Fact-checked by Grok 2 weeks ago

AlphaFold

AlphaFold is an (AI) system developed by that predicts the three-dimensional () structures of proteins from their sequences with high accuracy. Introduced in during the Critical Assessment of Structure Prediction (CASP) competition, the initial version of AlphaFold demonstrated significant advances in using techniques. AlphaFold 2, unveiled in 2020 and detailed in a 2021 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 . In July 2021, DeepMind open-sourced the code for AlphaFold 2, and in 2022, in collaboration with the (EMBL), they launched the AlphaFold containing predicted models for over 200 million protein structures across organisms. 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. 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. The development of AlphaFold has profoundly impacted scientific research, accelerating , , and understanding of biological processes, and was recognized by the 2024 , awarded to and John Jumper of DeepMind (shared with David Baker for related work in computational ).

Background

The Protein Folding Problem

The protein folding problem refers to the challenge of predicting the three-dimensional () structure of a protein from its one-dimensional sequence, a core unsolved issue in since the 1970s. According to , established through experiments on ribonuclease A in the late and , the native of a protein is uniquely determined by its sequence under physiological conditions, as the folding process is thermodynamically driven to minimize . However, Cyrus Levinthal highlighted a in 1969: if a typical protein of 100 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 to find the native state, yet proteins fold in milliseconds to seconds, implying guided pathways rather than exhaustive search. 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 disorders, where aberrant conformations lead to toxic aggregates. Furthermore, structure prediction accelerates by enabling the design of molecules that target specific protein binding sites. Prior to the widespread adoption of , efforts to solve the problem relied on computational methods like , which constructs 3D models by aligning the target sequence to experimentally determined structures of homologous proteins in databases such as the (PDB). methods, in contrast, attempt prediction using physical principles, simulating energy minimization and to explore conformational space without templates, though they are computationally intensive and limited to small proteins. A notable tool in this era was , developed in the 1990s by David Baker's group, which employs fragment assembly and sampling to generate low-energy decoys, achieving successes in design but struggling with accuracy for larger or novel folds. Key metrics for evaluating structure prediction accuracy include the (RMSD), which quantifies the average atomic distance (in angstroms, ) between superimposed predicted and native structures after optimal , with lower values indicating better agreement (e.g., <2 for high-quality models). 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 , scaled from 0 to 100, providing a more robust assessment for partial similarities as it is less sensitive to outliers than RMSD. 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.

DeepMind's Development

In 2016, shortly after the success of DeepMind's 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. 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. 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 to bridge AI techniques with structural biology knowledge. Key researchers like joined in 2017, bringing a physics and chemistry background to contribute to the project's core development. 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. 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. Early collaborations focused on assembling domain expertise, with the team iterating internally on prototypes that combined evolutionary data analysis and geometric modeling approaches. 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 competition in December 2018. 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. 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. 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. 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. 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. In the CASP13 assessment, AlphaFold 1 demonstrated strong performance, particularly in the free-modeling category for novel folds.

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. 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. 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. 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. 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. Training of AlphaFold 2 involved self-distillation on structures from the , where synthetic MSAs were generated from predicted structures to expand the training data beyond naturally occurring alignments. 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. 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.

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 and , 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 to align with UniProt sequences as of 2025. 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 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. 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. 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.

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. 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. 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. 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. 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. 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 for final outputs. Assessors, through visual inspection and quantitative metrics, praised AlphaFold's consistency, noting it provided the visually best or near-best model for nearly every , often outperforming competitors by wide margins in topology and domain packing. 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.

CASP14 (2020)

The Critical Assessment of Structure Prediction () 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 , which underscored the urgency of accurately modeling viral proteins. 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. This timing heightened interest, as the competition included challenging targets like the ORF8 protein (target T1064), a viral accessory protein implicated in immune evasion, allowing predictors to contribute to pandemic-related structural biology. 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. 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. 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. In the free modeling category, where targets lacked templates, AlphaFold 2 topped rankings, solving structures that had eluded traditional methods for decades. Key highlights included AlphaFold 2's accurate prediction of the ORF8 structure, submitted blindly during the competition and later validated against experimental data, aiding early insights into viral mechanisms. 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. 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. 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.

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 's prior successes. These additions reflected the transition from '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. 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. 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 —a fourfold increase from the 8% success rate two years earlier in CASP14. For multi-chain assemblies, default AlphaFold applications yielded average TM-scores of approximately 0.7 on challenging , indicating reliable and fold capture, while Cα RMSD values below 3 Å were obtained for over 90% of overall. 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. 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.

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. AlphaFold3 was benchmarked in CASP16 through an automatic server submission (AF3-server) and manual predictions by participating groups, performing comparably to top methods overall. 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 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.

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. 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. 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. For AlphaFold 3, introduced in May 2024, DeepMind initially provided access through a non-commercial web-based via the AlphaFold Server, which allows users to submit sequences for predictions without local installation. In November 2024, the inference code and pre-trained weights for AlphaFold 3 were made available for non-commercial use on , again under restrictive terms that prohibit commercial applications and exclude training procedures. 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. 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. 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. This expansion was built using the open-sourced AlphaFold 2 models to generate the predictions. 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. Additional data include predicted aligned error (PAE) matrices for assessing inter-domain flexibility and multiple sequence alignments (MSAs) used in the predictions. Structures are integrated with , allowing seamless cross-referencing to protein sequences and functional annotations directly from database entries. Access to the database is freely available under a 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 . Interactive tools enable on-site exploration of models, while an supports programmatic batch queries and data retrieval for large-scale analyses. As of October 2025, the database had attracted over three million users, facilitating research on previously understudied "" proteins lacking experimental structures by providing accessible 3D models to guide hypothesis generation and experimental design.

