AlphaGenome AI: What It Predicts and How to Run It on Your Own Genome

AlphaGenome is a deep learning model from Google DeepMind that takes up to 1 million base pairs of DNA as input and predicts thousands of molecular measurements along that sequence, at up to single base-pair resolution, for hundreds of human and mouse tissues and cell types ¹. You give it a reference interval and a variant. It returns predicted tracks for the reference allele and the alternate allele, plus scalar scores summarizing the difference. For someone holding their own whole-genome sequencing data, it is the most capable available tool for asking what a non-coding variant might do in a specific tissue. The question can concern gene expression, splicing, or chromatin. It is a predictor of molecular consequence. It is not a clinical interpreter.

What the model outputs

AlphaGenome is a sequence-to-function model. It reads DNA and predicts functional genomics assay readouts as if those assays had been run on that sequence. The output modalities cover RNA-seq coverage, CAGE and PRO-cap for transcription initiation, ATAC-seq and DNase for chromatin accessibility, ChIP-seq for transcription factors and histone marks, splice site probabilities, splice junction counts, splice site usage, and 3D contact maps ¹.

Each track is tied to an ontology term, so you request predictions for a specific biological context. Lung is UBERON:0002048, whole blood is UBERON:0000178, and cell lines use EFO identifiers. This is the part that matters most in practice. A variant in a liver enhancer will produce a strong signal in liver tracks and nothing in brain tracks, and asking the wrong tissue is the most common way to get an uninformative answer.

Resolution differs by modality. Most coverage tracks come back at 1 bp resolution across the requested window. Contact maps are binned much more coarsely, at 2,048 bp, because chromatin interaction data is not meaningful at base resolution. The model was trained jointly on human (hg38) and mouse (mm10) data, and reported state-of-the-art performance on the large majority of the variant effect benchmarks it was evaluated against ¹.

Getting access and making the first call

The model runs as a hosted gRPC service. You register for an API key through DeepMind, install the client, and call it from Python. Free-tier access is limited to non-commercial use, and there are request quotas that make genome-wide scoring impractical.

pip install alphagenome

A minimal variant prediction looks like this:

from alphagenome.data import genome
from alphagenome.models import dna_client

model = dna_client.create(API_KEY)

variant = genome.Variant(
    chromosome='chr22',
    position=36201698,
    reference_bases='A',
    alternate_bases='C',
)

interval = variant.reference_interval.resize(
    dna_client.SEQUENCE_LENGTH_1MB
)

output = model.predict_variant(
    interval=interval,
    variant=variant,
    ontology_terms=['UBERON:0000178'],
    requested_outputs=[
        dna_client.OutputType.RNA_SEQ,
        dna_client.OutputType.SPLICE_JUNCTIONS,
    ],
)

Two details cause most first-attempt failures. The interval length must be one of the supported values (2 kb, 16 kb, 100 kb, 500 kb, or 1 Mb), and the variant must sit near the center of it. Using variant.reference_interval.resize(...) handles both. The other is coordinate convention: positions are 1-based and must match hg38, so a VCF lifted from hg19 will silently point at the wrong base.

The result object has .reference and .alternate attributes. Each holds arrays shaped (sequence_length, num_tracks) with track metadata in a companion DataFrame. For a quick look, subtract the two and find where the difference is largest.

Scoring variants from your own WGS

A human genome carries roughly four to five million variants against the reference, and the overwhelming majority are common and non-coding. Feeding all of them through the API is neither possible nor useful. We would filter hard before scoring.

The filter we use has three stages. First, drop common variants using a gnomAD allele frequency annotation and keep those below about 0.1%. Common variants with large regulatory effects are mostly already characterized. Second, restrict to variants within gene bodies or within 100 kb of a transcription start site, because the model’s signal falls off with distance and scoring intergenic deserts wastes quota. Third, prioritize deep intronic and promoter-proximal variants, which is where AlphaGenome adds information that a coding-variant annotator like VEP cannot provide.

bcftools view -i 'INFO/AF<0.001 && TYPE="snp"' sample.vcf.gz \
  | bcftools view -R gene_windows_100kb.bed -Oz -o candidates.vcf.gz

For scoring, use the packaged scorer set rather than hand-rolling summary statistics:

from alphagenome.models import variant_scorers

scorers = list(variant_scorers.RECOMMENDED_VARIANT_SCORERS.values())

scores = model.score_variant(
    interval=interval,
    variant=variant,
    variant_scorers=scorers,
)

df = variant_scorers.tidy_scores(scores)

tidy_scores returns a long DataFrame with one row per track per scorer. Two columns matter. The raw score is the signed effect size in the model’s units. The quantile score places that effect against the distribution of scores for common variants, which is the number to use when you want to know whether an effect is unusually large. A raw log-fold-change of 0.3 means little on its own. A quantile score above 0.99 in the relevant tissue means the model considers the effect large relative to background.

