276: AlphaGenome: 1-Mb multimodal deep model predicts regulatory variant effects including splicing and TAL1 mechanisms

Avsec et al., Nature, doi:10.1038/s41586-025-10014-0 - AlphaGenome, a 1 Mb DNA deep‑learning model, predicts base‑pair‑resolution genome tracks (RNA‑seq, splicing, chromatin) and scores variant effects, achieving state‑of‑the‑art performance across modalities. Key terms: AlphaGenome, splicing, eQTL, chromatin-accessibility, 1Mb-sequence. Study Highlights:AlphaGenome is a unified sequence‑to‑function deep learning model trained on human and mouse genomes that consumes 1 Mb of DNA and predicts 5,930 human genome tracks across 11 modalities using a U‑Net‑inspired encoder, transformer tower and decoder. The model was pretrained with fold splits and distilled into a single student model for efficient variant scoring, enabling base‑pair resolution outputs and splice junction prediction alongside splice site usage and RNA‑seq coverage. Quantitatively, AlphaGenome outperformed or matched external models on 22 of 24 genome track tasks and on 25 of 26 variant effect benchmarks, improving eQTL sign prediction and QTL effect correlations. The multimodal outputs enable mechanistic interpretation of variants, for example recapitulating oncogenic TAL1 enhancer mutations and identifying splice‑disrupting variants. Conclusion:AlphaGenome provides a unified 1‑Mb multimodal, base‑resolution sequence model that substantially improves genome track and regulatory variant effect prediction and enables mechanistic, cross‑modality interpretation. QC:This episode was checked against the original article PDF and publication metadata for the episode release published on 2026-01-30. QC Scope:- article metadata and core scientific claims from the narration- excludes analogies, intro/outro, and music- transcript coverage: Audited the transcript portions presenting AlphaGenome architecture, multimodal outputs, benchmarking, splicing, TAL1 mechanism, 3D genome, in silico mutagenesis, and limitations.- transcript topics: AlphaGenome architecture (1 Mb input, U‑Net backbone, transformers, sequence parallelism); Multimodal genome tracks (5,930 human tracks, 1,128 mouse tracks across 11 modalities); Variant effect benchmarking (26 tasks; 25/26 ahead); Splicing variant predictions (splice sites, splice junctions, usage); TAL1 oncogene mechanism (neo-enhancer, MYB motif); 3D genome contact maps predictions QC Summary:- factual score: 10/10- metadata score: 10/10- supported core claims: 8- claims flagged for review: 0- metadata checks passed: 4- metadata issues found: 0 Metadata Audited:- article_doi- article_title- article_journal- license Factual Items Audited:- AlphaGenome takes input of 1 Mb DNA and predicts thousands of genome tracks across 11 modalities at base-pair resolution- 5,930 human genome tracks and 1,128 mouse genome tracks predicted- AlphaGenome matched or outperformed strongest external models on 25 of 26 variant-effect benchmarks- Splicing variant predictions include splice sites, splice junctions, and splice site usage- TAL1 oncogene mechanism recapitulated via a neo-enhancer and MYB motif- 3D genome folding predicted via contact maps with improved performance over specialized tools QC result: Pass.