Recent Developments (2025)

In October 2025, the AlphaFold Protein Structure Database underwent a major update to version 6, synchronizing it with release 2025_03 to incorporate the latest protein sequences and reference proteomes. 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. Metadata for existing entries was also enhanced with current annotations, improving accessibility and relevance for applications. Coinciding with this update, the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI) and renewed their partnership on October 7, 2025, committing to ongoing maintenance, expansion, and development of the database to advance global protein science. 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. In July 2025, announced a one-time funding gift to the competition, providing support for approximately 12 months after the U.S. (NIH) grant expired in early August. This commitment, the amount of which was not disclosed, aims to sustain CASP's operations, prevent layoffs at the organizing , and ensure the biennial event's continuity as a benchmark for structure prediction advancements. Throughout 2025, the AlphaFold Server received several enhancements for non-commercial users, including updates to template and files in June and July, improved visualization tools in March, and expanded template customization options. These improvements bolster support for interactions in AlphaFold 3 predictions, enabling more accurate modeling of protein-small molecule complexes for academic research. 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.

Performance and Limitations

Accuracy Benchmarks

AlphaFold's prediction accuracy is assessed using established metrics in , 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 (TM-score), ranging from 0 to 1 with values above 0.5 indicating similar folds; (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 competitions and the Continuous Automated Model Evaluation () benchmark. In the CASP13 competition (2018), 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 GDT-TS scores on similar difficult cases. , evaluated in CASP14 (2020), revolutionized the field with a GDT-TS of 92.4 across all targets and an backbone RMSD of 0.96 at 95% coverage, far surpassing the next-best method's 2.83 RMSD. On the 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 RMSD for approximately 70% of tested cases, enabling reliable hypothesis generation. 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.
CASP EditionAlphaFold VersionMedian GDT-TS (All Targets)Key Competitor (e.g., trRosetta)Notes
CASP13 (2018)AlphaFold 1~60 (hard targets)~50 (hard targets)Topology-level success on folds
CASP14 (2020)AlphaFold 292.4~70-80Near-atomic precision; top-ranked
CASP15 (2022)AlphaFold-MultimerN/A (TM-score ~0.72 for complexes)~0.5-0.6 TM-scoreFocus on multimer assemblies
CASP16 (2024)AlphaFold 387.1 (average, proteins)MassiveFold (similar DockQ for multimers)Top-ranked overall; challenges on hard targets, ,

Experimental Validations

Experimental validations of AlphaFold predictions have been conducted through various laboratory techniques, including , cryo-electron microscopy (cryo-EM), () spectroscopy, and functional assays, confirming the reliability of the models in real-world biological contexts. In one notable from 2021, AlphaFold2 predictions facilitated the cryo-EM structure determination of components in bacterial ion channels, such as the proton-activated GLIC from Gloeobacter violaceus, where the resolved experimental structures aligned closely with predictions, achieving an all-atom (RMSD) of approximately 1.5 Å compared to experimental data. This integration of computational predictions with cryo-EM not only accelerated the modeling process but also validated the atomic accuracy of AlphaFold2 for membrane proteins in bacterial systems. High-throughput experimental validations have further substantiated AlphaFold's predictive power. A 2023 Nature Methods study analyzed over 200 experimental structures, finding a median Cα RMSD of 1.0 between AlphaFold2 models and deposited models, with high-confidence regions (pLDDT >90, comprising 86% of residues) showing a median error of 0.6 . For AlphaFold3, validations have extended to biomolecular complexes, particularly drug-target interactions. The 2024 release of AlphaFold3 was benchmarked against a large set of protein-ligand complexes from the , primarily resolved by , where high-confidence predictions achieved a median ligand RMSD of 1.3 , confirming accurate modeling of inhibitor binding poses for several 2024-developed small-molecule inhibitors targeting kinases and proteases. Broader evidence supporting AlphaFold's predictions includes strong correlations with evolutionary conservation patterns and mutagenesis outcomes. AlphaFold models incorporate multiple sequence alignments that reflect evolutionary constraints, leading to predictions that align with conserved structural motifs observed in phylogenetic analyses of protein families. Additionally, saturation mutagenesis experiments guided by AlphaFold2 structures have verified predicted stability changes, with computational ΔΔG values correlating well (R² > 0.7) with experimental measurements for variants in proteins like neurexins, underscoring the models' utility in interpreting evolutionary and functional data.

Known Limitations

AlphaFold produces static structural predictions, limiting its ability to model dynamic aspects such as conformational changes or pathways. The model outputs a single, equilibrium-like conformation rather than ensembles or trajectories, which hinders applications requiring insights into transient states or allosteric effects. In edge cases, AlphaFold exhibits poor performance on (IDPs), where predictions often yield low-confidence scores, such as pLDDT values below 50 for many residues in these regions. This stems from the scarcity of homologous contacts in training data, affecting approximately 20% of cases involving significant disorder. Similarly, the model struggles with very large complexes exceeding 10^6 atoms, such as viral capsids or supramolecular assemblies, due to computational constraints and training data limitations that cap practical predictions at around 1,500-4,000 residues. The reliance on the Protein Data Bank (PDB) for training introduces biases, leading to inaccuracies for non-eukaryotic proteins or those with novel folds underrepresented in the dataset. For instance, prokaryotic or archaeal proteins, which constitute a smaller fraction of PDB structures, often receive lower accuracy because the model extrapolates from predominantly eukaryotic templates. Novel folds absent or rare in the training set similarly challenge the system's prediction capabilities. Specific to AlphaFold 3, predictions show lower confidence for rare or novel ligands, with increased risks of hallucinations or stereochemical errors like violations occurring in about 4.4% of cases. While it supports basic post-translational modifications such as or , more complex or context-dependent modifications remain inadequately modeled, limiting reliability for diverse biomolecular interactions. CASP16 evaluations further highlighted limitations in predicting the of protein complexes, with correct predictions in only 34% of heteromeric targets using the first model, and lower accuracy for structures (average TM-score of 0.46 for single ).

