype were used to predict subtypes of TNAC and LK-TNBC. The cluster with high expression levels of genes related to the Burstein LAR subtype included most TNACs previously assigned to the intrinsic luminal A and Lehmann LAR subtypes. In FUSCC TNBC subtyping, many TNACs (34/73, 46.6%) and TCGA TNBCs with the FUSCC LAR subtype were sorted into the same cluster, whereas only four LK-TNBCs (4/32, 12.5%) were included in the FUSCC LAR subtype. Fig. 4, Supplementary Table 3). The level of expression of CTNNB1, which was enriched in basal-like breast cancer and associated with a poor prognosis 43 , significantly decreased in TNAC. However, apocrine carcinoma-associated genes such as ACSM1, FABP7, and HMGCS2 44 showed significant upregulation in TNAC compared to LK-TNBC (Wilcoxon signed-rank test, p < 0.05) ( Supplementary Fig.  5, Supplementary Table 4). GSEA revealed that metabolic processes of hormones and acids were enriched in TNAC, while epigenetic regulation of gene expression and chromatin organization-related pathways were significantly enriched in LK-TNBC (Supplementary Table 5).

Differentially expressed genes in TNAC Differential gene expression (DGE) analysis between TNAC and LK-TNBC showed that TNAC had 2082 upregulated genes and 162 downregulated genes compared with LK-TNBC (Supplementary
DGE analysis and GSEA of the intrinsic subtypes of TNAC showed that epithelial mesenchymal transition (EMT) and TNFalpha signaling pathways were upregulated, but E2F targets and mTORC1 signaling pathways were downregulated in luminal A TNAC samples ( Supplementary Fig. 6a, Supplementary Table 6). These results were consistent with previous studies reporting that EMT was involved in the pathogenesis of apocrine carcinoma 45 and that the E2F target gene set was more enriched in the basal subtype than in the luminal or normal-like subtype 46 . Enriched KEGG pathways and GO biological processes are described in Supplementary Fig. 6b, c and Supplementary Table 6.
Tumor microenvironment of TNAC Different tumor microenvironments (TMEs) in TNAC and LK-TNBC were identified by the CIBERSORTx algorithm 33 . Macrophages, CD4+ T cells, cancer HER2_sc, and mature luminal cell types were significantly enriched in TNAC compared to LK-TNBC (Wilcoxon signed-rank test, p < 0.05, Supplementary Fig. 7). In contrast, cancer cycling and cancer myeloid cell types were more prominent in LK-TNBC than in TNAC (Wilcoxon signed-rank test, p < 0.05).
Gene fusion FGFR2-TACC2 fusion, which was discovered as a novel rearrangement in TNAC 39 , was found in one TNAC patient in our cohort. The Fig. 1 Somatic Point Mutations in TNAC and LK-TNBC. A The top 16 significantly mutated genes as determined by the dNdScv algorithm (p < 0.001, TNAC: 13 genes, LK-TNBC: 5 genes) in TNAC and LK-TNBC were sorted by their mutation frequency. The right bar plot shows the mutation frequencies of the 16 genes in our cohorts and TCGA cohorts. The red dotted vertical line represents a 5% mutation frequency. The top bar plot indicates tumor mutational burden (TMB), which represents the number of mutations per megabase (Mb) for each sample. The samples with TMB values greater than 10 were considered hypermutated (black dotted horizontal line). The clinical and pathological characteristics were also annotated. The significant differences between TNAC and LK-TNBC according to baseline characteristics were calculated by Fisher's exact test. B Schematic representation of somatic point mutations in TP53, PIK3CA, and PIK3R1 in our TNAC and LK-TNBC cohorts.
list of gene fusion events identified from our TNAC and LK-TNBC cohorts is described in Supplementary Table 7.
The frequency of gene fusion events varies significantly depending on tumor type and intrinsic subtype ( Supplementary  Fig. 8). TNAC showed fewer gene fusion events than LK-TNBC (Wilcoxon signed-rank test, p < 0.05). Interestingly, the luminal A subtype had the lowest frequency of gene fusions, and the frequency was significantly increased in the order of luminal B, HER2-E, and basal subtype (Kruskal-Wallis, p < 0.05).
Survival analysis DFS and BC-specific OS were analyzed in both TNAC and LK-TNBC. The median follow-up duration was 71.3 months (interquartile range: 41.2, 90.0). In terms of both DFS and OS, TNAC had superior survival outcomes compared with LK-TNBC. Thus, the five-year DFS rate was 92.2% vs. 59.1% in TNAC vs. LK-TNBC (P = 0.001), and the five-year OS rate was 95.3% vs. 74.6% (P = 0.0099) (Fig. 5A). Among other clinical characteristics, stage and Ki-67 were associated with BC prognosis (Supplementary Fig. 9a, b). Additional survival analyses were performed on TNAC and LK-TNBC of the same stage. In stage II, TNAC had better DFS than LK-TNBC ( Supplementary Fig. 9c).

