AlphaFold
Why it matters
Deep Dive
AlphaFold takes a single input — the amino-acid sequence of a protein — and returns 3D coordinates for its atoms along with per-residue confidence scores. The AlphaFold 2 pipeline first searches sequence databases to build a multiple sequence alignment (MSA) of evolutionarily related proteins and looks for known structures that can serve as templates. Both feed into a neural network built around a block called the Evoformer, which exchanges information between the alignment and a pairwise representation of residue relationships, followed by a structure module that outputs atomic coordinates. The network is trained with supervised learning on the fewer than 200,000 experimentally determined structures in the Protein Data Bank, and the pipeline recycles its own output through the network a few times to refine the result. Running a prediction takes minutes to hours on a single GPU, depending on sequence length.
The MSA Is the Secret Weapon
AlphaFold 2's single most important input is not the sequence itself but the multiple sequence alignment: hundreds or thousands of related protein sequences pulled from public databases. Residues that mutate together over evolution are usually close together in the folded structure, and this covariation signal is what lets the network infer 3D contacts. Deep alignments produce the most reliable predictions; shallow ones are the main reason a prediction comes back untrustworthy, which is common for viral proteins, de novo designed proteins, and 'orphan' sequences with few known relatives. Follow-up systems such as Meta AI's ESMFold replace the database search with a protein language model and run orders of magnitude faster, at some cost in accuracy on well-aligned families. Teams that need maximum accuracy still run full AlphaFold 2 and reserve the language-model variants for screening millions of sequences.
Inside the Evoformer
The core of AlphaFold 2 is the Evoformer, a stack of 48 blocks that passes information back and forth between two representations: the rows and columns of the alignment, and a 'pair representation' that tracks the network's current belief about the relationship between every pair of residues. The machinery is built from attention, the same basic operation that powers the Transformer, adapted with custom operations called triangular updates that push the pair representation toward geometric consistency: if residue A is near residue B and B is near C, then A should probably be near C as well. A structure module then turns the refined representation into coordinates using invariant point attention, a geometry-aware variant that behaves correctly under 3D rotations and translations. The design lesson that surprises many practitioners is how little of the accuracy comes from raw scale: AlphaFold 2 has roughly 93 million parameters, tiny by modern language-model standards.
Reading the Confidence Scores
Every AlphaFold prediction ships with a per-residue confidence score called pLDDT on a 0–100 scale. Regions scoring above 90 are generally reliable down to side-chain placement, scores of 70–90 indicate a confident backbone, and anything below 50 should be read as 'probably not folded' — low pLDDT is often a correct prediction of intrinsic disorder rather than a model failure. For complexes, the predicted aligned error (PAE) matrix reports how confidently pairs of domains or chains are placed relative to each other, which is the difference between a usable docking hypothesis and a pretty picture. The working habit to build is to look at the pLDDT plot before the 3D view, and to treat any prediction as a hypothesis that still needs experimental support before it anchors a publication or a drug program.
It Predicts Structures, Not Folding
A persistent misconception is that AlphaFold 'solved the protein folding problem.' What it solved is structure prediction: given a sequence, produce the shape the protein most likely adopts. It says nothing about how the protein gets there — the pathway, kinetics, and intermediate states that the classic folding paradox is about — and it mostly outputs one static conformation, while real proteins flex, breathe, and switch between states. It also does not directly tell you what a mutation will do to stability or function, and its handling of ligands, membranes, and post-translational modifications varies in reliability. The deep learning here is best understood as an exceptionally good pattern matcher over evolutionary and structural data, not a physics engine; for dynamics, you still need molecular simulation and wet-lab experiments.
AlphaFold 3, the Database, and the Nobel Prize
Around the model grew the AlphaFold Protein Structure Database, a collaboration between Google DeepMind and EMBL-EBI that publishes free precomputed predictions: more than 200 million structures covering nearly every cataloged protein. AlphaFold 3, released in 2024, generalizes beyond single protein chains to complexes involving DNA, RNA, ligands, and ions, and it replaces the structure module with a diffusion model that generates atomic coordinates directly, which is what lets it handle chemistry beyond amino acids. Access runs through a free hosted server, and the code and weights were later released for non-commercial use, a notable open-weights gesture for a system of this strategic value, echoing a research lineage that started with AlphaGo. In 2024 Demis Hassabis and John Jumper received the Nobel Prize in Chemistry for protein structure prediction, shared with David Baker for computational protein design, cementing AlphaFold as the standard example of AI doing original science.