Applications

In Structural Biology Research

AlphaFold has significantly advanced the study of orphan proteins, which are uncharacterized or poorly understood members of the , by providing structural predictions for the entirety of the human comprising approximately 20,000 proteins. These predictions fill critical gaps for thousands of uncharacterized human proteins, many of which lack experimental structures, thereby enabling researchers to generate hypotheses about their roles in cellular processes based on structural similarities to known proteins. For instance, AlphaFold models have illuminated structural features in over half of the "dark" human proteins—those without prior structural —facilitating initial insights into their potential functions and interactions. In , AlphaFold predictions have revealed conserved protein folds across diverse species, allowing researchers to trace deep phylogenetic relationships that sequence-based methods often miss due to evolutionary divergence. By comparing predicted structures from the AlphaFold database, studies have identified ancient structural motifs preserved in proteins from to humans, providing a more robust framework for understanding protein family evolution and divergence. This structural conservation aids in reconstructing phylogenies at the fold level, highlighting how shared three-dimensional architectures underpin functional similarities despite sequence drift. AlphaFold has accelerated functional annotation by linking predicted structures to (GO) terms through structural searches, streamlining the curation of protein function databases. Methods like FASSO utilize AlphaFold models to identify structural orthologs and transfer GO annotations, expanding coverage for under-annotated proteins across and reducing reliance on sequence alone. This approach has proven effective in annotating and microbial proteomes, where experimental data is scarce, thereby enhancing the accuracy and completeness of functional databases. Representative examples include 2022 studies employing AlphaFold2 to elucidate the mechanisms of , such as those in nepoviruses, by modeling their active sites and substrate interactions to reveal catalytic residues and evolutionary adaptations. These structural insights have clarified how protease folds enable polyprotein cleavage essential for , guiding mechanistic hypotheses without initial experimental structures. The AlphaFold serves as a key resource for such investigations, offering accessible models to structural biologists.

In Drug Discovery

AlphaFold has significantly advanced target identification in drug discovery, particularly for neglected diseases. In January 2022, the AlphaFold Protein Structure Database was updated to include nearly 200,000 predicted protein structures from 27 organisms, with 17 of these linked to neglected tropical diseases prioritized by the World Health Organization, such as those causing Chagas disease and leishmaniasis. This expansion, developed in partnership with EMBL-EBI and aligned with WHO priority pathogens, has enabled researchers to explore ~200,000 potential drug targets by providing high-accuracy structural models that were previously unavailable or costly to obtain experimentally. These predictions facilitate the prioritization of viable therapeutic targets, accelerating the identification of novel intervention points for diseases affecting millions in low-resource settings. In virtual screening, AlphaFold structures have streamlined docking simulations, allowing computational evaluation of compound libraries against predicted protein conformations. Case studies demonstrate that integrating AlphaFold3-generated structures into virtual screening workflows can reduce wet-lab validation requirements, with structure-based drug discovery approaches achieving up to 50% reductions in costs and timelines by focusing experimental efforts on high-confidence hits. For instance, during the COVID-19 pandemic, AlphaFold2 models of the SARS-CoV-2 main protease were validated against experimental data, supporting their use in inhibitor design through molecular docking and facilitating the development of candidate antivirals. This has lowered the barrier for early-stage screening, enabling more efficient hit identification and optimization. As of 2025, pharmaceutical firms are building proprietary AI models trained on AlphaFold-derived data to enhance drug discovery, while open-source tools like OpenFold3 enable broader access to advanced predictions of protein-ligand interactions. Representative examples highlight AlphaFold's practical impact in pharma applications. , leveraging AlphaFold3, has applied the model to predict small-molecule binders for proteins like TIM-3, identifying novel binding pockets that align closely with experimental structures and expediting rational for targets. AlphaFold3's prediction capabilities have been instrumental in these efforts, offering accurate insights into protein-small molecule interactions. Additionally, industry partnerships, such as those between and established in 2024, utilize AlphaFold technologies to accelerate lead optimization across multiple therapeutic areas, potentially shortening development cycles from years to months by integrating AI-driven structural predictions into proprietary pipelines. These collaborations underscore AlphaFold's role in transforming pharmaceutical R&D from empirical to predictive paradigms.

