AlphaFold

1 Overview and History

AlphaFold is an artificial intelligence system developed by DeepMind, a subsidiary of Alphabet, that predicts the three‑dimensional structure of proteins from their amino acid sequences. First unveiled in 2018 and dramatically improved in 2020, the system uses deep learning techniques—especially attention‑based neural networks—to achieve near‑experimental accuracy. It emerged as the top performer in the Critical Assessment of Structure Prediction (CASP) competitions and has been hailed as a major breakthrough in computational biology.

1.1 The protein folding problem

Proteins are linear chains of amino acids that fold into specific three‑dimensional shapes essential for their function. The “protein folding problem” refers to the challenge of predicting a protein’s native structure solely from its sequence. This problem has been a central goal of molecular biology for decades, because structure determines function and enables rational drug design, enzyme engineering, and understanding of disease.

1.2 Early computational approaches (homology modeling, ab initio)

Before deep learning, computational methods fell into two categories: homology modeling, which builds a structure based on known templates from similar sequences; and ab initio (or de novo) prediction, which attempts to fold a protein from physical principles alone. Homology modeling succeeded only when a close relative had been solved experimentally. Ab initio methods were computationally expensive and often inaccurate for larger proteins, limiting their practical use.

1.3 DeepMind’s entry into protein structure prediction

DeepMind, best known for its game‑playing AI, applied deep learning to the protein folding problem in 2017. The company recognized that large‑scale sequence databases and advances in neural network architectures could capture evolutionary constraints and spatial relationships between amino acids.

1.3.1 CASP13 (2018): AlphaFold 1

At the 13th Critical Assessment of Structure Prediction (CASP13) in 2018, AlphaFold 1 debuted and achieved the highest accuracy among all participants. It used a combination of convolutional neural networks and distance‑based potentials to predict inter‑residue distances and torsion angles, outperforming traditional methods by a clear margin.

1.3.2 CASP14 (2020): AlphaFold 2

AlphaFold 2, entered in CASP14 in 2020, represented a leap in performance. It introduced an end‑to‑end architecture that directly predicted atomic coordinates, achieving a median backbone accuracy of roughly 0.96 Å root‑mean‑square deviation (RMSD) for single‑domain proteins—comparable to experimental structures. The system correctly folded the majority of targets, solving the protein folding problem for many practical purposes.

1.4 Public release and AlphaFold Protein Structure Database

In July 2021, DeepMind released the AlphaFold 2 source code and a database of over 350,000 predicted structures, covering nearly all proteins from the human proteome and 20 other model organisms. This database, hosted by the European Bioinformatics Institute (EMBL‑EBI), continues to expand and now includes hundreds of millions of structures from across the tree of life.

2 Technical Architecture

2.1 Input features: multiple sequence alignment (MSA) and templates

AlphaFold takes as input the amino acid sequence and a multiple sequence alignment (MSA) built by searching sequence databases for homologous proteins. The MSA captures evolutionary covariation, which indicates residues that are in physical contact. Additionally, the system can use experimental structures of related proteins (templates) as starting hints, but AlphaFold 2 learned to rely less on templates than earlier methods.

2.2 Neural network design

The core of AlphaFold 2 is a deep neural network with two main components: the Evoformer and the structure module.

2.2.1 Evoformer (pairwise representation)

The Evoformer is a transformer‑based architecture that processes the MSA and a pairwise representation of residue‑residue relationships. It uses axial attention and triangular multiplicative updates to refine a rich representation of the protein’s geometry and evolutionary context, effectively “reasoning” about how residues interact.

2.2.2 Structure module (iterative refinement)

After the Evoformer, the structure module converts the pairwise representation into explicit three‑dimensional coordinates. It applies iterative refinement using a variant of the SE(3) equivariant transformer, ensuring that the predicted structure rotates and translates correctly as a rigid body. This module produces the final atomic positions for the backbone and side chains.

2.2.3 Attention mechanisms and equivariant representations

AlphaFold 2 relies on several specialized attention mechanisms. Pairwise attention captures interactions between residues, while MSA attention integrates information across sequences. Equivariant representations guarantee that the network’s predictions are physically consistent under rotations and translations, a key innovation that improved accuracy.

