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

The proposed pipeline is designed for target-aware generation of ligand SMILES conditioned on a given protein target. It first performs sequence-based retrieval of the most similar protein–ligand–affinity examples from BindingDB to condition an open-weight LLM (GPT-OSS-20b) via retrieval-augmented generation. It then executes a multi-turn, feedback-driven refinement loop in which successive SMILES candidates are iteratively optimized using predicted pKi from a pretrained drug–target interaction model (DeepPurpose) as the optimization signal while enforcing chemical validity.