AlphaFold
AlphaFold is an artificial intelligence (AI) system developed by Google DeepMind that predicts the three-dimensional (3D) structures of proteins from their amino acid sequences with high accuracy.[1] Introduced in 2018 during the Critical Assessment of Structure Prediction (CASP) competition, the initial version of AlphaFold demonstrated significant advances in protein structure prediction using deep learning techniques.[2] AlphaFold 2, unveiled in 2020 and detailed in a 2021 Nature publication, achieved breakthrough performance by predicting protein structures with atomic-level precision, even for proteins without known homologs, effectively addressing a 50-year challenge in biology.[3] In July 2021, DeepMind open-sourced the code for AlphaFold 2, and in 2022, in collaboration with the European Molecular Biology Laboratory (EMBL), they launched the AlphaFold Protein Structure Database containing predicted models for over 200 million protein structures across organisms.[4][5] AlphaFold 3, released on May 8, 2024, expands capabilities to model joint structures and interactions of biomolecular complexes, including proteins with DNA, RNA, ligands, and ions, using a diffusion-based architecture for enhanced accuracy; in November 2024, its code and weights were made available for non-commercial academic research.[6][7] In October 2025, DeepMind and EMBL-EBI renewed their partnership, updating the database to align with UniProt release 2025_03 by adding protein isoforms and downloadable multiple sequence alignments.[8] The development of AlphaFold has profoundly impacted scientific research, accelerating drug discovery, enzyme design, and understanding of biological processes, and was recognized by the 2024 Nobel Prize in Chemistry, awarded to Demis Hassabis and John Jumper of DeepMind (shared with David Baker for related work in computational protein design).[9][10]Background
The Protein Folding Problem
The protein folding problem refers to the challenge of predicting the three-dimensional (3D) structure of a protein from its one-dimensional amino acid sequence, a core unsolved issue in biology since the 1970s. According to Anfinsen's dogma, established through experiments on ribonuclease A in the late 1950s and 1960s, the native structure of a protein is uniquely determined by its amino acid sequence under physiological conditions, as the folding process is thermodynamically driven to minimize free energy.[11] However, Cyrus Levinthal highlighted a paradox in 1969: if a typical protein of 100 amino acids were to randomly sample all possible conformations (3^{100} (approximately 5 \times 10^{47}) possibilities, assuming three states per residue), it would take longer than the age of the universe to find the native state, yet proteins fold in milliseconds to seconds, implying guided pathways rather than exhaustive search.[12][13] Accurate prediction of protein structures is essential for elucidating biological functions, as the 3D arrangement of atoms dictates how proteins interact with other molecules, perform enzymatic reactions, and maintain cellular processes. It also informs evolutionary studies by revealing conserved structural motifs across species and aids in understanding diseases, such as those involving misfolding like Alzheimer's or prion disorders, where aberrant conformations lead to toxic aggregates.[14][12] Furthermore, structure prediction accelerates drug discovery by enabling the design of molecules that target specific protein binding sites.[15] Prior to the widespread adoption of artificial intelligence, efforts to solve the protein folding problem relied on computational methods like homology modeling, which constructs 3D models by aligning the target sequence to experimentally determined structures of homologous proteins in databases such as the Protein Data Bank (PDB).[16] Ab initio methods, in contrast, attempt de novo prediction using physical principles, simulating energy minimization and molecular dynamics to explore conformational space without templates, though they are computationally intensive and limited to small proteins.[17] A notable tool in this era was Rosetta, developed in the 1990s by David Baker's group, which employs fragment assembly and Monte Carlo sampling to generate low-energy decoys, achieving successes in de novo design but struggling with accuracy for larger or novel folds.[18][19] Key metrics for evaluating structure prediction accuracy include the root-mean-square deviation (RMSD), which quantifies the average atomic distance (in angstroms, Å) between superimposed predicted and native structures after optimal alignment, with lower values indicating better agreement (e.g., <2 Å for high-quality models).[20] Another is the Global Distance Test Total Score (GDT-TS), which measures the percentage of residues aligned within distance cutoffs of 1, 2, 4, and 8 Å, scaled from 0 to 100, providing a more robust assessment for partial similarities as it is less sensitive to outliers than RMSD.[21] These metrics have been central to benchmarking progress through initiatives like the Critical Assessment of Structure Prediction (CASP) competitions, held biennially since 1994 to blindly evaluate methods on novel targets.[22]DeepMind's Development
