Advancing regulatory variant effect prediction with AlphaGenome 2026 Avsec et al

[jnmaciuch]

maybe we can give it all the potentially causal SNPs in the locus separately to test each one.

If you can use a raw sequence of DNA as input it should be possible to just manually change all the base pairs to match the SNPs from DecodeME in a certain region and then run that string as your input. Might be too labor intensive, though

 

[forestglip]

The batch variant scoring tool looks like it'll be useful to test many variants at once: https://www.alphagenomedocs.com/colabs/batch_variant_scoring.html

I think the goal is basically looking for variants that produce a large (or small if negatives are possible) quantile or raw score. For RNAseq, I think for each variant, the model creates one score for each gene in the interval. The quantile score would go from -1 to 1, which is a measure of how large of an effect the variant has on the gene. (0 is no effect.)

Edit: So I think what I'll try to get to doing is:

1. Filter DecodeME SNPs to log10p greater than something like 7 (produces about 300 variants).
2. Do batch scoring on all these variants with the RNAseq predictor on brain tissue.
3. Find the largest magnitude quantile scores to identify which variant-gene combinations might be important.