In Other Scientific Fields

AlphaFold has facilitated advancements in materials design by enabling the prediction of enzyme structures critical for biocatalysts in applications. For instance, researchers utilized AlphaFold to model the structures of polyester hydrolases, accelerating the discovery and engineering of enzymes capable of degrading (PET) plastics at industrial scales. This approach has led to variants that operate efficiently under ambient conditions, supporting sustainable recycling processes and reducing plastic waste accumulation. In , AlphaFold supports the engineering of novel protein folds, particularly for developing biosensors with enhanced stability. By predicting how designed sequences fold into target architectures, the tool allows optimization of protein backbones to withstand diverse environmental stresses, enabling the creation of responsive biosensors for detecting specific molecular signals. This has streamlined the design cycle, reducing reliance on trial-and-error while ensuring functional reliability in engineered systems. AlphaFold's structural predictions have also extended to , aiding the analysis of microbial proteins involved in . For example, it has been applied to model the domains of reductive dehalogenases, enzymes from pollutant-degrading that selectively interact with halogenated contaminants like chlorinated solvents. These insights guide the modification of such proteins to improve affinity and efficiency in breaking down environmental toxins, promoting cleaner remediation strategies. Notable 2023 studies demonstrated the integration of AlphaFold with -based tools for , yielding stable novel folds applicable across these fields. One effort fine-tuned RoseTTAFold—a network akin to AlphaFold—for generative modeling, producing diverse protein structures that informed biocatalyst and prototypes. Another combined AlphaFold predictions with Rosetta energy functions to refine binder designs, achieving high success rates in experimental validation for custom protein scaffolds. These works underscore AlphaFold's role in bridging computational design with practical interdisciplinary outcomes, though challenges persist in regions of intrinsic disorder that may impact certain engineered constructs.

Impact and Reception

Scientific and Community Reception

The performance of AlphaFold 2 in the 2020 Critical Assessment of Structure Prediction (CASP14) elicited widespread acclaim within the community, marking it as a transformative breakthrough in . Assessors and experts described the results as revolutionary, with evolutionary biologist Andrei Lupas, an official evaluator, stating, "The game has changed," highlighting how AlphaFold's unprecedented accuracy shifted paradigms in the field. This competition success served as a key catalyst, propelling immediate interest and validation from peers who had long viewed as a grand challenge. Following its CASP14 triumph, AlphaFold saw explosive adoption, evidenced by the seminal 2021 paper amassing over 10,000 citations by 2023, underscoring its influence across disciplines. Integration into established software tools further accelerated its uptake; for instance, ChimeraX incorporated AlphaFold predictions and prediction capabilities, enabling seamless visualization and analysis for researchers worldwide. However, alongside the praise, scholarly critiques arose concerning the potential overhyping of AlphaFold's scope, particularly its emphasis on static conformations over protein essential for . A EMBO Reports cautioned that AlphaFold's predictions yield only a single structural state, even when input data suggest multiple conformations or flexible regions, thus limiting insights into dynamic behaviors critical for biological processes. AlphaFold's open-source availability and the public database of over 200 million predicted structures have profoundly democratized protein , empowering low-resource laboratories in developing countries to tackle region-specific health challenges without reliance on costly experimental . This has fostered inclusive scientific progress, allowing diverse global teams to leverage high-accuracy predictions for innovative studies previously inaccessible due to resource constraints. In October 2025, EMBL-EBI and renewed their partnership, releasing a major update to the AlphaFold to align it more closely with , further enhancing its utility and global research impact.

Awards and Recognition

In October 2024, the was awarded to David Baker of the for computational , with the other half jointly to and John Jumper of for their development of AlphaFold, which enabled highly accurate predictions of . This recognition highlighted AlphaFold's transformative role in solving the long-standing problem, building on enthusiastic scientific reception of its 2020 performance and 2021 public release. Earlier accolades included the 2022 , a $3 million award shared by the DeepMind team behind AlphaFold for advancing through . In 2023, Hassabis and Jumper received the Albert Lasker Basic Medical Research Award for inventing AlphaFold, praised for revolutionizing by predicting nearly all known protein structures. Demis Hassabis, a key figure in AlphaFold's development, was elected a Fellow of the Royal Society in 2018 and later knighted in 2024 for services to artificial intelligence, reflecting broader institutional honors tied to his contributions including AlphaFold. In July 2025, Google DeepMind provided interim funding to sustain the Critical Assessment of Structure Prediction (CASP) competition, the benchmark where AlphaFold first demonstrated superiority, as U.S. National Institutes of Health support lapsed; this move underscored the model's enduring legacy in the field. AlphaFold's impact extended to public recognition, such as its inclusion in TIME magazine's 2022 list of the 200 Best Inventions for predicting structures of 200 million proteins and making the database freely available to accelerate global research.