In 2016, shortly after the success of DeepMind's AlphaGo program in mastering the complex game of Go, CEO and co-founder Demis Hassabis initiated a new project to apply artificial intelligence to the longstanding protein folding problem, a fundamental challenge in biology that had eluded scientists for over 50 years.[23][24] This effort marked DeepMind's pivot toward scientific applications of AI, aiming to predict protein structures from amino acid sequences with unprecedented accuracy to accelerate discoveries in medicine and biology.[25] The protein folding team was formed that year as a small, interdisciplinary group, drawing on expertise from computer science, machine learning, and biology; Hassabis led the initiative, recruiting a handful of biologists and consultants such as David Jones to bridge AI techniques with structural biology knowledge.[24][23] Key researchers like John Jumper joined in 2017, bringing a physics and chemistry background to contribute to the project's core development.[26] The team's initial goals were ambitious: to leverage AI's pattern-recognition capabilities, honed in game-playing systems, to tackle the "grand challenge" of protein structure prediction, potentially unlocking insights into diseases and drug design.[24][25] DeepMind provided substantial resources to support the effort, including access to powerful GPU and TPU clusters for training models and integrating advanced machine learning methods such as deep neural networks trained on vast datasets of known protein structures.[3] Early collaborations focused on assembling domain expertise, with the team iterating internally on prototypes that combined evolutionary data analysis and geometric modeling approaches.[23] The project progressed through internal milestones from its 2016 inception, including the development of preliminary systems by 2017 and rigorous testing against experimental benchmarks, culminating in the team's first public demonstration at the CASP13 competition in December 2018.[24][25] This timeline reflected a deliberate build-up, allowing the team to refine their AI-driven methodology before broader exposure.Algorithm and Versions
AlphaFold 1 (2018)
AlphaFold 1, developed by DeepMind and presented at the CASP13 competition in 2018, introduced a pioneering deep learning approach to protein structure prediction by leveraging multiple sequence alignments (MSAs) to infer evolutionary co-evolution signals. The system's architecture centered on deep residual neural networks, which processed input MSAs to extract features indicative of residue-residue interactions. These networks employed dilated convolutions to capture long-range dependencies within the sequence alignments, enabling the model to predict inter-residue geometric relationships directly from evolutionary data.[27] The training dataset comprised experimentally determined protein structures from the Protein Data Bank (PDB) up to March 2018, augmented with evolutionary information derived from MSAs generated using UniRef90 sequences clustered at 30% identity (Uniclust30 release from October 2017). By analyzing deep MSAs, the model identified co-evolution patterns, where correlated mutations between residues suggested spatial proximity in the folded structure, providing a rich source of implicit structural information without relying on explicit templates. This approach allowed AlphaFold 1 to learn directly from the vast evolutionary record, bypassing traditional physics-based simulations.[27] A key innovation of AlphaFold 1 was its end-to-end differentiable neural network model, which predicted 3D structures in a single forward pass, eliminating the need for fragment assembly or sampling methods common in prior techniques. Instead of piecing together short structural fragments, the system directly outputted probabilistic representations of the full protein fold, enabling seamless integration of prediction and refinement stages. This unified framework improved efficiency and accuracy by propagating gradients through the entire pipeline.[27] The model generated outputs in the form of distance maps, representing probability distributions over possible distances between pairs of residues (binned into 64 categories from 2 to 22 Å), alongside predictions for torsion angles that define the protein backbone. These predictions served as constraints for structure generation, where an initial atomic model was refined using gradient-based optimization to minimize violations of the predicted distances and angles. The refinement process employed a differentiable energy function, allowing iterative adjustments via backpropagation to produce a final 3D coordinate model.[27] Central to the training of the distance prediction component was a supervised loss function focused on inter-residue distances, using cross-entropy loss between the predicted distogram and true distance bins, with auxiliary losses for torsion angles and overall structure quality. By optimizing this loss during training, AlphaFold 1 achieved high-fidelity distance predictions that translated to reliable 3D structures.[27] In the CASP13 assessment, AlphaFold 1 demonstrated strong performance, particularly in the free-modeling category for novel folds.[27]AlphaFold 2 (2020)
