Advancing regulatory variant effect prediction with AlphaGenome (www.nature.com)

AlphaGenome has been unveiled as a revolutionary deep learning model designed to predict functional genomic measurements from 1 Mb of DNA sequences, significantly enhancing the field of regulatory variant effect prediction. By integrating the capabilities to predict thousands of genomic tracks across 11 modalities—including gene expression, splicing patterns, chromatin accessibility, and more—AlphaGenome achieves unprecedented single-base-pair resolution. This model outperforms existing state-of-the-art models in 25 out of 26 evaluations, showcasing its ability to accurately interpret the complex effects of non-coding variants, which comprise over 98% of genetic variation in humans. The significance of AlphaGenome lies in its capacity to unify long-sequence context modeling with detailed resolution without sacrificing performance across different biological modalities. Its innovative architecture employs both convolutional and transformer layers, enhancing the understanding of distant regulatory impacts while maintaining the precision required for fine-scale features like splice sites. In addition, it offers streamlined predictions through a two-stage training process, enabling high efficiency for large-scale analyses. Given its robust predictive power and the extensive range of outputs, AlphaGenome is poised to advance genomic research and improve our understanding of the regulatory code underlying genetic variation in health and disease.