AlphaFold 2

Overview

AlphaFold 2 is a deep learning system developed by Google DeepMind that predicts three-dimensional protein structures from amino acid sequences with near-experimental accuracy. Its release in 2021 represented a decisive solution to the protein folding problem — a challenge that had stood at the center of structural biology for over five decades. At the 14th Critical Assessment of Protein Structure Prediction (CASP14), AlphaFold 2 achieved a median Global Distance Test (GDT) score of 92.4 across all targets, a performance level that assessors described as comparable to low-resolution experimental methods, far exceeding all prior computational approaches.

The model introduced several architectural innovations that jointly account for its accuracy. Rather than treating folding as a sequential prediction task, AlphaFold 2 frames the problem geometrically: it learns representations of pairwise residue relationships and iteratively refines both sequence and structure representations through a coupled attention mechanism. This allowed the system to capture the long-range dependencies that govern tertiary structure in a way that earlier energy-minimization or homology-based methods could not.

AlphaFold 2 was developed under the leadership of John Jumper and Demis Hassabis and published in Nature in July 2021. The accompanying release of the AlphaFold Protein Structure Database, in partnership with EMBL's European Bioinformatics Institute (EMBL-EBI), made over 200 million predicted structures freely available, covering virtually every protein in UniProt.

Key Features

• Atomic-level accuracy: Predicts backbone and side-chain coordinates with confidence levels competitive with X-ray crystallography and cryo-EM for many protein families, as measured by GDT and Local Distance Difference Test (lDDT) scores.
• Evoformer architecture: A dedicated transformer-based module processes multiple sequence alignments (MSAs) and pairwise residue representations jointly, enabling the model to extract co-evolutionary signals that encode structural constraints.
• Structure module with invariant point attention: A geometric deep learning component operates directly in 3D space using rigid-body frames, enforcing physical plausibility of the predicted backbone without explicit energy functions.
• Per-residue confidence (pLDDT): Each prediction includes a per-residue confidence score scaled 0–100, allowing users to identify well-folded regions versus disordered or uncertain segments.
• Recycling inference: Predictions are iteratively refined by feeding the output back as input for multiple rounds, progressively improving accuracy without retraining.
• Open access: The model weights, code, and a precomputed database of 200M+ predictions are freely available for non-commercial research use.

Technical Details

AlphaFold 2 is a 93-million parameter neural network built around two core components: the Evoformer stack and the Structure Module. The Evoformer consists of 48 transformer-like blocks that jointly update an MSA representation (capturing evolutionary information across hundreds of related sequences) and a pairwise residue representation (capturing spatial relationships). Attention operations over both axes allow long-range residue interactions to propagate through the network.

The Structure Module takes the resulting embeddings and uses Invariant Point Attention (IPA) — an attention mechanism that is equivariant to rotations and translations — to place each residue in 3D space as a rigid frame. Side-chain torsion angles are predicted separately. Training used structures from the Protein Data Bank (PDB; approximately 170,000 structures as of training time), supplemented by self-distillation on the 350,000+ sequences in UniRef90 with no known experimental structure. Inference on a single protein typically completes in minutes on a single GPU, compared to weeks or months for traditional experimental structure determination.

Applications

AlphaFold 2 is used across virtually every domain of protein science. Drug discovery teams use predicted structures as starting points for virtual screening and structure-based drug design, particularly for targets without experimental structures. Structural biologists use it to aid molecular replacement in X-ray crystallography and to interpret cryo-EM density maps. Evolutionary biologists use the database to compare structural conservation across distant homologs. Protein engineers use pLDDT scores to identify structured versus disordered regions before designing mutations or fusion constructs. Academic researchers use the database directly as a reference for hypothesis generation, reducing the need for computationally expensive experimental characterization early in a project.

Impact

AlphaFold 2 is widely regarded as one of the most significant advances in computational biology of the past decade, and its 2021 Nature paper has accumulated tens of thousands of citations. John Jumper and Demis Hassabis were awarded the 2024 Nobel Prize in Chemistry in recognition of the work. The AlphaFold Protein Structure Database surpassed 1 million users within its first year and now hosts predictions for over 200 million proteins. The model's release accelerated downstream research into protein design, drug target identification, and structural genomics, and directly inspired a new generation of structure prediction systems including ESMFold, RoseTTAFold, and Boltz-1. A notable limitation is that AlphaFold 2 predicts static monomeric structures and does not natively model conformational flexibility, protein-ligand interactions, or large multi-chain assemblies — limitations partially addressed by the subsequent AlphaFold-Multimer and AlphaFold 3 systems.