Comparison of somatic mutation landscapes in Chinese versus European breast cancer patients

Summary

Recent genomic studies suggest that Asian breast cancer (BC) may have distinct somatic features; however, most comparisons of BC genomic features across populations did not account for differences in age, subtype, and sequencing methods. In this study, we analyzed whole-exome sequencing (WES) data to characterize somatic copy number alterations (SCNAs) and mutation profiles in 98 Hong Kong BC (HKBC) patients and compared with those from The Cancer Genome Atlas of European ancestry (TCGA-EA, N = 686), which had similar distributions of age at diagnosis and PAM50 subtypes as in HKBC. We developed a two-sample Poisson model to compare driver gene selection pressure, which reflects the effect sizes of cancer driver genes, while accounting for differences in sample size, sequencing platforms, depths, and mutation calling methods. We found that somatic mutation and SCNA profiles were overall very similar between HKBC and TCGA-EA. The selection pressure for small insertions and deletions (indels) in GATA3 (false discovery rate (FDR) corrected p < 0.01) and single-nucleotide variants (SNVs) in TP53 (nominal p = 0.02, FDR corrected p = 0.28) was lower in HKBC than in TCGA-EA. Among the 13 signatures of single-base substitutions (SBS) that are common in BC, we found a suggestively higher contribution of SBS18 and a lower contribution of SBS1 in HKBC than in TCGA-EA, while the two APOBEC-induced signatures showed similar prevalence. Our results suggest that the genomic landscape of BC was largely very similar between HKBC and TCGA-EA, despite suggestive differences in some driver genes and mutational signatures that warrant future investigations in large and diverse Asian populations.

Introduction

Despite being lower than in North America and Europe,¹ the incidence rates for invasive female breast cancer (BC) [MIM: 114480] have been increasing rapidly in many Asian populations. Moreover, Asian women seem to have a distinct profile of BC, such as earlier age at onset and higher frequencies of luminal B and HER2-enriched tumors, compared with European populations.² Recent genomic studies based on a limited number of Asian subjects suggest that Asian breast tumors may display a higher frequency of somatic mutations in TP53[MIM: 191170]3, 4, 5 and distinct immune gene expression profiles.4, 5, 6 In particular, a germline APOBEC3B deletion polymorphism [MIM:607110], which has been associated with increased BC risk⁷ and increased APOBEC-associated mutation signatures in BC,^8,9 is much more common in East Asians (31.2%) than in Europeans (9.0%) and West Africans (4.2%). These previous studies suggest that East Asian breast tumors may exhibit a distinct somatic profile compared with other BC populations.

Cancer develops as a result of somatic mutations and clonal selection. The selection pressure of somatic mutations in cancer driver genes reflects the adaptiveness or effect size of clonal selection.^10,11 The positive selection pressure indicates that the corresponding cancer subclone has a growth advantage over normal cells and other cancer subclones. Selection pressure is frequently measured by dN/dS ratio, which represents the mutation rate of nonsynonymous (NS) mutations (including missense, nonsense, splicing site mutations, and insertions and deletions (indels)) versus the mutation rate of synonymous mutations.¹¹ In previous studies, population differences in the driver gene landscape were compared primarily based on mutation frequencies among study populations. However, these comparisons are sensitive to differences in sequencing platforms and bioinformatic analyses.¹² On the other hand, by leveraging mutation rates of synonymous mutations as the reference group, the dN/dS ratio is less sensitive to variations in sequencing and somatic mutation calling, assuming the technical variations are not discriminative against the function of mutations. In addition, although the same set of cancer driver genes may present in different populations, the effect size of selection can be different. Cannataro et al. estimated the selection pressure of all recurrent single-nucleotide variants (SNVs) in 22 cancer types in The Cancer Genome Atlas (TCGA) and found that selection pressures varied considerably across cancer types, even within the same cancer type.¹³ For example, although TP53 is a driver gene for all BC subtypes, TP53 mutations present with higher selection pressure in ER-negative than ER-positive BCs. Comparing the selection pressure of driver genes in different populations may therefore provide a more accurate and quantitative measure to evaluate racial heterogeneity of somatic mutations than comparing mutation frequencies alone. In this study, we compared driver gene selection pressure, as well as mutational signatures and somatic copy number alteration (SCNA) profiles, between Chinese and TCGA BC patients of European ancestry that had similar distributions of age at diagnosis and PAM50 subtypes.

