REINVENT

REINVENT generates novel molecular structures using recurrent neural networks and reinforcement learning to optimize compounds toward specified properties for de novo molecular design.

Key Features:

• Sequence-Based Generative Model: Employs a sequence-based generative model implemented with recurrent neural networks to produce molecular structures.
• Augmented Episodic Likelihood Fine-Tuning: Fine-tunes the generative model using augmented episodic likelihood to bias generation toward desired properties.
• Reinforcement Learning Integration: Integrates reinforcement learning to optimize molecular designs toward predefined objectives such as activity against a biological target or structural similarity to a known compound.
• Analogue Generation: Generates analogues to a query structure for scaffold exploration and lead diversification.
• Target-Specific Compound Generation: Produces compounds predicted to be active against specified biological targets.
• Training Data Filtering: Supports training on datasets with explicit exclusions of elements or substructures (e.g., molecules without sulphur) prior to refinement.

Scientific Applications:

• Scaffold Hopping and Library Expansion: Enables scaffold hopping and expansion of chemical libraries starting from known compounds such as Celecoxib.
• Target-Specific Compound Generation: Generates compounds with high predicted activity against specific targets; for dopamine receptor type 2 tuning produced over 95% predicted actives including experimentally confirmed actives absent from the training data.
• Lead Optimization: Produces analogues with tailored properties to support lead optimization campaigns.
• Novel Compound Discovery: Aids discovery of novel compounds that may serve as potential therapeutic agents.
• Biological Target Exploration: Facilitates exploration of novel interactions between molecules and biological targets.

Methodology:

Uses a sequence-based RNN generative model trained on curated datasets (with possible exclusions such as molecules without sulphur), fine-tuned via augmented episodic likelihood, and refined through reinforcement learning to optimize for target objectives.