AlphaGenome's Research Value Depends on What Happens After Prediction

AlphaGenome's useful business question is not whether an AI model can generate more genomic predictions, but which research decisions those predictions can improve. Google DeepMind's documentation describes molecular outputs and practical scoring tools while setting boundaries on clinical use. Our assessment is that value should be tested through an explicit path from prediction to experiment.

Start with the experimental question

AlphaGenome documentation describes predictions covering gene expression, splicing patterns, chromatin features and contact maps (the model overview). DeepMind explains that variant scoring compares predictions for mutated DNA sequences with predictions for unmutated sequences (DeepMind's research explanation).

A research team should choose a biological question before choosing an output. For example, ask whether a proposed experiment could test the predicted direction of a molecular change. Define in advance which result would support the hypothesis and which would contradict it. That would make prediction quality part of a research decision rather than a standalone presentation.

Related reading: drug-discovery analysis.

Keep clinical decisions outside the pilot

The FAQ states that predictions are for theoretical modelling and research and must not be used for clinical decision-making (the research-use FAQ). The official repository separately points to commercial access through Google Cloud, alongside the API usage terms (the official repository).

Commercial access should not be interpreted as clinical validation. Procurement teams should check the terms for their intended access route and keep patient-care decisions outside this proposed research workflow. A sensible approval gate would specify permitted outputs, downstream users and the evidence required for experimental follow-up, rather than treating a licence as proof of medical suitability.

Related reading: biotechnology coverage.

Make inputs and scale reproducible

The quick-start guide says predictions cover multiple tracks across tissues and cell types, with subsets selectable for efficiency (the quick-start guide). The batch-scoring tutorial specifies chromosome, position, reference and alternate sequence information when preparing variants for scoring (the batch-scoring tutorial).

Our recommendation is to save input coordinates, reference sequences, selected outputs and configuration with every result. A proposed comparison should use the same settings for the baseline and the model-assisted workflow. Ask whether another analyst could reconstruct why a variant was prioritised. Reproducibility should be a pilot deliverable, not an assumption attached to automated scoring.

Related reading: medical technology coverage.

Match the score to the biological context

The ontology tutorial demonstrates finding tissue and output metadata by searching the available cell or tissue names (the ontology tutorial). The splicing tutorial combines splice-site, splice-site-usage and splice-junction scorers into a merged approach (the splicing tutorial).

These tools suggest different checks for different hypotheses. A splicing question should not be assessed solely through an unrelated output, and the chosen tissue context should be recorded explicitly. We would ask the laboratory and computational teams to agree on the relevant readout together, then inspect disagreements rather than allowing one aggregate score to settle the question.

Related reading: biological AI coverage.

Measure the next decision, not prediction volume

The haplotype tutorial considers combined variant effects and compares them with individual effects as an indication of predicted cooperation (the haplotype tutorial). The single-variant tutorial demonstrates predicting effects on modalities including gene expression and chromatin accessibility (the single-variant tutorial).

The next evidence to seek is whether the workflow improves a defined experimental choice under a fixed research budget. Track which predictions were tested, which failed and whether the outcome changed the next experiment. A larger scored dataset would not answer that question by itself. Expand only when the research team can explain what decision the model helped make.

Related reading: research infrastructure coverage.