Variant scorer quantiles not implemented in ag_research

valeh · March 2, 2026, 7:07pm

Hi,

I am using alphagenome_research to use the model for variant scoring. In the FAQ I see that some variant scorers return a quantiles score. But it looks like this functionality is not implemented in alphagenome_research (only in the alphagenome API):

github.com/google-deepmind/alphagenome

src/alphagenome/models/dna_client.py

2a5318c60


      
          
          var = track_data_utils.metadata_from_proto(
              dna_model_pb2.TracksMetadata(metadata=track_metadata)
          )
          var.index = var.index.map(str)
          
          uns = {'interval': interval}
          if variant is not None:
            uns['variant'] = genome.Variant.from_proto(variant)
          layers = None
          if values.shape[0] > 1:
            layers = {'quantiles': values[1]}
          
          return anndata.AnnData(
              X=values[0],
              obs=obs,
              var=var,
              uns=uns,
              layers=layers,
          )

Is there a reason this is not implement in alphagenome_research?

Thanks.

Tumi_Makgatho · March 4, 2026, 1:55pm

Hi There!

Thanks for reaching out.

Regarding your question about the quantiles score functionality, it is currently only available in the alphagenome API and has not yet been implemented in alphagenome_research.

However, this is on our roadmap, and we are planning to add this feature to alphagenome_research.

Kind regards,
Tumi

liandong_zhang · April 8, 2026, 1:28am

Hi !

I don’t know how to understand quantiles score,the quantiles score is tissue-weight or track-specific ,or the backgroud variants are random or fixed,how to use quantiles score to verify eQTLs ?

Thanks in advaced!

Tumi_Makgatho · April 8, 2026, 9:25am

Hi There!

Thanks for reaching out.

The quantile score is track-specific; it is computed independently for each unique combination of variant scorer and output track. The background variants used to build this null distribution are fixed, consisting of 348,126 common human SNPs (MAF > 0.01) from chromosome 22 in gnomAD v3.

To verify eQTLs, you can filter for variants with extreme quantile scores (e.g., greater than the 99th percentile, approaching +1.0 or -1.0). AlphaGenome predictions exceeding these high-confidence thresholds highly correlate with observed eQTL effect sizes and accurately predict the direction of the gene expression change.

Please see Avsec et al. 2026 , Extended data Fig.5 for a more detailed breakdown of eQTL performance at different quantile score thresholds.

Kind regards,
Tumi

Topic		Replies	Views
Eqtl analysis using alphagenome Help & Support	3	1243	October 15, 2025
How should I summarize thousands of per-variant quantile scores into one prioritization score? Help & Support	1	927	November 6, 2025
Versioning for AlphaGenome API and model Feedback & Feature Requests	2	914	September 26, 2025
Question about the difference between RNA-SEQ Quantile Score and predict_variant visualization Help & Support testing	1	914	September 17, 2025
101 video feedback Feedback & Feature Requests	3	225	January 30, 2026

Variant scorer quantiles not implemented in ag_research

Related topics