Breast cancer patient samples from minority patients were collected (Table I). Patients belong to the Los Angeles SPA6 (Service Planning Area 6) region, which traditionally shows significant health disparity. A total of thirteen African-American and fifteen Hispanic patients were analyzed for the study with confirmed invasive ductal carcinoma, except for one African-American sample with a Ductal Carcinoma in situ (DCIS) diagnosis. We also utilized matched normal (tumor-adjacent) for eight samples from each ethnicity. Patient samples were used from our ongoing Breast Cancer Study in the Division of Cancer Research and Training at the Charles R. Drew University of Medicine and Science in collaboration with Martin Luther King Ambulatory Care Center (formerly known as King-Drew Medical Center, #IRB 00-06-041) and the protocol has been approved since 1999 and continuing review approved annually (recent continuing review approval was August 18, 2021. The patient demographics and breast cancer subtypes are listed in Table I.

Total DNA was isolated from fresh-frozen breast cancer biopsies and matched normal tissues using the QIAGEN QiaAmp DNA purification kit and Quantified using Nanodrop (Thermo Fischer Scientific, USA). 250 to 500 ng DNA was used to prepare the library using Nextera Flex for Enrichment (Illumina, USA, Cat No 20025524). Libraries were run on a Bioanalyzer DNA 1000 chip (Agilent, USA) to assess quality. Library quantitation was done using the qubit 3.0 fluorometer (Thermo Fischer Scientific, USA) using the dsDNA high sensitivity kit (ThermoFisher, USA, Cat No. Q32851). The exome capture probes cover about 45 Mb of the primarily protein-coding region of the human genome (hg19 assembly). The bed file with the chromosomal coordinates is available at: (https://support.illumina.com/content/dam/illumina-support/documents/downloads/productfiles/nextera/nextera-flex-for-enrichment/TruSeq_Exome_TargetedRegions_v1.2.bed).

250 to 500 ng DNA was subjected to library preparation using the Nextera Flex for Enrichment (Currently, DNA Prep for Enrichment, Illumina, USA) following manufacturer-recommended protocol. Briefly, DNA was ‘tagmented’ Tagmentation is the initial step in library prep where genomic DNA is cleaved and tagged for analysis, cut into small pieces of 300–400 bp length by a transposase, and bead-linked transposes ligated adaptor for sequencing in the same process. The tagmentation was followed by captured by biotinylated oligonucleotides covering approximately 45 Mb of human genomic exons (sequence version UCSC hg19). Finally, exome library capture hybridization was performed for 1.5 h. Twelve patient pools were run on a NextSeq 550 sequencer (Illumina, USA) with paired-end 70 bp reads.

