Hello everyone,
I wanted to share a preprint I have just published that evaluates the calibration of AlphaGenome regulatory effect scores against experimental saturation mutagenesis data.
The study analyzes 20,549 SNVs across 15 regulatory elements and compares AlphaGenome predictions with experimentally measured regulatory effects.
The main observation is that the raw effect-score scale appears to be distance-dependent:
- Promoter variants tend to have over-dispersed predicted effects.
- Distal enhancer variants tend to have compressed predicted effects.
- As a consequence, raw scores from different loci may not be directly comparable using a single genome-wide threshold.
Importantly, this is not intended as a criticism of AlphaGenome, which I believe represents a major advance for regulatory genomics. Rather, the work suggests that local calibration may be necessary before comparing scores across different genomic regions.
The paper also includes:
- comparison against saturation mutagenesis experiments
- per-locus calibration analysis
- coordinate/allele verification pipeline
- released analysis code and validation framework
I would greatly appreciate any feedback, comments, or suggestions from the community.
Preprint:
https://www.preprints.org/manuscript/202607.1417
Thank you to the AlphaGenome team for making such an exciting model publicly available.