Researchers built GPN-Star, a DNA model that reads alignments of many species at once. It uses the evolutionary family tree of those species to weigh what each contributes. The team trained three human versions, covering vertebrate, mammal and primate time spans. GPN-Star scored best on a range of tasks that predict the effect of genetic variants. It worked in both protein-coding and non-coding parts of the human genome. It also ranked disease-causing variants better than earlier methods. Which time span helped most depended on the task. The team adjusted scores for background mutation rates to isolate selection. Training used 8 graphics chips over several days, far less than comparable models. They also trained versions for mouse, chicken, fruit fly, worm and a plant.
What changed
Earlier DNA language models needed far more compute and still trailed classical evolutionary models on some tasks.
What it unlocks
Scoring the likely harm of DNA variants across the whole genome using a modest compute budget.
- 200 million parameters
- 8 NVIDIA A100 GPUs, several days
- 5 model organisms
- 3 alignment timescales
What you need to act on it
- whole-genome alignment data for the species of interest
Sources