Novel discoveries and enhanced genomic prediction from modelling genetic risk of cancer age-at-onset

Genome-wide association studies seek to attribute disease risk to DNA regions and facilitate subject-specific prediction and patient stratification. For later-life disease, inference from case-control studies is hampered by the uncertainty that control group subjects might later be diagnosed. Time-to-event analysis treats controls as right-censored, making no additional assumptions about future disease occurrence and represent a more sound conceptual alternative for more accurate inference. Here, using data on 11 common cancers from the UK and Estonian Biobank studies, we provide empirical evidence that discovery and genomic prediction are greatly improved by analysing age-at-diagnosis, compared to a case-control model of association. We replicate previous findings from large-scale case-control studies and find an additional 59 previously unreported independent genomic regions, out of which 16 replicated in independent data (an increase of 18% and 6% over current findings). Our novel discoveries provide new insights into underlying cancer pathways, and our model yields a better understanding of the polygenicity and genetic architecture of the 11 tumours. We find that heritable germline genetic variation plays a vital role in cancer occurrence, with risk attributable to many thousands of underlying genomic regions. Finally, we show that Bayesian modelling strategies utilising time-to-event data increase prediction accuracy by an average of 20% compared to a recent summary statistic approach (LDpred-funct). As sample sizes increase, incorporating time-to-event data should be commonplace, improving case-control studies by using richer information about the disease process.