Fastq generated by the NextSeq 550 run was mapped to the hg19 human genome using the Burrows-Wheeler Aligner (BWA, BWA mem), and variants were identified using the Genome Analysis Toolkit (GATK). To this end, we utilized Illumina BaseSpace ‘BWA Enrichment’ pipeline (https://www.illumina.com/products/by-type/informatics-products/basespace-sequence-hub/apps/bwa-enrichment.html). Mapping was restricted to the chromosomal regions mentioned in ‘TruSeq_Exome_TargetedRegions_v1.2.bed’. This ‘bed’ file specifies the coordinates of the regions used to generate probes targeting all known protein-coding genes in the human genome. Variants were annotated using Illumina Variant Annotator. For the figures shown below, the variant call format (VCF) file generated by the Illumina BWA enrichment pipeline was filtered for the common/germline variants in the QIAGEN QCI software platform using the following strategy. The following criteria were used to exclude variants from the cancer exome VCFs. Variants that are present >=1% of Allele Frequency Community (QIAGEN Database) or >=3% in the following: the Genome Aggregation Database (gnomAD http://gnomad.broadinstitute.org) or ExAC (https://exac.broadinstitute.org, currently merged with gnomAD) or of NHLBI GO Exome Sequencing Project (NHLBI ESP Exomes: http://evs.gs.washington.edu/EVS/) Or 1000 Genomes project (https://www.internationalgenome.org). Variants were also excluded if present in the dbSNP database. However, common germline variants were kept for analysis if established as pathogenic variants with support from literature published. Matched normal samples were also used to filter out germline and common variants. However, variants with known pathogenicity in any disease were included in the analysis. The QIAGEN Clinical Insight contains gnomAD (v2.1.1), Exome Variant Server (EVS, vESP6500SI–V2), 1000 Genome Frequency (phase3v5b), Single Nucleotide Polymorphism Database (dbSNP v154), Combined Annotation Dependent Depletion (CADD v1.6), Sorting Intolerant from Tolerant (SIFT4G v2016-02-23), BSIFT (2016-02-23), Polymorphism Phenotyping (PolyPhen-2 v2.2.2) versions (PhyloP (2009–11). The version information was obtained from the release notes available at the https://variants.ingenuity.com/qci/website. The filtered VCF was annotated using Ensemble Variant Effect Predicted (VEP) (9), which was then converted to Mutation Annotation Format (MAF) using vcf2maf script (https://github.com/mskcc/vcf2maf). All figures were generated using the Maftools R package [R version 4.10 (2021-05-18) and maftools 2.8.0] (10). Mutational signatures were determined by maftools ‘extractSignature’ and Plot Signature functions. Finally, mutational profiles were compared with the current Single base substitutions (SBS) signature from the Catalogue Of Somatic Mutations In Cancer (COSMIC) database (https://cancer.sanger.ac.uk/signatures/sbs/). The overall survival analysis was conducted using maftools mafSurvival function and comparison between patients with high or low DFS was carried out with mafCompare function for Fisher's exact test. The survival analysis utilized Cox proportional hazard function and correlated gene mutations individually or in groups on the overall survival of the patients.

Tumor Mutational Burden (TMB) was calculated as total nonsynonymous mutations per mb (megabases, log2 transformed, per mb is calculated from the capture size of the exome capture baits). The statistical test was performed with Graphpad Prism 9 for the Mann-Whitney U test or maftools for the pairwise t-test. All differences in variant comparison between groups of the sample were calculated using 2×2 Fisher's Exact Test in maftools. In all cases, P-value <=0.05 was considered significant. We have used MutSig v1.4 to analyze significantly mutated genes in African-American and Hispanic samples. The genes with a q-value less than 0.0001 were analyzed for enrichment using the ShinyGO v.0741 web portal (http://bioinformatics.sdstate.edu/go/) against the Gene Ontology Biological pathways and the Hallmark Dataset form The Molecular Signatures Database (MSigDB) (11,12). For MutSig cancer driver identification, we used variants with the ‘PASS’ criteria attached to the variants of interest in the final filtered variant list for our samples utilizing the maftools prepareMutsig function. For the TCGA cohorts, we directly exported the MutSig Compatible variant list file for analysis from the GDC mc3 maf file. We did the comparative analysis in the ‘Ingenuity QCI’ (app.ingenuity.com/). The result includes SIFT (https://sift.bii.a-star.edu.sg/) and Polyphen (http://genetics.bwh.harvard.edu/pph2/) scores (13,14). These scores predict the functional consequences of a variation for a protein. We filter the variants to predict detrimental ‘damaging’ mutations and ‘activating’ mutations.

TCGA was accessed via cBioPortal using the general web interface (cbioportal.org). We also downloaded the publicly available ‘mc3.v0.2.8.PUBLIC.maf.gz ‘from the https://gdc.cancer.gov/about-data/publications/mc3-2017 website and used clinical data available from cBioPortal for the TCGA Breast Cancer (TCGA-BRCA) cohort to subset the maf file into African-American and Hispanic categories. We used the clinical category ‘Race’ as ‘Black or African-American’ (Total 162 extracted) and the ‘Ethnicity’ category ‘Hispanic’ (total 33 extracted) to create the ethnicity-specific mafs using the subset Maf function in maftools. We analyzed these ethnic categories for comparative analysis with our cohort of patients. COSMIC Census genes were downloaded from Sanger's COSMIC site (https://cancer.sanger.ac.uk/census). Venn diagram comparison of somatic mutations was conducted using the web portal http://genevenn.sourceforge.net/vennresults.php.