AlphaFold 2, developed by DeepMind, marked a significant advancement in protein structure prediction by achieving unprecedented accuracy in the CASP14 competition, where it outperformed all other methods with a median GDT_TS score of 92.4.[3] This version shifted the paradigm from predicting inter-residue distances, as in AlphaFold 1, to directly outputting 3D atomic coordinates through an end-to-end differentiable neural network.[3] The core architecture of AlphaFold 2 comprises two primary modules: the Evoformer and the structure module. The Evoformer processes input representations derived from multiple sequence alignments (MSAs) and structural templates, updating single representations for each residue and pairwise representations between residues across 48 stacked blocks.[3] It employs attention mechanisms to capture long-range dependencies, enabling the model to infer evolutionary and spatial relationships among residues. Specifically, the attention operation in the Evoformer uses the scaled dot-product formula: \text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right) V where Q, K, and V are query, key, and value matrices derived from residue features, and d_k is the dimension of the keys, preventing vanishing gradients in high-dimensional spaces.[3] This mechanism allows the model to query interactions between residues, effectively modeling how distant parts of the protein influence each other based on co-evolutionary signals in the MSA. The structure module then takes the Evoformer's output and iteratively refines 3D coordinates using invariant point attention (IPA), which predicts rigid-body transformations for each residue while preserving rotational invariance. A key innovation is the iterative recycling mechanism and the use of invariant point attention (IPA) in the structure module, which refines 3D coordinates over multiple cycles using frame-aligned point error (FAPE) loss.[3] Training of AlphaFold 2 involved self-distillation on structures from the Protein Data Bank (PDB), where synthetic MSAs were generated from predicted structures to expand the training data beyond naturally occurring alignments.[3] The process incorporated template features from homologous structures and employed a recycling mechanism, in which predictions from previous iterations were fed back as additional inputs to refine outputs over multiple cycles, enhancing accuracy for challenging targets.[3] Attention mechanisms throughout the network emphasized long-range dependencies, allowing the model to learn complex folding patterns without relying on traditional physics-based simulations. Unlike AlphaFold 1, which relied on distance map predictions followed by constrained optimization for folding, AlphaFold 2's direct coordinate regression bypasses such post-processing, integrating all aspects of structure prediction into a unified deep learning framework.[3]AlphaFold 3 (2024)
AlphaFold 3, released in 2024, extends biomolecular structure prediction to joint modeling of multi-component complexes, including proteins with DNA, RNA, ligands, ions, and modified residues, surpassing the protein-only focus of prior versions. Developed in collaboration between Google DeepMind and Isomorphic Labs, the model features a revamped architecture centered on the Pairformer module, which processes pairwise representations from multiple sequence alignments and structural templates to capture inter- and intra-molecular relationships. This trunk module employs triangular self-attention mechanisms for geometric consistency across all input molecules, generating refined embeddings that feed into a diffusion-based structure module for atomic coordinate prediction. The diffusion approach enables generative modeling of complex assemblies, starting from randomized atomic positions and iteratively refining them to form accurate 3D configurations. In October 2025, DeepMind and EMBL-EBI renewed their partnership, releasing a major update to the AlphaFold Database to align with UniProt sequences as of 2025.