MetaPredictor

MetaPredictor predicts possible human metabolites of small molecules including drugs as a sequence translation problem using deep language models with prompt engineering.

Key Features:

• Rule-free prediction: MetaPredictor operates without restrictions on metabolic transformation rules or specific enzyme families.
• Prompt engineering: Prompts that specify sites of metabolism (SoMs) steer the model to propose more accurate metabolite predictions.
• Two-stage schema: The method consists of automatic identification of SoMs followed by metabolite prediction.
• Transfer learning: A transfer learning strategy is utilized to tackle the limited availability of metabolic data.

Scientific Applications:

• Drug metabolite prediction: MetaPredictor predicts metabolites catalyzed by major enzyme families and less common enzymes for comprehensive drug metabolism profiling.
• Drug discovery: Investigation of the metabolic fate of drug candidates supports efficacy and safety assessment.

Methodology:

MetaPredictor formulates metabolite prediction as a sequence translation problem using deep language models with prompt engineering. Prompts specifying sites of metabolism guide prediction accuracy. A transfer learning strategy addresses limited metabolic data availability. The two-stage schema first automatically identifies sites of metabolism, then predicts metabolites.