De Novo Molecule Design Using Molecular Generative Models Constrained by Ligand-Protein Interactions

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Artificial intelligence in drug discovery

Background:

• Deep generative models are increasingly used for *de novo* drug design.
• Current models primarily utilize 2D ligand information, limiting 3D structural considerations.
• Generating molecules with specific binding modes remains a challenge.

Purpose of the Study:

• To develop a novel molecular deep generative model incorporating 3D ligand-protein interaction fingerprints.
• To evaluate the effectiveness of interaction fingerprint constraints in guiding molecular generation.
• To demonstrate the model's utility for targeted molecule generation and chemical space exploration.

Main Methods:

• Developed a recurrent neural network (RNN) architecture.
• Integrated ligand-protein interaction fingerprints, derived from docking poses, as constraints.
• Trained and compared generative models with and without interaction fingerprint constraints.

Main Results:

• Models trained with interaction fingerprints showed a propensity for generating compounds with similar binding modes.
• Constrained models demonstrated improved control over molecular structure generation based on 3D interactions.
• The approach successfully guided the exploration of drug-like chemical space.

Conclusions:

• Interaction fingerprint-constrained generative models offer a promising approach for targeted drug design.
• This method enhances the generation of molecules with specific ligand-protein binding characteristics.
• The model facilitates guided exploration of chemical space for novel therapeutic agents.