Pocket2Mol
E(3)-equivariant autoregressive 3D molecular generator inside protein binding pockets, which samples ligands atom by atom inside the pocket; docking (QVina) scores evaluate finished molecules.
The loop
- Proposes
An E(3)-equivariant graph neural network places frontier atoms, 3D coordinates, and covalent bonds step by step inside a protein pocket.
- Judged by
Docking (QVina) scores evaluate finished molecules.
- Steers
Ligands are sampled atom by atom inside the pocket.
Headline result
Generated drug-like 3D molecules with more realistic torsion angles and stronger Vina binding scores than 3D voxel baselines on CrossDocked2020.
What this does not prove
Autoregressive atom-by-atom placement can accumulate geometric error if an early frontier atom is poorly positioned.
How it generalizes the harness
An E(3)-equivariant 3D geometric network samples ligands atom by atom inside the pocket, and docking (QVina) scores evaluate finished molecules.
Classification
| What it proposes | Molecule |
|---|---|
| Generative model | E(3)-equivariant autoregressive 3D GNN |
| Algorithmic authority | QVina docking score, used to evaluate finished molecules |
| Feedback | Score |
| What gets updated | Search tree and population of candidates |
| When it runs | At inference time |
| Validity | Approximate physics (simulation, not lab results) |
| Evidence | Code released |
| Lineage | Original system |