Optimizing Multi-Task Learning with Evolutionary Relatedness Metrics for Enhanced QSAR-Based Natural Product Activity...

Area of Science:

• Biochemistry
• Computational Chemistry
• Drug Discovery

Background:

• Natural products possess complex structures crucial for drug-target interactions.
• Limited bioactivity data hinders quantitative structure-activity relationship (QSAR) predictions for natural products.
• Multitask learning (MTL) is a promising strategy to overcome data scarcity in QSAR.

Purpose of the Study:

• To optimize MTL by integrating protein evolutionary relatedness for enhanced natural product bioactivity prediction.
• To identify conditions where MTL is most effective for sparse datasets.
• To improve drug discovery efforts for natural products.

Main Methods:

• A curated dataset of natural products and their bioactivity against enzymes was constructed from ChEMBL.
• Single-task learning (STL) was used as a baseline.
• Feature-based MTL (FBMTL) and instance-based MTL (IBMTL), incorporating evolutionary relatedness, were applied.
• Performance was evaluated across different protein groups.

Main Results:

• Instance-based MTL (IBMTL) outperformed STL and FBMTL in predicting natural product bioactivity.
• Evolutionary relatedness significantly improved prediction performance, particularly for kinase and cytochrome P450 protein groups.
• Optimal performance in the kinase group was observed at the target parent level, indicating a balance between relatedness and data size.

Conclusions:

• Leveraging protein evolutionary relatedness within an MTL framework enhances QSAR predictions for natural products, even with limited data.
• MTL, particularly IBMTL, shows significant potential for advancing natural product-based drug discovery.
• The study highlights the importance of considering protein hierarchy and evolutionary context in predictive modeling.