AlphaFold
Also known as: AlphaFold 2, AlphaFold 3
AlphaFold is an artificial intelligence system from Google DeepMind that predicts the 3D shape of proteins from their amino acid sequences, a problem biologists had worked on for 50 years.
Knowing a protein's shape helps scientists understand how it works and design drugs that target it. While AlphaFold earned a share of the 2024 Nobel Prize in Chemistry, a predicted structure is still a hypothesis, and drug discovery remains slow, expensive work.
AlphaFold, explained at three levels
At home
Think of a protein as a long beaded necklace that crumples into a specific shape when you drop it, always the same way. That shape decides what the protein does in your body. Scientists used to spend months or years working out each shape with lab equipment. AlphaFold is an AI system that predicts the final crumpled shape just from the order of the beads.
At work
AlphaFold is Google DeepMind's AI system for predicting the 3D structure of proteins from their amino acid sequences. Its free database holds predictions for over 200 million proteins. For research organizations and biotech companies, it means structural hypotheses in minutes instead of months, faster target selection, and better-informed experiments, though predictions still need lab confirmation.
At home
AlphaFold starts with the protein's sequence and searches databases for related proteins in other species. Where two positions mutate together across evolution, they are probably touching in the folded shape. A neural network combines those clues with patterns learned from about 170,000 known structures to build a 3D model. It also scores its own confidence for each part, so you can see which regions are reliable and which are closer to guesses.
At work
AlphaFold 2 uses multiple sequence alignments and a transformer-based network, the Evoformer, to reason about pairs of residues, then a structure module that outputs atomic coordinates. Confidence comes as pLDDT per residue and PAE for the relative positions of domains. AlphaFold 3 switches to a diffusion-based approach and predicts complexes with DNA, RNA, ligands, and ions. Its code is available for academic use, with tighter terms than AlphaFold 2's open release, so commercial teams must check access before relying on it.
At home
AlphaFold predicts one most likely shape, but many proteins move, flex, or have floppy regions that only settle when they bind a partner. Low-confidence regions often reflect that real disorder, not a failure. It also learned from structures solved in labs, so it is weaker on truly novel folds. While its accuracy is remarkable, in my opinion it gives a very good photograph when biology is often a film.
At work
In a drug-discovery pipeline, AlphaFold predictions inform target selection, binding-site analysis, and construct design, but they are static and can be wrong about side-chain positions that matter for docking. Teams combine them with experimental methods such as cryo-EM, molecular dynamics, and generative design tools such as RFdiffusion from Baker's lab. Isomorphic Labs, DeepMind's spinout, applies these methods commercially. Check AlphaFold 3's licence terms before using outputs in commercial work.
Real-world examples
At home
Fatima, a biology undergraduate in Montreal, looked up the predicted structure of a protein linked to her grandmother's disease in the free AlphaFold database. Seeing the 3D shape made her lecture material suddenly concrete.
At work
Sophie leads a small structural biology team at a Canadian biotech company. AlphaFold predictions let them skip months of failed attempts to crystallize a target protein, so they moved straight to designing experiments around its most likely binding pocket.
How it compares
| feature | AlphaFold | Experimental structure determination |
|---|---|---|
| Method | AI prediction from sequence | X-ray crystallography or cryo-EM |
| Time per protein | Minutes to hours | Months to years |
| Cost | Low | High |
| Reliability | Often high, varies by region | Direct measurement |
AlphaFold gives fast, strong hypotheses. Experiments still confirm them.
For analysts
AI-driven drug discovery claims often lean on AlphaFold's reputation. I'd ask companies which steps AI actually shortened and what experimental validation followed, because predicted structures don't guarantee working medicines.
History & origin
In the 1960s, Christian Anfinsen, who won the 1972 Nobel Prize in Chemistry for the work, showed that a protein's shape is determined by its sequence, launching the 'protein folding problem.' Determining structures by experiment, through X-ray crystallography or cryo-electron microscopy, could take months or years per protein. The biennial CASP competition, started in 1994, tested prediction methods. AlphaFold placed first at CASP13 in 2018, and AlphaFold 2 achieved accuracy comparable to experiments at CASP14 in 2020. DeepMind and EMBL's European Bioinformatics Institute released a database of predicted structures in 2021, expanding it to more than 200 million proteins in 2022. AlphaFold 3, announced in 2024, extended predictions to interactions with DNA, RNA, and drug-like molecules. Demis Hassabis and John Jumper shared the 2024 Nobel with David Baker. In my view, AlphaFold is the strongest evidence so far that AI can accelerate science itself.