BOADICEA breast cancer risk prediction model: updates to cancer incidences, tumour pathology and web interface.

BACKGROUND: The Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm (BOADICEA) is a risk prediction model that is used to compute probabilities of carrying mutations in the high-risk breast and ovarian cancer susceptibility genes BRCA1 and BRCA2, and to estimate the future risks of developing breast or ovarian cancer. In this paper, we describe updates to the BOADICEA model that extend its capabilities, make it easier to use in a clinical setting and yield more accurate predictions.
METHODS: We describe: (1) updates to the statistical model to include cancer incidences from multiple populations; (2) updates to the distributions of tumour pathology characteristics using new data on BRCA1 and BRCA2 mutation carriers and women with breast cancer from the general population; (3) improvements to the computational efficiency of the algorithm so that risk calculations now run substantially faster; and (4) updates to the model's web interface to accommodate these new features and to make it easier to use in a clinical setting.
RESULTS: We present results derived using the updated model, and demonstrate that the changes have a significant impact on risk predictions.
CONCLUSION: All updates have been implemented in a new version of the BOADICEA web interface that is now available for general use: http://ccge.medschl.cam.ac.uk/boadicea/.

BOADICEA model

The Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm (BOADICEA) (Antoniou , 2008a) is a risk prediction model for familial breast and ovarian cancer. The model is used to compute BRCA1 and BRCA2 mutation carrier probabilities and age-specific risks of breast and ovarian cancer. It was developed using complex segregation analysis of breast and ovarian cancer based on a combination of families identified through population-based studies of breast cancer, and families with multiple affected individuals who had been screened for BRCA1 and BRCA2 mutations. The latest version of the model was based on 2785 families, of which 537 segregated BRCA1 and/or BRCA2 mutations. BOADICEA models the simultaneous effects of BRCA1 and BRCA2 mutations and assumes that the residual familial clustering of breast cancer is explained by a polygenic component (a large number of genes each of small effect on risk) with a variance that decreases linearly with age.

BOADICEA has been validated in a large series of families from UK genetics clinics (Antoniou ). In the United Kingdom, it is recommended as a risk assessment tool in the National Institute for Health and Care Excellence clinical guideline CG164 (National Institute for Health and Care Excellence, 2013) and has been incorporated in the guidelines of several countries for the management of familial breast cancer (Ontario Breast Screening Program, 2012; Riley ; Smith ).

Cancer incidences

To obtain risk predictions, BOADICEA considers the occurrence of breast, ovarian, pancreatic and prostate cancers in families (Antoniou ). To provide a consistent model, the breast and ovarian cancer incidences over all assumed genetic effects are constrained to agree with population incidences (Antoniou ). The old implementation of the model assumes calendar period- and cohort-specific incidences for the United Kingdom that span the period 1960–1997, taken from the Cancer Incidence in Five Continents (CI5) publications (Ferlay ). These incidences were the most up-to-date and relevant cancer incidences available to us when the BOADICEA model was initially developed. However, breast, ovarian and prostate cancer incidences have increased over time (Hayat ). For example, UK age-specific breast cancer incidences for females shown in Figure 1A (Ferlay ) show that breast cancer incidence has increased over the time period between 1960 and 2010. As BOADICEA is used to predict future risks of developing breast or ovarian cancer, it is essential to consider the most up-to-date cancer incidences. Cancer incidences from recent calendar periods (1992–2010) are now available and have been included in this version of BOADICEA.

Figure 1

(A) The age-specific female breast cancer incidences per 100 000 for the United Kingdom for the periods 1960–1963 and 1973–1977, and the years 1993 and 2010. (B) The age-specific female breast cancer incidences per 100 000 for 2007 for the United Kingdom, Australia, Canada, Sweden and the United States. (C) The old, non-smoothed and smoothed updated incidence per 100 000 for females born in 1975 used in the BOADICEA code. (D) The BOADICEA predicted risk for a 30-year-old female born in 1975 with no family history information for the United Kingdom, updated and old, and other countries.

Furthermore, incidences of breast, ovarian, pancreatic and prostate cancers vary widely by geographical region (Parkin ). Figure 1B demonstrates the differences in age-specific female breast cancer incidence between various countries in the year 2007. As BOADICEA is now used in over 45 countries, in the new version of BOADICEA we have included incidences specific to other countries.