DISCUSSION
In this study, we revealed that TNAC had different genomic characteristics compared to TNBC with low Ki-67. TP53 and PIK3CA mutations were less frequently observed in TNAC than in LK-TNBC, and other mutations were found at different frequencies between TNAC and LK-TNBC. PIK3R1 mutation was observed only in TNAC. The APOBEC signature was more dominant in LK-TNBC, whereas high TMB was more frequently observed in TNAC and was associated with mutation of MMRassociated genes. Considering intrinsic subtypes, the luminal A subtype was more frequent in TNAC, while the basal subtype was more common in LK-TNBC. DFS and OS were different for TNAC and LK-TNBC. In addition, TNBC subtypes differed between TNAC and LK-TNBC.
In TNBC, TP53 was the most commonly detected mutation. In the TCGA dataset,~80% of TNBCs harbored TP53 mutations, while only 9% harbored PIK3CA mutations 5 . Another study suggested that TP53 mutation was found in 80% of ER-negative BCs, with PIK3CA mutation found at a frequency of 20% 47 . In our study, TP53 mutation was observed in 28.5% of TNACs and 70.4% of LK-TNBCs. PIK3CA mutation was observed in 44.4% of LK-TNBCs, similar to the rate in luminal A type samples in the TCGA cohort 5 . PIK3R1 mutation was found in~1-3% of BCs regardless of HR and HER2 status 5 . Our study suggested that~10% of PIK3R1 mutations were exclusively observed in TNAC.
Ki-67 is one of the most important prognostic biomarkers in TNBC. A previous study suggested that high Ki-67 indicated a poor prognosis compared to low Ki-67 in TNBC 48 . Recent studies reported that TNAC had a lower Ki-67 score and better prognosis than TNBC 3,49 . However, the researchers did not perform multivariate analysis or evaluation of factors affecting prognosis between TNAC or Ki-67 score. Therefore, we analyzed the genetic characteristics and prognosis of TNAC and TNBC while considering Ki-67 expression. We chose LK-TNBC because TNAC generally has a low Ki-67 score. TNBC with low Ki-67 was not frequently observed, and only 25% of TNBCs had a Ki-67 value less than 30% 48 . In our study, 72% and 92% of TNBC and TNAC patients had Ki-67 less than 30%.
Mutation type The proportions of mutational signatures also differed between TNAC and LK-TNBC. MMR-associated signatures were dominant in TNAC, whereas APOBEC signatures were more prevalent in TNBC. Hypermutation was also frequently observed in TNAC with genetic alteration of MMR-associated genes. In previous study, the MMR signature is rarely detected in BC, but the APOBEC signature is common 50 . The MMR signature of TNAC showed unique genetic characteristics, suggesting this as a surrogate biomarker for the response to ICIs 51 .
Immune checkpoint inhibitor (ICI) treatment combined with chemotherapy has been approved as a standard treatment strategy for TNBC, and an understanding of the TME is important to treat TNBC patients [52][53][54] . Currently, programmed death-ligand 1 status and tumor-infiltrating lymphocyte (TIL) levels are   predictive biomarkers for ICI treatment response in TNBC. However, additional predictive biomarkers for ICI are needed 54,55 .
In our TME analysis, the CD4+ T-cell subset and macrophages were enriched in TNAC compared to TNBC. This suggests the potential applicability of ICI treatment for TNAC, as CD4+ T-cell infiltration is associated with the antitumor immune response and the ICI response 56 . The intrinsic subtype is a traditional biomarker of BC 6 and has been used as a surrogate biomarker for adjuvant chemotherapy and BC-specific survival 57,58 . Generally, TNBC is categorized into the basal-like subtype, and hormone receptor+HER2-BC is the luminal A subtype 57   were strongly related, TNAC remained an important prognostic factor for DFS after statistical adjustment. Treatment guideline indicates chemotherapy for early TNBC as systemic induction therapy or postoperative chemotherapy 59 . They recommended no systemic therapy for apocrine tumors because of low-risk endocrine nonresponsive histology. However, apocrine carcinomas are generally categorized into TNBC or HER2 + BC according to IHC, and they need to be treated with chemotherapy and/or anti-HER2 treatment 3 . In our study, TNAC, which was categorized into normal-like and luminal A subtypes, had very good DFS and OS compared to that of other intrinsic subtypes. Therefore, we suggest that early-stage TNAC does not need adjuvant chemotherapy if it is the normal-like or luminal A intrinsic type. This would prevent unnecessary chemotherapy, which induces several toxicities, including alopecia, nausea/ vomiting, cytopenia and severe infections 60 .
This study has limitations. First, this study was not a prospective clinical trial with intervention; therefore, we could not evaluate specific drug responses according to TNAC and LK-TNBC. In addition, we generated sequencing data using FFPE and FF tissue samples because of the limited availability of TNAC and LK-TNBC samples. To overcome the batch effect between FFPE and FF tissues, we analyzed sequencing data after batch correction. This is the first attempt to comprehensively assess the genomic characteristics of TNAC. Moreover, this study evaluates the genetic signatures of TNAC and could inform treatment decisions regarding adjuvant chemotherapy in early TNBC. Further prospective clinical trials and parallel translational research are warranted.