The Role of Genome Sequencing in Personalized Breast Cancer Prevention

Abstract

Background

There is uncertainty about the benefits of using genome-wide sequencing to implement personalized preventive strategies at the population level, with some projections suggesting little benefit. We used data for all currently known breast cancer susceptibility variants to assess the benefits and harms of targeting preventive efforts to a population subgroup at highest genomic risk of breast cancer.

Methods

We used the allele frequencies and effect sizes of 86 known breast cancer variants to estimate the population distribution of breast cancer risks and evaluate the strategy of targeting preventive efforts to those at highest risk. We compared the efficacy of this strategy to that of a “best-case” strategy based on a risk distribution estimated from breast cancer concordance in monozygous twins, and to strategies based on previously estimated risk distributions.

Results

Targeting those in the top 25% of the risk distribution would include approximately half of all future breast cancer cases, compared to 70% captured by the best-case strategy and 35% based on previously known variants. In addition, current evidence suggests that reducing exposure to modifiable nongenetic risk factors will have greatest benefit for those at highest genetic risk.

Conclusions

These estimates suggest that personalized breast cancer preventive strategies based on genome sequencing will bring greater gains in disease prevention than previously projected. Moreover these gains will increase with increased understanding of the genetic etiology of breast cancer.

Impact

These results support the feasibility of using genome-wide sequencing to target the women who would benefit from mammography screening.

Introduction

Describing an article in Science Translational Medicine entitled “The Predictive Capacity of Personal Genome Sequencing” by Roberts (1) an editor wrote:

Imagine that everyone at birth could have their whole genome sequenced at negligible cost. Surely, this must be a worthwhile endeavor, given the list of luminaries who have already had this sequencing completed. But how well will such tests perform?

The authors address this question by using the incidence data of twins for a range of diseases to estimate the percent of the population and the percent of diseased cases whose genomes would classify them at high risk, and the risk among those whose genomes put them at low risk relative to the population. They conclude that the genomes of most people would classify them at low risk for most of the diseases, and that the predictive value of such a negative test result would generally be small, because the disease risks among those who test negative would be similar to those in the general population. These conclusions have been criticized as deriving from assumptions that involve no information about the genetic factors relevant to the diseases studied (2). Here we compare the breast cancer findings of Roberts (1) to those obtained using information on the currently known breast cancer susceptibility loci.

To be specific, consider the strategy of sequencing the genomes of all young women, and using their genotypes at breast cancer loci to construct for each a genomic risk score, which specifies the rank of her inherited breast cancer risk relative to that of others in the population. These ranks can then be used to classify women into high- and low-risk categories, with high-risk women targeted for more intensive screening and preventive efforts. At present we don’t fully know a person’s breast cancer risk score, but genome sequencing coupled with current knowledge would allow us to assign her a partial score by combining her genotypes at all known breast cancer loci with the effect sizes of the risk alleles at these loci. Here we shall evaluate and compare the efficacy of this strategy to that of: a) the best-case classification theoretically possible if we knew the full risk scores; and b) the estimates of Roberts (1).

Materials and Methods

Population studied

Since the frequencies and effect sizes of breast cancer susceptibility loci may differ by race/ethnicity, we restrict analyses to women of European ancestry. This study did not require approval from an ethical review board, as it involved only the use of published summary data.

Statistical analysis

We modeled the lifetime probability of developing breast cancer (i.e., the absolute risk) for an individual with genomic risk score s as R = 1 − exp (−ce^s), where c is a positive constant. This model specifies a monotonically increasing relationship between risk R and risk score s. Therefore the population percentile of an individual’s risk equals that of her risk score, and the efficacy of percentile-based stratification depends on the variances of the risk score distributions in the population and among future breast cancer cases. For the theoretically derived distribution of fully known risk scores, these variances can be estimated using the arguments of Pharoah (3,4) and Begg and Pike (5), as described in the Supplementary Materials and Methods.

To estimate the variance of the partially known risk scores, we modeled them as linear combinations of genotypes at a set of uncorrelated breast cancer loci from the literature, with coefficients given by their estimated effect sizes. Specifically, we listed all loci with validated breast cancer association in women of European ancestry, and then selected a subset of 86 uncorrelated loci. We chose 93 breast cancer susceptibility loci by reviewing the literature for established, replicated associations (6–15). We then selected a subset of 86 uncorrelated loci by computing all 4,278 pairwise correlation coefficients using the SNP Annotation and Proxy Search tool <http://www.broadinstitute.org/mpg/snap/ldsearch.php> for the CEU population from the 1000 Genomes Project. We found seven pairs of loci with squared correlation coefficient exceeding 0.25 (16), and for each of these we selected the locus with the largest value of βp(1-p), where β is the log relative risk associated with the variant and p is its allele frequency. The seven SNPs we excluded were: rs10069690, rs3215401, rs2943559, rs10759243, rs11199914, rs494406, and rs75915166.

