AlphaFold Explained: How AI Cracked a Decades-Old Biology Problem

For 50 years, biologists could read a protein’s amino acid sequence but could not reliably predict the 3D shape it folds into. Experiments took months and cost six figures per structure. In 2020, DeepMind’s AlphaFold 2 predicted shapes at near-experimental accuracy in minutes. By 2024 it had earned a Nobel Prize, and AlphaFold 3 had extended it to DNA, RNA, ligands, and ions. This guide explains the architecture, shows how to use it, and covers where it fails.

Quick Takeaways

• AlphaFold 2 (2020) predicts single protein chains from sequence. It uses an Evoformer (48 blocks) and an SE(3)-equivariant Structure Module.
• AlphaFold 3 (2024) predicts whole complexes: proteins, nucleic acids, small molecules, ions, and modified residues. It swaps the Evoformer for a lighter Pairformer and adds a diffusion module that outputs raw atom coordinates.
• Confidence scores (pLDDT, PAE, pTM, ipTM) tell you which regions to trust. Always check them before using a model.
• Access: the AlphaFold Database (200M+ monomer structures), the free AlphaFold Server, and academic-licensed AlphaFold 3 code.

The Problem AlphaFold Solved

A protein is a chain of amino acids. The chain folds into a specific 3D shape, and that shape determines its function. Get the fold wrong and you get a different protein, or a misfolded one linked to diseases like Alzheimer’s.

Predicting the fold from sequence alone is the protein structure prediction problem. Levinthal’s paradox shows why brute force fails: a chain of 300 residues has more possible conformations than atoms in the universe. Real proteins fold in milliseconds, so physics clearly has shortcuts. Researchers just could not find them computationally.

The field measured progress through CASP (Critical Assessment of protein Structure Prediction), a blind competition running since 1994. Scores use GDT_TS (Global Distance Test), where 100 is a perfect match. For two decades, top methods stalled well below the roughly 90 threshold that experts treated as “solved.” At CASP14, AlphaFold 2 hit a median GDT around 92.

Experimental methods (X-ray crystallography, cryo-EM, NMR) remain the gold standard. But they are slow. The Protein Data Bank holds roughly 200,000+ experimental structures, while UniProt catalogs hundreds of millions of sequences. AlphaFold closed much of that gap computationally.

How AlphaFold 2 Works

AlphaFold 2 treats structure prediction as an attention-based inference problem over evolutionary and geometric information.

Step 1: Build the inputs

For a target sequence, the pipeline searches large sequence databases (UniRef90, MGnify, BFD) and assembles:

• Multiple Sequence Alignment (MSA): homologous sequences from other species. Residues that mutate together often touch in 3D space. This co-evolution signal is the main clue.
• Pair representation: an N×N matrix encoding relationships between every residue pair.
• Templates: known structures of related proteins, when available.

Step 2: The Evoformer

The Evoformer is a stack of 48 blocks that passes information back and forth between the MSA representation and the pair representation. Two ideas matter:

• Triangle updates: if residue A is near B, and B is near C, then A–C distance is constrained. The network enforces this geometric consistency through triangular multiplicative updates and triangle attention.
• Row and column attention: the MSA is processed along sequences and along positions, letting the model exchange evolutionary information in both directions.

Step 3: The Structure Module

The Structure Module converts the refined pair and single representations into 3D coordinates. It uses Invariant Point Attention (IPA), which operates in 3D space and is invariant to rotation and translation. Each residue is modeled as a rigid frame (a backbone triangle) plus side-chain torsion angles.

Step 4: Recycling

The output is fed back into the network as input for several iterations. The default is 3 recycles. This simple loop improves accuracy noticeably, especially on hard targets.

Training

AlphaFold 2 trained on experimental PDB structures plus a self-distillation set: it predicted structures for unlabeled sequences, kept high-confidence ones, and trained on them. A masked-MSA auxiliary loss pushed the network to learn evolutionary patterns.