Breast tumour pathology

The BOADICEA has recently been extended to incorporate breast tumour pathology information, where breast cancer subtypes are treated as distinct disease end points (Mavaddat ). In particular, oestrogen receptor (ER) status, triple-negative (TN) status (ER, progesterone (PR) and HER2 negative) and expression of basal markers (CK5/6 and CK14) are taken into account. Initially, tumour subtype distributions for BRCA1 and BRCA2 mutation carriers were based on data from the Breast Cancer Linkage Consortium (BCLC) (Lakhani ), and data from the Surveillance, Epidemiology and End Results (SEER) Program (Surveillance Epidemiology and End Results (SEER) Program, 2006) were used to obtain the ER distribution in the general population. However, because of the relatively small number of mutation carriers in the BCLC study (182 BRCA1 and 64 BRCA2), these distributions were imprecise, particularly for BRCA2 tumours. Recent analyses based on the much larger collaborative data set from the Consortium of Investigators of Modifiers of BRCA1/2 (CIMBA) demonstrated additional differences in the characteristics of BRCA1 and BRCA2 tumours, such as an increasing proportion of ER-negative tumours with age among BRCA2 carriers (Mavaddat ).Up-to-date data from CIMBA, in combination with data from the Breast Cancer Association Consortium (BCAC) on the age-specific frequencies in the general population, by tumour type, allow us to incorporate more accurate distributions of ER, PR and HER2 status into BOADICEA.

Computational requirements

In order to exploit fully the predictive potential of genetic models in the future, it will be necessary to add the explicit effects of known breast cancer susceptibility variants such as common SNPs (Michailidou ) and rare moderate-risk variants, such as in ATM, CHEK2, PALB2 and BRIP1 (Meijers-Heijboer ; Renwick ; Seal ; Rahman ). However, this presents a substantial problem in terms of the future computational requirements of the BOADICEA program. As the old implementation of the model uses the Elston–Stewart algorithm (Elston and Stewart, 1971) to compute pedigree likelihoods, incorporating additional genetic effects will result in an exponential increase in runtime. This problem is exacerbated by the fact that we may not be able to rely on future increases in single-processor performance (Fuller and Millett, 2011). In order to use these more sophisticated models in a clinical setting, they must run faster.

BOADICEA web application

We have implemented the BOADICEA model as a standalone FORTRAN program, termed here the BOADICEA FORTRAN program (BFP). In recent years, scientists have used the BFP as a research tool. However, in practice, it is time consuming to set up and run calculations using the BFP alone, which makes it inappropriate for use in a clinical setting. To address this problem, we developed the BOADICEA web application (BWA; http://ccge.medschl.cam.ac.uk/boadicea/), (Cunningham ), a user-friendly web interface to the BFP that makes it much easier for health-care professionals and members of the public to run BOADICEA calculations. The first version of the BWA was released for general use in November 2007. Since then, the number of BWA users has grown, and the purposes for which they use it have diversified. The BWA is now widely used for genetic counselling purposes with more than 3000 registered users based in >50 countries.

This report

In this paper, we first describe the update to the UK cancer incidences and the extensions to the BOADICEA model to include cancer incidences from other regions. We then describe updates to the distributions of tumour pathology characteristics using new data on BRCA1 and BRCA2 mutation carriers and women with breast cancer from the general population obtained from the largest data sets currently available. We also describe improvements to the computational efficiency of the algorithm so that risk calculations now run substantially faster, and a new version of the BWA (termed here BWA v3) that incorporates these new features and additional updates to make it easier to use in a clinical setting. We discuss cancer risk predictions and mutation carrier probabilities generated using the updated model. Finally, we summarise conclusions and prospects for future work.

Materials and methods

The underlying genetic model

Details of the underlying statistical model in BOADICEA have been described previously (Antoniou ; Mavaddat ). Briefly, the breast cancer incidence for individual i at age t was assumed to be birth cohort specific and to depend on the underlying BRCA1 or BRCA2 genotype and phenotype through a model in the form , where is the baseline incidence for the cohort, represents the major gene effect at age t (i.e., the age-specific log-relative risks associated with BRCA1 and BRAC2 mutations, with for non-mutation carriers), and where is the polygenic effect assumed to be normally distributed with mean 0 and variance . The polygenic variance and BRCA1 and BRCA2 log-relative risks were estimated previously (Antoniou ) and are assumed to remain the same in all the extensions presented. To obtain the baseline incidences, the cohort-specific breast cancer incidences over all assumed genetic effects are constrained to agree with the assumed population incidences (Antoniou ) that have been updated as described below. In all our extensions we assumed that the BRCA1 and BRCA2 mutation frequencies remain as previously estimated (Antoniou ). The genetic model can therefore be fully specified. BOADICEA incorporates tumour phenotypes by treating breast cancer subtypes as different disease end points (Mavaddat ). For example, in the case of ER status, breast cancer is divided into ER-negative and ER-positive disease with incidences that depend on the underlying genetic effects. To obtain the ER-specific incidences for each genotype in the model an extra constraint is imposed such that the overall incidence over ER status, major gene (BRCA1 carriers, BRCA2 carriers and noncarriers) and polygenic effects agrees with the population breast cancer incidences (Mavaddat ). This requires knowledge of the age-specific distributions of breast cancer subtypes in BRCA1 and BRCA2 mutation carriers and the general population.