References

  1. [1]
    AlphaFold Protein Structure Database
    AlphaFold is an AI system developed by Google DeepMind that predicts a protein's 3D structure from its amino acid sequence. It regularly achieves accuracy ...AlphaFold · Downloads · View protein · FAQs
  2. [2]
    Highly accurate protein structure prediction with AlphaFold - Nature
    Jul 15, 2021 · Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure ...
  3. [3]
    google-deepmind/alphafold: Open source code for ... - GitHub
    This is not an officially supported Google product. Copyright 2022 DeepMind Technologies Limited. AlphaFold 2 and its output are for theoretical modeling only.
  4. [4]
    AlphaFold reveals the structure of the protein universe
    Jul 28, 2022 · The AlphaFold DB serves as a 'google search' for protein structures, providing researchers with instant access to predicted models of the proteins they're ...
  5. [5]
    Accurate structure prediction of biomolecular interactions ... - Nature
    May 8, 2024 · Here we describe our AlphaFold 3 model with a substantially updated diffusion-based architecture that is capable of predicting the joint structure of complexes.Nobel Prize in Chemistry 2024 · Nature Machine Intelligence
  6. [6]
    Demis Hassabis & John Jumper awarded Nobel Prize in Chemistry
    Oct 9, 2024 · John Jumper were co-awarded the 2024 Nobel Prize in Chemistry for their work developing AlphaFold, a groundbreaking AI system that predicts ...
  7. [7]
    Press release: The Nobel Prize in Chemistry 2024 - NobelPrize.org
    Oct 9, 2024 · The Royal Swedish Academy of Sciences has decided to award the Nobel Prize in Chemistry 2024 with one half to David Baker.Popular information · 2024 · Nobelpriset i kemi 2024
  8. [8]
    The Anfinsen Dogma: Intriguing Details Sixty-Five Years Later - NIH
    Jul 14, 2022 · The pioneering experiments of Anfinsen on the oxidative folding of RNase have been revisited discovering some details, which update the statement of his dogma.
  9. [9]
    Levinthal's Paradox Solution & Protein Folding Theory
    Feb 6, 2020 · We will discuss in this review the key ideas and discoveries leading to the current understanding of folding kinetics, including the solution of Levinthal's ...Missing: dogma | Show results with:dogma
  10. [10]
    Protein Misfolding and Degenerative Diseases - Nature
    In 1968, Levinthal proposed that a protein folds rapidly because its constituent amino acids interact locally, thus limiting the conformational space that the ...The Energetic Funnel · Protein Conformation And The... · Misfolding In...
  11. [11]
    The Protein Folding Problem - PMC - PubMed Central
    The protein folding problem is the question of how a protein's amino acid sequence dictates its three-dimensional atomic structure.Missing: dogma | Show results with:dogma
  12. [12]
    Protein Structure Prediction: Challenges & Advances
    In 1969, Levinthal posed a paradox: a protein might have an astronomical number of possible structural conformations due to its massive number of degrees of ...Homology Modeling Methods · Figure 2 · Physicists' Research...
  13. [13]
    Homology modeling in the time of collective and artificial intelligence
    Homology modeling is a method for building protein 3D structures using protein primary sequence and utilizing prior knowledge gained from structural ...
  14. [14]
    AI-Driven Deep Learning Techniques in Protein Structure Prediction
    The three main methods for protein structure prediction [20] are homology modeling [22], protein threading [23], and ab initio modeling [24].
  15. [15]
    The Rosetta all-atom energy function for macromolecular modeling ...
    Rosetta traditionally models the solvent surrounding the protein using the Lazaridis-Karplus (LK) model, which assumes a solvent environment made of pure water.Missing: attempts | Show results with:attempts
  16. [16]
    The trRosetta server for fast and accurate protein structure prediction
    Nov 10, 2021 · The trRosetta (transform-restrained Rosetta) server is a web-based platform for fast and accurate protein structure prediction, powered by deep learning and ...
  17. [17]
    Methods of protein structure comparison - PMC - NIH
    Unlike the global backbone RMSD, the GDT measure recognizes structural similarity very well for the absolute majority of experimental pairs (GDT-TS > 50%).
  18. [18]
    Finding Nearly Optimal GDT Scores - PMC - NIH
    Global Distance Test (GDT) is one of the commonly accepted measures to assess the quality of predicted protein structures. Given a set of distance thresholds, ...
  19. [19]
    Protein Structure Prediction Center
    CASP aims to establish the current state of the art in protein structure prediction and identify progress, using blind prediction to test methods.CASP15 · CASP13 · CASP16 · CASP_Commons
  20. [20]
    How AI Revolutionized Protein Science, but Didn't End It
    Jun 26, 2024 · In 2016, when Jones started working as a consultant for Google DeepMind on a project that would later be known as AlphaFold, John Jumper was ...Part 1 · Part 2 · Part 3
  21. [21]
    This is the reason Demis Hassabis started DeepMind
    Feb 23, 2022 · Hassabis wants researchers to break AlphaFold—to find what doesn't work—and share the results with his team so that they can make the next ...
  22. [22]
    AlphaFold: a solution to a 50-year-old grand challenge in biology
    Nov 30, 2020 · In July 2022, we released AlphaFold protein structure predictions for nearly all catalogued proteins known to science. Read the latest blog ...
  23. [23]
    AlphaFold developers Demis Hassabis and John Jumper share the ...
    Sep 21, 2023 · The 2023 Albert Lasker Basic Medical Research Award has been given to Demis Hassabis and John Jumper of DeepMind for the invention of AlphaFold.<|control11|><|separator|>
  24. [24]
    Improved protein structure prediction using potentials from deep ...
    Jan 15, 2020 · Improved protein structure prediction using potentials from deep learning. Andrew W. Senior,; Richard Evans,; John Jumper ...Missing: DeepMind | Show results with:DeepMind
  25. [25]