How AlphaFold 3 Changes the Architecture

AlphaFold 3 targets a broader goal: joint structure prediction for almost any biomolecular complex. According to the Nature paper, the system reduces MSA processing by replacing the AlphaFold 2 Evoformer with the simpler Pairformer module, and directly predicts raw atom coordinates with a Diffusion Module that replaces the Structure Module.

Why diffusion?

A diffusion model learns to denoise. At inference, AlphaFold 3 starts with random atom positions and iteratively refines them into a structure. The multiscale behavior helps: low noise levels induce the network to improve local structure, which let the authors drop stereochemical losses and most special handling of bonding patterns. That is how one model handles proteins, nucleic acids, and drug-like molecules without separate machinery for each.

Diffusion is also generative. You can sample multiple plausible structures with different seeds rather than getting one fixed answer. The tradeoff is hallucination: diffusion models can invent plausible-looking structure in disordered regions. The AF3 authors counter this with cross-distillation from AlphaFold-Multimer predictions, which teaches the model to render unstructured regions as extended loops.

Reported accuracy

Per the Nature paper, AlphaFold 3 showed far greater accuracy for protein–ligand interactions than state-of-the-art docking tools, much higher accuracy for protein–nucleic acid interactions than nucleic-acid-specific predictors, and substantially higher antibody–antigen accuracy than AlphaFold-Multimer v.2.3. On the PoseBusters benchmark, roughly three in four ligand poses landed within 2 Å RMSD of the experimental structure, without requiring the binding pocket as input.

Reading AlphaFold’s Confidence Scores

A predicted structure without its confidence data is half an answer. Use these metrics.

Low pLDDT often signals an intrinsically disordered region, not a model failure. Many real proteins have floppy segments with no single fixed shape.

The AlphaFold Database and Scale

The AlphaFold Protein Structure Database, built with EMBL-EBI, launched in July 2021 with human and model-organism proteomes. By July 2022 DeepMind released predicted structures for nearly all catalogued proteins known to science, expanding the database over 200x, from nearly 1 million structures.

The 2026 update matters because it moves the database from single chains to quaternary structures. Many proteins only function in pairs or larger assemblies.

Hands-On: Using AlphaFold

Option 1: Query the AlphaFold Database (fastest)

No GPU needed. Fetch a predicted structure by UniProt ID and read pLDDT from the B-factor column.

import requests
from Bio.PDB import PDBParser

uniprot_id = "P69905"  # Human hemoglobin subunit alpha
meta = requests.get(
    f"https://alphafold.ebi.ac.uk/api/prediction/{uniprot_id}", timeout=30
).json()[0]

pdb_url = meta["pdbUrl"]
with open(f"{uniprot_id}.pdb", "wb") as f:
    f.write(requests.get(pdb_url, timeout=60).content)

# AlphaFold stores per-residue pLDDT in the B-factor field
structure = PDBParser(QUIET=True).get_structure("af", f"{uniprot_id}.pdb")
plddt = [res["CA"].get_bfactor() for res in structure.get_residues() if "CA" in res]
print(f"Residues: {len(plddt)}  Mean pLDDT: {sum(plddt)/len(plddt):.1f}")

Option 2: Predict a new sequence with ColabFold (AlphaFold 2)

ColabFold swaps the slow MSA search for the fast MMseqs2 server, cutting runtime from hours to minutes.

pip install "colabfold[alphafold]"

# input.fasta: one or more sequences (use ":" between chains for a complex)
colabfold_batch input.fasta out_dir \
  --num-models 5 \
  --num-recycle 3 \
  --amber \
  --use-gpu-relax

Option 3: Run AlphaFold 3 locally

AlphaFold 3 is released under terms limited to non-commercial use, and model weights require a request to Google DeepMind. Inputs are JSON:

{
  "name": "protein_ligand_demo",
  "modelSeeds": [1, 2, 3],
  "sequences": [
    {"protein": {"id": "A", "sequence": "MKTAYIAKQRQISFVKSHFSRQ..."}},
    {"ligand":  {"id": "L", "ccdCodes": ["ATP"]}}
  ],
  "dialect": "alphafold3",
  "version": 1
}

python run_alphafold.py \
  --json_path=input.json \
  --model_dir=/models/af3 \
  --db_dir=/databases \
  --output_dir=/results

Run multiple seeds (the example uses three) and compare ranking scores. Seed variance is a useful uncertainty signal.