We analyzed variants in total of 13 African-American and 15 Hispanic samples. We also utilized TCGA breast cancer data set with 163 African-American and 33 Hispanic breast cancer samples. The comparative analysis was conducted using an unpaired Mann-Whitney U test, one-way ANOVA with Tukey's multiple comparison or Fisher's exact test (comparison between high and low disease-free survival groups). The P-value generated by MutSig for a gene is a combination of three P-values for the mutation abundance, location in the genome and the conservation of genetic sequence across species. Details of the MutSig algorithm are available at https://www.broadinstitute.org/cancer/cga/mutsig. The data used for ANOVA and Mann-Whitney U test are available at the synapse project page (project ID, syn42137028). GraphPad Prism version 9, GraphPad Inc., San Diego, USA was used for one-way ANOVA with Tukey's multiple comparison and Mann-Whitney U test. Fisher exact test was carried out using maftools in R [Maftools R package (R version 4.10 (2021-05-18, http://cran.r-project.org/) and maftools 2.8.0, https://github.com/PoisonAlien/maftools/]. Kaplan-Meier plots were generated with univariate cox proportional hazard model analysis using maftools. In all cases, P-value <=0.05 was considered significant.

Our current study examines patient samples from a small cohort of ethnic minority groups, with an overall of 13 African-American (AA) and 16 Hispanic/Latinx Tumor samples (Table I Method section, 15 samples for the Hispanic group analyzed). Overall, our exome analysis recovered single nucleotide variations (SNVs) and a small number of insertions and deletions. The average coverage for African-American and Hispanic tumor samples was ~52X and ~48X, respectively, excluding one Hispanic sample with low coverage (excluded from analysis). The median Tumor mutational burden (calculated a log2 (missense mutations/Mb of capture size) were 0.98 and 0.89 in African-American and Hispanic samples, respectively (Fig. 1). The TMB values were slightly lower than the TCGA cohorts. However, we did not observe any statistically significant difference between our and TCGA samples. After filtering for common variants, the African-American samples had 647 mutations (SNVs including insertion and deletion). On the other hand, Hispanic samples had 594 mutations (Table SI: Summary of mutation after applying common variant filters).

In either case, while including known pathogenic mutations, the significantly mutated genes were F5 (Coagulation Factor V, p.Q534R) and PRSS1 (Serine Protease 1) in the African-American samples. On the other hand, Both African-American and Hispanic samples had MTHFR (methylenetetrahydrofolate reductase, p.E470A, and p.A263V) variants. Both samples showed variants in histone lysine methyltransferase gene SETD8. SETD8 had two in-frame deletions, namely p.A20_A21del and p.L181Hfs*20. One Hispanic and 3 AA samples showed insertion in the ARID1B (p.H1534Q and p.Q130_Q131dup). PRSS1 mutations were observed in 11 Hispanic and 8 AA samples (p.N29I). The p.N29I polymorphism in PRSS1 is possibly pathogenic (e.g., ClinVar Accession RCV000012652.31). Due to the higher frequencies observed in our samples and associated dbSNP identifiers, these genetic variants could be due to germline contribution and are predicted to be benign.

We examined single nucleotide variants in our dataset for known breast cancer genes. These genes are reported to be frequently mutated in breast cancer [Online Mendelian Inheritance of Man (OMIM) entry for breast carcinoma: https://omim.org/entry/114480 and Human phenotype ontology (HPO): https://hpo.jax.org/app/browse/term/HP:0003002] (15). We also compared the gene list we obtained for each cohort with the COSMIC Cancer Gene Census (CGC) genes, which in many cases, have experimental evidence as oncogenes and tumor suppressors (Tier 1) (16).

