aF2

AF2, short for AlphaFold 2, is a deep learning system developed by DeepMind for predicting the three-dimensional structures of proteins from their amino acid sequences. Publicly introduced in 2020, AF2 achieved unprecedented accuracy in the CASP14 competition, signaling a major advance in computational structural biology. The system uses an end-to-end neural network that integrates evolutionary information from multiple sequence alignments, residue contact maps, and investigative templates to predict atomic coordinates for proteins, including side chains.

A key feature of AF2 is its per-residue confidence estimate, known as pLDDT, which helps users assess which regions of a prediction are reliable. AF2 predictions are typically highly accurate for well-studied, single-chain proteins with sufficient evolutionary information, often approaching near-atomic resolution for many targets. The method can generate models rapidly at scale, enabling large-scale exploration of protein structure space.

Impact and access: AF2 has spurred substantial advances in biology and drug discovery by enabling the rapid generation of structural hypotheses for proteins lacking experimental structures. In collaboration with EMBL-EBI, DeepMind and partners established the AlphaFold Protein Structure Database, which houses predicted structures for numerous organisms, including the human proteome. The code and models have been released to the research community, and community-driven resources such as ColabFold provide more accessible ways to run AF2 predictions.

Limitations: AF2 is not designed to model dynamic conformations, protein–protein complexes, or proteins with large post-translational modifications or cofactors. Predictions may be less reliable for intrinsically disordered regions, membrane protein assemblies, or metalloproteins without explicit cofactors. Predicted structures should be interpreted with appropriate caution and, when possible, validated experimentally.