AlphaGenome Atlas delivers nine billion variants on a petabyte-scale scientific platform

The atlas aims to solve a classic bottleneck. Only about 2% of the genome encodes proteins; the remaining 98% regulates when and where genes are activated and contains many variants associated with traits and diseases. Testing every mutation in a laboratory is impossible. AlphaGenome predicts thousands of molecular effects per variant across hundreds of cell types and tissues, from chromatin accessibility to RNA expression and splicing. The new AVI score combines these predictions with AlphaMissense into a single value to rank variants, and then allows users to expand the score to see which features contributed. Added to this are more than 2,500 recurring DNA motifs and their locations.

Its utility comes from the ranking guiding an experiment, not replacing it. Researchers at the Broad Institute used AVI to prioritise variants for rare diseases and detected one related to DNM1 that created an incorrect splice site; subsequent assays validated the mechanism and nearby variants. In another study, data from more than 54,000 UK Biobank participants led to the discovery of 22% more non-coding genetic associations by grouping variants by their predicted effect. By focusing on the 1% that the atlas deemed most relevant, the analysis of body mass index pinpointed 19 regions for further research. These are promising examples, not authorisation to diagnose an individual.

Operationally, it is useful to distinguish between three products. The portal is for human exploration and lowers the barrier to entry. The API fits into reproducible pipelines but requires quotas, query versioning, and cost control. The cloud model allows for inference or integration with proprietary data, in exchange for managing compute, storage, and security. Model Garden provides common patterns for deployment, evaluation, and serving, as well as container scanning and organisation policies to allow or deny models. This convenience does not transfer responsibility: even the documentation itself warns that certain suspect models can still be deployed if the customer so chooses.

A sensible enterprise design does not move entire genomes unnecessarily. It might keep pseudonymised identifiers within a restricted perimeter, query only justified regions, store the model version and the hash of each input, and separate the predictive result from the clinical record. The output must include the score, attributions, date, region, parameters, and subsequent experimental evidence. For batch research, partitioning, compression, computation close to the data, and egress budgets are essential; for interactive inference, latency, availability, and API limits are key. In both cases, the cost per validated hypothesis is more useful than the cost per query.

The scientific and legal limits are strict. A molecular prediction does not establish causality or clinical utility. Models can inherit biases from unevenly represented populations and tissues, and combining them with clinical histories turns apparently technical data into highly sensitive health information. In Europe, GDPR, legal basis, minimisation, access control, retention and, depending on the use case, health regulations all come into play. Furthermore, the free atlas is advertised for non-commercial use; a company must distinguish this licence from the commercial access to the model or the future Cloud service and record which modality was used.

Adoption should start with a small set of variants whose mechanism is already known. The ranking, explanations, cost, time, and false-positive rate should be compared against the current process; this is then repeated with a second team that has not seen the results. If the atlas improves prioritisation without degrading traceability, it can be integrated as a filter, but never as a final arbiter. It is advisable to monitor the terms and pricing for commercial access, regional localisation, quotas, the versioning of the petabyte dataset, and the publication of independent validations. AlphaGenome Atlas makes genomic research much more navigable. Success will be measured by better-chosen experiments, not by the number of predictions stored.