Target-aware molecule SMILES generation using a large language model with retrieval-augmented generation, multi-turn memory, and a predictive model.

This work evaluates a target-aware ligand SMILES generation framework in which an open-weight LLM (GPT-OSS-20b) is conditioned on protein-specific retrieval examples and iteratively refined through multi-turn memory. For a given protein target, the system retrieves the most similar protein–ligand examples from BindingDB, Davis, or KIBA to provide retrieval-augmented context, and then generates successive SMILES candidates that are optimized using predicted pKi from a pretrained drug–target interaction model (DeepPurpose) while enforcing chemical validity. This study examines the feasibility, behavior, and trade-offs of combining retrieval-augmented generation, conversational memory, and predictive feedback for target-conditioned molecular generation.