Although

We used data from 1548 patients with OvC and their relatives from a population-based study, with known

The most parsimonious model included the effects of

The resulting model can be used to obtain the risk of developing OvC on the basis of

Ovarian cancer (OvC) is the third most common gynaecological cancer (

Risk models that incorporate both

We have used data from a large, population-based series of cases diagnosed with OvC, the Studies of Epidemiology and Risk factors in Cancer Heredity (SEARCH), and segregation analysis methods to develop genetic models for OvC that incorporate the effects of

We used data on 1548 OvC cases (probands) recruited between 1999 and 2010, along with information on their first-degree and second-degree relatives ascertained through an epidemiological questionnaire. The probands were drawn from SEARCH, a large population-based study with cases ascertained through the Eastern Cancer Registration and Information Centre.

SEARCH OvC probands were screened for

Complex segregation analysis was used to fit genetic models to the occurrence of OvC in families, incorporating the explicit effects of

To incorporate the effects of _{i} is the logarithm of the relative risk associated with a third hypothetical major gene and P_{i} is the polygenic component. P_{i} is assumed to have a normal distribution with variance σ^{2} and mean zero and is approximated by the hypergeometric distribution to make it amenable to ‘peeling’.

Since

In our analyses, we considered models with just the

All the families used in the analysis consisted of women ascertained on the basis of OvC. Thus, to adjust for ascertainment bias,_{i} is the phenotype of the proband. A sensitivity parameter was introduced, giving the probability of detecting a mutation if one existed, to take account of the fact that mutation screening methods used cannot detect large rearrangements in

Maximum-likelihood estimates of the gene frequencies, polygenic standard deviation and the log relative risk for the hypothetical major gene were calculated using pedigree analysis software MENDEL.

We used each of the models fitted to predict ^{2} goodness-of-fit tests that compared the observed and expected number of mutations and used these as an indicator of the model fit to the data. The expected number of mutation carriers was computed as the sum of the predicted

We used the most parsimonious model to estimate risk of developing OvC for a 50-year-old woman to demonstrate the possible clinical implications for different scenarios of BRCA1 and BRCA2 carrier status and extent of family history. The results were compared with the corresponding predictions from the current BOADICEA model.

We extended the most parsimonious model to also incorporate the explicit effects of the known common OvC susceptibility alleles following the methodology already published in the context of prostate cancer._{i} for each individual was divided into two parts for this purpose: a known-variant polygenic component P_{k,i} reflecting the polygenic risk score (PRS) due to 17 SNPs known to be associated with OvC_{U,i}. The two components were assumed to be independent and normally distributed with mean 0 and variance

The OvC risk associated with any individual common genetic variant is very small compared with rare variants like

Data from 1548 OvC cases recruited into the SEARCH study were used for our analyses. Female relatives of probands included 1340 mothers, 1404 sisters and 1144 daughters, of whom 80 were also diagnosed with OvC and 191 with breast cancer. The numbers of probands and their first-degree relatives, the number of OvCs diagnosed in each group and other sample characteristics are summarised in online supplementary table S1. All probands were screened for

The results for the seven models that incorporate the explicit effects of ^{−5}). The worst-fitting model for the residual familial aggregation of OvC, other than

Parameter estimates, goodness-of-fit measures and likelihood ratio tests (LRTs) of the seven cohort-specific models for breast and ovarian cancer

Model | Major gene frequency (95% CI) | Major gene log relative risk (95% CI) | Polygenic SD (95% CI) | Log-likelihood | AIC | LRT p value | ||
---|---|---|---|---|---|---|---|---|

Base | 0.00081 (0.00061 to 0.0011) | 0.0026 (0.0020 to 0.0033) | – | – | – | −2892.237 | 5788.474 | 5.11E-06 |

Major dominant | 0.00079 (0.00060 to 0.0011) | 0.0026 (0.0020 to 0.0032) | 0.00025 (0.000041 to 0.0015) | 4.8 (3.3 to 6.2) | – | −2880.343 | 5768.686 | 0.047 |

Major recessive | 0.00080 (0.00060 to 0.0011) | 0.0026 (0.0020 to 0.0032) | 0.085 (0.017 to 0.33) | 4.0 (2.0 to 6.0) | – | −2882.122 | 5772.244 | 0.0079 |

