How AlphaFold Works — The AI That Solved Protein Folding

The breakthrough that changed biology

In December 2020, DeepMind's AlphaFold2 achieved something that had eluded scientists for 50 years: predicting protein structures from amino acid sequences with near-experimental accuracy. At the CASP14 competition (the Olympics of protein structure prediction), AlphaFold achieved a median GDT score of 92.4 out of 100 — a level of accuracy that many thought was years away.

By July 2022, DeepMind and EMBL-EBI had released predicted structures for over 200 million proteins — nearly every known protein sequence. This database is what powers our structure predictor tool.

The input: sequence and evolution

AlphaFold takes a protein's amino acid sequence as input. But sequence alone isn't enough — the key insight is using evolutionary information. The system searches protein databases to find related sequences from other organisms, creating a Multiple Sequence Alignment (MSA). If a position in the sequence is always occupied by a hydrophobic amino acid across thousands of species, that position is probably buried in the protein core. If two positions always mutate together (coevolution), they're likely in contact in the 3D structure.

AlphaFold also uses structural templates — known protein structures with similar sequences — as additional hints about the fold.

The architecture: Evoformer and Structure Module

AlphaFold's neural network has two main components. The Evoformer processes the MSA and pairwise distance predictions through 48 transformer blocks, alternating between updating the MSA representation and the pair representation. This is where the system learns which residues are in contact and how the sequence relates to structure.

The Structure Module then converts these abstract representations into actual 3D coordinates. It represents the protein backbone as a set of rigid-body frames (one per residue), and iteratively refines their positions and orientations. The module uses an invariant point attention (IPA) mechanism that works in 3D space — unlike standard attention which operates on abstract features.

The recycling trick

AlphaFold runs its entire network three times, feeding the output of each round back as input to the next. Each round refines the prediction, like an artist sketching a rough outline, then adding detail, then polishing. This "recycling" step significantly improves accuracy.

Confidence scores: pLDDT

For each residue, AlphaFold predicts a confidence score called pLDDT (predicted Local Distance Difference Test), ranging from 0 to 100. Scores above 90 indicate very high confidence — the predicted position is likely within 1 Angstrom of the true position. Scores below 50 suggest the region may be intrinsically disordered (naturally flexible and lacking a fixed structure).

On our viewer, you can color proteins by confidence to instantly see which regions are reliable and which are uncertain.

What AlphaFold cannot do

Despite its power, AlphaFold has limitations. It predicts static structures — it doesn't capture protein dynamics or conformational changes. It struggles with proteins that are intrinsically disordered (about 30% of the human proteome). It cannot predict the effects of post-translational modifications, ligand binding, or mutations. And its predictions for protein-protein interfaces in complexes are less reliable than single-chain predictions.