Incidence update

We have updated the BOADICEA model to include UK population cancer incidences for more recent calendar periods using data from two sources: (1) data for the period 1960–1992 were obtained from CI5 (Ferlay ) (reported for periods: 1960–1962, 1963–1967, 1968–1972, 1973–1977, 1978–1982, 1983–1987 and 1988–1992); and (2) data for the years 1992–2010 were obtained from the Office for National Statistics (2011). Both sets of incidences were reported in 5-year age intervals up to the age of 84 years (0–4, 5–9, 10–14 and so on). As in Antoniou ), the cancer incidences were assumed to be calendar period and cohort specific. In previous versions of the BOADICEA model, we assumed only five birth cohorts. We have now extended this to eight birth cohorts in order to capture the incidences more accurately for those born more recently. For each birth cohort (<1919, 1920–1929, 1930–1939, 1940–1949, 1950–1959, 1960–1969, 1970–1979 and >1980) we derived lifetime incidences by assuming that each individual was born at the midpoint of the relevant birth cohort (1915 for the first cohort and 1985 for the last cohort) and that at each age they experience the relevant calendar period incidences. Incidences before 1960 were assumed to be the same as for the period 1960–1962. Incidences post 2010 were assumed to be the same as for 2010. As in the original BOADICEA model (Antoniou ), we assumed that the relative risks associated with BRCA1 and BRCA2 mutations and the polygenic variances were the same for each birth cohort. However, the absolute risks of disease for each underlying genotype in the model were not the same because the incidences over all genetic effects were constrained to agree with the population incidences.

Published incidences are reported in 5-year age intervals that can result in large variations between adjacent age intervals. As it is plausible that incidences vary continuously with age, we smoothed the population incidences using a locally weighted regression technique (LOWESS) (Cleveland, 1979) with a bandwidth of 0.2. We smoothed the population incidences using the STATA statistical software (StataCorp, College Station, TX, USA, Stata Statistical Software: Release 11, 2009), as described by Antoniou ).

To make the model more specific for populations outside the United Kingdom, we compiled incidences for other geographical regions, including Australia, Canada, Denmark, Finland, Iceland, New Zealand, Norway, Sweden and the United States. Data for these regions were obtained from a combination of CI5 and CI5 plus (Ferlay ), NORDCAN (Engholm , 2013) as well as national bodies (Australian Institute of Health and Welfare (AIHW), 2011; Surveillance Epidemiology and End Results (SEER) Program, 2011a, 2011b, 2011c; Statistics Canada, 2012; Lewis, 2013). The BOADICEA model was originally developed using data reported for families of European ancestry. As cancer incidences vary by ethnic background, we derived incidences for the United States using only data reported for US whites, and for New Zealand using data from the non-Maori population. In all cases incidences were reported in 5-year age intervals.

We derived smoothed calendar period- and cohort-specific incidences for these regions in the same manner as for the United Kingdom, for the same 8 birth cohorts. We assumed that relative risks of all cancers conferred by BRCA1 and BRCA2, relative to the population incidences, was the same for each country. However, this still allowed for the absolute cancer risks conferred by BRCA1 and BRCA2 to vary between countries. We also assumed that the polygenic variance was the same for each country. As part of these updates, we also modified the BFP so that we can now easily extend the BOADICEA model to include cancer incidences from other countries at the request of users, if appropriate data are available.

Pathology proportions

We obtained age-specific proportions of ER and TN tumour status for BRCA1 and BRCA2 mutation carriers from CIMBA data previously described by Mavaddat ). For this purpose, we used an updated data set from CIMBA that included ER status information on 3832 BRCA1 mutation carriers and 2399 BRCA2 mutation carriers. Of those who had ER-negative tumours, a total of 1582 BRCA1 mutation carriers and 231 BRCA2 mutation carriers, had information on breast cancer TN status.