Methods

Participants and samples

We analyzed data and biospecimens collected from a hospital-based BC case-control study in Hong Kong BC (HKBC), as previously described.¹⁴ In brief, fresh frozen breast tumors and paired normal tissues were collected from newly diagnosed BC patients of Han Chinese ancestry in two Hong Kong hospitals between 2013 and 2016. Patients with pre-surgery treatment were excluded from the study. Clinical characteristics and BC risk factors were obtained from medical records and a questionnaire. The study protocol was approved by ethics committees of the Joint Chinese University of Hong Kong, New Territories East Cluster, the Kowloon West Cluster, and the National Cancer Institute (NCI). Written informed consent was obtained prior to the surgery for all participants.

Bioinformatic analyses

Paired tumor and histologically normal breast tissue samples were processed for pathology review at the Biospecimen Core Resource (BCR), Nationwide Children's Hospital, using modified TCGA criteria.⁸ Specifically, only tumors with >50% tumor cells and normal tissue with no detected tumor cells were included for dual DNA/RNA extraction.

Whole-exome sequencing (WES) was performed on 98 paired tumor and normal samples at the Cancer Genomics Research Laboratory (CGR), NCI, using SeqCAP EZ Human Exome Library v3.0 (Roche NimbleGen, Madison, WI) for exome sequence capture. The captured DNA was then subjected to paired-end sequencing utilizing Illumina HiSeq2000. The average sequencing depth was 106.2x for tumors and 47.6x for the paired blood or normal tissues. Somatic mutations were called using four different algorithms (MuTect,¹⁵ MuTect2 (GATK tool), Strelka,¹⁶ and TNScope by Sentieon)¹⁷ and the final variant calls were based on mutations called by three or more of four established callers. Variants were excluded if they did not pass the pipeline quality control metrics, had variant allele fraction (VAF) < 0.07 in tumor, VAF >0.02 in normal, alternative allele read count <3 or total read count <8 in tumor, total read count <6 in normal. In addition, variants would be excluded if its minor allele frequency (MAF) was >0.1% in reference germline variant databases including 1,000 Genomes Project,¹⁸ the ESP6500 dataset from University of Washington's Exome Sequencing Project (http://evs.gs.washington.edu/EVS/), or ExAC.¹⁹

SNP rs12628403, which is a proxy for the APOBEC3B deletion (r² = 1.00 in Chinese from Beijing (CHB) in HapMap samples), was genotyped in germline DNA with a custom TaqMan assay, as previously described.²⁰

RNA sequencing (RNA-Seq) data were generated in these tumors at Macrogen Corporation on Illumina HiSeq4000 using a TruSeq stranded RNA kit with Ribo-Zero for rRNA depletion and 100-bp paired-end method. Gene expression was quantified as transcript per million (TPM) using RSEM,²¹ and log2TPM was used for statistical analyses. PAM50 subtype, which is a 50-gene signature that classifies BC into five molecular intrinsic subtypes, was defined by an absolute intrinsic subtyping (AIMS) method²² using RNA-Seq data.⁶ For patients without RNASeq data, subtype was defined using immunohistochemical status of estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor-2 (HER2).

TCGA dataset

We included 686 BC patients of European ancestry in TCGA (TCGA-EA, defined by the study information) with both WES and RNA-Seq as a comparison dataset. WES calls for 779 BC tumors, which were downloaded from Multi-Center Mutation Calling in Multiple Cancers (MC3),²³ and tumors from non-European ancestry patients were excluded. RNA-Seq data were downloaded from the NCI's Genomic Data Commons (GDC) (legacy archive) and processed and quantified using similar methods as in HKBC. PAM50 was called using the same AIMS method for each TCGA sample, as it was used for HKBC. The somatic mutation files (including called variants and SCNAs) were obtained from the pan-cancer TCGA study,²³ and variants were processed and filtered using the same criteria as in HKBC.

