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

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Fluorination is crucial for optimizing drug properties.
• Identifying optimal fluorination sites and introducing fluorinated groups presents significant challenges, often requiring extensive experimentation.

Purpose of the Study:

• To develop an efficient computational tool to predict suitable fluorination sites in organic molecules.
• To accelerate the design of novel fluorinated drug candidates.

Main Methods:

• Developed F-site, a transformer-based Seq2Seq framework.
• Trained the model on a large dataset of preclinical fluorinated compounds from ChEMBL.
• Validated the model on independent test sets of clinical-stage and approved fluorinated drugs.

Main Results:

• Achieved over 98% token-level accuracy on internal validation.
• Successfully recovered validated fluorination patterns in approximately 80% of cases on independent test sets.
• Demonstrated the model's ability to generate relevant modification hypotheses, narrowing the experimental search space.

Conclusions:

• The F-site model offers a powerful computational approach to guide fluorination strategies in early-stage drug discovery.
• This tool can significantly enhance the efficiency of designing novel fluorinated small molecules.
• F-site has the potential to reduce the time and resources required for drug development.