AlphaFold 3

AlphaFold 3 extended the model that reshaped structural biology from protein folding to the prediction of complexes — proteins with nucleic acids, ligands, ions and modified residues. For a great many structural questions it is now the first thing a researcher tries, before any experiment is designed.

What it does

Access comes in two forms: the AlphaFold Server for non-commercial use through a browser, and released code for local deployment under licence conditions. The server handles the common cases without infrastructure; local deployment matters for confidential targets and for pipelines that need to run at volume.

Strengths

• State-of-the-art accuracy across a wide range of target classes.
• Handles complexes and ligands, not only single-chain proteins.
• Browser-based server requires no local compute.
• Per-residue confidence scores that meaningfully flag unreliable regions.

Limitations

Licensing is the sore point: non-commercial terms and restricted code release drove much of the field toward open alternatives such as Boltz and Chai-1. Predictions are models, not structures — confidence scores flag uncertainty but do not eliminate it, and disordered regions and novel folds remain weak spots.