This is a developmental repo for adding physics-informed features to the Latent3D Graph Diffusion (You et al., ICLR 2024) for docking problems. The original paper and implementation repo can be found here: https://github.com/Shen-Lab/LDM-3DG
The Latent3D Graph Diffusion model is having issues when applying to docking problems. Similar to related research, the model may generate some unrealistic docking outcomes (e.g. collision). We suggest that adding physics-informed features into the model would help improve the accuracy and alleviate the collision issue.
We use the exact same dataset as the DiffDock.
Including:
- PDBBind: download the processed complexes from zenodo.
- BindingMOAD: download the processed complexes from zenodo under
BindingMOAD_2020_processed.tar. - DockGen: download exclusively the complexes of the DockGen benchmark already processed (e.g. chain cutoff) from zenodo downloading the
DockGen.tarfile. - PoseBusters: download the processed complexes from zenodo.
- van der Mers: the protein structures used for the van der Mers data augmentation strategy were downloaded here.
The Latent3D Graph Diffusion model is using a pre-trained Autoencoder to project input ligand to a latent space. After that, a diffusion model is trained in the latent space for generating ligand in the inference mode. The protein is used as a condition for decoder, which consists of 2 parts, giving the topology and 3D coordinates outcomes seperately describing the generated ligand.
The overview of the Latent3D Graph Diffusion model is illustrated below:
Note that the physics-informed features have not integrated, and not mentioned above since it is still UNDERDEVELOPMENT.