All mutation carriers were of self-reported European ancestry. We derived the corresponding distributions of tumour characteristics in the general population using country-matched data from the Breast Cancer Association (Blows ; Broeks ; Garcia-Closas ), based on participants of European ancestry. We used the distributions of the CK5/6 and CK14 tumour markers among TN tumours reported by Mavaddat ) because of the lack of information on these markers in both CIMBA and BCAC. We used 5-year age intervals for ER status in the general population. However, we used longer age intervals for BRCA1 and BRCA2 mutation carriers and for the TN distributions to ensure that we obtained robust estimates of these proportions. As it is reasonable to assume that the age-specific proportions of ER-positive and TN tumours vary continuously with age, we smoothed these parameters using LOWESS (Cleveland, 1979) with a bandwidth of 0.2.

Optimisation

Previous versions of the BOADICEA model have used the MENDEL pedigree analysis software package (Lange ) to calculate likelihoods on pedigrees. In order to exploit fully the predictive potential of genetic models, it is necessary to add the explicit effects of known breast cancer susceptibility variants. However, if we were to use MENDEL to compute the pedigree likelihoods under these circumstances, it would result in an exponential increase in runtime. This presents a substantial problem in terms of the future computational requirements of the BOADICEA program. In addition, the MENDEL FORTRAN source codes used in previous versions of the BOADICEA model were implemented in accordance with the FORTRAN 77 standard that has now been superseded by more recent standards.

In order to improve the computational efficiency of the BFP and thereby facilitate future extensions of the model, we have optimised the algorithm in a number of ways:

(1) We have converted the source code from FORTRAN 77 to FORTRAN 95, and restructured it into modules. The restructured code is easier to read, and simpler to maintain and extend. It has also been designed to allow easy parallelisation of the code in future versions. The FORTRAN 95 module structure allows data abstraction and enforces automatic interfacing, reducing the possibility of programming errors. The conversion of the source codes to FORTRAN 95 has made dynamic memory allocation possible that has enabled us to implement the algorithm in a more resource-efficient manner.

(2) We profiled the code and found that the majority of the runtime was consumed by array multiplication. To address this, where possible we now perform array multiplication using the Basic Linear Algebra Subprograms (BLAS) libraries (Lawson ; Dongarra , 1990b). This has reduced the runtime of the program by a factor of 3.5. We also noted that a significant proportion of the runtime of the code was spent repeatedly calculating a small number of variables. As a result, we modified the code so that these variables were calculated once at the beginning of the program, stored in a lookup table and retrieved when required. This further reduced the runtime of the code by a factor of 1.7, giving an overall decrease by a factor of approximately 6. The relative change in runtime is dependent on the specific computer being used and the pedigree being processed. Here we used a desktop computer with an Intel Xeon E5630 processor and a test pedigree consisting of a dendritic three generational family with 12 people.

(3) Furthermore, we found that these modifications to the FORTRAN source code also resulted in more effective compiler optimisation. In recent years, we have compiled previous versions of the BOADICEA program with the GNU FORTRAN compiler (http://www.gnu.org) (version 4.4.3). However, as part of this study, we conducted benchmark tests to compare the performance of BOADICEA program executables compiled with the GNU FORTRAN compiler and Intel FORTRAN compiler (http://www.intel.co.uk) (version 11.1). These tests showed that when we compiled the previous version of the program with the Intel FORTRAN compiler, the runtime was reduced by a factor of approximately 1.05 (relative to that of an equivalent executable built by compiling the same code with the GNU FORTRAN compiler). However, when we compiled the latest version of the program (with the modifications described above) using the Intel FORTRAN compiler, the runtime of the code was reduced by a factor of approximately 1.37 (relative to that of an equivalent executable built by compiling the same code with the GNU FORTRAN compiler). Relative runtime changes are for the same pedigree and computer as in (2) above.

Web interface updates

We have now implemented BWA v3 in order to accommodate extensions to the BOADICEA model presented here, and to implement modifications requested by users to make the program easier to use in a clinical setting. BWA v3 makes the latest version of the BOADICEA model easily accessible to health-care professionals and members of the public.

Use of pathology and cancer incidence data

BWA v3 now enables users to include pathology information in risk calculations. The BWA enables users to either build a pedigree online for processing or to upload a text file containing one or more pedigrees for processing. When users build an input pedigree online, the program now prompts for details of ER, PR, HER2, CK14 and CK5/6 status. Similarly, we have extended the BOADICEA import/export format (described in Appendix A of the BWA v3 user guide: https://pluto.srl.cam.ac.uk/bd3/v3/docs/BWA_v3_user_guide.pdf) so that users can include the new pathology parameters in the data files that they upload for processing.