    AlphaFold 3 predicts the structure and interactions of all of life's ...
    May 8, 2024 · Our new AI model AlphaFold 3 can predict the structure and interactions of all life's molecules with unprecedented accuracy.
  26. [26]
    AlphaFold 3 predicts the structure and interactions of all of life's ...
    May 8, 2024 · In a paper published in Nature, we introduce AlphaFold 3, a revolutionary model that can predict the structure and interactions of all ...
  27. [27]
    google-deepmind/alphafold3: AlphaFold 3 inference pipeline. - GitHub
    AlphaFold 3's release was made possible by the invaluable contributions of the following people: ... NT: 2023_02_23 (modified) See the Supplementary Information ...Issues 19 · Pull requests 1 · Actions · Security
  28. [28]
    Home - CASP13 - Protein Structure Prediction Center
    CASP13 is a community experiment to assess protein structure modeling methods, where participants submit models and independent assessors compare them with ...
  29. [29]
    AlphaFold at CASP13 - PMC - NIH
    DeepMind's entry, AlphaFold, placed first in the Free Modeling (FM) category, which assesses methods on their ability to predict novel protein folds.
  30. [30]
    Home - CASP14 - Protein Structure Prediction Center
    CASP (Critical Assessment of Structure Prediction) is a community wide experiment to determine and advance the state of the art in modeling protein structure ...
  31. [31]
    Critical Assessment of Methods of Protein Structure Prediction (CASP)
    The CASP14 conference, usually a very intense in-person event, was held virtually. The CASP community also responded to the emergency by working together to ...Missing: context | Show results with:context
  32. [32]
    AlphaFold2 examples from CASP 14 - Proteopedia, life in 3D
    Sep 29, 2023 · Among the 14 FM targets, the second-longest has 276 residues, the median 132, and the shortest, 92. SARS-CoV-2 ORF8. Our first example is SARS- ...
  33. [33]
    AlphaFold2 @ CASP14: “It feels like one's child has left home.”
    Dec 8, 2020 · This time, AF2 did best for 88 out of 97 targets! Below is the comparison between the 2nd and 3rd best methods to illustrate this point.
  34. [34]
    CASP14 press release - Protein Structure Prediction Center
    *AlphaFold produced models for about two-thirds of the CASP14 target proteins with global distance test scores above 90 out of 100. Above the 90-score ...
  35. [35]
    CASP14: what Google DeepMind's AlphaFold 2 really achieved ...
    Dec 3, 2020 · The first target I am going to look at the ORF8 protein, a viral protein involved with the interaction between SARS-CoV-2 and the immune ...
  36. [36]
    Applying and improving AlphaFold at CASP14 - Wiley Online Library
    Oct 2, 2021 · In this paper, we describe the entry from team AlphaFold2 to the “human” category in the 14th Critical Assessment of Protein Structure ...
  37. [37]
    The breakthrough in protein structure prediction | Biochemical Journal
    May 24, 2021 · The structure prediction problem for single protein chains has been solved. Here, we will review the path to CASP14, outline the method employed by AlphaFold2.<|control11|><|separator|>
  38. [38]
    New prediction categories in CASP15 - PMC - NIH
    In CASP15, four new prediction categories were introduced: RNA structure, ligand‐protein complexes, accuracy of oligomeric structures and their interfaces.Missing: binders | Show results with:binders
  39. [39]
    Home - CASP15 - Protein Structure Prediction Center
    CASP15 is a community experiment to assess protein structure modeling methods, where participants submit models of unknown structures.Rankings: RNA prediction · Target List · Zscores · Domain Definition
  40. [40]
    [PDF] Critical assessment of methods of protein structure prediction (CASP)
    The Cα RMSD is below 3 Å for over 90% of CASP15 targets, and less than 1 Å for 40% of the targets. These are impressive results, even though only modest ...
  41. [41]
    Impact of AlphaFold on structure prediction of protein complexes
    We present the results for CAPRI Round 54, the 5th joint CASP-CAPRI protein assembly prediction challenge. The Round offered 37 targets.
  42. [42]
    Progress at protein structure prediction, as seen in CASP15
    DeepMind presented AlphaFold v2.0 at the CASP14. The results of AlphaFold were impressive. The average GDTts was close to 90 for individual domains, i.e. it was ...Missing: median | Show results with:median
  43. [43]
    Tertiary structure assessment at CASP15 - PMC - PubMed Central
    Sep 25, 2023 · The results of tertiary structure assessment at CASP15 are reported. For the first time, recognizing the outstanding performance of AlphaFold 2 (AF2) at CASP14.
  44. [44]
    ColabFold: making protein folding accessible to all | Nature Methods
    May 30, 2022 · ColabFold offers accelerated prediction of protein structures and complexes by combining the fast homology search of MMseqs2 with AlphaFold2 or RoseTTAFold.
  45. [45]
    sokrypton/ColabFold: Making Protein folding accessible to all! - GitHub
    It supports wide range of operating systems, such as Windows 10 or later (using Windows Subsystem for Linux 2), macOS, and Linux. Is there a way to amber-relax ...Wiki · Issues 372 · Pull requests 10 · ActionsMissing: extension | Show results with:extension
  46. [46]
    DeepMind and EMBL release the most complete database of ...
    Jul 22, 2021 · DeepMind and EMBL release the most complete database of predicted 3D structures of human proteins.
  47. [47]
    AlphaFold Protein Structure Database: massively expanding the ...
    Nov 17, 2021 · Powered by AlphaFold v2.0 of DeepMind, it has enabled an unprecedented expansion of the structural coverage of the known protein-sequence space.
  48. [48]
    AlphaFold Protein Structure Database in 2024 - Oxford Academic
    Nov 2, 2023 · Data archiving for AlphaFold DB began with an initial release in July 2021, housing over 360 000 structures for 20 model organism proteomes with ...
  49. [49]
    UniProt: the Universal Protein Knowledgebase in 2023
    Nov 21, 2022 · This interface includes access to AlphaFold structures for more than 85% of all entries as well as improved visualisations for subcellular ...