The cancer genome atlas lists PIK3CA (45%) as the most frequently mutated gene, followed by MAP3K1, GATA3, TP53, CDH1, and MAP2K [5]. Although TP53 was the most frequently mutated gene in African-American breast cancer patients in our cohort, we found missense mutations in ARID1B (5 samples), BRCA1 (4 samples), BRCA2 (4 samples), and RAD51B (3 samples) (Fig. 2, Table II). The details of example genes from the top 50 frequently mutated genes [Online Mendelian Inheritance of Man (https://omim.org/) and Human Phenotype Ontology] and the COSMIC Census (updated as of February 2022) are shown in Table II. To capture genetic variants in these genes, we included variants with dbSNP ids while comparing the tumor with matched normal (Table III). As a result, BRCA2 (6 samples), ARID1B (AT-rich interactive domain, 5 samples), and BARD1 (BRCA1 Associated RING Domain 1B, 2 samples) showed variants in this cohort. In addition, a single sample in each cohort showed variants in the TP53, ERBB2, and HELQ (Helicase, POLQ Like, a single-stranded DNA-dependent ATPase, and DNA helicase) genes. On the other hand, Hispanic patients also exhibited similar mutational profiles in the top frequently mutated breast cancer or DNA damage response genes (Fig. 2, Table II).

As mentioned earlier, we compared the somatic mutation data from our cohort with the COSMIC CGCs, for variants causally linked with cancer. Overall, 47 genes from the African-American and 52 in the Hispanic cohorts overlap with the CGC lists (Table SII: COSMIC Census genes found in our patient cohort). Most of the genes were found to be mutated in a single sample. In the African-American cohort, both KDM6A [Lysine (K)-specific demethylase 6A, c.2703-5dup/del, Intronic] and PABPC1 [poly (A) binding protein cytoplasmic 1, p.K254Nfs*24] showed variants in 4 samples. In the Hispanic group, KMT2C [lysine (K)-specific methyltransferase 2C, p.A30P; p.M1774T], EIF1AX (eukaryotic translation initiation factor 1A; X-linked, c.337+1G>C; C256-3A>C, Splice Site), and SGK1 (serum/glucocorticoid regulated kinase 1 c.285+50T>C; c.362-3230T>G; Intronic) each showed variants in 2 samples. ZNF384 (p.Q547del) showed the same variants in 3 Hispanic samples.

The Cancer Genome Atlas (TCGA, http://www.cancer.gov/tcga.) provides a large dataset on various cancer types to querying for mutations and gene expression with samples from multiple ethnicities. We queried the TCGA data set via cBioPortal (cbioportal.org) (17). We utilized the TCGA Pan-Can Atlas 2018 dataset to this end as a maf (mutation annotation format) file from the NCI Genomic Data Common (GDC). We compared the mutational profile of African-American (Race Category) and Hispanic patients (Ethnicity Category) with our cohorts. According to the mutations listed in the TCGA data, the most frequently mutated gene is TP53 (accessed via cBioPortal). TP53 was the most significantly mutated gene in African-American samples (42.7%, 72 patients out of 182 on cBioPortal), followed by FBXW7 (F-box and WD repeat domain containing 7) at 8%. This finding is in concordance with other reported studies, which found TP53 mutation to be highly enriched in African-American patients (18). The GDC maf file also listed PIK3CA, TTN, GATA3, and KMT2C, MAP3K1 as frequently mutated in the Black or African-American cohort. TP53 mutations were found in our patient set in two African-American and only one Hispanic sample. A somatic FBXW7 variant was not observed in our cohorts (Filtered) except for one Hispanic patient (p.I394*). In white patients, on the other hand, the mutation level of TP53 was 29.23% (216 patients out of 752). PIK3CA was the most frequently mutated gene in white patients (34.64% compared to 19.66% in African-American patients).

The genetic variants in the TCGA cohort are in well-known tumor suppressors or oncogenes. However, excluding variants in databases such as gnomAD or dbSNP and applying the ‘PASS’ filter to the vcf files, the variants in genes such as TP53 and FBXW7 were filtered out from most of our samples. We only observed TP53 p.R175H in one African-American and p.R213* in one Hispanic patient. Therefore, we examined the tumor and normal samples for known variants in those genes to capture the possible contribution of tumor matched-normal tissues for TP53 variants and other gene mutations in our samples. In our patient samples, TP53 mutations that were most frequent were p.P72R (COSMIC id: COSV52666208), p.R273H (COSMIC id: COSV52660980), p.R342* (stop codon, COSMIC id: COSV52665487). The Hispanic group only showed the P72R mutations (Fig. 3). All of these TP53 variants are implicated in cancer (19).