Mutational signatures

Given the limited sample size of the HKBC study, we did not perform de novo mutation signature analysis. Instead, we evaluated contributions of 13 COSMIC single-base substitution (SBS) signatures that were previously reported to be common in BC (SBS1, SBS2, SBS3, SBS5, SBS8, SBS9, SBS13, SBS17a, SBS17b, SBS18, SBS37, SBS40, and SBS41).²⁴ SBS signatures were defined by 96 mutational catalogs, each of which refers to a mutated pyrimidine (C or T) in the center and two flanking nucleotides (flanking 5′ AND-3′ bases). SignatureEstimation (version 1.0.0),²⁵ SIGNAL (downloaded around Nov 2021),²⁶ and YAPSA (version 1.19.0)²⁷ were used to evaluate the contribution of each COSMIC SBS signature (version 3.2) with confidence for each tumor sample.

Driver gene detection

dNdScv¹¹ (version 0.1.0) with the default arguments was used to identify driver and significantly mutated genes (SMGs) (the significance was defined as false discovery rate (FDR) q < 0.01), which were compared with previously identified BC driver genes.²³ For this analysis, we further filtered out somatic mutations with VAF <0.1. After the filtering step and restricting the variants to those in protein-coding regions, 39,439 mutations from 686 TCGA-EA samples (98.7% are singletons) and 5,116 mutations from 98 HKBC samples (98.8% singletons) were used for the selection pressure comparison. To examine if the sample size would impact the number of detected driver genes, we randomly down-sampled the TCGA-EA samples and reran dNdScv.

SCNAs

Allele-specific SCNA analysis was performed using subHMM²⁸ (https://dceg.cancer.gov/tools/analysis/subhmm). An arm-level SCNA was defined if the SCNA event covered 90% of the p or q arm of each chromosome. In addition, subHMM also estimated subclonal genotypes of SCNA events based on the distribution of variant allele fractions. GISTIC2.0²⁹ was used to detect significant focal recurrent SCNAs. Following the methods in previous reports,30, 31, 32 the SCNA-based homologous recombination deficiency (HRD) was quantified based on loss of heterozygosity (LOH), telomere allelic imbalance (TAI), and large-scale state transition (LST).

Two-sample Poisson model to compare driver gene selection pressure in the two populations

We developed a new statistical test, dNdScf, to compare the selection pressure of 23 known BC driver genes between the two populations, while adjusting for study sample size, background mutation rate, and mutable bases. Specifically, we first estimated the background mutation rates for missense, nonsense, splicing site single-nucleotide variants (SNVs), and indels, using the statistical framework adapted from dNdScv. We then compared selection pressure for each specific driver gene in a two-sample test using a Poisson regression model, which adjusts the sample size of each population and accounts for genome-wide mutation rates, 192 tri-nucleotide context-specific mutation rates, and a gene-specific mutation rate.

Let $n_{ij}^g$ be the number of NS mutations for the $i$ th population and $j$ th mutation category in the $g$ th driver gene, $i=\mathrm{1,2},$ $j=\mathrm{1,2},\dots ,96,$ and $g=\mathrm{1,2},\dots ,23.$ We modeled the number of NS mutations following a Poisson distribution as $n_{ij}^g\sim Poisson\left(n_i{m}_j^g{\tau }_i^g{\beta }_i^g\right)$, for which $n_i$ is the sample size of the $i$ th population, $m_j^g$ is the mutable base per subject for the $j$ th mutation category in the $g$ th driver gene, ${\tau }_i^g$ is the background mutation rate for the $g$ th driver gene in the $i$ th population, and ${\beta }_i^g$ is the selection pressure, measured as dN/dS, for the $g$ th driver gene in the $i$ th population; $n_i$ and $m_j^g$ are given based on the study sample size and human reference genome; ${\tau }_i^g$ is a plugged-in estimate from dNdScv¹¹; we parameterized ${\beta }_2^g={\omega }^g{\beta }_1^g$. A Poisson regression model was fitted to estimate ${\beta }_1^g$ and to test the null hypothesis that dN/dSs of two populations are the same, i.e., ${\omega }^g=1$. Rejecting the null hypothesis indicates the disparity of selective pressure between the two populations. Both the p value and FDR adjusted p value were reported in the results.

