Molecular AffinityHKMU AI4S Lab
← All four methods

03 / ETDock

Protein–ligand docking

Predict how a ligand might fit within a protein.

ETDock starts with protein and ligand structures and predicts a three-dimensional binding pose. It learns from chemical connectivity and geometry to reconstruct the ligand’s position and orientation.

Try docking Explore the method

Predict a binding pose

Supply a protein and ligand, choose a binding region, then explore and download predicted poses. The live calculation uses AutoDock Vina through Webina on your computer. ETDock’s geometric approach is described in the research below.

01 / Your input

AutoDock Vina runs on your computer through Webina. Your structures stay in this browser; loading a PDB ID contacts RCSB. Keep this tab open while docking. Use a prepared protein with resolved alternate locations and complete binding-site atoms. Protein hydrogens are rebuilt using Open Babel defaults without pH optimization. Waters are excluded. Sites requiring cofactors, metals, modified residues or covalent docking are unsupported.

Load a structure from the Protein Data Bank, then choose protein chains and supply the ligand to dock. The same preparation requirements apply.

One connected ligand, 5–80 heavy atoms, at most 15 rotatable bonds. Elements: H, C, N, O, F, P, S, Cl, Br and I. Specify stereochemistry and protonation. A fresh conformer is generated from bonds; SDF starting coordinates are not used as a predicted pose.

Box centre (Å)
Box dimensions (Å)

Use the protein’s coordinate frame. Each dimension: 8–40 Å, total volume at most 32,000 ų. Place the box around the intended binding pocket.

Search settings

Higher values search more thoroughly and take longer.

02 / Result

Your predicted poses will appear here

Load or upload a protein, supply a ligand and define its binding region. AutoDock Vina searches for poses in the protein’s coordinate frame.

From separate structures to a binding pose

A ligand’s chemical identity does not specify where it sits or how it is oriented in a protein. ETDock predicts distances between the binding partners, then uses those distances and constraints on the ligand’s internal geometry to reconstruct a pose.

The model combines atom-level and whole-molecule ligand features with a geometric protein representation. Its TAMformer module exchanges information through triangle, attention and message layers. Scalar features describe chemical properties; equivariant vectors preserve how spatial relationships change when a structure is rotated.

ETDock architecture. Ligand and protein features enter TAMformer, which combines triangle, attention and message layers. The learned protein–ligand distances guide reconstruction of a ligand binding pose. Figure 1, Yi et al. (2027 issue; online 2026).

ETDock architecture

ETDock Figure 1: ligand atom-level and graph-level features are fused with protein features; triangle, attention and message layers produce a distance matrix, docking pose and model confidence. Detail panels show attention and scalar/vector message passing.

Ligand and protein features enter TAMformer, which combines triangle, attention and message layers. The learned protein–ligand distances guide reconstruction of a ligand binding pose. Yi et al. (2027 issue; online 2026).

How closely do the predicted poses match?

The paper evaluates ETDock on a time split of PDBbind v2020. Ligand RMSD measures how closely predicted atom positions match the reference pose; lower values indicate a closer match.

Median ligand RMSD
3.8 Å
Poses within 2 Å
23.2%
Poses within 5 Å
61.1%

Reported in Table 1 of the ETDock paper. These metrics describe pose accuracy in that evaluation; they do not measure binding affinity.

From a pose to an affinity estimate

A pose describes an arrangement of atoms. Affinity concerns the strength of association between the protein and ligand. The arrangement establishes the contacts that a structure-based affinity model can examine.

Study the binding contacts

AGIMA-Score investigates how graphs of protein–ligand contacts can inform affinity estimation from a complex.

Explore AGIMA-Score

Explore a complex of your own

Already have a bound complex? The live GNINA tool scores the pose in your PDB structure and lets you explore its spatial occlusion response.

Score a bound complex

Paper & resources

Yiqiang Yi, Yuting Huang, Xu Wan, Yatao Bian, Debby D. Wang, Peilin Zhao and Le Ou-Yang. ETDock: an equivariant transformer for protein-ligand docking. Frontiers of Computer Science 21(5), 2105903 (2027). Published online on 5 January 2026; assigned to the 15 May 2027 journal issue.

Yuri Kochnev, Erich Hellemann, Kevin C. Cassidy and Jacob D. Durrant. Webina: an open-source library and web app that runs AutoDock Vina entirely in the web browser. Bioinformatics 36(16), 4513–4515 (2020).

Oleg Trott and Arthur J. Olson. AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of Computational Chemistry 31(2), 455–461 (2010).

Andrew T. McNutt, Yanjing Li, Rocco Meli, Rishal Aggarwal and David Ryan Koes. GNINA 1.3: the next increment in molecular docking with deep learning. Journal of Cheminformatics 17, 28 (2025).

Article and supplementary materials ↗ETDock source code ↗Webina browser engine ↗
← All four methods