Compare · 20 models
Protein structure prediction models you can run online
These tools predict three-dimensional coordinates for proteins, complexes, and nucleic acids from sequence data or retrieve solved structures, serving researchers in structural biology and molecular design.
What protein structure prediction models do
Protein structure prediction determines the three-dimensional atomic coordinates of biological macromolecules from their primary amino acid or nucleotide sequences. Traditional computational approaches, such as AlphaFold2, rely on constructing multiple sequence alignments (MSAs) to extract evolutionary co-evolution patterns across homologous sequences. More recent approaches employ protein language models, such as ESMFold and OmegaFold, to translate sequence representations directly into Cartesian coordinates without querying alignment databases.
Beyond single protein chains, modern architectures co-fold biomolecular complexes comprising proteins, nucleic acids, post-translational modifications, and small-molecule ligands. While most tools output static all-atom models alongside per-residue confidence scores like pLDDT, specialised models also evaluate dynamic conformational ensembles or simulate binding affinity directly from sequence and chemical inputs.
Compared
- AlphaFold DB LookupDeepMind / EMBL-EBIRetrieve pre-computed AlphaFold protein structures and confidence scores across 247 million UniProt entries.Commercial OKInstant
- ESMFoldMeta AI (FAIR)Predicts all-atom 3D protein structures directly from single amino acid sequences without alignments in seconds.Commercial OKInstant
- PDB Structure FetchRCSB Protein Data BankRetrieves experimentally determined protein coordinates and metadata for any PDB entry.Commercial OKInstant
- AlphaFold2DeepMind (served as BioNeMo NIM)Predicts all-atom 3D protein structures from sequence data using multiple sequence alignments.Check termsReady
- Boltz-2MIT × RecursionPredicts 3D complex structures and binding affinities for protein-ligand pairs in a single forward pass.Commercial OKReady
- Chai-1Chai DiscoveryPredicts 3D structures of multimodal biochemical complexes, including proteins, nucleic acids, and small molecules.Check termsOn demand
- ESMFold2biohubPredicts all-atom 3D structures for proteins, nucleic acids, and ligands from sequence or sequence alignments.Check termsOn demand
- ESMFold2-ExperimentalbiohubPredicts all-atom protein structures and confidence metrics from sequences for experimental evaluation.Check termsOn demand
- ESMFold2-Experimental-FastbiohubA fast, experimental protein structure prediction model generating all-atom coordinates from amino acid sequences.Check termsOn demand
- ESMFold2-FastbiohubPredicts all-atom 3D protein structures rapidly from single amino acid sequences for high-throughput screening.Check termsOn demand
- ImmuneBuilderOxford Protein Informatics GroupFast structure prediction for antibodies, nanobodies, and T-cell receptors across full repertoires.Commercial OKOn demand
- BioEmumicrosoftGenerative model predicting protein equilibrium structural ensembles and thermodynamic stabilities from sequence.See its pageNot yet
- ColabFoldSteinegger LabMMseqs2-accelerated AlphaFold2 protein monomer and complex structure prediction pipeline.Commercial OKNot yet
- IgFoldRuffolo et al., Johns HopkinsPredicts antibody 3D structures from heavy and light chains in sub-second runtimes.Non-commercialNot yet
- OmegaFoldHeliXonPredicts protein structures from single sequences without multiple sequence alignments, suited for orphan proteins.Commercial OKNot yet
- OpenFoldAlQuraishi Lab / ColumbiaPredicts three-dimensional protein structures from amino acid sequences using a retrained PyTorch implementation.Commercial OKNot yet
- OpenFold3OpenFoldPredicts 3D structures of proteins, nucleic acids, ligands, and complexes for structural biology and drug discovery.See its pageNot yet
- Promerabjing-mitGenerative model for biomolecular co-folding, binder filtering, and miniprotein and nanobody design.See its pageNot yet
- Protenix-RNALiteFoldA Protenix-lineage PyTorch model fine-tuned for RNA three-dimensional structure prediction.See its pageNot yet
- RoseTTAFold All-AtomBaker Lab / UWFull-atom structure prediction for proteins, non-canonical residues, glycans, metals, and small-molecule cofactors.Non-commercialNot yet
How to choose
For unstudied sequences, the choice depends on whether evolutionary alignments exist and how quickly coordinates are required. When deep sequence alignments are available and maximum accuracy on monomers is paramount, AlphaFold2 serves as the benchmark, while ColabFold replaces Jackhmmer with MMseqs2 to reduce MSA generation time to seconds. OpenFold offers a PyTorch reproduction with matching accuracy for workflows requiring retraining. Conversely, for orphan sequences lacking known homologs, de novo designs, or high-throughput screens, single-sequence models like ESMFold, OmegaFold, and ESMFold2-Fast generate structures in seconds without requiring MSAs, though ESMFold trails AlphaFold2 when deep alignments exist.
