Wiki · Bio × AI Methods · concept
AlphaFold
A model that predicts a protein's 3D structure directly from its amino-acid sequence.
AlphaFold predicts a protein’s three-dimensional structure from its amino-acid sequence. A protein is a chain of amino acids that folds into a specific shape, and that shape largely determines what the protein does. Working out the shape used to mean months or years of painstaking lab experiments. AlphaFold, from DeepMind, reads the sequence and outputs where every atom lands.
How it works
AlphaFold2 — the version that stunned the CASP14 structure-prediction contest in 2020 — starts by gathering a multiple sequence alignment (MSA): the same protein across many species, whose patterns of co-variation hint at which residues touch. A module called the Evoformer (a transformer-style network) passes information back and forth between that alignment and a map of residue pairs, and a final structure module turns the result into 3D coordinates. Crucially, it reports a per-residue confidence score, pLDDT, so you know which parts to trust. AlphaFold3 (2024) rebuilt the back end around a diffusion model: instead of a dedicated structure module, it denoises atom positions into place, and this more general approach lets it handle not just proteins but ligands, nucleic acids, and their complexes.
Why it matters (for bio × AI)
AlphaFold turned protein structure from a years-long experiment into something close to an API call — feed in a sequence, get a credible structure back in minutes. That shift underpins much of modern computational biology: it is the foundation beneath co-folding (predicting how proteins bind partners) and a routine first step in drug discovery. It also showed a field what an AI breakthrough looks like when a hard, expensive measurement becomes a fast prediction.
Related
See protein language model for models that read sequences without an alignment, diffusion model for the generative method behind AF3, and transformer for the architecture at the Evoformer’s core.