Compare · 14 models
Molecular docking and binding affinity models you can run online
Tools in this category predict three-dimensional binding poses and calculate protein–ligand binding affinities, helping researchers evaluate candidate molecules and understand macromolecular interactions.
What molecular docking and binding affinity models do
Molecular docking and binding affinity tools evaluate how chemical compounds interact with biological targets. These systems determine where and in what spatial orientation a ligand binds to a receptor, estimate the strength of the resulting interaction, or perform both tasks concurrently. Predictions range from binary classification of binder status to quantitative continuous metrics in kilocalories per mole or relative free energy differences.
Methodologies divide into structural search algorithms, generative models, language-based architectures, and thermodynamic simulations. Classical tools explore candidate conformations within receptor search boxes, while generative diffusion models sample translational and torsional space. Newer co-folding and sequence-based models predict complex coordinates or evaluate binding scores directly from primary protein sequences and small-molecule SMILES strings without requiring pre-resolved crystal structures.
Compared
- Boltz-2MIT × RecursionPredicts 3D complex structures and binding affinities for protein-ligand pairs in a single forward pass.Commercial OKReady
- DiffDockMIT / Barzilay LabDiffusion-based generative model predicting 3D poses of small-molecule ligands bound to proteins.Commercial OKReady
- AutoDock VinaScripps ResearchPredicts small-molecule binding poses and affinity scores against protein structures.Commercial OKOn demand
- MAMMALIBM ResearchA biomedical foundation model predicting binding affinity, blood–brain barrier permeability, and solubility.Check termsOn demand
- SMINAKoes LabA molecular docking tool optimised for fragment-sized ligands and custom scoring functions.Commercial OKOn demand
- AQAffinitySandboxAQAn open-source protein–ligand binding affinity prediction head built on OpenFold3 for drug discovery.See its pageNot yet
- AutoDock-GPUScripps ResearchGPU-accelerated molecular docking tool for high-throughput virtual screening of compound libraries.Commercial OKNot yet
- CoLiPRImicrosoftA 3D vision–language encoder that jointly represents chest CT scans and radiology reports for medical imaging research.See its pageNot yet
- DiffDock-PPMITA diffusion-based model for docking protein–protein interfaces.Commercial OKNot yet
- GninaKoes Lab / University of PittsburghRescores protein–ligand docked poses using convolutional neural networks to improve virtual screening enrichment.Commercial OKNot yet
- LULA-1.1omtxA sequence-only model predicting protein-ligand binding scores from amino-acid sequences and SMILES strings.See its pageNot yet
- MM/GBSAAmber consortiumEstimates binding free energies to rank congeneric small-molecule series against protein targets.Check termsNot yet
- NessorecursionpharmaPredicts protein–ligand binding affinities and binder status from protein sequences and small molecules.See its pageNot yet
- OpenFE / LomapOpen Free Energy consortiumPlans and executes alchemical free energy perturbation networks to calculate relative ligand binding affinities.Commercial OKNot yet
How to choose
To dock small molecules when a target structure or sequence is available, DiffDock uses generative diffusion to place ligands, achieving 38% top-1 accuracy under 2 Å on PDBBind compared to 23% for AutoDock Vina. However, DiffDock does not calculate absolute binding affinities. If simultaneous all-atom structure generation and binding affinity quantification are required from sequence and SMILES inputs, Boltz-2 predicts both in a single pass at roughly 1,000 times the speed of free-energy perturbation methods, though it requires 40 GB of GPU memory. For protein–protein interfaces rather than small molecules, DiffDock-PP docks pairs of pre-existing PDB structures.
For conventional docking pipelines using crystallographic structures, AutoDock Vina remains a standard benchmark that executes on CPU hardware and outputs scores in kcal/mol within a designated search box. When working specifically with fragment-sized compounds or needing to adjust scoring function terms, SMINA provides a CPU-based AutoDock Vina fork tailored for fragments. If virtual screening involves libraries of up to one million compounds, AutoDock-GPU applies a massively parallel Lamarckian genetic algorithm across 40 GB GPU hardware, prioritising throughput over per-pose accuracy. To improve hit enrichment from existing AutoDock Vina poses, Gnina rescores docked complexes using convolutional neural networks on a 16 GB GPU to yield CNNscore and CNNaffinity metrics.
When experimental-grade accuracy is needed to distinguish closely related compounds in a congeneric series, physics-based thermodynamic tools are preferable to single-pass machine learning models. OpenFE and Lomap plan perturbation networks and calculate alchemical relative binding free energies (ΔΔG) with statistical error estimates, requiring 80 GB of GPU memory. MM/GBSA offers a faster end-point alternative on a 16 GB GPU that provides per-residue free energy breakdowns for congeneric series, though it does not yield reliable absolute binding affinities.
For workflows that avoid 3D structural inputs altogether, LULA-1.1 predicts binding scores directly from protein amino-acid sequences and ligand SMILES strings without any docking or folding step. Nesso accelerates continuous affinity and binary binder predictions by using coarse-grained co-folding from sequences and SMILES, CCD, or SDF inputs. For multi-task screening that also includes properties like solubility and blood–brain barrier permeability, MAMMAL processes sequence and molecular tokens on a 24 GB GPU, while AQAffinity provides an open-source affinity prediction head integrated with OpenFold3 representations.
Questions
Can I run these models without a GPU or installing anything?
5 of the 14 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.
Which tools predict binding affinity directly from protein sequences without an existing 3D structure?
Boltz-2, Nesso, and LULA-1.1 can all process primary amino-acid sequences paired with ligand SMILES strings. Boltz-2 predicts both all-atom complex coordinates and affinity metrics concurrently. Nesso uses coarse-grained co-folding to output binary binder status and continuous affinity, whereas LULA-1.1 generates a binding score without calculating any 3D coordinates, docking poses, or folded structures.
Does DiffDock estimate how tightly a small molecule binds?
No. DiffDock is designed to predict the 3D binding pose of a ligand using generative diffusion. While it produces an internal confidence score indicating the reliability of the generated pose, it does not calculate or predict absolute binding affinity.
How do AutoDock Vina, SMINA, and AutoDock-GPU differ?
AutoDock Vina is an established CPU-based tool that docks ligands within a user-specified search box. SMINA is a CPU-based fork of AutoDock Vina modified to handle fragment-sized ligands and allow custom scoring functions. AutoDock-GPU adapts classical docking to run massively in parallel on GPU hardware (requiring 40 GB of GPU memory), enabling high-throughput screening of up to one million compounds by trading per-pose accuracy for speed.
When should alchemical free energy tools like OpenFE be used instead of docking scores?
OpenFE and Lomap are designed for situations where low-resolution docking scores are insufficiently discriminative, such as ranking close analogues within a congeneric series. It executes alchemical free energy perturbation (FEP) calculations to provide high-precision relative binding free energy differences (ΔΔG) and statistical error estimates, requiring 80 GB of GPU memory.
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.