When using the 3D Ligand Based Score in a generator, the generated molecules are optimized with respect to a crystallized reference molecule using two 3D scores: the 3D Shape Similarity and 3D Pharmacophore Similarity.
The model generates initial poses using our Flow Matching algorithm. These poses are then refined and optimized under shape, pharmacophore, and internal-energy constraints.
Pharmacophore Definition
We use RDKit to identify the following pharmacophore types:
- H-bond donor
- H-bond acceptor
- Anion
- Cation
- Metal binder
- Aromatic ring
- Hydrophobe
-
Halogen
Pharmacophore Score
The 3D Pharmacophore Score is calculated as the degree of overlap between the pharmacophoric volumes of the generated molecule and the reference compound.
Each pharmacophore is modelled as a spherical Gaussian function. We then compute the volume overlap between Gaussian pharmacophores of the same type.
Finally, the relative contribution of each pharmacophore to this Pharmacophore Score is weighted, based on the weights provided by the user during the set-up.
References
For more information, please check the following NeurIPS paper:
Template-Guided 3D Molecular Pose Generation via Flow Matching and Differentiable Optimization
Also available as a preprint: https://arxiv.org/abs/2506.06305