Deep generative design with 3D pharmacophoric constraints.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence in Chemistry

Background:

• Generative models are proposed for molecular design but struggle with control and incorporating prior knowledge.
• Current methods often neglect crucial three-dimensional (3D) structural information vital for molecular binding.
• Limited use of 3D structural data hinders practical applications in drug discovery.

Purpose of the Study:

• To develop a generative model that effectively incorporates 3D structural information for molecular design.
• To demonstrate the benefits of using 3D pharmacophoric information in generative models.
• To enhance control and quality in molecular generation for drug discovery.

Main Methods:

• Combined DeLinker, a graph-based deep generative model, with a convolutional neural network.
• Utilized physically-meaningful 3D representations of molecules and target pharmacophores.
• Applied the developed model, DEVELOP, to linker and R-group design tasks.

Main Results:

• Incorporating 3D pharmacophoric information improved molecular generation quality and control.
• Achieved over 300% improvement in generating molecules with high 3D similarity on a PDBbind test set.
• The DEVELOP model recovered 10x more original molecules compared to the baseline DeLinker method.

Conclusions:

• The developed method, DEVELOP, effectively integrates 3D structural information into generative molecular design.
• 3D pharmacophoric constraints substantially enhance the quality and similarity of generated molecules.
• The general-purpose approach is adaptable and can be integrated into various generative frameworks for drug discovery.