Deriver

Deriver generates novel small-molecule structures for de novo drug design by integrating modular generation, discrimination, and exploration methods in a Python package.

Key Features:

• Python package: Implemented as a Python package for programmatic de novo molecular generation.
• Modular framework: Integrates multiple tunable generative methods that can be combined to target specific design objectives.
• Separation of concerns: Explicitly separates generation, discrimination, and exploration processes so each can be independently configured.
• Chemical space traversal: Independent parameter adjustments across generation, discrimination, and exploration improve traversal of chemical space and reduce entrapment in local minima.
• Optimization of molecular properties: Integration of diverse generative methods facilitates optimization of molecular properties and mitigates convergence on suboptimal solutions.
• Drug-likeness filtering: Supports pre-scoring filtering of molecules for drug-likeness using physicochemical property thresholds and SMARTS pattern matching, with an acknowledged trade-off between filter stringency and exploration breadth.
• Exploration algorithms: Supports multiple exploration strategies, including linear probabilistic sampling, Monte Carlo sampling, and greedy sampling, with linear probabilistic sampling reported to perform favorably in many cases.

Scientific Applications:

• De novo drug design: Generation of novel small molecules for lead discovery and hit-to-lead campaigns.
• Molecular property optimization: Optimization of physicochemical and other scored properties through combined generative approaches.
• Drug-likeness prioritization: Filtering and prioritization of candidate molecules based on physicochemical criteria and SMARTS pattern matching prior to scoring.
• Chemical space exploration: Systematic traversal and sampling of chemical space to identify diverse candidate structures and avoid local minima.
• Comparative algorithm assessment: Evaluation and comparison of exploration strategies such as linear probabilistic sampling, Monte Carlo, and greedy sampling for molecule generation tasks.

Methodology:

Deriver integrates multiple tunable generative methods within a modular Python framework, separates computation into generation, discrimination, and exploration components with independent parameter adjustment, applies pre-scoring drug-likeness filtering using physicochemical property criteria and SMARTS pattern matching, and supports exploration via linear probabilistic sampling, Monte Carlo sampling, and greedy sampling.