The remaining 86 loci consist of: 1) genes containing rare variants of high and moderate penetrance; and 2) single nucleotide polymorphisms (SNPs) identified in genome-wide association studies of breast cancer. We used the cumulative allele frequency and took the relative risk estimates for rare variants in breast cancer susceptibility genes to be the midpoints of the ranges spanned by the published studies. We also used the averages of the risk allele frequencies and relative risk (odds-ratio) estimates for SNPs that were associated with breast cancer in multiple genome-wide association studies.

Table 1 shows the 86 loci, and the frequencies and effect sizes of their risk alleles (6-15). We modeled the combined effects of these 86 loci by assuming that they act multiplicatively on a woman’s cumulative hazard for breast cancer. As shown in the Supplementary Materials and Methods, this implies that her partially known risk score has the additive form s = β₁g₁ + … + β₈₆g₈₆, where g_k = 0,1,2 denotes her count of risk alleles at locus k, and β_k denotes the effect size of the risk allele at locus k as obtained from the literature, k = 1, …, 86. Determining the variance of the resulting partial scores is infeasible, since it would require summing over all 3⁸⁶ = 10⁴¹ possible multi-locus genotypes. However it can be approximated by random genotype sampling as described in the Supplementary Materials and Methods.

Table 1.

Risk-allele frequencies and relative risks for breast cancer susceptibility loci among women of European-American ancestry

	Locus	Chromosome	Gene	Risk allele frequency (%)	Relative risk
1	rs11249433	1p	NOTCH2/FCGR1B	39	1.14
2	Multiple variants	1p	MUTYH	0.5	1.4–2.2
3	rs616488	1p	PEX14	67	1.06
4	rs11552449	1p	TPN22/BCL2L15	17	1.07
5	rs4245739	1q	MDM4	26	1.14
6	rs4849887	2p	None	90	1.10
7	Multiple variants	2p	MSH6	0.9	4.9–4.9
8	Multiple variants	2p	MSH2	0.01	2.4–2.4
9	rs12710696	2p	None	36	1.11
10	rs2016394	2q	METAP1D	52	1.05
11	rs1550623	2q	CDCA7	84	1.06
12	rs1045485	2q	CASP8	0.85	1.03
13	rs13387042	2q	IGFBP2, IGBP5, TPN2	52	1.12
14	rs16857609	2q	DIRC3	26	1.08
15	rs4973768	3p	SLC4A7/NEK10	46	1.11
16	rs12493607	3p	TGFBR2	35	1.06
17	rs6762644	3p	ITPR1/EGOT	40	1.07
18	rs9790517	4q	TET2	23	1.05
19	rs6828523	4q	ADAM29	87	1.11
20	rs10941679	5p	MRPS30/HCN1	26	1.19
21	rs7734992	5p	TERT	43	1.05
22	rs889312	5q	MAP3K1/MEIR3	28	1.13
23	rs10472076	5q	RAB3C	38	1.05
24	rs1353747	5q	PDE4D	90	1.09
25	rs1432679	5q	EBF1	43	1.07
26	rs204247	6p	RANBP9	43	1.05
27	rs17530068	6q	None	22	1.09
28	rs2046210	6q	ESR1	34	1.13
29	rs3757318	6q	ESR1	7	1.21
30	rs11242675	6q	FOXQ1	61	1.06
31	rs720475	7q	ARHGEF5/NOBOX	75	1.06
32	Multiple variants	8q	NBN	0.9	1.3–3.1
33	rs9693444	8p	None	32	1.07
34	rs6472903	8q	None	82	1.10
35	rs2943559	8q	HNF4G	7	1.13
36	rs13281615	8q	MYC	40	1.08
37	rs11780156	8q	MIR1208	16	1.07
38	rs1011970	9p	CDKN2A/B	17	1.09
39	rs865686	9q	KLF4/RAD23B	61	1.12
40	rs10759243	9q	None	39	1.06
41	rs2380205	10p	ANKRD16	56	1.02
42	rs7072776	10p	MLLT10/DNAJC1	29	1.07
43	rs11814448	10p	DNAJC1	2	1.26
44	rs10995190	10q	ZNF365	85	1.16
45	rs704010	10q	ZMIZ1	39	1.07
46	Multiple variants	10q	PTEN	0.01	2.0–10.0
47	rs7904519	10q	TCF7L2	46	1.06
48	rs11199914	10q	None	68	1.05
49	rs2981579	10q	FGFR2	38	1.26
50	rs3817198	11p	LSP1/H19	30	1.07
51	rs3903072	11q	OVOL1	53	1.05
52	rs614367	11q	CCND1/FGFs	15	1.15
53	rs494406	11q	CCND1	26	1.07
54	Multiple variants	11q	ATM	0.3	2.0–3.0
55	rs11820646	11q	None	59	1.09
56	rs10771399	12p	PTHLH	88	1.19
57	rs12422552	12p	None	26	1.05
58	rs17356907	12q	NTN4	70	1.10
59	rs1292011	12q	TBX3/MAPKAP5	58	1.10
60	Multiple variants	13q	BRCA2	0.1	9.0–21.0
61	rs2236007	14q	PAX9/SLC25A21	79	1.08
62	rs999737	14q	RAD51B	78	1.09
63	rs2588809	14q	RAD51L1	16	1.08
64	rs941764	14q	CCDC88C	34	1.06
65	rs3803662	16q	TOX3/LOC643714	25	1.20
66	rs17817449	16q	MIR1972-2-FTO	60	1.08
67	rs11075995	16q	FTO	24	1.10
68	Multiple variants	16q	CDH1	0.01	2.0–10.0
69	Multiple variants	16p	PALB2	0.01	2.0–4.0
70	rs13329835	16q	CDYL2	22	1.08
71	Multiple variants	17q	BRCA1	0.06	5.0–45.0
72	Multiple variants	17q	BRIP2	0.1	2.0–3.0
73	Multiple variants	17p	TP53	0.01	2.0–10.0
74	rs6504950	17q	STXBP4/COX11	73	1.05
75	Multiple variants	17q	RAD51C	0.5	3.2–3.5
76	rs527616	18q	None	62	1.05
77	rs1436904	18q	CHST9	60	1.04
78	Multiple variants	19p	STK11	0.01	2.0–10.0
79	rs8170	19p	MERIT40	19	1.25
80	rs4808801	19p	SSBP4/ISYNA1/ELL	65	1.06
81	rs3760982	19q	KCNN4/ZNF283	46	1.06
82	rs2284378	20q	RALY	33	1.16
83	rs2823093	21q	NRIP1	73	1.09
84	Multiple variants	22q	CHEK2	0.4	2.0–3.0
85	rs132390	22q	EMID1/RHBDD3	3.6	1.12
86	rs6001930	22q	MKL1	11	1.12

