Expanding the Landscape of Breast Cancer-Associated Rare Variants and Combining with Polygenic Risk Score

Investigating the association at variant-level of rare variants(RVs, MAF $< 0.01$) with breast cancer (BC) risk in population studies poses challenges due to low statistical power and multiple testing burdens. To increase power, current approaches often aggregate RVs into genetic units, such as genes or gene sets. However, these strategies typically focus on high-penetrance genes and pathways already known to be involved in cancer, limiting their capacity to identify novel contributors. Likewise, most existing methods fail to provide insights into the individual contributions of specific RVs, reducing the interpretability and clinical utility of the findings. The work described in the present thesis aimed at addressing these two gaps by firstly providing a more systematic and scalable method to comprehensively analyze RV impact to BC and secondly introducing an alternative approach (Bayesian Hierarchical Generalized Linear Model, BhGLM) to investigate the single variant association to BC risk. We trained and test the methods using the UK Biobank (UKBB) cohort (15868 BC cases, 165067 controls). The Burden test assessed the cumulative association of RVs in aggregated genetic units, while BhGLM accounted for complex relationships within the data to identify BC-associated variants. First, we applied the Burden Test to different lists of genes and different RV masks combining Loss Of Function (LoF) and missense to determine the impact of the multiple testing burden on the detection of BC-related RVs and the contribution of different type of variants to BC susceptibility. Second, we exploited the BhGLM to assess single RV association. We evaluated the quality of the retrieved RVs using the American College of Medical Genetics (ACMG) and ClinVar annotation, and by comparing their impact with the effect of unselected RVs using odds ratio (OR) across different PRS classes. Finally, we built two different RVScores by combining RVs significantly associated to BC by the Burden Tests and the BhGLM model. These scores allowed us to explore the cumulative impact of RVs in BC risk stratification in combination with PRS. The findings were assessed using OR on a distinct test set. Through the application of the classical Burden test approach we underscored the importance of gene list selection in detecting associations at gene level. We showed how smaller curated lists were more effective at identifying weaker, yet meaningful associations, while larger lists provided a broader view of potential contributors but were less sensitive to subtle signals. Strong associations were consistently observed in well-established BC susceptibility genes like BRCA1, BRCA2, ATM, CHEK2, and PALB2. Notably, we identified two new potential risk genes, ASPRV1 and ADGRA3, that showed a strong relation with BC risk. Weaker associations emerged for further 7 genes (BARD1, MAP3K1, PLCG1, LZTR1, POLD2, DDX1 and NDFUS4), highlighting the need for a balanced approach to gene selection. The RVScore, computed on Burden Test results, showed stable performance across different variant masks, underscoring its robustness as a tool for patient stratification. When calculated using only LoF RVs, the RVScore enabled more precise stratification of BC risk across PRS classes for 2.5\% of the population compared to the presence of RVs in high- and moderate-risk genes. Furthermore, when combining LoF and missense RVs, high levels of RVScore yielded higher OR across PRS classes than the sole presence of RVs in high- or moderate-risk genes. At the same time, the BhGLM approach demonstrated high specificity levels in simulation settings, translating in low false-positive rate (FPR, average $\leq$ 0.001). Conversely, sensitivity assumed considerably low values, which highlights an overall conservative trend in the model's classification strategy. When evaluated in a controlled setting of a short list of genes, BhGLM mostly selected pathogenic RVs with higher OR than non the selected ones on the same genes. When extended to the ClinicalExome (5369 genes), we identified a total of 550 LoF RVs, of which 40.2\% annotated as Uncertain Significance, and the 24.74\% as Pathogenic. Notably around 80\% of the annotated Pathogenic RVs are associated with a positive effect size. The comparison the ORs for the selected RVs with the one of unselected RVs across PRS classes reveals their significant contribution to amplifying BC risk. The comparison between the two approaches revealed notable differences in the number and characteristics of selected variants: BhGLM identified a larger set of RVs on a broader set of genes, likely due to its capacity to model complex relationships. Nonetheless, high levels of both the RVScores were associated to an increment risk of BC with respect to PRS alone. Furthermore, the here proposed approach based on the a Bayesian hierarchical model, not only introduces a novel methodological framework in the context of BC studies, but enables also the quantification of the collective impact of RVs on BC risk while preserving the capacity to interpret the contribution of individual variants. However, the reduced discriminatory power at lower BhGLM RVScore levels in the test set suggests that the score’s ability to provide finer stratification of cancer risk is influenced by sample size. Similarly, The Burden-derived RVScore showed variability in his distribution between the training and test sets, likely reflecting the smaller size of the test set.