On the Value of Using 3D Shape and Electrostatic Similarities in Deep Generative Methods.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in chemistry

Background:

• Multiparameter optimization in drug design remains a significant challenge.
• Automated methods for designing compounds with multiple controlled properties are needed.
• Previous fragment-based reinforcement learning (DeepFMPO) generated optimized molecules.

Purpose of the Study:

• To present an extended fragment-based reinforcement learning method (DeepFMPO v3D) for novel molecule generation.
• To enable the replacement of molecular parts with fragments of similar 3D shape and electrostatics.
• To improve automated compound design for multiple controlled properties.

Main Methods:

• Developed and benchmarked the ESP-Sim Python package for comparing electrostatic potential and molecular shape.
• Calculated high-quality partial charges using quantum chemistry program Psi4 (e.g., RESP with B3LYP/6-31G**).
• Implemented DeepFMPO v3D, a generative method incorporating 3D fragment comparisons to simulate properties and overcome conformation challenges.

Main Results:

• The ESP-Sim package enables accurate comparison of 3D molecular fragments based on shape and electrostatics.
• DeepFMPO v3D successfully generated novel molecules by replacing structural parts with similar 3D fragments.
• Demonstrated scaffold-hopping for identifying CDK2 bioisosteres using the new method.

Conclusions:

• DeepFMPO v3D represents a significant advancement in automated molecule design for multiparameter optimization.
• The integration of 3D fragment similarity improves the generation of molecules with desired properties.
• The open-source code and ESP-Sim package facilitate further research in computational drug design.