[8][6] The core of the structure generation relies on a denoising diffusion process, where noise is progressively added to ground-truth structures during training and reversed during inference. The forward noise addition step follows the standard diffusion equation: \mathbf{x}_t = \sqrt{\alpha_t} \, \mathbf{x}_0 + \sqrt{1 - \alpha_t} \, \boldsymbol{\epsilon}, with \boldsymbol{\epsilon} \sim \mathcal{N}(\mathbf{0}, \mathbf{I}) representing Gaussian noise, \mathbf{x}_0 the original structure, and \alpha_t a time-dependent variance schedule controlling the noise level at step t. The model, comprising a diffusion Transformer with pair-biased attention and transition blocks, learns to predict the noise at each denoising iteration, converging over multiple steps to output precise all-atom coordinates for the entire complex. This framework supports flexible tokenization for diverse biomolecule types, scaling representations appropriately for backbones, side chains, or small molecules.[6] Training leveraged expanded datasets drawn primarily from the Protein Data Bank, incorporating experimentally determined structures of protein-nucleic acid complexes, protein-ligand bindings, and ion interactions to enhance prediction of binding sites and overall complex stability. Emphasis was placed on learning interaction energies implicitly through structural fidelity, with loss functions optimized for coordinate accuracy across covalent bonds and non-bonded contacts. Innovations include specialized handling of covalent interactions, such as disulfide bridges and ligand conjugations, alongside non-covalent ones like electrostatics and van der Waals forces; improved modeling of modified residues, including post-translational modifications like phosphorylation and glycosylation; and uncertainty quantification via ensemble predictions, where variability across multiple inference runs informs confidence metrics such as interface predicted TM-score (ipTM) and predicted TM-score (pTM). These advances enable robust predictions for previously challenging multi-molecule systems.[6] The model was released with partial open-sourcing through the AlphaFold Server API, offering free access for non-commercial research to predict structures of user-specified biomolecular inputs. Full code and model weights became available to academics via GitHub in November 2024, facilitating broader adoption while supporting Isomorphic Labs' drug discovery applications. AlphaFold 3 achieved superior accuracy on complex benchmarks, including those from CASP15, particularly for protein-protein and protein-ligand interactions.[6][7][28][29]Competitions
CASP13 (2018)
The Critical Assessment of Structure Prediction (CASP13) was the 13th biennial community experiment, held in 2018, designed to evaluate the progress of computational methods for predicting protein structures from amino acid sequences alone, with independent assessors comparing submitted models against experimentally determined structures.[30] This edition featured over 100 participating groups worldwide, focusing on challenging targets without close homologs in existing databases, particularly in the free-modeling (FM) category for novel folds.[27] DeepMind's AlphaFold, in its inaugural competition entry, dominated the FM category, achieving the top ranking across 25 of 43 domains and second place in 11 others, with a median global distance test total score (GDT-TS) of 58.9 on these hard targets—a substantial improvement over prior CASP editions and roughly double the recent rate of progress.[27][31] The system, which relied on multiple sequence alignment-based co-evolutionary analysis to train deep neural networks for predicting inter-residue distance distributions as statistical potentials, enabled accurate modeling without relying on fragment assembly or physics-based simulations typical of traditional approaches.[27] Specific successes included precise predictions of novel protein architectures, such as outer membrane beta-barrels, where AlphaFold's models captured intricate strand topologies that eluded human expert-guided methods.[27] Submissions for CASP13 involved teams uploading up to five ranked models per target via the official prediction center portal, with AlphaFold generating ensembles refined using Rosetta energy minimization for final outputs.[27] Assessors, through visual inspection and quantitative metrics, praised AlphaFold's consistency, noting it provided the visually best or near-best model for nearly every FM target, often outperforming competitors by wide margins in topology and domain packing.[27] This debut marked the first instance where an AI-driven system decisively challenged and surpassed longstanding homology- and physics-based techniques, signaling a paradigm shift in de novo protein structure prediction.[31]CASP14 (2020)
