Ovarian and Breast Cancer Risks Associated With Pathogenic Variants in RAD51C and RAD51D

Abstract Background The purpose of this study was to estimate precise age-specific tubo-ovarian carcinoma (TOC) and breast cancer (BC) risks for carriers of pathogenic variants in RAD51C and RAD51D. Methods We analyzed data from 6178 families, 125 with pathogenic variants in RAD51C, and 6690 families, 60 with pathogenic variants in RAD51D. TOC and BC relative and cumulative risks were estimated using complex segregation analysis to model the cancer inheritance patterns in families while adjusting for the mode of ascertainment of each family. All statistical tests were two-sided. Results Pathogenic variants in both RAD51C and RAD51D were associated with TOC (RAD51C: relative risk [RR] = 7.55, 95% confidence interval [CI] = 5.60 to 10.19; P = 5 × 10-40; RAD51D: RR = 7.60, 95% CI = 5.61 to 10.30; P = 5 × 10-39) and BC (RAD51C: RR = 1.99, 95% CI = 1.39 to 2.85; P = 1.55 × 10-4; RAD51D: RR = 1.83, 95% CI = 1.24 to 2.72; P = .002). For both RAD51C and RAD51D, there was a suggestion that the TOC relative risks increased with age until around age 60 years and decreased thereafter. The estimated cumulative risks of developing TOC to age 80 years were 11% (95% CI = 6% to 21%) for RAD51C and 13% (95% CI = 7% to 23%) for RAD51D pathogenic variant carriers. The estimated cumulative risks of developing BC to 80 years were 21% (95% CI = 15% to 29%) for RAD51C and 20% (95% CI = 14% to 28%) for RAD51D pathogenic variant carriers. Both TOC and BC risks for RAD51C and RAD51D pathogenic variant carriers varied by cancer family history and could be as high as 32–36% for TOC, for carriers with two first-degree relatives diagnosed with TOC, or 44–46% for BC, for carriers with two first-degree relatives diagnosed with BC. Conclusions These estimates will facilitate the genetic counseling of RAD51C and RAD51D pathogenic variant carriers and justify the incorporation of RAD51C and RAD51D into cancer risk prediction models.


Variant frequencies
We estimated the RAD51C and RAD51D pathogenic variant frequencies in the population using the UK Biobank exome sequencing dataset (http://www.ukbiobank.ac.uk). Specifically, among the 49,960 available subjects, we selected cancer-free individuals (either self-reported or medical records) and removed relatives up to second degree, leaving 42,325 individuals for the variant frequency estimation. The pathogenic variants within RAD51C and RAD51D were extracted.
Variants in the last exon were excluded. The pathogenic variant frequencies were estimated and were used as input parameters in the segregation analysis.

Missing age at cancer diagnosis
Individuals with missing age at cancer diagnosis but other age information available were assumed to develop the corresponding cancer at the minimum available age.
For those without any age information available, we assigned the age at cancer diagnosis to be the "average cancer-specific age at diagnosis" obtained from: the family, within the study group and within the country, whichever was available in this order. A summary of the number of individuals with missing age is shown in Supplementary Table 13.

Statistical models
Two main genetic models were fitted: (1) a major-gene model that assumed all familial aggregation of tubo-ovarian carcinoma (TOC) and breast cancer (BC) to be due to where 0 ( , , ) is the baseline incidence for non-RAD51C/D carriers at age t for cohort k and country c, Gi is an indicator variable taking values 1 for RAD51C/D pathogenic variant carriers and 0 for non-carriers, and is the polygenic component which was set to 0 under the single-gene models and was assumed to be normally distributed with mean 0 and variance 2 under the polygenic models (3,4). ( ) is the log-risk ratio for RAD51C/D pathogenic variant relative to non-carriers. To ease interpretation, the models were parameterised in terms of the cancer-specific logrelative risk (log-RR) for RAD51C and RAD51D pathogenic variant carriers relative to the population incidences for TOC and BC. Specifically, the RR at age t was defined as: where i RAD51C/D+ (t, k, c) denotes the average cancer incidence for RAD51C/D pathogenic variant carriers at age t born in cohort k from country c (over all polygenic effects) and i pop (t, k, c) denotes the population incidence at age t for cohort k and country c.
We constrained the total genetic variance ( 2 ), which was defined as the sum of the variance due to RAD51C/D pathogenic variant ( 2 ) and the residual polygenic variance ( 2 ), to agree with external estimates of the total polygenic variance. This was assumed be equal to 2.06 for TOC and 1.66 for BC, based on estimates from previously published segregation analyses (1, 5-7).
When the logRR for RAD51C/D pathogenic variant carriers relative to the population incidences was assumed to be a piecewise linear function of age, the logRR(t) was modelled as: where, t is the age, is the age-breakpoint where the slope changes to 2 . We optimised by fitting a series of models in which took values from age 55 to 65 (the plausible age range from the age-specific logRR models).

Cancer incidences
Country-and cohort-specific population cancer incidences (Cancer incidence in five continents, http://ci5.iarc.fr/CI5plus/Default.aspx) were used here to take into account differences in incidences by study group, study location and changes in incidences over time. The overall cancer incidences were constrained over all assumed genetic effects in the model to agree with the population incidences (5). The reported 5-year interval constant incidences were smoothed using the locally weighted regression LOWESS approach (8,9). A total of eight cohort-specific incidences (<1920, 1920-1929, 1930-1939, 1940-1949, 1950-1959, 1960-1969, 1970-1979 and >1980) were used in the model by assuming each individual was born at the midpoint of each assumed cohort period (1915 for the first cohort and 1985 for the last cohort).

Ascertainment adjustment
We adjusted for ascertainment for each family separately by employing an assumption-free approach (10)(11)(12). We divided the data for each family into two parts depending on whether the data could be relevant to the ascertainment (F1) or not (F2).  Table 4).

Variant screening sensitivity
Four studies (ICR, UKFOCSS, UKFOCR and SEARCH) provided data on all families screened for RAD51C or RAD51D variants, irrespective of the mutation search result.
Details of these studies and methods have been published elsewhere (13)(14)(15). In these families only the proband was screened for RAD51C/D mutations. To maximise the number of informative families included in the analysis (after ascertainment adjustment), for these four studies, the analysis included also the families in which the proband was found not to carry a pathogenic variant in RAD51C or RAD51D and these probands were treated as non-carriers in the analyses. However, this assumes that the variant screening sensitivity, describing the probability of detecting a variant given it exists, is 100%, which may not be necessarily true given the variant screening was carried in research setting in those studies. In practice variant screening sensitivity could be lower and some of the non-carrier families may carry pathogenic variants in RAD51C or RAD51D. To assess the impact of a reduced variant screening sensitivity on the risk estimates we extended the models to allow for a reduced variant screening sensitivity parameter (16) which was assumed to range from 0.6 to 0.9.