molclaw-similarity-optimization

What problem does it solve?

This skill generates molecular analogs that remain similar to a reference molecule while actively optimizing multiple physicochemical properties using reinforcement learning, addressing cases where simple sampling plus filtering fails to find viable candidates under tight multi-objective constraints.

Core Features & Use Cases

• Reinforcement-learning guided generation (REINVENT4 staged_learning) that steers molecule generation toward high-scoring candidates rather than relying on post-hoc filtering.
• Multi-objective weighted scoring with configurable components such as Tanimoto similarity, QED, MW, LogP, TPSA, and MMP similarity, including transform functions and normalization.
• User recipes and runtime guidance for quick tests, standard optimization, thorough training, and solubility-focused searches; outputs best-found SMILES and a CSV file for downstream evaluation.
• Use Cases: hit-to-lead analog generation, constrained property optimization for lead candidates, and targeted improvement of drug-likeness metrics while preserving chemical similarity.

Quick Start

Start a similarity-guided RL optimization on the target SMILES with similarity_weight 0.6, qed_weight 0.3, and max_steps 100 to produce optimized analogs.