Improved breast cancer risk prediction using chromosomal-scale length variation

Early diagnosis of breast cancer leads to higher long-term survival rates. The development of a germline genetic test, or polygenic risk score, to identify women at high risk of breast cancer holds the potential to reduce cancer deaths. However, current tests based on SNPs do not perform much better than predictions based on family history and perform significantly worse in populations with non-European ancestry. We have developed an alternative method to characterize a genome, called chromosomal-scale length variation, which can be applied to polygenic risk scores.

The objective of this paper is to characterize a breast cancer genetic risk score based on chromosomal-scale length variation using the NIH All of Us dataset in different self-identified racial groups when trained on different populations.

We used the NIH All of Us dataset to compile a dataset with 4,533 women who have been diagnosed with breast cancer (including 440 who self-identified as Black) and 44,518 women who have not. We acquired, through All of Us, genetic information for each of these women. We computed a set of 88 values for each woman in the dataset, representing the chromosomal-scale length variation parameters. These numbers are average log R ratios for four different segments from each of the 22 autosomes. We used machine learning algorithms to find a model that best differentiates the women with breast cancer from the women without breast cancer based on the set of 88 numbers that characterize each woman's germline genome.

The best model had an AUC of 0.70 (95% CI, 0.67-0.73) in the All of Us population. Women who scored in the top quintile by this model were nine times more likely to have breast cancer when compared to women who scored in the lowest quintile.

In conclusion, we found that this method of computing genetic risk scores for breast cancer is a substantial improvement over SNP-based polygenic risk scores. In addition, we compared models trained on populations of only White women and only Black women. We found that the models trained only on White women performed better than models trained only on Black women when tested on only White women. We did not see a significant difference between the two models when tested on only Black women.