The Critical Assessment of Structure Prediction (CASP) 14 experiment, held in 2020, emphasized blind predictions of novel protein structures that had no close homologs in existing databases, providing a rigorous test of computational methods amid the global COVID-19 pandemic, which underscored the urgency of accurately modeling viral proteins.[32][33] CASP14 was postponed from its original April start due to the pandemic's impact on the research community, beginning instead in late May and running through August, with targets released progressively to simulate real-world discovery scenarios.[32] This timing heightened interest, as the competition included challenging targets like the SARS-CoV-2 ORF8 protein (target T1064), a viral accessory protein implicated in immune evasion, allowing predictors to contribute to pandemic-related structural biology.[3][34] AlphaFold 2, developed by DeepMind, dominated CASP14 by achieving a median Global Distance Test Total Score (GDT-TS) of 92.4 across all structure categories, far surpassing the previous competition's top scores and demonstrating unprecedented accuracy for blind predictions.[3] It provided the best models for 88 out of 97 targets, with many achieving near-atomic resolution, including a median backbone root-mean-square deviation (RMSD) of 0.96 Å from experimental structures.[3][35] This performance was enabled by AlphaFold 2's Evoformer architecture, which processed evolutionary and spatial information to generate highly precise multiple sequence alignments and structure representations.[3] In the free modeling category, where targets lacked templates, AlphaFold 2 topped rankings, solving structures that had eluded traditional methods for decades.[36] Key highlights included AlphaFold 2's accurate prediction of the SARS-CoV-2 ORF8 structure, submitted blindly during the competition and later validated against experimental data, aiding early insights into viral mechanisms.[3][37] The system's rapid turnaround—generating and submitting high-quality models within the tight 2-3 day windows per target—demonstrated its practical utility under blind testing conditions, where no prior knowledge of structures was allowed.[3][38] The structural biology community reacted with astonishment to AlphaFold 2's leap forward, viewing it as a paradigm shift that established AI's supremacy in protein structure prediction and rendered many prior approaches obsolete overnight.[39] Organizers and participants noted its top rankings across all major categories, including human and server predictions, prompting widespread acclaim for solving a 50-year-old grand challenge.[36]CASP15 (2022)
The Critical Assessment of Structure Prediction (CASP15), held in 2022, broadened its scope beyond traditional single-domain protein folding to include new prediction categories for oligomeric structures and protein-ligand complexes (binders), addressing the growing need to model multi-domain assemblies and interaction interfaces in light of AlphaFold2's prior successes.[40] These additions reflected the transition from AlphaFold2's focus on monomeric proteins to enhanced capabilities for multimers, with nearly 100 groups submitting over 53,000 models across five main categories, including human-expert, server-automated, and assembly predictions.[41] Although DeepMind did not submit official entries, AlphaFold2 and its specialized variant, AlphaFold-Multimer, were widely adopted by participants for both protein monomer and complex predictions, forming the backbone of top-performing methods.[42] In the assembly category, which encompassed oligomers and binders, AlphaFold-Multimer-based approaches achieved substantial progress, producing high-quality models (defined by DockQ > 0.23 or equivalent interface metrics) for about 40% of the 37 targets—a fourfold increase from the 8% success rate two years earlier in CASP14.[43] For multi-chain assemblies, default AlphaFold applications yielded average TM-scores of approximately 0.7 on challenging targets, indicating reliable topology and fold capture, while Cα RMSD values below 3 Å were obtained for over 90% of tertiary structure targets overall.[44][42] Despite these advances, notable challenges emerged in modeling dynamic elements, such as flexible loops and conformational changes in binders, where predictions often underperformed compared to rigid core structures.[45] This outcome underscored a field-wide evolution toward prioritizing protein-protein and protein-ligand interactions, influencing subsequent benchmarks by emphasizing multimer accuracy and experimental validation of interfaces.[43]CASP16 (2024)
The Critical Assessment of Structure Prediction (CASP16), held in 2024, expanded further to evaluate predictions of complex biomolecular structures, including proteins with nucleic acids, ligands, and modifications, reflecting ongoing advancements in AI-driven modeling.[46] AlphaFold3 was benchmarked in CASP16 through an automatic server submission (AF3-server) and manual predictions by participating groups, performing comparably to top methods overall.[47] For protein domains, the AF3-server achieved an average GDT-TS of 87.1, outperforming AlphaFold2's 85.5 on first-ranked models. In protein complexes, it attained an average DockQ score of 0.53 (vs. AlphaFold2's 0.46), demonstrating slight improvements, particularly on easier targets. However, performance was lower for nucleic acids (TM-score 0.46) compared to specialized methods and for stoichiometry prediction, correctly identifying it in 34% of top-ranked models across 41 targets. These results highlighted AlphaFold3's strengths in atomic-level precision for proteins and complexes while identifying areas for refinement in non-protein components.[47][48]Release and Accessibility
