Advancing regulatory variant effect prediction with AlphaGenome

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Deep Learning Applications

Background:

• Predicting functional genomic measurements from DNA sequences is crucial for understanding the genetic regulatory code.
• Current deep learning models face limitations due to a trade-off between input sequence length and prediction resolution, restricting their scope and performance.
• Diverse functional genomic modalities, including gene expression, chromatin accessibility, and transcription factor binding, require integrated analysis.

Purpose of the Study:

• To develop a unified deep learning model, AlphaGenome, capable of predicting numerous functional genomic tracks from long DNA sequences at high resolution.
• To overcome the limitations of existing methods by integrating diverse genomic modalities within a single framework.
• To provide accurate variant effect predictions across multiple functional genomic layers.

Main Methods:

• Developed AlphaGenome, a deep learning model that accepts 1 megabase (Mb) of DNA sequence as input.
• Trained AlphaGenome on human and mouse genomes to predict thousands of functional genomic tracks, including gene expression, chromatin accessibility, transcription factor binding, and more, at single-base-pair resolution.
• Evaluated AlphaGenome's performance on variant effect prediction against existing state-of-the-art models across 26 different evaluations.

Main Results:

• AlphaGenome successfully predicts thousands of functional genomic tracks across diverse modalities with single-base-pair resolution.
• The model matches or surpasses the performance of the strongest external models in 25 out of 26 variant effect prediction evaluations.
• AlphaGenome accurately recapitulates the mechanisms of clinically relevant variants, demonstrated by its analysis of variants near the TAL1 oncogene.

Conclusions:

• AlphaGenome represents a significant advancement in deep learning for genomics, offering a unified approach to predict functional genomic information from long DNA sequences.
• The model's high resolution and broad modality scope enable more accurate and comprehensive variant effect predictions.
• Tools are provided to facilitate the use of AlphaGenome for genome track and variant effect predictions, promoting wider research applications.