Performance of risk-score-based classification

We estimated how well we can identify future breast cancer cases by classifying women into high-risk (targeted) and low-risk (untargeted) subgroups based on the percentiles of their fully and partially known risk scores (see Supplementary Materials and Methods for details). Specifically, we estimated the sensitivity (Sn), specificity (Sp), positive predictive value (PPV), negative predictive value (NPV), and risk in untargeted women relative to that of the population.

Results

We estimated the population variance of the risk scores based on the 86 currently known breast cancer susceptibility variants to be 0.35. This variance, while lower than the estimate of 1.44 for the variance of the fully known risk scores determined using the arguments of Pharoah (3,4), is nevertheless considerably higher than the value 0.07 obtained for the risk scores based on the seven loci known in 2008 (4).

Figure 1 shows the percentage of breast cancer cases included among women having the highest 100 (1- α)^th percent of risk scores, for 0 <α< 1. The curves correspond to the best-case classification with risk score variance equal to 1.44 (solid curve), the currently feasible classification based on partially known risk scores, with variance equal to 0.35 (dashed curve), and the classification based on the seven loci known in 2008 (4) with variance equal to 0.07 (dotted curve). Since the efficacy of risk stratification for prevention depends on the population variance of the risk scores, these results indicate that: a) current genetic knowledge far exceeds that in 2008 (4); and b) despite these gains, considerably better stratification should still be possible in the future, as we better understand the etiology of this disease.

Figure 1. Percent of all breast cancer cases explained by those at highest risk for the disease.

Curves are shown for the best-case scenario when all breast cancer susceptibility alleles are known (solid curve), currently known susceptibility alleles (dashed curve), and seven alleles known in 2008 (4) (dotted curve).

