1 h
Seminar, Hosted by Google Developer Groups — GDG AI for Science Australia
AlphaGenome: a deep learning model that advances our understanding of the regulatory genetic code
Google DeepMind
Recording
Abstract
More than 98% of the genetic differences observed between humans do not directly code for proteins, and much of this variation underlies human disease, so deciphering the molecular function of this "dark DNA" is a major frontier in human genetics. Deep-learning models are emerging as powerful tools for reading the regulatory genetic code. AlphaGenome takes one million letters of DNA sequence as input and predicts thousands of experimental measurements at up to single-base-pair resolution across diverse regulatory mechanisms: gene expression, transcription initiation, chromatin accessibility, histone modifications, transcription-factor binding, chromatin contact maps, splice-site usage, and splice-junction coordinates and strength. The model generalises to predicting the effects of genetic variants, matching or exceeding the strongest available alternatives on the great majority of evaluations, and by predicting variant effects across several molecular phenomena at once it recapitulates the mechanisms of clinically relevant variants, such as those near the TAL1 oncogene; an API provides predictions of experimental assays and variant effects from DNA sequence. Clare Bycroft gives an overview of the model and its performance across tasks, highlights key use cases and outlines the practical steps for applying it in research.