Major general | 0.00079 (0.00060 to 0.0011) | 0.0026 (0.0020 to 0.0032) | 0.00025 (0.00020 to 0.0033) | 4.8 (3.3 to 6.3) | – | −2880.335 | 5770.67 | 0.013 |

7.4 (−14.1 to 28.8) | ||||||||

Polygenic | 0.00079 (0.00060 to 0.0011) | 0.0026 (0.0020 to 0.0033) | – | – | 1.43 (1.10 to 1.86) | −2879.186 | 5764.372 | 0.28 |

Mixed dominant | 0.00079 (0.00059 to 0.0011) | 0.0026 (0.0020 to 0.0032) | 0.00023 (0.000023 to 0.0022) | 4.7 (2.8 to 6.6) | 1.09 (0.64 to 1.86) | −2877.289 | 5764.576 | 0.91 |

Mixed recessive | 0.00079 (0.00060 to 0.0011) | 0.0026 (0.0020 to 0.0032) | 0.076 (0.020 to 0.25) | 3.7 (1.5 to 5.9) | 1.19 (0.74 to 1.91) | −2878.374 | 5768.806 | 0.14 |

Mixed general | 0.00079 (0.00059 to 0.0011) | 0.0026 (0.0020 to 0.0032) | 0.00023 (0.000023 to 0.0023) | 4.7 (2.8 to 6.6) | 1.09 (0.64 to 1.86) | −2877.283 | 5766.566 | |

9.4 (−20.5 to 39.3) |

AIC, Akaike's information criterion; LRT p value, probability of the difference between log-likelihoods comparing each model against the mixed general model.

The expected numbers of

Number of mutation carriers predicted by each model and comparison with observed numbers

Model for the residual familial aggregation | Observed | Expected | Observed | Expected | χ^{2} value* |
---|---|---|---|---|---|

Baseline | 44 | 56.95 | 62 | 63.59 | 2.98 |

Polygenic | 44 | 49.32 | 62 | 61.98 | 0.57 |

Dominant major | 44 | 55.62 | 62 | 63.08 | 2.45 |

Recessive major | 44 | 55.97 | 62 | 63.11 | 2.58 |

General major | 44 | 55.62 | 62 | 63.08 | 2.45 |

Dominant mixed | 44 | 48.07 | 62 | 61.01 | 0.36 |

Recessive mixed | 44 | 49.08 | 62 | 61.10 | 0.54 |

General mixed | 44 | 48.05 | 62 | 61.02 | 0.36 |

BOADICEA | 44 | 45.76 | 62 | 23.03 | 66.01 |

*χ^{2} value, value of χ^{2} goodness-of-fit test.

BOADICEA, Breast and Ovarian Analysis of Disease Incidence and Carrrier Estimation Algorithm.

Similarly, when computing the expected number of families with a mother, a sister or mother and sister diagnosed with OvC, the predicted numbers were closer to that observed for the polygenic and mixed models of inheritance (see online supplementary table S3).

We estimated the probabilities of developing OvC for a 50-year-old woman born in 1940, with the following family histories: (i) no information on relatives; (ii) having a mother and sister cancer free at ages 65 and 50; (iii) mother and sister diagnosed with OvC at ages 65 and 50; and (iv) and (v) with both combinations of one diagnosed and one cancer free at these same ages. We compared these estimates with the risk estimates from the current version of BOADICEA.

Predicted risks of ovarian cancer over time to a woman born in the 1940 birth cohort without a

The loci, minor allele frequencies and ORs for the 17 SNPs considered are displayed in online supplementary table S2. Under the assumptions that the effects of the SNPs on OvC are all mutually independent and the same for

The lifetime risks of OvC to a 20-year-old non

Lifetime risks of ovarian cancer to a 20-year-old born in the 1940 birth cohort without a

Examples of age-specific risks for a 50-year-old woman at the 5th and 95th percentiles of the PRS and by different family history assumptions are shown in online supplementary figures S3–S5

For a polygenic log-risk with the SD of 1.434, estimated under the best-fitting segregation analysis model, and assuming a baseline population OvC risk of 0.02 by age 80, the half of the population at higher risk accounts for 92% of all OvCs.

Proportion of population above a specified absolute risk of ovarian cancer and proportion of cases occurring in that fraction of the population. Half the population have an absolute risk of ovarian cancer greater than 0.72% by age 80 and 92% of all cases occur in this half of the population. Half of all cancers occur in the 7.7% of the population with risk higher than 5.6%.

