Application of transfer learning for cancer drug sensitivity prediction

Area of Science:

• Pharmacogenomics
• Computational Biology
• Precision Medicine

Background:

• Scarcity of biological data hinders predictive model design in precision medicine.
• Large-scale pharmacogenomics datasets (e.g., CCLE, GDSC) offer potential but suffer from distribution shifts.
• Transfer learning is a promising strategy to integrate multi-source data.

Purpose of the Study:

• To develop novel transfer learning approaches for improving drug sensitivity prediction.
• To address the challenge of distribution shift when combining data from different pharmacogenomics databases.
• To enhance the accuracy of predictive models in precision medicine.

Main Methods:

• Proposed two novel transfer learning approaches: latent variable cost optimization and polynomial mapping.
• Utilized the Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC) databases.
• Evaluated model performance against existing methods and database-specific models.

Main Results:

• The proposed approaches significantly improved drug sensitivity prediction compared to existing methods.
• Both latent variable cost optimization and polynomial mapping demonstrated effectiveness.
• The nonlinear (polynomial) mapping model exhibited the best overall performance across various scenarios.

Conclusions:

• The developed transfer learning methods outperform individual database models and current transfer learning techniques.
• Novel approaches effectively integrate data from disparate sources like CCLE and GDSC.
• The nonlinear mapping model shows particular promise for accurate anti-cancer compound sensitivity prediction in precision medicine.