AlphaGenome Atlas: DeepMind Precomputes 9 Billion DNA Variants to Rank Rare-Disease Candidates

Benchmark evidence for the base AlphaGenome model is relatively mature.

The model is published in Nature. It beat external models in 22 of 24 sequence-track evaluations and matched or exceeded the best external model in 25 of 26 variant-effect evaluations. Its clearest advantage is unifying one-megabase context, single-base resolution and multimodal output—not dominating every individual task.^[S1][S9]

AVI does not lead every task.

AVI improves clearly on ClinVar intronic, synonymous and 3′ UTR categories, but is comparable to or slightly below the best method on 5′ UTR, non-coding Mendelian disease, fine-mapped GWAS and rare-complex-trait tasks. For example, its AUPRC is 0.76 versus 0.77 for GPN-Star-M on a non-coding Mendelian benchmark, and 0.28 versus 0.27 on fine-mapped GWAS. The accurate claim is strong or near-best performance across multiple tasks, with clear gains in some non-coding categories—not universal superiority.^[S6]

DNM1 is a relatively complete but still local evidence chain.

AVI ranked the variant first for one patient, AlphaGenome proposed a testable abnormal-splicing mechanism, and minigene experiments supported it and the authors' Likely Pathogenic recommendation. This shows usefulness for candidate ranking in a specific case, not prospective clinical effectiveness across undiagnosed patients.^[S6]

The UK Biobank 22% is a statistical discovery rate, not additional causal truth.

Across 2,028 circulating-protein analyses in 54,189 participants, Atlas-filtered aggregates increased significant-association discovery by 22%; 177 of 241 associations that remained significant after conditional analysis came from Atlas features. These are mainly statistical-genetic signals without variant-by-variant molecular or clinical validation, and the analysis used an earlier Atlas version. Of 25 of 31 complex-trait associations with sufficient All of Us counts, only four reached nominal replication and none passed Bonferroni correction.^[S6]

AVI's unified score deliberately compresses context.

It takes maximum effects across tissues and compresses them into one score. That improves genome-wide ranking while sacrificing some cell and tissue context. SHAP attribution explains which inputs drive AVI but cannot establish in-vivo causality.

AlphaGenome can model only biology represented in its training data.

The technical report notes that poly(A)-selected RNA-seq misses non-polyadenylated RNA, relevant cell types are absent, modality coverage is uneven and the model mainly learns cis-regulatory grammar rather than trans effects from changes such as transcription-factor expression. Distal regulation beyond effective context, personal-genome effects, genetic background, development and environment remain difficult.^[S1][S6]

Existing independent evaluation mainly targets the base model, not the Atlas.

An independent preprint using GTEx data from 953 people found AlphaGenome clearly better than Enformer yet behind Elastic Net and Random Forest models trained directly on individual data. Different objectives and available data prevent a simple “traditional models beat AlphaGenome” conclusion; the result shows that reference-sequence molecular prediction does not automatically transfer to personal expression prediction.^[S5]

External experts praise unified multitask capability while warning that one score can be misread.

Experts assembled by the Science Media Centre praised AlphaGenome's engineering while noting that individual tasks often merely match specialist tools and are not ready for patient care. IEEE Spectrum quoted UBC genomicist Carl de Boer warning that AVI is useful but easy to misread because it compresses a complex system. Nature's launch coverage likewise describes the Atlas as a predictive resource rather than an experimental or clinical verdict.^{[S9][S10][S11]}

The ‘genomic AlphaFold Database’ analogy has clear limits.

AlphaFold Database maps known protein objects; Atlas maps counterfactual consequences of possible mutations. Protein structures can be compared with relatively clear experimental coordinates, while genomic function is distributed across cell types, tissues, modalities and time, with multiple causal layers between molecular effect and disease. Most hypothetical Atlas variants may never be observed in people. DeepMind explicitly states that AlphaGenome and Atlas are research-only, not a substitute for medical advice and not validated or approved for clinical decisions.^[S2][S3][S6]