MegaSyn: Integrating Generative Molecular Design, Automated Analog Designer, and Synthetic Viability Prediction

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Machine learning

Background:

• Generative machine learning models are used in drug discovery to create novel molecules.
• Existing models often struggle with synthetic feasibility and converge to similar molecular spaces.
• These limitations hinder the real-world application of generated compounds.

Purpose of the Study:

• To introduce MegaSyn, an automated tool suite for generative molecular design.
• To address limitations in synthetic feasibility and molecular space exploration.
• To generate novel, synthesizable, and druglike compounds.

Main Methods:

• Developed MegaSyn, incorporating a hill-climb algorithm with SMILES-based recurrent neural network (RNN) models.
• Integrated analog generation software and retrosynthetic analysis for synthetic feasibility scoring.
• Utilized substructure deconstruction and an ensemble of generative models.

Main Results:

• MegaSyn effectively generates new molecular scaffolds and targeted analogs.
• The generated compounds demonstrate improved synthetic feasibility.
• The tool suite shows promise for optimizing molecules and prioritizing lead compounds.

Conclusions:

• MegaSyn offers a robust solution for generative molecular design in drug discovery.
• The approach enhances the generation of synthesizable and druglike molecules.
• This toolset can aid in optimizing lead compounds and accelerating drug development.