Open Alternatives to AlphaFold 3

The AF3 license limits commercial use, so the community built alternatives.

License terms change often. Check each repository before commercial use.

Real-World Use Cases

1. Drug discovery. Teams predict protein–ligand poses to prioritize virtual screening hits. Isomorphic Labs, the DeepMind spin-off, builds its drug design engine on this lineage. In February 2026 it claimed its IsoDDE system more than doubled AlphaFold 3’s accuracy on a generalization benchmark for difficult protein-ligand structure predictions. Treat that as a company-reported result.

2. Solving experimental puzzles. Researchers use predictions as search models for molecular replacement in X-ray crystallography, and to interpret cryo-EM density. AlphaFold predictions helped complete the structure of the nuclear pore complex, a giant assembly of hundreds of protein parts.

3. Variant interpretation. Clinical geneticists map missense variants onto predicted structures. A mutation buried in a high-confidence core is more suspicious than one in a disordered loop.

4. Antibody and vaccine design. AlphaFold 3 raised antibody–antigen docking accuracy over AlphaFold-Multimer. Teams screen candidate binders computationally before synthesis.

5. Neglected disease research. Structure coverage for pathogens like leprosy bacteria and Chagas parasites gives under-funded labs a starting point they could not afford experimentally.

A practical validation workflow

1. Predict with 5+ seeds.
2. Filter by pLDDT and ipTM; discard low-confidence interfaces.
3. Inspect the PAE plot for domain placement.
4. Compare against known homolog structures in the PDB.
5. Confirm key claims experimentally (mutagenesis, SAXS, cryo-EM).

Limitations You Need to Know

• Static snapshots. AlphaFold predicts one or a few conformations, not dynamics. It does not model folding pathways or conformational ensembles well.
• Disordered regions. Low pLDDT segments are not reliable structures.
• Mutations. It is poor at predicting how a single point mutation changes stability.
• Novel folds and orphan proteins. Shallow MSAs reduce accuracy.
• Ligand generalization. Performance drops on chemistry far from the training set, and the model may memorize common poses.
• Not a physics engine. It learned patterns from data. It does not compute free energy.

The State of AlphaFold in 2026

The technology is still in wide use, but its institutional home shifted. The Financial Times reported on July 29, 2026 that Google DeepMind had dissolved AlphaFold’s dedicated team and moved most of the researchers behind the original paper to different roles. The database, AlphaFold Server, and academic AlphaFold 3 remain available. Core researchers moved to other projects, and DeepMind has been shifting toward Gemini-based agentic systems that call specialist models like AlphaFold as tools. Expect development to continue through the open-model ecosystem and Isomorphic Labs.

FAQ

Did AlphaFold really solve the protein folding problem?

Largely, for single-chain structure prediction of well-behaved proteins. AlphaFold 2 reached near-experimental accuracy at CASP14. Folding dynamics, disordered proteins, mutation effects, and conformational change remain open problems.

What is the difference between AlphaFold 2 and AlphaFold 3?

AlphaFold 2 predicts protein structures using the Evoformer and a Structure Module. AlphaFold 3 predicts complexes of proteins, DNA, RNA, ligands, and ions using a Pairformer and a diffusion module that generates raw atom coordinates.

Is AlphaFold free to use?

Yes, with limits. The AlphaFold Database is free. AlphaFold Server is free for non-commercial use with daily job limits. AlphaFold 3 code and weights are available for academic use only. AlphaFold 2 is open source under Apache 2.0.

How accurate are AlphaFold predictions?

Residues with pLDDT above 90 are typically accurate to near-experimental resolution. Below 50, treat the region as likely disordered. For complexes, use ipTM and PAE to judge interface reliability.