Interpreting AlphaGenome and Atlas scores

Distilled from the AlphaGenome and Atlas papers, the official FAQ, and the interpretation guide DeepMind ships with its own AlphaGenome agent skill. Reviewed against official scoring definitions and model source on 2026-09-30. These rules describe predictions, not established effects in a patient.

The frame

AlphaGenome predicts what a genomics assay would read out from a DNA sequence. An Atlas or model score is therefore a predicted molecular effect in a biosample the model was trained on, nothing more:

• It is not pathogenicity. AVI is trained to separate rare from common variants; that correlates with impact and is validated on ClinVar, but a high AVI is one line of evidence in a chain, "not sufficient evidence on its own" (Atlas report, Discussion).
• It is not a measurement. Predictions are correlative and can be wrong in either direction; independent epigenomic or functional data decides.
• It is not clinical. Terms of use and the report both exclude diagnostic use. Never turn a score into a diagnosis, penetrance, prognosis, or treatment statement. If the question is clinical, answer about molecular mechanism and say so.

Raw score and quantile: always both

For RNA_SEQ, the scorer is ln(mean ALT + 0.001) - ln(mean REF + 0.001). Thus exp(raw_score) is the ratio of pseudocount-adjusted means: -1 is about 0.37x, not 0.5x; ln(2) is about 0.693. At very low expression the pseudocount matters, so inspect baseline tracks before translating a score into a biological fold change. Recommended center-mask accessibility/TF/TSS scorers instead use a log2 ratio of summed signal with a +1 pseudocount. RNA_SEQ_ACTIVE is the larger mean activity, with no log transform. Never apply one scorer's scale to another.

The scripts' --min-abs-quantile 0.99 keeps signed quantiles at or beyond +/-0.99 (the two 0.5% CDF tails), or unsigned quantiles >=0.99 (the upper 1%). It does not retain small unsigned scores near zero. Missing calibration does not pass this filter; mouse and custom scorers may have only raw scores. AVI has a separate genome-wide SNV CDF and Phred transformation. Quantiles are ranks, not significance tests or evidence of benignity. See the official scoring definitions and calibration FAQ.

The common trap: high quantile, tiny raw score. In low-expression genes and quiet regions the background distribution is so narrow that a raw change of 0.05 ranks above 0.99. Report a small predicted change in the named track and inspect the baseline signal; this does not establish a statistically nonsignificant or biologically absent effect. Conversely a raw -1.5 with quantile 0.9 in a highly variable track is still a large predicted effect worth mentioning.

Unsigned scorers (POLYADENYLATION, SPLICE_*, CONTACT_MAPS, all *_ACTIVE) have no direction; *_ACTIVE scores are the activity of the stronger allele and cannot be read as a difference at all.

AVI thresholds

The report recommends ranking over hard cut-offs and thresholds that are "genomic region- or application-aware": pathogenic regulatory variants sit in lower AVI bins than pathogenic protein-truncating or splice-motif variants, so a single genome-wide Phred cut-off systematically under-calls regulatory hits. Practical pattern:

1. Rank the candidate set by Phred and report the ranks.
2. State the top-percentile the Phred implies (Phred 20 = top 1 % of SNVs).
3. Read the feature attribution. MERGED_SPLICING or ALPHAMISSENSE dominant means a coding or splice mechanism; MAX_ABS_DNASE / MAX_ABS_CHIP_TF / MAX_ABS_RNA_SEQ mean a regulatory mechanism you can drill into by track; CACTUS_241_WAY / PHASTCONS_470_WAY dominant means conservation is carrying the score and the molecular mechanism is not resolved by AVI.
4. Only then drill into the per-track scores for the matched tissue.

Tissue matching

• Choose biosamples by the disease's organ system and cell type, resolved to ontology terms, not by name similarity. Cardiac muscle is not smooth muscle; a fibroblast line is not brain.
• Report both the tissue with the largest predicted effect and the disease-relevant tissue, even when the latter shows nothing. An unexpected top tissue is a finding; a silent relevant tissue is also a finding.
• If the specific cell type has no track, fall back to the organ and say so.
• Evidence from an unrelated tissue is evidence about a different question.

Negative results

A flat prediction does not establish that a variant is benign. "AlphaGenome predicts no appreciable change in the queried DNase and RNA-seq tracks" is a complete, valuable answer. Do not invent a cryptic splice site or an enhancer disruption that the tracks do not show; do not infer disruption from location in a peak when REF and ALT tracks are identical.

What the model cannot see

• Trans effects. Only cis-regulatory grammar in a 1 Mb window; nothing about TF abundance, signalling, or the rest of the genome.
• Training-data gaps. RNA-seq training includes poly(A)+ and total RNA assays, with uneven coverage of non-polyadenylated RNAs (including snRNAs); inspect the relevant assay before claiming the model captures them. Many cell types are absent; coverage is uneven across assays. A flat prediction in an untrained cell type is absence of data, not absence of effect.
• Protein-level consequences. Missense stability, catalysis, folding: use AlphaMissense (already an AVI feature) or protein tools.
• Post-transcriptional biology beyond splicing and polyadenylation: RNA structure, miRNA processing, localisation, translation.
• Diploidy. One haplotype per prediction; no heterozygous dosage, no compound effects unless you build the haplotype sequence yourself (the "haplotype workaround" tutorial).
• Developmental time and stimulus context. Static biosample profiles only.
• Indels in the Atlas. Scored for the paper, not yet served; use the model. Very large structural variants are unreliable everywhere.
• Other species. Only human and mouse were trained; nothing else is benchmarked.

Coordinate hygiene before any lookup

1. Assembly must be GRCh38 for the Atlas (hg38 for the human model, mm10 for mouse). Lift over GRCh37 sources first and re-check REF.
2. REF must equal the reference base; the Atlas key ignores REF, so a swapped allele can miss; a mismatched REF/build can misidentify a record. These helpers compare the returned variant to the requested variant but do not perform a FASTA reference check.
3. Variants are 1-based, genome.Interval is 0-based half-open. Converting a 1-based closed chr:start-end means start - 1, end.
4. Left-normalise indels before scoring with the model.
5. Keep the chr prefix (chrM for the mitochondrion).

The genomic-coordinates skill covers all of this with scripts.

Reporting checklist

• Name the source: "AlphaGenome Atlas (AVI, precomputed)" or "AlphaGenome model (on-demand, 1 Mb window)", with the client version and the date of the query.
• For every variant: raw score and quantile or Phred, the scorer, the track or biosample and its ontology CURIE, and the gene for gene-centric scorers.
• Direction as words ("predicted 2.3-fold lower HBB expression in erythroblast RNA-seq"), never a bare number.
• Say which claims rest on the model alone and which have independent support.
• Link each variant to the Atlas website so a reader can inspect the tracks (scripts/atlas_link.py).
• Close with the limitation that applies (untrained tissue, conservation-driven AVI, non-polyadenylated gene, ...), and the standard research-only statement.
• Cite the AlphaGenome paper and the Atlas report.