Descriptions:
Google DeepMind researchers — including Ziga Avsec (genomics team lead), Natasha Latysheva, Tom Ward, and Jun Cheng — sit down for a roundtable on AlphaGenome, their newly published unified DNA sequence-to-function prediction model featured in Nature. The conversation covers the scientific motivation, architectural decisions, and real-world implications of the work.
AlphaGenome builds on prior DeepMind efforts like AlphaMissense, which addressed only 2% of the genome (the protein-coding region), and extends prediction to the non-coding 98% — a region implicated in a large fraction of rare genetic diseases still lacking diagnosis. The model integrates multiple output modalities including 1D genomic expression tracks, splicing predictions, and 2D contact maps that capture how DNA folds in three-dimensional space inside the nucleus.
The team discusses a longstanding architectural tradeoff between sequence length and resolution, explaining how AlphaGenome manages to handle both simultaneously. Evaluation strategy is covered in depth: the model is tested both on unseen DNA sequences and on variant effect prediction — where two DNA strings differing by a single mutation are compared — the latter being the most clinically relevant benchmark. The researchers note that adding contact map modeling did not degrade performance on existing modalities, a result they describe as non-obvious and significant.
📺 Source: Google DeepMind · Published January 28, 2026
🏷️ Format: Interview