On the other hand, one additional FBXW7 mutation was found in 2 African-American patients. The FBWX7 p.P160L missense mutations might be a loss-of-function implicated in cancer (COSMIC ID: COSV55920521). In addition, the Hispanic breast cancer cohort also showed p.D600N mutation, which might be a somatic loss of function of the protein (COSMIC ID: COSV55951044).

In recent years, it has become apparent that certain mutational processes are causative for a specific type of single nucleotide variations in the genome. These processes involving the APOBEC3 (Apolipoprotein B Editing Complex) enzyme or DNA damage response protein will leave a ‘signature’ behind, which can be revealed by WES or WGS (20). This mutational signature is displayed using a 96-mutational profile classification. The signature is calculated using the substitution class (A>G, T>C) and three nucleotides 3′ downstream or 5′ upstream to the mutated base. We applied the signature algorithms on the exome profile of our patient data. We did not find APOBEC mutational pattern enrichment in our patient samples. However, all breast cancer patients in our cohort showed similar mutational profiles. Furthermore, we discovered that Signatures 3 and 5 are enriched in African-American and Hispanic samples (Fig. 4). Signature 3 is associated with DNA double-strand Homologous Recombination repair. We compared the signature profile with TCGA African-American and Hispanic cohorts. Signature 3 was found to be over-represented in all samples. The TCGA cohorts showed Spontaneous or Enzymatic deamination of 5-methylcytosine (SBS1) and APOBEC Cytidine Deaminase (SBS2). The cytosine deaminase APOBEC3 mediated C to T mutation is prevalent in breast cancer, and generally, multiple cancer shows mutational patterns indicative of APOBEC activity (20,21). However, we did not observe these signatures in our datasets. Only one African-American patient and three Hispanic patients had APOBEC enrichment (>2) in our cohort (Table SIII: APOBEC Enrichment scores for patient samples). APOBEC enrichment score profiles are shown as a boxplot in Fig. 4D. The TCGA-AA group has a higher APOBEC value than our cohort (DCRT-AA) (Mann-Whitney U test, Median TCGA-AA=1.439 and DCRT-0.954, two-tailed P=0.039). However, all other comparisons between AA and Hispanic categories did not significantly differ.

We also examined somatic variants in the signaling pathways associated with cancer. Ten most affected pathways were examined using maftools. The pathways are derived from Sanchez-Vega et al (22). In analyzing African-American and Hispanic tumor samples, we found that both have similar profiles regarding affected genes in the pathways. For example, African-American and Hispanic tumors have RTK-RAS, NOTCH, and WNT as the top three pathways, although each category's number of affected genes varies. For example, the RTK-RAS pathway showed 5 and 3 genes mutated for African-American and Hispanic patients, respectively (Fig. 5). Hispanic patients had more variants in The TGFβ signaling pathway-associated genes than African-American patients. Similarly, the Notch pathway showed variants in 5 and 4 genes out of 71 in African-American and Hispanic patients. The TCGA data for African-American and Hispanic cohorts showed similar patterns in the oncogenic pathways. Hippo and PI3K pathways showed similar profiles in all cohorts analyzed.

However, individual subtype-based analysis shows other genes and the genes described above. For example, we compared the tumor of three African-American patients with three Hispanic patients with triple-negative (TNBC) subtype in QCI interpretation. In this case, we observed enrichment for variants in IGSF3 (p.R456C, p.D254N), ZNF717 (p.E370Q; p.Y499*, p.K622fs*79, p.K790*, p.S861fs*?) KIR3DL3 (p.V324A) and KMT2C (p.W858L, p.P860S) while selecting pathogenic or likely pathogenic variants. A G6PD p.V68M (known loss of function) was found in one African-American sample. However, this variant has a high prevalence in the African-American population (gnomAD 11.64% in Africans, Table SIV: Details of all genetic variants).

