How AlphaFold Cracked the 50-Year Protein Folding Problem

300 Million Protein Structures Predicted in Under Two Years

For 50 years, the protein folding problem stood as one of the central unsolved challenges in biology: given only the sequence of amino acids in a protein, predict its three-dimensional shape. The structure determines function — enzymes, antibodies, receptors, and structural proteins all derive their activity from their precise geometry. Solving a single protein structure by X-ray crystallography can take years and hundreds of thousands of dollars. In 2021, DeepMind's AlphaFold2 achieved atomic-level accuracy on the majority of proteins in the Critical Assessment of Structure Prediction (CASP14) benchmark, matching the accuracy of experimental methods. By 2022, AlphaFold had predicted structures for over 200 million proteins — essentially the entire catalogued proteome of life on Earth.

The achievement did not just solve a computational puzzle. It transformed structural biology, accelerated drug discovery, and demonstrated that deep learning could tackle problems at the frontier of science.

Why Protein Folding Was So Hard

Proteins are chains of amino acids — 20 types, each with a distinct side chain. A typical protein contains hundreds to thousands of amino acids; the largest exceed 10,000. Anfinsen's dogma (Nobelist Christian Anfinsen, 1972) established that the amino acid sequence determines the three-dimensional structure: folding is thermodynamically driven, seeking the lowest-energy conformation.

The computational obstacle is combinatorial. Levinthal's paradox, proposed in 1969, illustrates it: if a 100-residue protein could adopt just two conformations per residue, it would have 2¹⁰⁰ possible configurations. Sampling all of them at nanosecond intervals would take longer than the age of the universe. Yet proteins fold reproducibly in milliseconds to seconds.

• Experimental methods — X-ray crystallography, cryo-electron microscopy, NMR spectroscopy — determine structure empirically but require purified protein, time, and expertise.
• By 2020, the Protein Data Bank contained about 170,000 experimentally determined structures — a tiny fraction of the estimated 200 million unique proteins in sequenced organisms.
• Computational approaches before AlphaFold used fragment assembly, molecular dynamics simulations, and evolutionary covariation analysis, but none achieved consistent near-experimental accuracy.

How AlphaFold2 Works

AlphaFold2 combines evolutionary information, geometric constraints, and deep learning in a novel architecture. The key insight: amino acids that co-evolved — that mutated together across thousands of related species — are likely in physical contact in the folded protein. Mutations at one position are compensated by mutations at a contacting position to preserve structure and function.

AlphaFold2 processes these inputs through an Evoformer module — a novel transformer architecture that jointly reasons over the MSA and pairwise distances between residues — and a Structure Module that produces actual 3D coordinates. The network outputs a predicted aligned error (PAE) score — a confidence estimate for the relative orientation of each residue pair — alongside per-residue local confidence scores (pLDDT, 0–100).

• At CASP14, AlphaFold2 achieved a median backbone RMSD of 0.96 Å across free-modelling targets — comparable to experimental accuracy.
• For proteins with pLDDT > 90, AlphaFold predictions are routinely used as starting models for X-ray crystallography and cryo-EM without experimental structure determination.
• AlphaFold correctly predicted the structure of human nuclear pore complex components that had resisted crystallisation for decades.

AlphaFold Database and Scientific Impact

DeepMind released AlphaFold2's source code and a precomputed database of predicted structures in July 2021 in collaboration with the European Bioinformatics Institute (EMBL-EBI). By 2022, the database contained predicted structures for 200 million proteins from UniProt — 98.5% of all known protein sequences, including every human protein.

• Within one year of release, the AlphaFold database had been accessed by over 1 million unique users from 190 countries.
• Drug discovery applications proliferated immediately: AlphaFold structures are used to identify binding pockets, design inhibitors, and model protein-drug interactions for disease targets previously lacking structural data.
• In parasitology, AlphaFold provided structures for Toxoplasma gondii, Trypanosoma brucei, and malaria parasite proteins — potential drug targets that had no solved structures.
• AlphaFold2 co-inventors Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry, along with David Baker for protein design work.

Protein Design: The Inverse Problem

AlphaFold solved the forward problem: sequence to structure. Baker's lab at the University of Washington tackled the inverse problem: design a sequence that folds into a desired structure. RoseTTAFold (2021) extended this capability. RFdiffusion and ProteinMPNN (2022–2023) enabled the de novo design of proteins with user-specified folds, binding interfaces, and enzymatic activities.

In 2022, Baker's group designed an enzyme — RoseTTAFold-designed enzyme — that catalyses a Diels-Alder reaction with no natural equivalent, demonstrating that computational protein design can create biochemical functions evolution never produced. In 2024, the team reported designed proteins capable of delivering mRNA into cells — a potential platform for non-viral gene therapy. The protein folding problem, once a benchmark for human understanding, has become a springboard for designing molecular machines with programmable function.