A Rag-Conditioned Model for De Novo Molecular Generation Using Safe

De novo molecular generation aims to produce drug-like compounds with desirable pharmacological properties.Most generative models, however, offer limited contextual conditioning or ignore fragment-fragment interactions during synthesis, so combinations of individually promising fragments often turn out to be chemically unsatisfactory.To address these issues, this paper presents a Retrieval-Augmented Generation (RAG) conditioned system that uses SAFE-GPT (Sequential Attachment-based Fragment Embedding) for SMILES-based molecular design.To balance novelty with chemical validity, the framework combines interaction-aware fragment scoring, conditional dataset sampling and controlled adjustment of the SAFE generation settings, and it uses beam search during generation to explore high-scoring chemical routes.A fragment vocabulary of 5,612 unique arm and linker fragments was built from the ZINC250K dataset and each fragment was scored by the average property value of the molecules containing it.The retrieved fragments have an average QED of 0.5688, an average logP of 0.6874 and an average molecular weight of 151.16 Da, which shows that property-based scoring ranks fragments in line with drug-like chemical space.Planned experiments will test whether interaction-aware scoring improves diversity, validity and property optimisation (QED, logP and synthetic accessibility) over fragment-based and non-retrieval baselines.