In diesem Sommersemester wird die im WiSe 2022/23 erfolgreich gestartete Vortragsreihe Medical Information Sciences fortgesetzt. Renommierte Wissenschaftlerinnen und Wissenschaftler unterschiedlicher Fachdisziplinen und Forschungsstandorte geben jeden Dienstag ab 17:30 Uhr Einblicke in aktuelle Fragestellungen und Anwendungsgebiete des breiten Forschungsfeldes Medical Information Sciences.

Rare genetic variants can strongly predispose to disease, yet accounting for rare variants is statistically challenging, and principled strategies for integrating possibly diverse types of variant annotations in a data-driven manner are lacking. Here, we present DeepRVAT (Deep Rare Variant Association Testing), a deep set model that learns gene impairment scores from rare variants, annotations, and phenotypes. DeepRVAT infers the relevance of different annotations and their combination directly from data, eliminating ad hoc modeling choices that characterize existing methods. DeepRVAT estimates a single, trait-agnostic gene impairment score for each gene in each sample, enabling both risk prediction and gene discovery in a unified framework and seamless integration into established association testing frameworks. We apply DeepRVAT on 34 quantitative and 63 binary traits across 469,382 whole-exome-sequenced individuals from the UK Biobank. We integrate state-of-the-art annotations, including AlphaMissense, PrimateAI, AbSplice, DeepRipe, and DeepSEA, and find a substantial increase in gene discoveries and improved replication rates on held-out individuals over previous methods.  We demonstrate the applicability of pre-trained DeepRVAT models to new traits, facilitating the study of disease cohorts with limited training data. Furthermore, we significantly improve the detection of individuals at high genetic risk by combining common variant polygenic risk scores with DeepRVAT.