AlphaFold 2 — AI Solves the Protein Folding Problem

On 30 November 2020, at the release of the CASP14 results, AlphaFold 2 achieved a median GDT score of 92.4 out of 100 — beyond the threshold of about 90 above which the problem is held to be "solved". Predicting the 3D structure of a protein from its sequence alone, an open question since Levinthal's paradox (1969), went from months of crystallography to a few minutes of computation. In 2022, 200 million predicted structures were released free of charge.

Source: deepmind.google

In plain terms

A protein is a long chain of links that folds back on itself to adopt a precise shape. Guessing that shape from the list of links alone amounts to searching for a needle in an astronomically large space of possibilities: Levinthal's paradox put that space at roughly 10³⁰⁰ conformations. For fifty years, obtaining the real shape required months, sometimes years of crystallography. AlphaFold 2 does the same job in a few minutes, with a score of 92.4 out of 100 at the reference competition CASP14 — 90 being the threshold above which the community speaks of a "solved" problem. The caveat fits in one sentence: the program delivers a model accompanied by a confidence index (pLDDT), not an experimental measurement.

Discovery

Technical explanation

1. Multiple sequence alignment (MSA) and co-evolutionary extraction. Starting from the target sequence, AF2 searches for homologous sequences in the UniRef90 and MGnify databases (~250 million sequences). Positions that co-evolve together (correlated columns in the MSA) indicate 3D contacts: if residues i and j always mutate as a pair, they are probably close in space (< 8 Å). A typical MSA contains 1,000 to 100,000 aligned sequences.

2. Evoformer: axial attention over the MSA and the pairs. The Evoformer alternates between (a) attention over the rows of the MSA (row attention), capturing relations between residues of a single sequence, and (b) attention over the columns (column attention), capturing co-evolution across sequences. The triangle updates and triangle attention enforce geometric consistency: if $d(i,j)$d(i,j) and $d(j,k)$d(j,k) are known, $d(i,k)$d(i,k) is constrained by the triangle inequality. 48 Evoformer blocks, each with 256 + 128 channels.

3. Structure Module: Invariant Point Attention (IPA). The structure module converts abstract representations into 3D coordinates. IPA operates in the space of frames (local SE(3) reference frames) attached to each residue — a local backbone defined by N, Cα, C. IPA attention is invariant under rotation and translation: the predicted structure does not depend on the initial orientation. 8 blocks with recycling (3 complete iterations of the network).

4. Loss functions and FAPE. The main loss is the Frame Aligned Point Error (FAPE): for each pair of residues, the error is measured in the local frame of one of the two residues. This ensures that local distances (bonds, angles) and global distances (tertiary contacts) are optimised simultaneously. Training combines this main loss with several auxiliary losses:

Why it worked

Two architectural innovations were decisive. First, the triangle updates enforce 3D geometric consistency directly in the space of representations — earlier methods (trRosetta, RaptorX) predicted 2D distance maps and then reconstructed the structure separately, losing information. Second, IPA makes it possible to reason in local frames invariant under SE(3), removing the need for rotational data augmentation.

Learning through recycling (3 complete passes of the network) allows iterative self-correction: errors from the first pass are detected and corrected by subsequent passes, mimicking refinement in crystallography.

Causal chain

Levinthal's paradox (1969, 10³⁰⁰ possible conformations) → Protein Data Bank created (1971) → CASP launched (1994, biennial benchmark) → use of MSAs to predict contacts (Marks et al., 2011) → deep learning on distance maps (trRosetta, 2020) → AlphaFold 2 (2020, GDT 92.4, Evoformer + IPA) → release of 200M structures (2022, AlphaFold DB) → AlphaFold 3 (2024, diffusion model, multi-target) → Nobel Prize in Chemistry 2024 (Hassabis, Jumper)

Anecdote

When the CASP14 results were announced on 30 November 2020, John Moult — the founder of CASP — commented: "This is a big deal, in some sense the problem is solved." Andrei Lupas, one of the assessors, told Nature: "This will change everything", and specified that AlphaFold's models had allowed his laboratory to solve a structure it had been stuck on for nearly ten years. The bioinformatics community, accustomed to incremental progress of a few GDT points per CASP cycle, saw an unprecedented leap in a single competition.

Legacy and current data

Since 2022, the AlphaFold DB database has been distributing free of charge the predicted structures of 200 million proteins. The approach was extended by AlphaFold 3 (2024), built on a diffusion model and broadened to multiple targets. In 2024, the Nobel Prize in Chemistry honoured Hassabis and Jumper.

Sources

References verified during the factual audit of August 2026: these are the pages
against which the claims in this bulletin were checked.

1. AlphaFold: a solution to a 50-year-old grand challenge in biology — Google DeepMind
2. Official press release of the CASP14 results — Prediction Center
3. Highly accurate protein structure prediction with AlphaFold — Nature 596, 583-589