BWA v3 allows users to select cancer incidences from a number of countries. The program also includes the most up-to-date UK incidences, but users can still select UK cancer incidences used in previous versions of the BWA for backward compatibility.

Batch processing

Previous versions of the BWA enabled users to either build or upload a single input pedigree. BWA v3 now includes a batch processing module that enables users to upload and process multiple pedigree data sets in a single processing run. When the user uploads a text file that contains multiple pedigrees, BWA v3 now initiates a batch processing job to validate and process the pedigrees sequentially. When the processing job has run to completion, the user can download the computed results across a secure web connection.

Displaying equivalent baseline predicted risks

In recent years, clinical geneticists have requested a means to plot BOADICEA breast and ovarian cancer risks with equivalent baseline cancer risks (i.e., the equivalent predicted cancer risks for a random female of the same age from the general population) to help them communicate the significance of the computed risks to patients. To meet this requirement, BWA v3 now plots BOADICEA breast and ovarian cancer risks and equivalent baseline cancer risks in graphs. BWA v3 also includes additional updates requested by clinical geneticists to make the program easier to use in a clinical setting (e.g., improved pedigree building and pedigree data validation functions).

Checking for breaks in pedigree trees

The BWA allows users to upload input pedigrees for processing. However, in the past, we have encountered problems when a single pedigree has included multiple disjoint family trees or disconnected individuals. As a result, BWA v3 now checks that all family members within a single pedigree are genealogically connected to the index before it is processed.

Results

Incidences

We derived up-to-date calendar- and cohort-specific incidences for breast, ovarian, prostate and pancreatic cancer for the United Kingdom and a number of other regions. For each cancer site, gender and region, we derived eight sets of incidences corresponding to different birth cohort periods, which were then smoothed. The effect of smoothing on breast cancer incidence for a UK female born in 1975 is demonstrated in Figure 1C. The resulting smoothed incidences capture the locality of age-specific incidences before smoothing. Figure 1C also shows a comparison of the updated incidences and previous BOADICEA incidences. The corresponding absolute breast cancer risks for a 30-year-old UK female born in 1975 are shown in Figure 1D. The updated breast cancer incidences are higher at all ages. The remaining risk by age 80 years is 11.4% under the updated model as compared with 9.2% based on the previous implementation.

As expected, differences in the population-specific cancer incidences result in differences in the absolute risks of developing breast or ovarian cancer predicted by the BOADICEA model. This is demonstrated in Figure 1D that shows how the risk of breast cancer predicted by BOADICEA varies by country (for a 30-year-old female born in 1975). Table 1 shows the population-based lifetime risks for all cancer sites considered in BOADICEA for some of the regions now available in BWA v3.

Table 1

Lifetime risk by age 80 years as a percentage, for each country, for each birth cohort