Results

This analysis included 98 Chinese BC patients in HKBC and 686 TCGA-EA BC patients. Distributions of age at diagnosis (mean = 57.9 in HKBC and 58.3 in TCGA-EA) and PAM50 subtypes (44.1% luminal A, 26.9% luminal B, 16.1% HER2-enriched, and 12.9% Basal-like in HKBC; 40.9% luminal A, 26.8% luminal B, 14.2% HER2-enriched, and 18.1% Basal-like in TCGA-EA) were similar in the two datasets (Table S1).

The tumor mutational burden in HKBC was 1.74 mutations/Mb. Figure 1A shows the somatic mutation landscape of the 23 known BC driver genes²³ in HKBC, in which somatic mutations in these genes were seen in 73.5% of patients. The prevalence of mutations in most driver genes was similar in the two datasets (Table S2), with the exception of CDH1 (MIM:192090; 4% in HKBC versus 13% in TCGA-EA, nominal p = 0.015, FDR adjusted p = 0.35) and GATA3 (MIM: 131320; 5% in HKBC versus 12% in TCGA-EA, nominal p = 0.075, FDR adjusted p = 0.74), which showed lower frequencies in HKBC than in TCGA-EA (Figure 1B). De novo identification of SMGs found five genes in HKBC versus 23 genes in TCGA-EA, using dNdScv. We down-sampled TCGA-EA samples to test if the disparity of the number of driver genes was due to the different sample size in HKBC (n = 98) and TCGA-EA (n = 686). Down-sampling the TCGA-EA dataset to 100 tumors also identified five SMGs, suggesting that the difference in the number of SMGs in the two datasets was largely due to different sample sizes (Figure S1).

Figure 1.

BC driver gene landscape in the HKBC study

(A) Genomic driver landscape of BC in HKBC. Top bar graph shows the number of NS mutations per tumor. The middle gene panel reports NS mutations in 23 known BC driver genes that were reported previously for the TCGA BC (BRCA) study. Bottom panels show age groups and PAM50 subtypes.

(B) Frequencies of mutations in 23 known BC driver genes in HKBC and TCGA-EA (European ancestry) studies.

To examine whether the selection pressure of BC driver genes varied by population/ethnicity, we applied a new statistical test, dNdScf (see Methods), to compare the selection pressure (ω) of the 23 BC driver genes in HKBC and TCGA-EA. These genes were selected because they were significant (FDR corrected p < 0.1) in the TCGA-EA study and were reported as BC driver genes.²³ Overall, we found that ω of most driver genes, including PIK3CA (MIM:171834; the most frequently mutated gene in BC), was similar in the two datasets (Figure 2, Table 1). However, ω of small indels in GATA3 was significantly lower in HKBC than in TCGA-EA (FDR corrected p < 0.01). In addition, ω of SNVs in TP53 was also suggestively lower in HKBC (nominal p = 0.02, FDR adjusted p = 0.28).

Figure 2.

Comparisons of selection pressure of BC driver genes between HKBC and TCGA-EA studies

(A) Selection pressure of SNVs; (B) selection pressure of small indels. Genes without mutations detected in either HKBC or TCGA-EA samples are not shown. The bar represents the 95% confidence interval of the estimate of selection pressure.

Table 1.

