Inferring multi-target QSAR models with taxonomy-based multi-task learning.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Multi-target drugs are crucial for complex diseases like cancer.
• Accurate quantitative structure-activity relationship (QSAR) models predict affinity profiles for lead optimization.
• Knowledge transfer between similar targets can improve QSAR model accuracy.

Purpose of the Study:

• To present and evaluate two transfer learning-based multi-task algorithms for QSAR modeling.
• To exploit target similarity for knowledge transfer and enhanced model performance.

Main Methods:

• Utilized two multi-task learning algorithms from transfer learning.
• Evaluated methods on simulated data and a dataset of 112 human kinases from ChEMBL.
• Derived target relatedness from the human kinome taxonomy.

Main Results:

• Multi-task learning improved QSAR model performance compared to separate models when tasks were sufficiently similar.
• The best multi-task approach reduced the mean squared error for QSAR models of 58 kinase targets.
• Performance gains were observed with sufficient task similarity.

Conclusions:

• Multi-task learning is a valuable strategy for inferring multi-target QSAR models in lead optimization.
• Knowledge transfer is most effective from similar, data-rich tasks to data-poor tasks.
• Benefits increase with decreasing overlap in the chemical space spanned by tasks.