AlphaFold, AlphaGenome, Evo 2 and State: what each model does

Choose the biological question before comparing model names. A researcher trying to locate a ligand in a protein complex needs a structural prediction. A researcher studying a regulatory DNA change needs a predicted molecular consequence of that change. A researcher choosing a cell intervention needs a response prediction in the relevant cells.

The four models below occupy different parts of that work. This comparison names the published versions and their output types. It gives no overall accuracy ranking: the models have not taken one common test with one common answer.

Match the output to the experiment

AlphaFold: positions of atoms

AlphaFold 2's 2021 paper concerned protein structure prediction. AlphaFold 3's 2024 paper expanded the scope to complexes containing proteins, nucleic acids, small molecules, ions and modified residues. Its output describes predicted atom positions. Confidence estimates help researchers inspect the prediction.

The chemical question often continues after the structure. A ligand's predicted position in a pocket is a proposal for its binding arrangement. Binding affinity asks how strongly the ligand associates under specified conditions; selectivity asks how it behaves against other targets. Both require their own measurements.

A structural prediction does not establish exposure, duration of action or effects in a tissue. Those properties need experimental measurements.

AlphaGenome: molecular effects of DNA changes

The January 2026 AlphaGenome paper describes a model taking a one-million-base DNA interval and predicting genome tracks. Those tracks represent measurements such as expression, splicing and chromatin accessibility. Comparing predictions for a reference sequence and an altered sequence estimates molecular effects of a variant.

A track associates positions in the genome with predicted assay values. It is therefore more specific than saying a stretch of DNA is useful or harmful. The model can suggest which measurement to make and where a regulatory effect might appear.

DeepMind's account states that AlphaGenome was not designed or validated for personal genome prediction or direct clinical purposes. Its molecular outputs also leave developmental and environmental contributions to disease outside their direct scope. A predicted splicing change deserves a splicing experiment; it does not by itself establish a person's eventual symptoms.

Evo 2: learning and generating sequence

Evo 2 learns from DNA sequences across diverse organisms. Its March 2026 Nature paper describes 7-billion-parameter and 40-billion-parameter versions, with long sequence context. The work examines variant-effect prediction and DNA generation; guided designs also underwent experimental tests of chromatin accessibility.

A sequence model assigns probabilities in context. Researchers can use those probabilities to compare sequence alternatives or generate new sequence. A generated string is the model's proposal. Any intended biological function must survive synthesis and measurement.

The paper's genome-scale generation results should not become a claim that every generated genome was assembled into a viable organism. Generation, assay validation of a designed element and construction of a functioning organism are distinct experimental achievements.

State: responses of cell populations

State's Cell paper, available online on 31 August 2026, reports a model of expression responses to perturbations. It operates on sets of cells, giving the prediction a way to account for population variation.

Its authors test transfer across datasets and cellular contexts using declared comparisons. A claim about an unfamiliar context needs the training boundary stated clearly. Seeing an intervention in another cell line, seeing some interventions in the target cell line and seeing no matching perturbations in that line provide different information.

The model can help propose a cellular experiment. The response assay tests its prediction; a separate functional assay tests whether the intervention accomplishes the research objective.

Connecting models in a research programme

A hypothetical programme could use a sequence model to propose regulatory DNA, a regulatory predictor to prioritise its expected effect and a cell-response model to suggest what to measure after introducing it. That is a proposed workflow. An actual paper needs to identify which tools it used, which candidates were physically made and what happened in the assay.

Predictions linked in this way carry uncertainty from one step into the next. A strong result in the final assay can justify the programme's choice, while an unsuccessful result can reveal which assumption failed. The candidate count and failed tests help show what the models contributed.