Predicting non-coding variant effects with AlphaGenome
Published in Cell Research (Research Highlight), 2026
Most trait- and disease-associated genetic variation lies in non-coding DNA and acts by altering gene regulation. Computational sequence-to-function models offer a scalable way to prioritise variants and reason about mechanism, but they face persistent tensions between integrating many molecular readouts versus specialising on one assay, and between modelling long-range cis-regulatory context versus resolving predictions at single-base resolution.
This Cell Research research highlight summarises recent work on AlphaGenome, which addresses these tensions through architectural choices such as sequence sharding, transformer layers for distal interactions, expanded outputs (including contact maps and splicing-related tasks), and ensemble distillation for accuracy without the full cost of ensembling at inference. We outline how the model was benchmarked on single-nucleotide variant effects across diverse assays, what ablations suggest about resolution and training context, and where important limitations remain — including complex alleles, personal-genome-style combinatorial effects, rigorous evaluation against distal perturbation data, and dependence on cellular context implicit in the training data.
Read the highlight at Cell Research
Recommended citation: Murphy, A.E., Nagai, M. & Koo, P.K. Cell Res. (2026). https://doi.org/10.1038/s41422-026-01249-1
Recommended citation: Murphy, A.E.*, Nagai, M.* & Koo, P.K. Cell Res. (2026). https://doi.org/10.1038/s41422-026-01249-1
Download Paper