Selection pressure for 23 BC driver genes in HKBC and TCGA-EA studies

	HKBC study					TCGA-EA
Gene	${\mathit{N}}_{\mathit{s}}$	${\mathit{N}}_{\mathit{n}}$	${\mathit{N}}_{\mathit{i}}$	${\hat{\mathit{\omega }}}_{\mathit{n},}$	${\hat{\mathit{\omega }}}_{\mathit{i}}$	${\mathit{N}}_{\mathit{s}}$	${\mathit{N}}_{\mathit{n}}$	${\mathit{N}}_{\mathit{i}}$	${\hat{\mathit{\omega }}}_{\mathit{n}}$	${\hat{\mathit{\omega }}}_{\mathit{i}}$	${\mathit{p}}_{\mathit{n}}$	${\mathit{q}}_{\mathit{n}}$	${\mathit{p}}_{\mathit{i}}$	${\mathit{q}}_{\mathit{i}}$
TP53	2	14	8	94.8 (56, 106)	642.5 (321, 1,285)	0	175	41	167.0 (144, 194)	864.6 (651, 1,201)	0.03∗	0.28	0.43	0.70
PIK3CA	0	36	1	100.9 (73, 140)	59.1 (8, 419)	4	256	12	97.9 (87, 111)	54.2 (31, 96)	0.87	0.93	0.94	0.98
CDH1	0	0	4	0 (NA)	138.6 (52, 369)	3	41	52	18.4 (14, 25)	340.5 (259, 447)	<0.01∗	0.03∗	0.05∗	0.37
MAP3K1	0	1	3	1.9 (0.3, 132)	104.7 (34, 325)	3	29	51	7.5 (5, 11)	165.0 (125, 217)	0.08	0.35	0.41	0.70
PTEN	0	5	4	36.5 (15.88)	413.0 (155, 1,100)	0	18	22	18.1 (11, 29)	260.8 (172, 396)	0.19	0.45	0.42	0.70
MAP2K4	0	0	1	0 (NA)	101.0 (14, 717)	1	17	13	16.7 (10, 27)	146.7 (87, 257)	0.04∗	0.28	0.70	0.96
FOXA1	0	2	2	10.8 (3, 43)	160.0 (40, 640)	1	18	7	13.8 (9, 33)	139.3 (87, 257)	0.73	0.93	0.86	0.96
KMT2C	1	7	6	4.1 (2, 9)	28.4 (13, 63)	0	26	21	2.1 (2, 3)	24.5 (16, 38)	0.14	0.44	0.75	0.96
GATA3	0	1	4	6.3 (0.9, 45)	157.0 (59, 418)	0	9	75	8.0 (4, 16)	998.3 (796, 1,252)	0.81	0.93	<0.01∗	<0.01∗
RUNX1	0	1	1	5.2 (0.7, 37)	76.7 (11, 544)	1	7	16	5.2 (3, 11)	188.2 (115, 307)	0.99	0.99	0.33	0.68
TBX3	0	0	0	0 (NA)	0 (NA)	0	7	16	3.5 (2, 7)	151.2 (93, 247)	0.17	0.44	0.02∗	0.22
PIK3R1	0	0	1	0 (NA)	45.2 (6, 321)	0	6	14	3.2 (2, 7)	83.8 (50, 142)	0.21	0.45	0.51	0.78
RB1	0	2	1	6.1 (2, 24)	49.1 (7, 349)	1	12	6	5.0 (3, 9)	28.0 (11, 75)	0.80	0.93	0.64	0.96
CBFB	0	2	0	25.2 (6, 101)	0 (NA)	1	12	4	20.8 (12, 37)	123.5 (46, 329)	0.81	0.93	0.32	0.68
NCOR1	0	1	0	1.1 (0.2, 8)	0 (NA)	2	24	9	3.8 (3, 6)	20.5 (11, 75)	0.15	0.44	0.10	0.37
AKT1	0	2	0	10.6 (4, 42)	0 (NA)	0	18	0	13.4 (8, 21)	0 (NA)	0.75	0.93	1.00	1.00
GPS2	0	0	0	0 (NA)	0 (NA)	0	2	7	2.5 (0.6, 10)	80.8 (39, 170)	0.46	0.82	0.23	0.68
ARID1A	0	1	2	1.2 (0.2, 9)	24.3 (6, 97)	0	9	3	1.6 (0.8, 3)	5.8 (2, 18)	0.80	0.93	0.15	0.56
CTCF	0	1	0	3.8 (0.5, 27)	0 (NA)	2	10	4	5.2 (3, 10)	35.8 (13, 95)	0.75	0.93	0.32	0.68
CDKN1B	0	0	2	0 (NA)	518.3 (130, 2,072)	1	3	5	5.6 (2, 17)	84.9 (35, 204)	0.37	0.72	0.06	0.37
NF1	0	2	1	2.2 (0.5, 9)	10.6 (2, 76)	1	13	5	1.9 (1, 3)	8.9 (4, 21)	0.89	0.93	0.87	0.96
BRCA1	0	0	1	0 (NA)	16.5 (2, 117)	0	11	4	2.5 (1, 5)	11.2 (4, 30)	0.09	0.35	0.73	0.96
ERBB2	0	0	1	0 (NA)	29.2 (4, 208)	0	11	1	3.6 (2, 7)	6.1 (1, 43)	0.09	0.35	0.28	0.68