  50. [50]
    Case study: AlphaFold uses open data and AI to discover the 3D ...
    Feb 9, 2023 · AlphaFold database in numbers. 200 million protein structure predictions; 1 million organisms; 2 million users in 190 countries; AlphaFold ...
  51. [51]
    AlphaFold Database release notes | Protein Data Bank in Europe
    ### Summary of AlphaFold DB v6 Update (October 2025)
  52. [52]
    EMBL-EBI and Google DeepMind renew partnership and release ...
    Oct 7, 2025 · To mark the partnership renewal, the AlphaFold Database (AFDB) has been synchronised with the UniProt protein sequence Knowledgebase (release ...
  53. [53]
    Google DeepMind to fund CASP, as NIH funding runs out | STAT
    Jul 21, 2025 · Protein structure prediction contest, CASP, gets temporary funding from Google DeepMind as NIH grant runs out.
  54. [54]
    Release Updates - AlphaFold Server
    Jul 5, 2025 · Recent updates include template files (2025.07.05), MSA files (2025.06.10), improved visualization (2025.03.18), and template customization ( ...
  55. [55]
    Highly accurate protein structure prediction for the human proteome
    Jul 22, 2021 · Here we markedly expand the structural coverage of the proteome by applying the state-of-the-art machine learning method, AlphaFold, at a scale ...
  56. [56]
    Enhancing AlphaFold-Multimer-based Protein Complex Structure ...
    May 18, 2023 · The average TM-score of the first models predicted by MULTICOM_qa for CASP15 assembly targets is ∼0.76, 5.3% higher than ∼0.72 of the standard ...
  57. [57]
    AlphaFold two years on: Validation and impact - PNAS
    Here, we discuss some of the latest work based on AlphaFold, with a particular focus on its use within the structural biology community.
  58. [58]
    Dynamic closed states of a ligand-gated ion channel captured by ...
    Jul 1, 2021 · We report cryo-electron microscopy (cryo-EM) structures of the proton-activated Gloeobacter violaceus ligand-gated ion channel (GLIC) under three pH conditions.<|control11|><|separator|>
  59. [59]
    AlphaFold predictions are valuable hypotheses and accelerate but ...
    Nov 30, 2023 · Protein structure predictions using AlphaFold, RoseTTAFold and related methods are far more accurate than previous generations of prediction ...
  60. [60]
    Using AlphaFold to predict the impact of single mutations on protein ...
    Mar 16, 2023 · Here we assay the ability of AlphaFold to predict the impact of single mutations on protein stability (ΔΔG) and function.
  61. [61]
    Computational Saturation Mutagenesis to Investigate the Effects of ...
    Apr 28, 2022 · We have demonstrated the usefulness of AlphaFold models to investigate the evolution of proteins such as NRXN1. Using the AlphaFold structures ...3. Results · 3.2. Laminin G-Like Domains... · 3.6. Disease-Causing...
  62. [62]
    What AlphaFold 3 struggles with - EMBL-EBI
    Nevertheless, AlphaFold 3 still has limitations. A key limitation of protein structure prediction models is that they typically predict static structures as ...Missing: PTM | Show results with:PTM
  63. [63]
    FAQs - AlphaFold Protein Structure Database
    Data is available for academic and commercial use, under a Creative Commons Attribution 4.0 (CC-BY 4.0) licence terms. Where can I deposit AlphaFold structure ...
  64. [64]
    Predicting the structure of large protein complexes using AlphaFold ...
    Oct 12, 2022 · Since AlphaFold-multimer has a limit of nine chains or 1536 residues20 in its training and testing data, and there is no available method ...
  65. [65]
    The power and pitfalls of AlphaFold2 for structure prediction beyond ...
    Furthermore, membrane proteins represent less than 3% of total structures in the PDB, meaning that the training set used by AF2 was highly biased towards ...
  66. [66]
    Proteins with alternative folds reveal blind spots in AlphaFold-based ...
    Here, we review three blind spots that alternative conformations reveal about AF-based protein structure prediction.
  67. [67]
    Assessing the relation between protein phosphorylation, AlphaFold3 ...
    Apr 18, 2025 · The biases and limitations identified in AlphaFold predictions have significant implications for researchers studying post-translational ...
  68. [68]
    AlphaFold illuminates half of the dark human proteins - ScienceDirect
    AlphaFold illuminates half of the dark human proteins. Author links open overlay panel. Jessica L. Binder 1 , Joel Berendzen 1 b , Amy O. Stevens 2 , ...
  69. [69]
    AlphaFold Illuminates Half of the Dark Human Proteins - PMC
    ... pLDDT<50) residues. The low confidence region at the center of the AF ... AlphaFold and Implications for Intrinsically Disordered Proteins. J Mol Biol ...
  70. [70]
    Structural phylogenetics unravels the evolutionary diversification of ...
    Oct 10, 2025 · In contrast, the fold of proteins is often conserved well past sequence signal saturation. Furthermore, because three-dimensional (3D) structure ...
  71. [71]
    Revealing the secrets of protein evolution using the AlphaFold ...
    Sep 13, 2023 · Researchers use the AlphaFold database and Foldseek Cluster algorithm to analyse millions of predicted protein structures and offer new insights into protein ...
  72. [72]
    3D protein shapes can resolve ancient evolutionary connections in ...
    Jan 15, 2025 · Proteins fold into complex shapes that determine a cell's function. These shapes are more conserved over evolutionary time than the ...
  73. [73]
    PANDA-3D: protein function prediction based on AlphaFold models
    Aug 6, 2024 · PANDA-3D predicts gene ontology (GO) terms from the predicted structures of AlphaFold and the embeddings of amino acid sequences based on a large language ...
  74. [74]
    AlphaFold modeling of nepovirus 3C-like proteinases provides new ...
    We used AlphaFold to generate models of twelve nepovirus proteinases. All models show close relationships to picornavirus 3C proteinases.
  75. [75]
    Using AlphaFold Predictions in Viral Research - PMC - NIH