The challenge with survival analysis is the low frequency of somatic mutations, significantly reducing the number of patients carrying a mutation. Therefore, we conducted the overall survival analysis on the TCGA cohorts due to the higher available number. In addition, we chose the genes that are frequently mutated in our cohorts. In our survival analysis for the TCGA African-American cohort, we observed a statically significant (P<0.05) reduction in the probability of overall survival with a mutation in F5 (Median Survival 174.5 days vs. 414.5 days in the WT, HR=2.96) or BRCA2 (HR=5.2). We also conducted survival analysis with gene sets instead of individual genes. We found that patients with mutations in at least two genes among XRCC3, TP53, BRCA1 and BRCA2 have reduced overall survival (Median survival 176 days vs. 413 in patients with Wt alleles, HR=2.37).

Additionally, patients with at least one mutation in either SETD8, PRSS1, ARID1B, F5, or CDC27 genes showed reduced survival to 176 days compared to 426 days in Wt patients (Wt patients) HR=3.06) (Fig. S1). Additionally, we also utilized the Disease-Free Survival (DFS) data (Table SV) for our patients (median=38 months) and compared the genetic variants in patients with Low DFS with patients who had higher DFS (Median Split). Even though not statistically significant, African-American patients with lower DFS had more PRSS1 and CDC27 frequently mutated. On the other hand, Hispanic patients had F5 and MTHFR more frequently mutated along with SETD8 and PRSS1 (Table SVI).

We wanted to understand the differentially mutated genes in African-American and Hispanic breast cancer patient samples. To this end, we utilized MutSig v1.4 to determine the novel driver genes in the patients in both African-American and Hispanic samples (23). MutSig algorithm frequently determines mutated genes while considering gene expression and chromatin state and has been tested to find novel driver mutations across 21 different tumor types (24). The somatic mutation rate is calculated with respect to a background mutation rate. Among the frequently mutated somatic genes, 97 genes are shared between African-American and Hispanic samples among the top 250 genes as determined by MutSig [Table SVII: MutSig scores for genes in the African-American and Hispanic sample and Table SVIII: Top mutation significance (MutSig) genes across various ethnicities]. After Filtering the dataset with P-values (<0.05) in African-American samples, the top ten mutated genes are SETD8, PRSS1, TMIE, PABPC1, OR8D4, OR6P1, EPHB6, BMP2K, MEF2A, and TPP1. However, in the TCGA African-American cohort, GATA3, TP53, PIK3CA, CDH1, PTEN, MAP3K1, MAP2K4, MUC4, RUNX1, and FBXW7 were the top ten candidate genes. We also conducted GO (Gene Ontology) analysis on the significant driver genes for both our and TCGA cohorts. Our data set did not show enrichment in any particular category. The TCGA African-American cohorts showed enrichment in the protein stabilization and intrinsic apoptotic signaling pathways (Fig. 6). Hallmark MSigDB Hypoxia, P53 Pathway, and E2F targets were also enriched in the candidate driver gene profiles from the TCGA African-American cohort set.

In the TCGA Hispanic group, the top ten genes found by MutSig were TP53, PIK3CA, GATA3, MAP3K1, TYW3, KHDC1, CEACAM8, TCP10L2, GPS2, and SNAP29. However, GO ontology analysis failed to show enrichment in any specific category. Our Hispanic patient cohort showed a similar candidate driver gene profile, with the top ten being SETD8, MTHFR, PRSS1, KIAA2018, CDC27, ZNF384, MEF2A, F5, and NUDT15. As a candidate driver gene, we observed SETD18, a histone lysine methyltransferase. SETD8 is known to play a role in breast cancer metabolism by stabilizing hypoxia-inducible factor 1α (HIF1α) and is also implicated in DNA damage response maintaining genomic integrity (25,26).

As with all pathways, the mutated genes in African-American samples were similar to Hispanic samples. 11 genes (including ATXN1, CMTM5, EIF1AX, GPRC6A, MEF2A, PRSS1, and SETD8) were found to be candidate drivers in both categories. However, some genes were only enriched in one or the other sample (72 in African-American, 42 in the Hispanic cohort, Fig. 6 Venn Diagram). There were only two genes common between our Hispanic and TCGA cohorts. Our analysis with the Hispanic samples observed significant mutations in genes such as CDC27 (cell division cycle 27), making it a candidate driver gene. CDC27 protein levels and polymorphism are associated with breast cancer mortality and risk (27,28).