The background mutation rates were estimated as ${\hat{\tau }}_{HK}=1.28\left(0.41,3.96\right){\hat{\tau }}_{WH}=1.13\left(0.74,1.73\right)(w.95\%CI)$, assuming the mutation counts following a Poisson distribution.

$N_s$: the number of synonymous mutations; $N_n$: the number of NS mutations; $N_i$: the number of indels.

${\omega }_i$: selection pressure among indels in cancer genes (relative to indel in non-cancer genes); ${\omega }_n$: selection pressure among NS mutations (relative to synonymous mutations); ${\omega }_i$: selection pressure among indels in cancer genes (relative to indel in non-cancer genes).

95% confidence intervals for ${\omega }_n$ and ${\omega }_i$ in brackets.

$p_n:$ p value from two sample Poisson likelihood ratio test for point mutations; $p_i:$ p value from two sample Poisson likelihood ratio test for indels; $q_n:$ q value for two sample Poisson likelihood ratio tests for point mutations; $q_i:$ q value for two sample Poisson likelihood ratio tests for indels.

A star mark (∗) indicates a test is significant at level 0.05.

Overall, the mutational spectrum showed similar patterns in the two datasets (Figures 3a and 3b). We further estimated the contributions of 13 known BC mutational signatures²⁴ in HKBC (Figure 3C) and TCGA-EA using three algorithms SignatureEstimation, SIGNAL, and YAPSA. Results based on SignatureEstimation and YAPSA were very similar. We observed the same eight SBS signatures in HKBC and TCGA-EA, each contributing to >0.5% mutations in each dataset. The most prevalent SBS signatures were SBS1, SBS2, SBS5, SBS3, SBS13, and SBS18 in both datasets (Table S3). SBS18, which is related to damage by reactive oxygen species (ROS),³³ appeared to show higher contribution in HKBC than in TCGA-EA (mean contribution by SignatureEstimation: 14.4% in HKBC versus 10.9% in TCGA-EA, p = 0.0064), whereas SBS1, considered as a clock-like signature, showed lower contribution in HKBC (8.2%) than in TCGA-EA (22.2%, p = 3.99x10⁻³⁰). The prevalence of these signatures estimated by SIGNAL was much lower as compared with the other two algorithms, resulting in insignificant results for most signatures (Table S3). Notably, contributions of SBS2 and SBS13, the two APOBEC signatures, were similar between HKBC and TCGA-EA across all algorithms (Table S3), despite a much higher frequency of the germline APOBEC3B deletion polymorphism in HKBC (40.4%) than in TCGA-EA (5.7%). We found the expected associations between the APOBEC3B deletion and decreased levels of APOBEC3B RNA expression in both tumor and normal tissue, validating SNP rs12628403 as a proxy for APOBEC3B deletion.⁶ In addition, the homozygous deletion of APOBEC3B was associated with higher contributions of SBS2 and SBS13 in HKBC (Figure S2).

Figure 3.

SBS mutational spectrum and prevalence of SBS signatures in HKBC

(A and B) Mutational spectrum in HKBC (A) and TCGA-EA studies (B).

(C) The contributions of COSMIC SBS signatures for each patient in HKBC.

We also compared SCNA profiles in HKBC and TCGA-EA and found similar patterns of gains, deletions, or LOH between the two datasets for both clonal and subclonal SCNAs²⁸ (Figure 4). The frequencies of those recurrent SCNA regions in BC, such as 1q gain, 8p loss, 8q gain, 16p gain, and 16q loss, were similar in the two datasets. Focal SCNA regions identified using GISTIC were also very similar in the two datasets (Figure S3), although a greater number of significant regions were identified in TCGA-EA because of the larger sample size.

Figure 4.

