Updating variant score quantiles

Announcements
1

Hi everyone,

From the 18th June 2026, we will update our recommended variant scorer quantiles.

Previously we computed the score quantiles against a background set of common genetic variants from chromosome 22 (see the “Calibration Methodology” methods section for more details). We have now run the same calibration procedure across the entire genome, to generate a more accurate background distribution.

We recommend using these new quantiles for all future variant interpretation, but if you want to use the original quantiles, you can use the following code snippet to reprocess raw scores to the old quantiles, using our open-source calibration code:

from alphagenome.data import genome
from alphagenome_research.model.variant_scoring.calibration import calibration


calibration_scorer = calibration.load(
    'gs:///alphagenome/data/hg38/calibration_scores.pb'
)

model = dna_client.create('API_KEY')
scores = model.score_variant(interval, variant)

# The new quantiles are stored in score.layers['quantiles']
# Add old quantiles as a new layer.
for score in scores:
  score.layers['old_quantiles'] = calibration_scorer.quantile_scores(
      score.uns['variant_scorer'].name, score
  )

If you have any further questions or concerns, please let us know in the comments below.

Thanks!

Tom (on behalf of the entire AlphaGenome team).

3

Hi! I am now hitting a new blocking issue when the model tries to load calibration_scores.pb:E0000 00:00:1782102492.940529 867376 curl_http_request.cc:625] The transmission of request 0x55f675f54670 (URI: https://www.googleapis.com/storage/v1/b/alphagenome/o/data%2Fhg38%2Fcalibration_scores.pb?fields=size%2Cgeneration%2Cupdated) has been stuck at 0 of 0 bytes for 61 seconds and will be aborted.
How may I fix it?

4

Hi @ZhongtingHuang05,

This looks like a networking issue, are you able to download https://storage.googleapis.com/alphagenome/data/hg38/calibration_scores.pb ?

if you don’t need calibrated scores when running the OSS model, you can disable by overriding the path on construction like so:

import dataclasses
from alphagenome_research.model import dna_model

organism_settings = dna_model.default_organism_settings()

# Set calibration path to None.
organism_settings[dna_model.Organism.HOMO_SAPIENS] = dataclasses.replace(
    organism_settings[dna_model.Organism.HOMO_SAPIENS], calibration_path=None
)

model = dna_model.create_from_huggingface(
    'all_folds', organism_settings=organism_settings
)
5

Lovely! The model is now currently initialized. Thank you for your kind reply. I also have downloaded the corresponding .pb file with the url.

tward <notifications@watermelon.discoursemail.com> 于2026年6月22日周一 16:44写道：