In

Proportion of cases accounted for by a given proportion of the population above a specified risk of ovarian cancer according to the total polygenic risk and the observed 17 SNP distribution. Under the total polygenic risk distribution, 50% of cancers occur in the 7.7% of the population at highest risk and 92.4% of cancers occur in the half of the population at greater-than-average risk, whereas under the 17 SNP only 62% of cancers occur in the 50% at higher risk and 50% of cases are spread among almost 40% of the population at highest risk.

We used complex segregation analysis to develop a risk-prediction model for familial OvC that incorporates the effects of

The most parsimonious model included the effects of

Previous OvC segregation analyses

Under the best-fitting model, the

In the long term, we expect that these differences will be resolved by fitting a single algorithm to all available data that models comprehensively both the genetic susceptibility to breast cancer and OvC. However, at this stage this is not feasible based on current technologies due to computational complexities (in particular, the number of underlying genotypes in the models). The current approach aims to develop separate algorithms for the susceptibility to breast cancer and OvC that individually incorporate the explicit effects of all observed and unobserved genetic variants such that we obtain accurate risks of each cancer. Validation studies in independent datasets will determine the most appropriate model for use in each context. As technologies evolve over time, in the long term we expect to synthesise the models into a single algorithm.

In our analyses, we took account of OvCs occurring after a breast cancer diagnosis, assuming the OvC incidence remains the same before and after the breast cancer diagnosis. Repeating the analysis but censoring at the first cancer yielded similar results (eg, under the polygenic model

In our analysis, we aimed to include only epithelial OvCs. However, subsequent to the model fitting process, additional pathology information became available, which revealed 41 of the probands’ tumours to be non-epithelial OvCs. This consisted of one

Our models assumed that the mutation testing sensitivities were 0.9 for both

One possible source of bias in our analysis is the possibility of errors in the reporting of family cancer history. However, previous studies have found reported OvC history in first-degree relatives to be reasonably accurate (83.3% probability of agreement between reported OvC status in first-degree relatives and established status).

Under our models, the probabilities of developing OvC increase with the number of OvCs in relatives, while under BOADICEA

In all models incorporating a polygenic component or known SNPs, the effects were assumed to be the same for carriers of a

Although we have incorporated the explicit effects of the common low-risk alleles, future efforts should focus on incorporating the explicit effects of other intermediate risk OvC susceptibility variants such as

We also used our models to investigate the possible implications for OvC risk stratification in the population. Using the parameters from the polygenic model, we estimate that 50% of OvCs occur within 7.7% of the population at highest risk. Meanwhile, half of the population at lower risk is forecast to contain only 1 in 13 cancer cases. Targeting the 10% at highest polygenic risk for preventative measures or excluding the low-risk half could therefore lead to a much more efficient distribution of resources. However, to achieve this will require that we identify all the genetic factors that contribute to polygenic inheritance. The almost flat curve in

Our model can be used in the genetic counselling process of women with family history of OvC as well as for counselling women both with and without

Although the mutation carrier probability algorithms produced very accurate estimates of the number of carriers in the SEARCH data, an external validation is needed to establish the performance of the model in independent datasets and to assess the model performance in predicting OvC risk in prospective studies. Future plans to extend the models include the addition of lifestyle and reproductive factors such as parity, breast feeding and oral contraceptive use,

We thank all the study participants who contributed to this study and all the researchers, clinicians and technical and administrative staff who have made possible this work. In particular, we thank Craig Luccarini, the SEARCH team and the Eastern Cancer Registration and Information Centre. ACA is a Cancer Research UK Senior Cancer Research Fellow. IJ is a National Institute for Health Research Senior Investigator.

The license of this article has changed since publication to CC BY 4.0.

Conception and design: ACA and PPDP. Analysis: SJ, HS, AL, ED, PH and CB. Interpretation of data: SJ, ACA, PPDP, RM, DFE and IJ. Acquisition of data: PPDP and DFE. Drafting the article: SJ, PPDP and ACA. Critical revision: all authors. Final approval: all authors.

This work has been supported by grants from Cancer Research UK (C1005/A12677, C12292/A11174, C490/A10119, C490/A10124) including the PROMISE research programme, the Eve Appeal and the UK National Institute for Health Research Biomedical Research Centre at the University of Cambridge.

IJ is a director of Abcodia.

Obtained.

Cambridgeshire 4 research ethics committee.

Not commissioned; externally peer reviewed.