When modelling assemblies that include small molecules, nucleic acids, or chemical modifications, multimodal models are required. Chai-1 folds assemblies of proteins, DNA, RNA, and ligands from sequence embeddings, supporting user-defined physical restraints and multiconformational sampling. Boltz-2 concurrently predicts protein-ligand 3D complex structures and binding affinities in a single forward pass. RoseTTAFold All-Atom models full assemblies incorporating non-canonical residues, glycans, and metals, while OpenFold3 replicates AlphaFold3 capabilities across nucleic acids and small molecules, matching benchmarks on monomeric RNA. For dedicated RNA modelling, Protenix-RNA is fine-tuned specifically on RNA MSAs up to 768 tokens, excluding protein templates.
Targeted biological workflows benefit from domain-specific predictors. In immunology, ImmuneBuilder delivers sub-second predictions across antibodies, nanobodies, and T-cell receptors, while IgFold predicts antibody structures from paired heavy and light chains roughly 1,000 times faster than AlphaFold2. For binder generation and filtering, Promera combines co-folding with masked-sequence design for miniproteins and nanobodies. If the objective is to capture conformational dynamics rather than a single static structure, BioEmu emulates Boltzmann equilibrium ensembles and free energy landscapes on a single GPU, outputting backbone frame representations without side chains.
Before computing novel predictions, researchers should consider retrieval tools. PDB Structure Fetch retrieves approximately 230,000 experimentally determined structures with experimental provenance and bound ligand details using CPU hardware. For catalogued proteins without experimental structures, AlphaFold DB Lookup provides instant access to 247 million pre-computed AlphaFold predictions and confidence scores using UniProt accessions, eliminating GPU compute requirements entirely.
Questions
Can I run these models without a GPU or installing anything?
11 of the 20 models here run on our servers from their page: open one, fill in the inputs and press Run. Those marked "Not yet" are listed with their licence and documentation but have no loader here yet.
When should an MSA-free predictor be chosen over an alignment-based model?
MSA-free models such as ESMFold, OmegaFold, and ESMFold2-Fast are preferred for high-throughput screening and for sequences with few or no evolutionary relatives, such as de novo designed proteins or orphan genes. Alignment-based predictors like AlphaFold2 achieve higher accuracy on targets that have deep, diverse multiple sequence alignments, but require minutes rather than seconds due to homology database searches.
Which models support complexes containing nucleic acids and small molecules?
Chai-1, OpenFold3, ESMFold2, Boltz-2, and RoseTTAFold All-Atom support non-protein entities. Chai-1, OpenFold3, and ESMFold2 co-fold proteins together with DNA, RNA, and ligands. Boltz-2 predicts protein-ligand complex geometries alongside binding affinity metrics in a single forward pass. RoseTTAFold All-Atom models full structures including non-canonical residues, cofactors, metals, and glycans.
How does BioEmu differ from static structure predictors like AlphaFold2?
Rather than producing a single static structural hypothesis, BioEmu models the Boltzmann equilibrium distribution of conformations that proteins adopt in solution, generating thousands of independent structures per hour to capture domain motions and cryptic pocket conformations. Unlike AlphaFold2, BioEmu outputs coordinates strictly in backbone frame representation and cannot model multimeric complexes or small-molecule ligands.
What tools are specialised for antibody and immune receptor structure prediction?
ImmuneBuilder and IgFold are specifically optimised for immune molecules. IgFold predicts antibody structures from paired heavy and light chains in under a second by focusing on CDR loop variability within conserved frameworks. ImmuneBuilder extends sub-second prediction across antibodies, nanobodies, and T-cell receptors to process entire immune repertoires on modest GPU hardware.
Written from each model's own documentation, as summarised on its page here. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.