AlphaGenome

AlphaFold solves protein structure prediction, while AlphaGenome Atlas further answers what functional consequences genetic changes will lead to.

DeepMind is developing an agent framework that integrates different specialized models such as AlphaFold and AlphaGenome.

AlphaGenome Atlas pre-computed AlphaGenome model predictions for all 9 billion possible mutations, resulting in a dataset of 1 PB (1,000 TB) in scale.

AlphaGenome outputs approximately 10,000 multidimensional predictions covering biological features such as gene expression, chromatin accessibility, and transcription factor binding across different cell types, compressed into a single AVI scalar value.

The AVI (AlphaGenome Variant Impact) score is a single numerical value measuring the degree of harmfulness of a genetic variant, with a higher score indicating a greater likelihood that the variant will have an impact on the phenotype.

AlphaGenome represents a new generation of variant effect prediction methods that directly learn 'grammatical rules' from DNA sequences, analogous to large language models in natural language processing.

The core model behind AlphaGenome Atlas is AlphaGenome, released by DeepMind in 2024.

AlphaGenome predicted that the variant in the epilepsy case creates an erroneous splice site leading to abnormal protein elongation, and laboratory experiments supported this prediction.

AlphaFold addresses structural questions, while AlphaGenome attempts to answer regulatory questions at the genomic level—namely, how DNA sequences determine when, where, and at what intensity a gene is expressed.

The core capability of AlphaGenome Atlas comes from DeepMind's AlphaGenome model, which can predict the effects of variants on molecular phenotypes such as gene expression regulation and splicing signals without requiring wet-lab experiments to be performed one by one.

Earlier functional prediction tools such as CADD, PolyPhen-2, and SIFT can also score large numbers of variants, but AlphaGenome has the ability to predict multi-dimensional molecular phenotypes from sequence.

The AlphaGenome model drew inspiration from earlier work such as Enformer.

For cell types or species underrepresented in the training data, the reliability of predictions from such models decreases significantly.

Such models fundamentally predict whether a variant affects a molecular phenotype rather than directly predicting clinical phenotypes, and additional biological inference is still required to go from molecular predictions to disease associations.

AlphaGenome attempts to establish a mapping from DNA sequences to molecular-level functional readouts, such as gene expression levels, chromatin accessibility, and splicing patterns.

The predictions made by AI models are computational hypotheses by nature and still require experimental validation.

AlphaGenome relies on publicly available datasets accumulated by large-scale genomic functional annotation projects such as ENCODE.

The prediction accuracy of such tools on specific tissue types or rare variants still needs to be confirmed through large-scale independent validation.

AlphaGenome adopts a Transformer architecture similar to that of large language models, treating DNA sequences as a kind of language and capturing long-range interactions between bases in a sequence through self-attention mechanisms.

AVI (AlphaGenome Variant Impact) is a single numerical value measuring the degree of harmfulness of a genetic variant, where a higher score indicates a greater likelihood that the variant affects the phenotype.