Table 2 shows additional measures of discrimination obtained by classifying women as high- or low-risk on the basis of their risk scores, where the high-risk group is defined as those whose scores exceed the 100 (1- α)^th percentile of the centered Gaussian risk score distribution. How do these results compare with the breast cancer predictions of Roberts (1)? The latter were based on a high-risk group defined as those whose risks exceed the 90^th–95^th percentile of the population distribution. The authors estimated that this classification would target between 10–35% of all future breast cancer cases. In contrast, Table 2 shows that the percentage of cases targeted would be approximately 47% using the best-case classification and 32% using the currently feasible classification. In addition, the authors estimated that the ratio of risk among women classified at low risk relative to the population would be as high as 0.72–0.90, indicating poor specificity. Yet Table 2 suggests that this relative risk is lower: 0.59 with the best-case classification and 0.75 with the currently feasible classification. Thus the present estimates provide more optimistic projections than those obtained using the theoretical model of Roberts (1).

Table 2.

Impact^a of targeting a subset of European-American women for breast cancer preventive strategies

Percent of population at high risk (α)	Lifetime risk in low-risk (%)	Risk for high-risk relative to low-risk	Risk for low-risk relative to population	Percent of cases who are high-risk (Sn)	Percent of non cases who are high-risk (1-Sp)	Percent of high-risk who are cases (PPV)	Percent of low-risk who are non cases (NPV)	NNTb
A. Current knowledge of genetic susceptibility alleles (risk score variance = 0.35)
100	—	—	—	100	100	12.68a	—	95
90	4.11	3.32	0.32	96.76	89.02	13.63	95.89	88
75	5.20	2.92	0.41	89.74	72.86	15.17	94.80	79
67	5.68	2.84	0.45	85.22	64.35	16.13	94.32	74
50	6.64	2.82	0.52	73.82	46.54	18.72	93.36	64
33	7.66	2.99	0.60	59.54	29.15	22.88	92.34	52
25	8.20	3.18	0.65	51.49	21.15	26.12	91.80	46
10	9.54	4.30	0.75	32.31	6.76	40.97	90.46	29
0	12.68	—	1.00	0	0	—	87.32	—
B. Best-case scenario (risk score variance = 1.44)
100	—	—	—	100	100	12.68a	—	95
90	0.83	16.85	0.07	99.34	88.64	14.00	99.17	86
75	1.55	10.60	0.12	96.95	71.81	16.39	98.45	73
67	1.94	9.24	0.15	94.94	62.94	17.97	98.06	67
50	2.92	7.68	0.23	88.48	44.41	22.44	97.08	53
33	4.24	7.04	0.33	77.61	26.52	29.82	95.76	40
25	5.07	7.00	0.40	69.99	18.47	35.50	94.93	34
10	7.51	7.88	0.59	46.67	4.68	59.18	92.49	20
0	12.68	—	1.00	0	0	—	87.32	—

^aAssuming that the lifetime risk of developing breast cancer is 12.68% based upon estimates for the U.S white female population (19).

^bNNT = number of women who must be targeted to avoid one breast cancer death, assuming one sixth of all breast cancers are fatal and targeting prevents death from half of them.

Discussion

We have estimated the variance of breast cancer risk scores among women of European ancestry by using a multiplicative model for the joint effects of currently known breast cancer loci. We found that the distribution of these partially known risk scores has variance 0.35, which is similar in magnitude to the estimate of 0.28 obtained independently by Pashayan (17). We have compared the performance of targeted preventive measures based on the currently feasible partially known risk scores to those obtained using the theoretical distribution of fully known risk scores derived by Pharoah (3,4). The results suggest that the predictive power of genome sequencing to determine breast cancer risk is considerably greater than that described by Roberts (1), and the estimates contradict the authors’ statement that “…our conclusions […] represent an absolute upper bound that cannot be improved by improvements in technology or genetic knowledge.”

To achieve the optimal predictive power represented by the “best-case” classification, we will need to identify the combined effects of all causal alleles for breast cancer. Moreover, better understanding of gene-environment interactions could further improve predictive power. For example, the performance measures described here underestimate the potential value of genomic-based risk classification. This is because a child’s lifetime breast cancer risk is determined not only by her genome, but also by her future levels of nongenetic (lifestyle, environmental and epigenetic) risk factors. Epidemiologic data support a multiplicative model for the joint effects of genetic and nongenetic factors on breast cancer risk (18). Under this model, a modifiable nongenetic factor associated with an overall 50% increase in risk would add considerably more to the absolute risk of a female whose lifetime genetic risk is 36% (increasing it to 54%) than to that of a female whose lifetime genomic risk is 4% (increasing it only to 6%). Thus high-risk women have considerably more to gain by appropriate choices of lifestyle factors than do low-risk women.

In conclusion, the data-based estimates presented here suggest that personalized breast cancer preventive strategies informed by genome sequencing may bring greater gains in cost-efficient disease prevention than previously projected. Moreover, these gains will increase as we gain increased understanding of the etiology of breast cancer.