AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determination

2023/11/30 by Thomas C. Terwilliger, Dorothée Liebschner, Tristan I. Croll +8 · 3 voices · 364 citations
Biochemistry, Genetics and Molecular Biology · Materials Science · #Artificial intelligence #Biological system #Biology #Computer science #Distortion (music) #Enzyme Structure and Function #Machine Learning in Bioinformatics #Physics #Protein Structure and Dynamics #Protein structure #Scale (ratio)

paper · pdf · doi:10.1038/s41592-023-02087-4

published in Nature Methods 21(1), 110-116 (Nature Portfolio)

openalex publication_date 2023/11/30 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/06

Abstract

Artificial intelligence-based protein structure prediction methods such as AlphaFold have revolutionized structural biology. The accuracies of these predictions vary, however, and they do not take into account ligands, covalent modifications or other environmental factors. Here, we evaluate how well AlphaFold predictions can be expected to describe the structure of a protein by comparing predictions directly with experimental crystallographic maps. In many cases, AlphaFold predictions matched experimental maps remarkably closely. In other cases, even very high-confidence predictions differed from experimental maps on a global scale through distortion and domain orientation, and on a local scale in backbone and side-chain conformation. We suggest considering AlphaFold predictions as exceptionally useful hypotheses. We further suggest that it is important to consider the confidence in prediction when interpreting AlphaFold predictions and to carry out experimental structure determination to verify structural details, particularly those that involve interactions not included in the prediction.

Cited by

• The wall confronting large language models
• Machine learning in NMR spectroscopy
• Protein Structural Phylogenetics
• UCSF ChimeraX: Tools for structure building and analysis
• Macromolecular crystallography from an industrial perspective – the impact of synchrotron radiation on structure-based drug discovery
• Smart Integration of Structural Biology and Medicinal Chemistry to Unlock Target‐Driven Drug Discovery
• _sampleworks:_ A Modular Platform for Experimentally Guided Biomolecular Ensemble Generation
• AI Needs Physics More Than Physics Needs AI
• Categorizing prediction modes within low-pLDDT regions of AlphaFold 2 structures: near-predictive, pseudostructure and barbed wire
• Impact of AlphaFold on structure prediction of protein complexes: The CASP15‐CAPRI experiment
• Learning the language of plant immunity: opportunities and challenges for AI-assisted modelling of fungal effector x host protein complexes
• The Age of AI Agents Demands A New Scientific Paradigm To Sustain Trustworthy Science
• Proteomics—The State of the Field: The Definition and Analysis of Proteomes Should Be Based in Reality, Not Convenience
• From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery
• AlphaFold as a prior: experimental structure determination conditioned on a pretrained neural network
• Replacement of a single residue in an antibody abolishes cognate antigen binding, as predicted by theoretical methods
• TBM preferred to AlphaFold 3 for functional models of insect odorant receptors
• Slice'N'Dice : maximizing the value of predicted models for structural biologists
• A systematic exploration of bacterial form I rubisco maximal carboxylation rates
• Structure is beauty, but not always truth
• The structural chemistry and biosynthesis of chlorophylls
• Integrative structural and physicochemical characterization of chalcone synthase enzymes from medicinal plants using AlphaFold, molecular docking, and molecular dynamics
• DomainFit: Identification of protein domains in cryo-EM maps at intermediate resolution using AlphaFold2-predicted models
• Sequence-based protein-protein interaction prediction and its applications in drug discovery
• Towards better structural models from cryo‐electron microscopy data with physics‐based methods
• Community recommendations on cryoEM data archiving and validation
• Utilizing anomalous signals for element identification in macromolecular crystallography
• CrystalBoltz: End-to-End Protein Structure Determination via Experiment-Guided Diffusion for X-Ray Crystallography
• On the Reliability of AI Methods in Drug Discovery: Evaluation of Boltz-2 for Structure and Binding Affinity Prediction

Discussions

• AlphaFold predictions accelerate protein structure determination but require experimental validation. Many align with crystallographic maps, but discrepancies occur, highlighting the importance of con [bsky, 15 points, 0 comments]
• So why is this happening? It could be structural inaccuracies - Terwilliger et al showed that even high-confidence residues can be wrong, by comparing deviations between models and structures to those [bsky, 5 points, 1 comments]
• I am seeing AlphaFold predictions being used more and more in papers coming across my desk. This is a good article looking at how good those predictions actually are doi.org/10.1038/s415... [bsky, 1 points, 0 comments]

Related

• Proteins with alternative folds reveal blind spots in AlphaFold-based protein structure prediction
• Protein Structure Prediction until CASP15
• A structural biology community assessment of AlphaFold2 applications
• Highly accurate protein structure prediction with AlphaFold
• Physics-Grounded Evaluation to Guide Accurate Biomolecular Prediction