AlphaGenome Can Rank Variants Fast but Validation Still Decides

AlphaGenome, Google DeepMind's DNA sequence model, is being positioned as a prioritisation engine across the genome. DeepMind's AlphaGenome page presents Atlas as a catalogue of predicted effects for every possible single-letter change. Its documentation FAQ is equally clear that these outputs are theoretical modelling tools, not clinical decisions. That gap is the real buying question for biotech teams weighing where prediction ends and validation begins.

What the model actually predicts

The Nature paper describes a model taking 1 Mb of DNA and predicting thousands of functional genomic tracks at up to single-base-pair resolution. It matched or exceeded the strongest external models in 25 of 26 variant effect evaluations, per the authors. The PubMed record confirms that scope. Our analysis: this is a triage instrument, useful for shrinking candidate lists before costly wet-lab work, not a substitute for it. A high score narrows the search space; it does not establish that a variant causes disease or that a gene is a tractable target.

Validation is the real gate

DeepMind cites a Broad Institute case in which AlphaGenome scores helped flag a DNM1 editing error, and lab experiments designed with the Broad team validated the prediction, reclassifying the variant. That is one validated example, not a general warranty. The documentation index states that the API is free for non-commercial use. Our assessment: research findings should not be treated as clinical approval. Teams should budget confirmatory assays into any programme that leans on these scores, and treat a high score as a hypothesis worth testing rather than an answer. Our proposed buyer question: which assays, in which tissue, would falsify a top-ranked variant before it enters a programme?

Read the limitations before the benchmarks

The FAQ limitations section is unusually candid. Tissue-specificity and long-range interactions remain challenging. The model was trained and evaluated on human and mouse DNA. It has not been benchmarked for predicting personal genomes, and it is not diploid-aware because training used unphased data. Direct applicability to complex traits is limited. The API docs also note the free service suits thousands of predictions, not more than a million.

Noncommercial API versus the Cloud route

The GitHub repository states the API is free for noncommercial use, and outputs must not train other machine learning models. The Atlas page repeats the noncommercial restriction. Commercial and enterprise workflows go through Google Cloud via Model Garden, per the FAQ commercial-use entry. Google Cloud's AlphaGenome page lists a paid subscription and separate hosting costs. It supports self-deployment on 80 GB NVIDIA A100 or H100 accelerators, inference rather than fine-tuning, and research rather than clinical diagnostic use. Our analysis: the licensing boundary is a procurement boundary, so confirm which route a given workflow sits on before publishing results.

What to watch next

The June 2025 launch blog described its initial non-commercial research API and plans to release the model. Our analysis: buyers should ask whether predicted scores have been prospectively validated in their target tissue, what the Cloud subscription and GPU hosting will cost, and who owns derivative work. For related reading on how peers are wiring AI into discovery pipelines, see our coverage of Novo Nordisk and AWS, Lilly's AI factory, the data reproducibility problem, enterprise AI in biopharma, and synthetic cells and genomics. The next signal is published prospective validation studies, not benchmark tables. The variant scoring guide shows how scores are computed; whether they hold in the clinic is a separate question.