Country/YOB	1915	1925	1935	1945	1955	1965	1975	1985
Female breast cancer
Australia	7.43	8.25	9.29	10.57	10.47	10.59	10.61	10.61
Canada	9.00	9.66	9.97	10.53	10.26	10.26	10.25	10.25
Sweden	7.53	8.49	10.06	12.15	12.33	12.57	12.73	12.77
United Kingdom	7.15	8.35	9.65	11.13	11.16	11.45	11.49	11.50
United Kingdom (old)	7.15	8.02	8.64	9.13	9.24	9.24	9.24	9.24
United States	10.79	12.24	12.29	12.85	12.32	12.31	12.31	12.31
Male breast cancer
Australia	0.11	0.11	0.11	0.11	0.11	0.11	0.11	0.11
Canada	0.09	0.11	0.10	0.10	0.10	0.10	0.10	0.10
Sweden	0.07	0.09	0.06	0.07	0.08	0.08	0.08	0.08
United Kingdom	0.09	0.09	0.10	0.12	0.12	0.12	0.12	0.12
United Kingdom (old)	0.09	0.08	0.09	0.09	0.09	0.09	0.09	0.09
United States	0.11	0.12	0.13	0.12	0.12	0.12	0.12	0.12
Female ovarian cancer
Australia	1.28	1.31	1.17	1.11	1.06	1.04	1.05	1.04
Canada	1.55	1.52	1.37	1.31	1.23	1.23	1.22	1.22
Sweden	1.91	1.89	1.58	1.42	1.17	1.10	1.07	1.07
United Kingdom	1.62	1.82	1.85	1.78	1.69	1.69	1.72	1.74
United Kingdom (old)	1.61	1.75	1.75	1.72	1.72	1.72	1.72	1.72
United States	1.82	1.70	1.54	1.40	1.33	1.29	1.29	1.29
Male prostate cancer
Australia	9.35	11.6	14.62	18.05	19.26	19.47	19.48	19.48
Canada	10.04	11.49	12.69	14.07	14.56	14.61	14.60	14.60
Sweden	10.03	11.83	14.19	16.05	16.51	16.59	16.59	16.59
United Kingdom	5.44	7.22	9.88	11.76	12.28	12.36	12.37	12.36
United Kingdom (old)	5.39	6.80	7.37	7.45	7.46	7.46	7.46	7.46
United States	13.12	15.10	15.02	15.08	15.07	15.11	15.12	15.12
Female pancreatic cancer
Australia	0.92	0.91	0.92	0.89	0.87	0.88	0.88	0.88
Canada	0.93	0.99	0.96	0.95	0.95	0.95	0.95	0.95
Sweden	0.97	0.81	0.86	0.90	0.85	0.86	0.86	0.87
United Kingdom	0.87	0.90	0.91	0.95	0.95	0.96	0.96	0.96
United Kingdom (old)	0.87	0.87	0.86	0.85	0.85	0.85	0.85	0.85
United States	1.04	1.02	1.06	1.07	1.08	1.08	1.08	1.08
Male pancreatic cancer
Australia	1.29	1.25	1.23	1.26	1.25	1.24	1.24	1.24
Canada	1.34	1.28	1.27	1.25	1.25	1.25	1.25	1.25
Sweden	1.28	1.08	0.95	1.03	1.00	1.00	1.01	1.01
United Kingdom	1.33	1.20	1.19	1.20	1.21	1.21	1.21	1.21
United Kingdom (old)	1.32	1.21	1.17	1.17	1.16	1.16	1.16	1.16
United States	1.45	1.39	1.40	1.42	1.44	1.44	1.44	1.45

Abbreviations: BOADICEA=Breast and Ovarian Analysis of Disease Incidence and Carrier Estimation Algorithm; YOB=year of birth.

United Kingdom (old) figures are calculated from the incidences used in the previous version of BOADICEA, whereas United Kingdom figures are from the updated incidences.

We derived age-specific pathology proportions for ER and TN status for the general population, BRCA1 mutation carriers and BRCA2 mutation carriers (Tables 2 and 3). The proportion of ER-negative tumours among BRCA1 mutation carriers decreased with increasing age at diagnosis, whereas the proportion of ER-negative tumours in BRCA2 mutation carriers increased with increasing age at diagnosis, in contrast to the decreasing trend in tumours in the general population seen in the BCAC data Table 1. The proportions of TN tumours among those individuals with an ER-negative tumour were estimated to be 88% and 75.8% in BRCA1 and BRCA2 mutation carriers, respectively. The corresponding proportion of TN tumours in the general population decreased with increasing age at diagnosis from 65.8% for ages <30 years to 57.8% for those diagnosed with ER-negative breast cancer at age ⩾40 years. In Figure 2, the updated distributions of ER and TN status for BRCA2 mutation carriers and the general population are compared with those described in the original publication.

Table 2

Age-specific proportions of oestrogen receptor (ER)-negative tumours and ER-negative tumours divided by all tumours for which information on ER status was available in the general population and BRCA1 and BRCA2 mutation carriers

Age group (years)	General population % ER -negative (no. of ER-negative tumours out of total no. of tumours)	BRCA1 mutation carriers % ER negative (no. of ER-negative tumours out of total no. of tumours)	BRCA2 mutation carriers % ER negative (no. of ER-negative tumours out of total no. of tumours)
<30	46.2 (162 out of 351)	82.0 (1422 out of 1734)	18.8 (158 out of 840)
30–34	43.8 (463 out of 1058)
35–39	34.0 (743 out of 2185)
40–44	27.3 (964 out of 3536)	74.0 (1015 out of 1372)	21.4 (191 out of 894)
45–49	23.4 (1250 out of 5332)
50–54	22.1 (1525 out of 6889)	70.0 (390 out of 557)	27.2 (121 out of 445)
55–59	21.0 (1370 out of 6522)
60–64	17.8 (1138 out of 6415)	55.6 (94 out of 169)	27.7 (61 out of 220)
65–69	15.9 (840 out of 5274)
70–74	16.5 (488 out of 2952)
>74	14.6 (328 out of 2254)

General population data were obtained from the Breast Cancer Association Consortium (BCAC) and BRCA1/2 mutation carrier data were obtained from the Consortium of Investigators of Modifiers of BRCA1/2 (CIMBA).

