A feature transferring workflow between data-poor compounds in various tasks

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning in toxicology

Background:

• In silico compound screening aids in identifying potent drug candidates and predicting safety.
• Existing models struggle with predicting drug activity and toxicity due to insufficient and imbalanced data across different targets.
• Understudied targets often lack adequate data, hindering accurate predictions.

Purpose of the Study:

• To develop a novel prediction model for accurate drug activity and toxicity prediction in targets with limited observations.
• To address the challenges posed by insufficient and imbalanced drug data in existing models.
• To improve the prediction accuracy for understudied targets in drug discovery.

Main Methods:

• Proposed a two-stage transfer learning workflow.
• Developed a drug activity and toxicity prediction model using Siamese networks and graph convolution for multitasking output.
• Created a balanced dataset based on the Tox21 dataset.
• Applied transfer learning from data-rich to data-poor targets.

Main Results:

• Achieved high accuracy in predicting compound activity and toxicity for both data-rich and data-poor targets.
• The prediction model achieved 0.877 AUROC for classification tasks on the Tox21 dataset.
• Transfer learning strategies significantly improved model accuracy for understudied targets across five unbalanced datasets.

Conclusions:

• The developed model effectively overcomes data imbalance issues in target data.
• It enables accurate prediction of compound activity and toxicity for understudied targets.
• This approach can help prioritize biological experiments in drug discovery.