AlphaFold 2 / 3
The model that solved the 50-year protein-folding problem.
Evidence4/7 evidence fields documented
AccessLimited open access
ComputeGPU / managed server
BenchmarkCASP14 / complex evaluations
Protein Structure Prediction AI Models
Protein-structure models are not interchangeable. Monomer prediction, multimolecular co-folding, ligand support, nucleic-acid support, access terms and evaluation protocols all change what a model can responsibly support.
Curated model passports
28 relevant BioAtlas records.
Open evidence passport →
RoseTTAFold / All-Atom
The three-track network that followed folding into all-atom space.
Evidence4/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkCASP14 / complex modelling
OpenFold
A fully open, trainable reproduction of AlphaFold 2.
Evidence3/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkNot yet curated
ESMFold
Structure from a single sequence — no alignment required.
Evidence3/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkNot yet curated
Chai-1 / Chai-2
An AlphaFold3-class complex predictor, made freely usable.
Evidence4/7 evidence fields documented
AccessLimited open access
ComputeGPU / hosted service
BenchmarkComplex and antibody-design evaluations
Boltz-1 / Boltz-2
Open-source AF3-quality structure — plus binding affinity.
Evidence4/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkPoseBusters and affinity benchmarks
Protenix
ByteDance's open reproduction of AlphaFold 3.
Evidence2/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkNot yet curated
RFdiffusion
Diffusion models that hallucinate brand-new proteins.
Evidence5/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkExperimental binder validation
ProteinMPNN / LigandMPNN
Given a shape, design the sequence that folds into it.
Evidence3/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkNot yet curated
ESM3
A generative model that reasons over sequence, structure & function at once.
Evidence4/7 evidence fields documented
AccessLimited open access
ComputeGPU / hosted service
BenchmarkGenerative protein evaluations
Chroma
Programmable protein generation with a diffusion 'grammar'.
Evidence3/7 evidence fields documented
AccessLimited open access
ComputeGPU recommended
BenchmarkNot yet curated
Latent-X
Push-button, all-atom binder design in the browser.
Evidence2/7 evidence fields documented
AccessProprietary
ComputeGPU recommended
BenchmarkNot yet curated
AlphaProteo
High-affinity binder generation from DeepMind.
Evidence2/7 evidence fields documented
AccessProprietary
ComputeGPU recommended
BenchmarkNot yet curated
NVIDIA BioNeMo
The GPU-accelerated toolkit that ships biology's foundation models.
Evidence2/7 evidence fields documented
AccessLimited open access
ComputeNVIDIA GPU / managed service
BenchmarkNot yet curated
NeuralPLexer / Enchant
Physics-aware structure + multi-task ADMET foundation models.
Evidence2/7 evidence fields documented
AccessLimited open access
ComputeGPU or managed service
BenchmarkNot yet curated
AlphaMissense
Classifying which missense mutations cause disease.
Evidence2/7 evidence fields documented
AccessLimited open access
ComputeGPU recommended
BenchmarkNot yet curated
IgLM / AntiBERTy
The antibody-specific language models many tools build on.
Evidence4/7 evidence fields documented
AccessOpen source
ComputeManaged platform or GPU
BenchmarkAntibody sequence evaluations
LabGenius — EVA
A robotic evolution engine for multi-specific antibodies.
Evidence1/7 evidence fields documented
AccessProprietary
ComputeManaged platform or GPU
BenchmarkNot yet curated
Basecamp Research (EDEN / BaseFold)
Biodiversity-scale data as the fuel for better bio-models.
Evidence1/7 evidence fields documented
AccessLimited open access
ComputePlatform dependent
BenchmarkNot yet curated
Isomorphic Labs (IsoDDE)
AlphaFold's spin-off, aiming to 'solve all disease'.
Evidence1/7 evidence fields documented
AccessProprietary
ComputeVendor managed
BenchmarkNot yet curated
Generate:Biomedicines
Generative biology turned into a clinical-stage pipeline.
Evidence1/7 evidence fields documented
AccessLimited open access
ComputeVendor managed
BenchmarkNot yet curated
BoltzGen
All-atom generative binder design across proteins, peptides and molecular partners.
Evidence2/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkBinder-design evaluations
RFantibody
De novo epitope-specific antibody design from the RFdiffusion lineage.
Evidence4/7 evidence fields documented
AccessLimited open access
ComputeGPU recommended
BenchmarkExperimental antibody design
SaProt
Protein language modelling with amino-acid and structural-alphabet tokens.
Evidence4/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkProtein understanding benchmarks
Uni-Mol / Uni-Mol2
3D molecular foundation models trained on atoms, topology and conformational geometry.
Evidence3/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkMolecular representation/property benchmarks
RiNALMo
A large RNA language model for transferable nucleotide representations.
Evidence4/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkRNA downstream and structure tasks
RNA-FM
A foundational BERT-style model for non-coding RNA sequence representations.
Evidence2/7 evidence fields documented
AccessOpen source
ComputeGPU recommended
BenchmarkNot yet curated
HelixFold3
An open AF3-class biomolecular complex predictor from the PaddleHelix ecosystem.
Evidence2/7 evidence fields documented
AccessLimited open access
ComputeGPU recommended
BenchmarkNot yet curated
How to interpret this topic
Choose by context of use.
Inputs matter
Sequence, structure, ligand, assay and single-cell inputs imply different data-quality and preprocessing assumptions.
Outputs are not interchangeable
A predicted pose, confidence score, affinity estimate, generated sequence and perturbation profile support different decisions.
Benchmarks are protocol-bound
Claims should only be compared when task, dataset, split, metric and evaluation protocol genuinely align.
Validation remains external
BioAtlas records evidence boundaries; prospective scientific and experimental validation is still required for consequential use.