Figure 2

(A) The age-specific proportions of ER-negative tumours used in BOADICEA for The previous version used data from SEER for the general population, whereas the new version uses data from CIMBA, specific to BRCA2 mutation carriers. (B) The age-specific proportions for TN tumours among ER-negative tumours for the general population. The previous version of BOADICEA used an age constant proportion derived from BCLC data, whereas the new version uses an age-specific proportion derived from BCAC data.

Discussion

In this study, we have described updates and further extensions to the BOADICEA model for genetic susceptibility to breast cancer. These included extensions to the input parameters of the underlying genetic model, to the algorithm and to the BWA. Changes to the model input parameters included the use of up-to-date cancer incidences, new population-specific incidences and updates to the distributions of tumour pathology characteristics using large data sets from CIMBA and BCAC. We have improved the computational efficiency of the algorithm so that BOADICEA risk calculations now run substantially faster. These updates also make possible future extensions of BOADICEA to model more complex genetic effects. In addition, we have implemented BWA v3 so that the latest version of the BOADICEA model is easily accessible to health-care professionals and members of the public. BWA v3 also includes several modifications requested by clinical geneticists to make the software easier to use in a clinical setting.

A consequence of including more up-to-date cancer incidences is that, given the same input, cancer risks predicted by the new BOADICEA model are higher than those predicted by previous versions. This is to be expected because population cancer incidences have increased over time (in particular, the incidence of breast cancer). The observed increase in breast cancer incidences over time is known to be driven by the increased prevalence of screening. These changes in incidences could explain results from two recently published studies that evaluated BOADICEA. A Swedish study reported a slight underestimation (which was not statistically significant) in the predicted number of invasive breast cancers for the first version of BOADICEA (Stahlbom ). However, a more recent prospective study that evaluated the BOADICEA model with the updated cancer incidences presented here has demonstrated that the updated version is well calibrated for predicting first invasive breast cancers overall and for most age and family history subgroups (Macinnis ). Furthermore, the same study suggested that the updated model has good discriminatory accuracy with an estimated area under the receiver-operating characteristic curve of 0.70 (95% CI: 0.66–0.75).

In extending the model, we are implicitly assuming that the genetic loci (both BRCA1/2 and the variants comprising the polygenic component) continue to confer the same relative risk in a screened population. There are no direct data to substantiate this, but there is evidence that the common SNPs associated with invasive breast cancer are also associated with DCIS (Easton ; Michailidou ). In the future, we hope to incorporate the effect of screening into BOADICEA more explicitly. Similarly, we assumed that the BRCA1 and BRCA2 relative risks associated with different cancer sites relative to the population incidences remained the same across birth cohorts and across populations. These assumptions are consistent with the observation that the absolute risk of breast cancer in carriers increases with more recent year of birth, in line with the increased risk observed for breast cancer in the general population, and with evidence of variation in the absolute cancer risks by country (Antoniou ; Simchoni ; Milne ). The assumptions are also consistent with a model in which the risks of cancer in carriers are modified by lifestyle risk factors in an approximately multiplicative model. A multiplicative model for the combined effects of lifestyle risk factors and the established common variants has been shown to fit well (i.e., no evidence for GxE interaction), but has not been shown directly for most breast cancer risk factors, although we have shown elsewhere that oral contraceptive use confers a similar (protective) association for ovarian cancer risk in BRCA1/2 mutation carriers as in the general population (Antoniou ). Moreover, the absolute risks of breast cancer in BRCA1/2 mutation carriers have been shown to increase with more recent year of birth (Antoniou ).

