Anticancer drug synergy prediction in understudied tissues using transfer learning.

Area of Science:

• Oncology
• Computational Biology
• Genomics

Background:

• Drug combination screening offers potential for enhanced cancer treatment efficacy and safety.
• A significant challenge in drug synergy prediction is the variable availability of in vitro drug response data across different cancer types.
• Understudied cancer tissues present data scarcity issues, hindering the development of effective treatment strategies.

Purpose of the Study:

• To develop a drug synergy prediction model specifically for understudied cancer tissues.
• To overcome data scarcity limitations in cancer research.
• To improve the identification of effective cancer treatment options.

Main Methods:

• Collected comprehensive genetic, molecular, and phenotypic features from cancer cell lines.
• Developed a multitask deep neural network model for integrating multimodal data and multiple outputs.
• Employed transfer learning techniques to leverage data from data-rich tissues for application to data-poor tissues.

Main Results:

• Achieved improved accuracy in predicting drug synergy for both data-rich and understudied tissues.
• Demonstrated high prediction accuracy in data-rich tissues (0.9577 AUROC for classification, 174.3 MSE for regression).
• Confirmed that transfer learning significantly enhances prediction accuracy in understudied tissues.

Conclusions:

• The developed synergy prediction model effectively ranks synergistic drug combinations for understudied tissues.
• The model aids in prioritizing future in vitro experimental investigations.
• Open-source code is available to facilitate further research and application.