Where it works and where it does not

Independent evaluation matters more than the developer’s own benchmarks. A benchmarking study in Cell Research assessed AlphaGenome on non-coding variant effect prediction and found that performance varies substantially by task and modality, with some settings where simpler baselines remain competitive ². The practical reading is that AlphaGenome is strong at ranking candidates within a locus and weaker at producing calibrated effect sizes you can compare across loci.

The successes are concrete. In a family with PLA2G6-associated neurodegeneration, AlphaGenome flagged a deep intronic variant that standard pipelines had passed over, closing a diagnostic gap ³. A study of non-coding regulatory variants controlling RHD expression used the model to generate hypotheses and then confirmed them with wet-lab experiments ⁴. Both follow the same pattern: narrow set of candidates, specific gene, orthogonal validation.

Larger-scale uses are more mixed. Work applying the model to the regulatory architecture of endometriosis shows the approach can be extended to GWAS loci across a disease ⁵. Cross-species work has examined how well sequence-to-function predictions transfer between organisms and how performance scales ⁶. And at least one study has used AlphaGenome to test a theoretical question, evaluating the mutation accumulation hypothesis in aging ⁷. Treat outputs from these broad applications as hypothesis generation.

Checking predictions against your own molecular data

This is the part most guides omit. If you have RNA-seq from the same person whose genome you scored, you can test a subset of predictions directly.

For a predicted expression change, look for allele-specific expression. Phase the variant with a heterozygous exonic SNP in the same gene, count reads carrying each haplotype with a tool like phASER or GATK ASEReadCounter, and check whether the allelic ratio deviates from 0.5 in the predicted direction. For a predicted splicing change, pull junction counts from your aligner’s SJ.out.tab (STAR) and check whether the novel or weakened junction appears at measurable frequency.

Agreement is meaningful evidence. Disagreement is common and does not necessarily mean the model is wrong, because your RNA-seq tissue may not be the tissue where the regulatory element is active, and coverage at the relevant junction may be too low to detect a modest shift.

None of this is a diagnosis. A high-scoring variant in a gene associated with disease is a finding to bring to a clinical geneticist, who can order confirmatory testing and interpret it against your family history and phenotype.

Questions people also ask

Is AlphaGenome free to use? Access to the hosted API is free for non-commercial research use with a registered key, subject to rate limits. Commercial use requires separate arrangement with DeepMind.

Can I run AlphaGenome locally? The model is served through the API rather than distributed as downloadable weights for general use, so you are dependent on network access and quota. This is the main constraint on large-scale scoring.

How does it compare to Enformer or Borzoi? AlphaGenome extends the same architectural lineage to a 1 Mb input window with single base-pair output resolution and a wider set of modalities, including splicing and contact maps in one unified model ⁸. It reported better performance than prior models on most benchmarks tested ¹.

Which variants should I prioritize scoring? Rare non-coding variants near genes you already have reason to care about, especially deep intronic variants and promoter-proximal changes. Coding variants are better served by established annotation tools first.

Does a high score mean a variant is pathogenic? No. The score estimates a molecular effect size, and molecular effect does not map directly onto clinical consequence. Interpretation requires a clinician.

Footnotes

1. Žiga Avsec, Natasha Latysheva, Jun Cheng, et al. Advancing regulatory variant effect prediction with AlphaGenome. Nature, 2026. https://doi.org/10.1038/s41586-025-10014-0 ↩ ↩² ↩³ ↩⁴

2. Alan E. Murphy, Masayuki Nagai, Peter K. Koo. Predicting non-coding variant effects with AlphaGenome. Cell Research, 2026. https://doi.org/10.1038/s41422-026-01249-1 ↩

3. Sarah J Eger, Greizy López, Luisa Fernanda Gómez Navarro, et al. AlphaGenome identifies a deep intronic variant in a family with PLA2G6-associated neurodegeneration: Closing the diagnostic gap in rare genetic diseases. 2026. https://doi.org/10.64898/2026.06.10.26355004 ↩

4. Miao Liu, Ziyang Shen, You Kyeong Jeong, et al. AlphaGenome-enabled analysis of non-coding regulatory variants underlying RHD expression with wet-lab validation. 2026. https://doi.org/10.64898/2026.01.21.700828 ↩

5. Sanu Bifal Maji, Apostol Apostolov, Alberto Sola-Leyva, et al. Decoding the regulatory genetic architecture of endometriosis using AlphaGenome. 2026. https://doi.org/10.64898/2026.06.27.26356730 ↩

6. Priya Ramarao-Milne, Suyu Ma, Letitia Sng, et al. Evaluating the cross-species transferability and scaling of sequence-to-function predictions in AlphaGenome. 2026. https://doi.org/10.64898/2026.07.10.737654 ↩

7. Arthur Fischbach. Testing the mutation accumulation hypothesis in aging with AlphaGenome. 2026. https://doi.org/10.64898/2026.05.10.724136 ↩

8. Žiga Avsec, Natasha Latysheva, Jun Cheng, et al. AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model. 2025. https://doi.org/10.1101/2025.06.25.661532 ↩