Open-Sourcing Efforts
DeepMind released the source code for AlphaFold 2 in July 2021 alongside the publication of its core methodology, making the inference pipeline available under the Apache 2.0 license via a GitHub repository.[3][4] This release included pre-trained model weights, scripts for generating multiple sequence alignments (MSAs) using tools like JackHMMER, and the structure prediction module implemented in JAX for efficient computation on GPUs.[4] However, the open-sourcing effort for AlphaFold 2 did not include the full training code or datasets, limiting reproducibility of the model's development process to inference only.[3] For AlphaFold 3, introduced in May 2024, DeepMind initially provided access through a non-commercial web-based API via the AlphaFold Server, which allows users to submit sequences for predictions without local installation.[7] In November 2024, the inference code and pre-trained weights for AlphaFold 3 were made available for non-commercial use on GitHub, again under restrictive terms that prohibit commercial applications and exclude training procedures.[29] These releases emphasized accessibility for academic research while protecting proprietary training details. The open-sourcing of AlphaFold has spurred significant community engagement, with numerous forks and extensions enhancing usability. A prominent example is ColabFold, a user-friendly adaptation that integrates faster MSA search via MMseqs2 and enables predictions directly in Google Colab environments, democratizing access for researchers without high-end hardware.[49][50] Such contributions have facilitated broader adoption and integration of AlphaFold into scientific workflows.AlphaFold Protein Structure Database
The AlphaFold Protein Structure Database was launched on July 22, 2021, through a collaboration between DeepMind and the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI), initially providing predicted 3D structures for approximately 20,000 human proteins along with those from 19 other key organisms, totaling over 350,000 models.[51] In July 2022, the database was expanded to encompass predictions for nearly all cataloged proteins known to science, reaching over 200 million structures across more than 1 million species.[5] This expansion was built using the open-sourced AlphaFold 2 models to generate the predictions.[52] The database hosts high-accuracy predicted protein structures in standard PDB and mmCIF formats, accompanied by per-residue confidence scores known as predicted Local Distance Difference Test (pLDDT) values, which range from 0 to 100 and indicate the reliability of each residue's position.[52] Additional data include predicted aligned error (PAE) matrices for assessing inter-domain flexibility and multiple sequence alignments (MSAs) used in the predictions.[53] Structures are integrated with UniProt, allowing seamless cross-referencing to protein sequences and functional annotations directly from database entries.[54] Access to the database is freely available under a Creative Commons license for both academic and commercial use, with users able to download individual or bulk files via the website alphafold.ebi.ac.uk, which features a search interface for querying by UniProt ID, gene name, or organism.[1] Interactive visualization tools enable on-site exploration of models, while an API supports programmatic batch queries and data retrieval for large-scale analyses.[52] As of October 2025, the database had attracted over three million users, facilitating research on previously understudied "orphan" proteins lacking experimental structures by providing accessible 3D models to guide hypothesis generation and experimental design.[55]Recent Developments (2025)
In October 2025, the AlphaFold Protein Structure Database underwent a major update to version 6, synchronizing it with UniProt release 2025_03 to incorporate the latest protein sequences and reference proteomes.[56] This release added over 65 million new predicted structures, bringing the total to more than 241 million entries, including predictions for 40,054 protein isoforms to better support research on genetic variants.[56] Metadata for existing entries was also enhanced with current UniProt annotations, improving accessibility and relevance for structural biology applications.[56] Coinciding with this update, the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI) and Google DeepMind renewed their partnership on October 7, 2025, committing to ongoing maintenance, expansion, and development of the database to advance global protein science.[8] The agreement emphasizes deeper collaboration to ensure the resource remains a cornerstone for researchers worldwide, building on AlphaFold 3's capabilities for multi-molecule predictions.