Comparisons of SCNAs between HKBC and TCGA-EA studies

(A) Main clones (left: HKBC, right: TCGA-EA). Each panel shows the frequency of copy number gain, loss, and copy number neutral LOH across the samples in each study.

(B) Subclones (left: HKBC, right: TCGA-EA).

Discussion

In this study, we compared genomic profiles between a clinically and molecularly well-annotated Asian population and TCGA BC patients of European ancestry that had similar age and subtype distributions. We found that somatic mutation and SCNA profiles were overall very similar in the two populations, which is consistent with the conclusion from a recent large BC genomic study conducted in Malaysia, where the majority of patients were of Chinese ancestry.⁵ However, in contrast to findings from previous studies, including the study by Pan et al.,⁵ we did not find higher frequencies of TP53 mutations and APOBEC signature mutations among Chinese BC patients in Hong Kong. Instead, we found that selection pressure for SNVs in TP53 seemed to be lower in HKBC than in TCGA-EA. The discrepancies might be due to younger age and more aggressive tumor subtypes in previous Asian studies, given that TP53 mutations occur more often in younger women and triple-negative tumors.³⁴ In our study, the age and subtype distributions were similar between HKBC and TCGA-EA. In addition, comparing mutation frequencies is subject to differences in sequencing coverage and mutation calling pipelines. For example, in the Chinese study by Zhang et al.,³⁵ the coverage of the targeted sequencing was about 20 times higher (∼1,200x) than that of TCGA (<60x), which likely contributed to the higher frequencies of mutations for most cancer driver genes among Chinese BC patients than in TCGA overall. In contrast, by leveraging mutation rates of synonymous mutations as the reference group, our method of comparing driver gene selection pressure is less sensitive to variations in sequencing and somatic mutation calling, assuming the technical variations are uniform regardless of mutation types. Further, Asians are an extremely heterogeneous population, comprising groups with diverse genetic background and sociodemographic characteristics. The variations in the BC genomic landscape across different Asian populations are therefore not surprising, highlighting the importance of conducting more genomic studies in different Asian populations to capture the complete spectrum of the BC genomic landscape in Asia. Nevertheless, our findings of lower selection pressure for TP53 in HKBC are in line with the highest survival rate among Asian Americans at all stages among BC patients of all race/ethnic groups in the United States,³⁶ given that TP53 mutations are known to be associated with poor prognosis in multiple cancer types.

Despite a much higher prevalence of the germline APOBEC3B deletion polymorphism in HKBC than in TCGA-EA and the expected associations between the APOBEC3B deletion and decreased levels of APOBEC3B RNA expression in both tumor and normal tissue, the fractions of APOBEC signatures (SBS2/13) were similar between HKBC and TCGA-EA. The percentage of samples with SBS2 was actually lower in HKBC, suggesting that this germline deletion polymorphism may not play a major role in APOBEC-related mutagenesis in our Hong Kong study.

Notably, we found a consistent increase in SBS18 in HKBC; both the percentage of patients with SBS18 and the mean fraction of SBS18 were higher in HKBC than in TCGA-EA. SBS18 has been associated with ROS damage and defective base excision repair due to germline MUTYH mutations.³⁷ However, a germline pathogenic variant in MUTYH was only seen in one patient in HKBC and zero in TCGA-EA. Given the sample size and variations in signature estimates across different algorithms, future studies are warranted to follow up these findings in additional Asian studies.

The major limitation of the HKBC study is the small sample size, which limited the power to agnostically identify HK-specific events, to extract de novo mutational signatures, and to compare selection pressure for less frequently mutated genes. In addition, lack of whole-genome sequencing data prohibited us from assessing HRD signatures based on structural variants. Nevertheless, taking advantage of the existing knowledge of BC genomics, we found suggestive differences in mutational signatures and selection pressure for several genes, suggesting some extent of racial heterogeneity in mutation generation and selection. Future large studies are warranted to confirm these findings and to relate these genomic differences with germline genetic variants and other etiologic factors.

Data and code availability

Sequencing data generated in the HKBC study has been deposited in the dbGaP database under Accession Code phs001870.v1.p1. at https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs001870.v1.p1.

Web resources

Online Mendelian Inheritance in Man: http://www.omim.org.