By incorporating cancer incidences from different countries, we also assumed that the distribution of the breast cancer polygenic component remains the same across populations. This assumption would be violated where polygenic effects exist because of loci conferring small effects on risk (e.g., the common breast cancer susceptibility alleles) were different across populations. However, recent results from large genome-wide association studies (based on data from international consortia) did not demonstrate any significant heterogeneity in the associations of common breast cancer susceptibility alleles across populations of European ancestry (Michailidou ). However, it is possible that the polygenic distribution differs between populations of different ethnic ancestry; it is known that some loci confer different risks in women of East Asian ancestry, and the frequencies of many of the risk SNPs differ markedly by ethnicity. In addition, BRCA1 and BRCA2 mutation frequencies also vary between populations, and previous studies have shown that BOADICEA may not be well calibrated in populations of non-European ancestry (Thirthagiri ). Extending the model to non-European populations, particularly from Asia and Africa, may therefore require additional modifications and perhaps re-estimation of the model parameters. Although not currently implemented, the use of different baseline incidences also provides a basis for incorporating the effect of nongenetic risk factors, such as reproductive history, HRT use and mammographic density into the model.

A previous version of the BOADICEA model incorporated tumour pathology information based primarily on data from BCLC that included only 182 BRCA1, 62 BRCA2 mutation carriers and 109 controls (Mavaddat ). Because of limited data, accurate estimates of the age-specific distributions of tumour marker information could not be obtained. Further information on the distribution of ER tumour status in BRCA2 mutation carriers had been scarcer. We had therefore assumed that the distribution of ER status for BRCA2 mutation carriers was the same as in the general population. Recent analyses based on CIMBA data (Mavaddat ) however suggest that the proportion of ER-negative tumours in BRCA2 mutation carriers increases with increasing age at diagnosis. Such differences could potentially improve the predictive ability of the model. Using data from CIMBA and BCAC, we were able to include age-specific distributions of ER and TN status for mutation carriers and the general population. A recent evaluation of BOADICEA for its ability to predict BRCA1 and BRCA2 carrier status showed that the updated BOADICEA model incorporating tumour pathology information, based on the CIMBA/BCAC data distributions, provided a significant improvement in discrimination and re-classification over the previous BOADICEA model without pathology, and that the updated BOADICEA model also performed better than the previous BOADICEA model that incorporated tumour pathology information based primarily on the BCLC data (Fischer ). These results suggest that the updated BOADICEA model is a valid tool for use in genetic counselling.

The inclusion of tumour pathology in BOADICEA affects both the predicted mutation carrier probabilities, as demonstrated by Mavaddat ), and the predicted risks of developing breast or ovarian cancer. Figure 3A shows the predicted remaining lifetime risks of breast cancer by age 80 years for a healthy female aged 30 years, born in 1975, depending on the age at breast cancer diagnosis and tumour characteristics of her mother. BOADICEA is the only model to: (1) use tumour pathology information from affected relatives of the proband and (2) use tumour pathology information for cancer risk predictions. This is a unique feature among current breast cancer risk models. (Tai ; Evans ).

Figure 3

Risk of (A) breast cancer and (B) ovarian cancer calculated by BOADICEA for a female aged 30 years, born in 1975, as a function of her mother's age at breast cancer and cancer subtype.

As part of this work, we have implemented BWA v3 to make the updated BOADICEA model easily accessible to health-care professionals and members of the public. During the course of the BOADICEA project, clinical geneticists have provided important feedback on the usability of the program. As a result, BWA v3 also includes additional modifications to make the software easier to use in a clinical setting.

In summary, we present updates to the BOADICEA breast and ovarian cancer risk prediction model that result in more accurate cancer risk predictions and that make the software easier to use in a clinical setting. However, it will be important to evaluate the updated BOADICEA model in prospective studies for its ability to predict future cancer risks. Our current research focusses on extending BOADICEA to include the explicit effects of SNPs known to be associated with breast or ovarian cancer risk, the effects of other breast or ovarian cancer risk factors and modelling the residual genetic susceptibility to ovarian cancer not due to BRCA1 or BRCA2.

Table 3

Proportion of triple-negative (ER-negative, PR-negative and HER2-negative) tumours as a proportion of all tumours for which information on ER, PR and HER2 status was available

Age group (years)	General population, % TN (no. of TN tumours out of total no. of tumours)	BRCA1 mutation carriers, % TN (no. of TN tumours out of total no. of tumours)	BRCA2mutation carriers, % TN (no. of TN tumours out of total no. of tumours)
<30	65.8 (52 out of 79)	88.0 (1392 out of 1582)	75.8 (175 out of 231)
30–39	61.7 (338 out of 548)
>39	57.8 (2394 out of 4140)

Abbreviations: ER=oestrogen receptor; HER2=human epidermal growth factor receptor 2; PR=progesterone; TN=triple negative.

General population data were obtained from the Breast Cancer Association Consortium (BCAC) and BRCA1/2 data were obtained from the Consortium of Investigators of Modifiers of BRCA1/2 (CIMBA).