Target Prediction Model for Natural Products Using Transfer Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Natural products are a significant source of lead compounds for drug development.
• Many natural products lack comprehensive target information, hindering their potential.
• Efficient target prediction is essential for accelerating natural product-based drug discovery.

Purpose of the Study:

• To develop a robust computational model for predicting the biological targets of natural products.
• To leverage transfer learning to overcome data limitations in natural product target prediction.
• To enhance the identification and utilization of natural products in drug discovery pipelines.

Main Methods:

• Utilized transfer learning by pre-training a model on the ChEMBL dataset and fine-tuning it on a natural product dataset.
• Employed a data balancing technique to improve model performance.
• Validated model reliability through embedding space analysis and case studies on drug datasets.

Main Results:

• Achieved a high Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.910, demonstrating strong predictive performance.
• The model successfully identified all targets for 62 drugs in case studies.
• Outperformed previous studies in AUROC validation and success rate for identifying active targets.

Conclusions:

• The transfer learning model provides a reliable and effective method for predicting natural product targets.
• This approach can significantly aid in discovering novel lead compounds from natural sources.
• The model has potential applications in drug repurposing and accelerating the drug discovery process.