2.3 Training data and loss functions

The network was trained on the Protein Data Bank (PDB), which contains tens of thousands of experimentally solved structures. Loss functions included a frame‑aligned point error (FAPE) that penalizes deviations in atomic coordinates, alongside auxiliary losses for distograms and torsion angles. Training required massive computational resources (128 TPUv3 cores for several days).

2.4 Inference pipeline and confidence metrics (pLDDT, PAE)

For inference, AlphaFold 2 runs repeatedly to generate multiple predictions (typically 5 models). It outputs two confidence metrics: the predicted local distance difference test (pLDDT) per residue, which measures local accuracy; and the predicted aligned error (PAE), which indicates relative domain orientation errors. These allow users to assess reliability and identify flexible regions.

3 Performance and Evaluation

3.1 CASP competition results

In CASP14, AlphaFold 2 achieved a median global distance test (GDT) score above 90 for most targets, far surpassing the next best method. For the hardest “free modeling” targets, it reached GDT scores of 80‑90, whereas earlier methods rarely exceeded 40‑50. The competition organizers declared the protein folding problem largely solved for single‑chain proteins.

3.2 Comparison with experimental methods (X‑ray crystallography, cryo‑EM)

AlphaFold 2 predictions are often comparable to medium‑resolution X‑ray crystallography (1.5–2.5 Å). For many targets, the predicted backbone is within experimental error. However, for side‑chain conformations and flexible loops, experimental methods remain more reliable. Cryo‑electron microscopy (cryo‑EM) maps often agree well with AlphaFold models, and the two techniques are increasingly combined.

3.3 Accuracy on different protein classes (globular, membrane, disordered)

Accuracy is highest for compact, globular domains with many homologous sequences. Membrane proteins, which are under‑represented in the PDB, show slightly lower but still useful accuracy. Intrinsically disordered regions are difficult to predict because they lack stable structure; AlphaFold often assigns such regions low pLDDT scores, correctly indicating disorder.

3.4 Strengths and limitations

3.4.1 High accuracy for single‑domain proteins

For proteins with a single, well‑folded domain, AlphaFold 2 achieves near‑experimental accuracy in most cases. This has enabled structural coverage of entire proteomes and revealed folds for previously unsolved families.

3.4.2 Challenges with multi‑domain assemblies and conformational flexibility

AlphaFold can struggle with large multi‑domain proteins, especially when domain orientations are not conserved across homologs. It also predicts a single static structure and does not directly capture conformational ensembles or functional movements. Predictions for protein complexes require specialized extensions.

4 Applications in Life Sciences

4.1 Structural biology and drug discovery

AlphaFold accelerates the determination of protein structures, providing models that guide experimental crystallography and cryo‑EM. In drug discovery, predicted structures are used for virtual screening, hit identification, and understanding binding sites, though careful validation remains necessary.

4.2 Understanding disease mutations and protein misfolding

By mapping missense mutations onto AlphaFold structures, researchers can assess whether a variant destabilizes a protein or alters a binding interface. This aids interpretation of genome‑wide association studies and helps explain diseases such as Alzheimer’s and cancer.

4.3 Enzyme design and protein engineering

AlphaFold enables the rational design of enzymes with novel activities. Computational redesign of active sites can be tested by predicting the resulting structure, reducing the need for extensive experimental screening.

4.4 Metagenomics and unknown protein function prediction

Thousands of metagenomic sequences code for proteins of unknown function. AlphaFold provides structural predictions that often suggest functional annotations—for example, identifying a fold associated with a specific enzymatic mechanism.

4.5 Impact on synthetic biology and vaccine design

In synthetic biology, AlphaFold helps design protein‑based materials, sensors, and signaling pathways. For vaccine development, predicting the structure of viral surface proteins allows rational antigen design, as demonstrated during the COVID‑19 pandemic.

5 Successors and Related Work

5.1 AlphaFold‑Multimer (protein complex prediction)

Released in 2022, AlphaFold‑Multimer extends the architecture to predict the structures of protein‑protein complexes. It modifies the input to include multiple sequences from different chains and adjusts the training loss to account for inter‑chain contacts. Accuracy for heterodimers is high, though larger complexes remain challenging.