[8] In July 2025, Google DeepMind announced a one-time funding gift to the Critical Assessment of Structure Prediction (CASP) competition, providing support for approximately 12 months after the U.S. National Institutes of Health (NIH) grant expired in early August.[57] This commitment, the amount of which was not disclosed, aims to sustain CASP's operations, prevent layoffs at the organizing University of California, Davis, and ensure the biennial event's continuity as a benchmark for structure prediction advancements.[57] Throughout 2025, the AlphaFold Server received several enhancements for non-commercial users, including updates to template and multiple sequence alignment files in June and July, improved visualization tools in March, and expanded template customization options.[58] These improvements bolster support for ligand interactions in AlphaFold 3 predictions, enabling more accurate modeling of protein-small molecule complexes for academic research.[58] In November 2025, the AlphaFold Protein Structure Database introduced a custom annotations feature, enabling users to integrate and visualize their own sequence annotations alongside predicted structures, further enhancing the tool's utility for personalized research analyses.[1]Performance and Limitations
Accuracy Benchmarks
AlphaFold's prediction accuracy is assessed using established metrics in structural biology, including the Global Distance Test Total Score (GDT-TS), which measures the fraction of residues under distance cutoffs of 1, 2, 4, and 8 Å between predicted and experimental structures (scaled to 0-100); the Template Modeling score (TM-score), ranging from 0 to 1 with values above 0.5 indicating similar folds; root-mean-square deviation (RMSD) for atomic-level precision; and, for later versions, the predicted TM-score (pTM), an estimate of TM-score confidence. These metrics highlight AlphaFold's progression from topology-level predictions to near-atomic accuracy, particularly in blind tests like CASP competitions and the Continuous Automated Model Evaluation (CAMEO) benchmark.[3] In the CASP13 competition (2018), AlphaFold 1 demonstrated substantial improvement on challenging free-modeling targets, achieving GDT-TS scores exceeding 60 on hard targets where previous methods averaged below 50, marking a doubling of progress in prediction quality compared to prior CASP rounds. This performance positioned AlphaFold 1 as the top-ranked method overall, outperforming competitors like trRosetta, which achieved lower median GDT-TS scores on similar difficult cases. AlphaFold 2, evaluated in CASP14 (2020), revolutionized the field with a median GDT-TS of 92.4 across all targets and an average backbone RMSD of 0.96 Å at 95% coverage, far surpassing the next-best method's 2.83 Å RMSD. On the CAMEO blind benchmark, AlphaFold 2 consistently outperformed template-based baselines, adding substantial accuracy even for targets with limited homology. Compared to experimental methods, AlphaFold 2 predictions aligned with cryo-EM structures within 2 Å RMSD for approximately 70% of tested cases, enabling reliable hypothesis generation.[25][3][59] AlphaFold-Multimer, an extension for complexes tested in CASP15 (2022), achieved an average TM-score of about 0.72 on assembly targets, improving to 0.76 with enhanced sampling, which represented a significant advance over prior multimer predictors like trRosetta. AlphaFold 3 (2024) further enhanced accuracy for biomolecular complexes, including ligands, with pTM scores exceeding 0.8 for over 50% of ligand-bound protein structures in benchmark tests—more than double the success rate of specialized tools like DiffDock. In protein-protein interfaces, AlphaFold 3 delivered median interface RMSD values below 2 Å, outperforming AlphaFold 2 by 50% on average for challenging interactions. In CASP16 (2024), the first blind test for AlphaFold 3, it ranked top overall for protein structure prediction with an average GDT-TS of 87.1 (slightly better than AlphaFold 2's 85.5) and DockQ of 0.53 for multimers (better than ColabFold's 0.46 but comparable to top methods like MassiveFold). However, it ranked lower (30-40%) on hard targets and showed limitations in nucleic acids (TM-score 0.46 for RNA). Ligands were not officially evaluated due to submission issues. These benchmarks underscore AlphaFold's evolution toward comprehensive, high-fidelity predictions across diverse molecular systems.[44][60][6][47]| CASP Edition | AlphaFold Version | Median GDT-TS (All Targets) | Key Competitor (e.g., trRosetta) | Notes |
|---|---|---|---|---|
| CASP13 (2018) | AlphaFold 1 | ~60 (hard targets) | ~50 (hard targets) | Topology-level success on novel folds |
| CASP14 (2020) | AlphaFold 2 | 92.4 | ~70-80 | Near-atomic precision; top-ranked |
| CASP15 (2022) | AlphaFold-Multimer | N/A (TM-score ~0.72 for complexes) | ~0.5-0.6 TM-score | Focus on multimer assemblies |
| CASP16 (2024) | AlphaFold 3 | 87.1 (average, proteins) | MassiveFold (similar DockQ for multimers) | Top-ranked overall; challenges on hard targets, RNA, stoichiometry |