molclaw-similarity-optimization

Molecular analog optimization while preserving target similarity is achieved by using reinforcement learning to steer generation toward high-scoring candidates. You provide a target SMILES and weighted scoring components like QED, LogP, and TPSA to guide the process.

You can weight Tanimoto similarity, QED, MW, LogP, TPSA, and MMP similarity during drug discovery optimization. These components act as configurable scoring functions with transform functions and normalization to control the multi-objective molecular generation.

To run similarity-guided optimization, input a target SMILES and configure parameters like similarity_weight, qed_weight, and max_steps. The process uses REINVENT4 staged_learning to produce optimized analogs and outputs a CSV file of the best SMILES for downstream evaluation.

Yes, this molecular optimization approach works for hit-to-lead analog generation and constrained library optimization. It applies reinforcement learning to improve drug-likeness metrics and physicochemical properties while preserving chemical similarity to your reference molecule.

Reinforcement learning generation actively steers molecule creation toward high-scoring candidates during the process, whereas post-hoc filtering samples broadly and discards unsuitable molecules afterward. RL-driven optimization is necessary when tight multi-objective constraints make simple filtering inefficient.