5.2 AlphaFold 3 (2024 update with broader capabilities)

In 2024, DeepMind released AlphaFold 3, which can predict structures of proteins interacting with DNA, RNA, small molecules, and other ligands. It uses a diffusion‑based generative head and integrates interaction confidence metrics for non‑protein components, broadening the system’s applicability to drug design.

5.3 Open‑source implementations (OpenFold)

OpenFold is an open‑source reproduction of AlphaFold 2, developed by academic and industry partners. It replicates the architecture and accuracy of the original, enabling researchers to modify and retrain the model on custom data.

5.4 Alternative approaches (RoseTTAFold, ESMFold)

Other deep learning methods include RoseTTAFold (from the Baker lab), which uses a three‑track neural network, and ESMFold, which replaces MSA with large language model embeddings. These offer different trade‑offs between speed and accuracy and have advanced the field collectively.

6 Data and Database

6.1 AlphaFold Protein Structure Database (hosted by EMBL‑EBI)

The AlphaFold Database is the largest repository of protein structure predictions. It provides freely downloadable PDB and mmCIF files, along with metadata and quality scores. The website allows users to search by accession number, gene name, or taxonomy.

6.2 Coverage: UniProt reference proteomes

The database covers the entire UniProt reference proteome for several hundred species, including all human proteins. As of 2025, it contains over 200 million predicted structures, covering a wide range of bacteria, archaea, eukaryotes, and viruses.

6.3 Data access and API

Users can access the data via a REST API for programmatic download, bulk retrieval, and integration with bioinformatics pipelines. The database is also mirrored through Google Cloud and other partners.

6.4 Quality indicators and metadata

Each entry includes per‑residue pLDDT scores, predicted aligned error (PAE) heatmaps, and packing quality metrics. Metadata fields include source organism, sequence length, and date of prediction, allowing users to filter by confidence.

7 Ethical and Practical Considerations

7.1 Data privacy and responsible use of predicted structures

AlphaFold predicts structures based solely on sequence data and does not reveal sensitive information about individuals. However, predicted structures could be used in bioengineering applications that require oversight. Researchers are encouraged to follow relevant biosafety guidelines.

7.2 Reproducibility and benchmarking

The AlphaFold 2 source code is publicly available, and the training procedure has been documented. Benchmarking against independent datasets (e.g., CAMEO) shows consistent performance. Reproducing full training requires large computational resources, but inference is accessible on typical GPU servers.

7.3 Integration with experimental structural biology

AlphaFold is best used alongside experimental methods. Structures predicted by AI can guide crystallographic phasing, model building into cryo‑EM maps, and NMR assignment. The two approaches are complementary, with experiments providing definitive validation.

7.4 Limitations in predicting dynamic or disordered regions

AlphaFold predicts a single static conformation and does not capture the full ensemble of states that a protein can adopt in solution. Regions with low pLDDT may be intrinsically disordered or subject to conformational change; users should not over‑interpret them. Further development is needed for dynamics and large‑scale flexibility.

8 Future Directions

8.1 Beyond static structures: dynamics, multi‑state predictions

Current methods are limited to static structures. Future versions may incorporate molecular dynamics simulation data or generate multiple plausible conformations to reflect functional motions, such as open‑closed transitions in enzymes.

8.2 Integration with small‑molecule docking and drug screening

Linking AlphaFold predictions with docking algorithms could accelerate virtual screening. Improvements in side‑chain accuracy and ligand‑binding site prediction are needed to reliably identify candidate drugs.

8.3 Whole‑cell structural modeling

An ultimate goal is to model the entire proteome of a cell, including interactions with nucleic acids, lipids, and small metabolites. This would require scaling complex prediction to thousands of proteins and accounting for cellular environment effects.

8.4 Continuous learning and community contributions

AlphaFold 2 was trained once on a fixed dataset. Future systems could incorporate new experimental structures as they are deposited, enabling continuous improvement. Community contributions, such as fine‑tuning on specific families or incorporating new loss terms, will also drive progress.