    Apr 21, 2023 · AlphaFold can be used for studies of molecular mechanisms of viral infection to facilitate several activities, including drug design.
  76. [76]
    AlphaFold update – neglected tropical diseases | EMBL-EBI
    Jan 28, 2022 · Latest AlphaFold database update adds 27 new organisms and almost 200000 new protein structure predictions relevant to neglected tropical ...Missing: collaboration | Show results with:collaboration
  77. [77]
    AlphaFold: Highly accurate protein structure prediction for WHO ...
    Jan 27, 2022 · Alphafold has recently added the proteome of nearly all the WHO priority pathogens to their library of highly accurate prediction of protein structures.Missing: collaboration | Show results with:collaboration
  78. [78]
    Structure and dynamics in drug discovery - Nature
    Nov 7, 2024 · ... cost of drug discovery and development can be reduced up to 50%. ... Once a well-defined drug target structure is obtained, virtual screening ...
  79. [79]
    Concordance of X-ray and AlphaFold2 Models of SARS-CoV-2 Main ...
    Nov 10, 2021 · We evaluate the concordance of models generated by the newly introduced AlphaFold2 structure prediction program with residual dipolar couplings (RDCs) measured ...
  80. [80]
    Rational drug design with AlphaFold 3 - Isomorphic Labs
    May 8, 2024 · AlphaFold 3 is an AI model that allows a scientist to input a description of a biomolecular complex that they are interested in, and predicts the 3D structure ...Alphafold 3 And The Research... · Settings · Looking Ahead
  81. [81]
    Isomorphic Labs kicks off 2024 with two pharmaceutical collaborations
    Isomorphic Labs announces strategic collaborations with two of the world's leading pharmaceutical companies, Eli Lilly and Company and Novartis.Missing: AlphaFold | Show results with:AlphaFold
  82. [82]
    Creating plastic-eating enzymes that could save us from pollution
    Jul 28, 2022 · The idea is to use enzymes to break down plastic polymers so that they can be 100% recycled back to their initial state – or even upcycling ...
  83. [83]
    From nature to industry: Harnessing enzymes for biocatalysis | Science
    Nov 24, 2023 · Engineered enzymes provide a green chemistry solution to the synthesis of complex organic molecules and can be used in the degradation of waste plastic and ...
  84. [84]
    De novo protein design by inversion of the AlphaFold structure ... - NIH
    De novo protein design aims to create stable, well‐folded proteins with sequences distant from those found in nature and potentially new functions. To date, de ...
  85. [85]
    Accurate prediction by AlphaFold2 for ligand binding in a reductive ...
    Mar 11, 2023 · We showed that AF2 provides processual (dynamic) predictions for the binding pockets of ligands (cofactors and/or substrates).
  86. [86]
    Data-driven synthetic microbes for sustainable future - PMC - NIH
    Jul 7, 2025 · From producing alternative fuels to degrading pollutants like plastic, microbes offer innovative solutions to a range of environmental concerns.
  87. [87]
    De novo design of protein structure and function with RFdiffusion
    Jul 11, 2023 · Here we show that by fine-tuning the RoseTTAFold structure prediction network on protein structure denoising tasks, we obtain a generative model ...
  88. [88]
    Improving de novo protein binder design with deep learning - Nature
    May 6, 2023 · Here, we explore the augmentation of energy-based protein binder design using deep learning. We find that using AlphaFold2 or RoseTTAFold to assess the ...
  89. [89]
    'The game has changed.' AI triumphs at solving protein structures
    Nov 30, 2020 · All of the groups in this year's competition improved, Moult says. But with AlphaFold, Lupas says, "The game has changed." The organizers even ...
  90. [90]
    AlphaFold with ChimeraX - UCSF RBVI
    Nov 9, 2021 · To predict the chicken sequence structure run AlphaFold by pressing the Predict button on the ChimeraX AlphaFold tool. This will run AlphaFold ...Output · Alphafold Ebi Database... · Example Alphafold-Multimer...Missing: integration | Show results with:integration
  91. [91]
    AI revolutions in biology: The joys and perils of AlphaFold
    Oct 20, 2021 · The most important limitation of AlphaFold predictions is that only a single state is predicted, even if hints for multiple states and dynamic ...
  92. [92]
    Leveraging AlphaFold for innovation and sustainable health ... - NIH
    Feb 4, 2025 · This democratization empowers scientists working on critical, region-specific health issues and exemplifies how technological innovations can ...Missing: low- | Show results with:low-
  93. [93]
    AlphaFold's Journey for an Inclusive Open Science Community
    Jul 16, 2025 · AlphaFold is transforming protein structure prediction by making cutting-edge science accessible to researchers worldwide.<|separator|>
  94. [94]
    Chemistry Nobel goes to developers of AlphaFold AI that predicts ...
    Oct 9, 2024 · The 2024 chemistry Nobel was awarded to John Jumper and Demis Hassabis at Google DeepMind in London, for developing a game-changing AI tool for predicting ...
  95. [95]
    AlphaFold Developers Win $3-Million Breakthrough Prize in Life ...
    Sep 22, 2022 · The researchers behind the AlphaFold artificial-intelligence (AI) system have won one of this year's US$3-million Breakthrough prizes—the most ...
  96. [96]
    AlphaFold—for predicting protein structures - Lasker Foundation
    Sep 2, 2025 · The group at Google DeepMind led by John Jumper and Demis Hassabis generated structure predictions that were several-fold better than the ...Missing: history | Show results with:history<|control11|><|separator|>
  97. [97]
    Sir Demis Hassabis CBE FREng FRS awarded Nobel Prize in ...
    Oct 9, 2024 · The Royal Society extends its warmest congratulations to Sir Demis Hassabis CBE FREng FRS on being awarded the Nobel Prize in Chemistry 2024.
  98. [98]
    DeepMind AlphaFold: The 200 Best Inventions of 2022 | TIME
    Nov 10, 2022 · In July, the company announced that AlphaFold had predicted the structures of 200 million proteins—nearly all known to humankind. And in what ...