F-Site: Recognizing Fluorination Patterns in Small-Molecule Drugs via a Two-Stage Transformer-Based Model

Fluorination is a powerful strategy for modulating drug properties. However, accurately identifying suitable fluorination sites and introducing appropriate fluorinated groups into organic molecules remain significant challenges, often relying on chemical intuition and iterative experimentation. To address this gap, we developed F-site, a transformer-based Seq2Seq framework trained on a large, structurally nonredundant data set of preclinical fluorinated compounds curated from ChEMBL. The model achieved over 98% token-level accuracy on internal validation, demonstrating robust learning of molecular sequence patterns. Notably, on a nonredundant independent test set of clinical-stage and approved fluorinated drugs, F-site recovered the validated fluorination patterns in approximately 80% of cases within the top-ranked candidates for a given input scaffold. This result highlights the model's capability to generate compact and highly relevant sets of modification hypotheses. In summary, this performance underscores the F-site model's potential to substantially narrow the experimental search space for fluorination design in small molecules, thereby providing an efficient computational tool to guide fluorination strategies in early-stage drug discovery.