A Comprehensive Investigation of Active Learning Strategies for Conducting Anti-Cancer Drug Screening.

Area of Science:

• Computational biology
• Pharmacogenomics
• Machine learning in drug discovery

Background:

• Cancer cell lines with the same histology can exhibit varied responses to treatments.
• Accurate computational drug response prediction is crucial for preclinical screening and clinical treatment strategies.
• Drug response prediction models require experimental screening data for training.

Purpose of the Study:

• To investigate active learning strategies for selecting experiments to generate drug response data.
• To improve the performance of drug response prediction models.
• To enhance the identification of effective cancer treatments.

Main Methods:

• Developed and compared various active learning strategies for cell line selection: random, greedy, uncertainty, diversity, and hybrid approaches.
• Evaluated strategies based on the number of identified hits (responsive experiments) and prediction model performance.
• Conducted analysis across 57 drugs for drug-specific response prediction models.

Main Results:

• Active learning approaches demonstrated significant improvement in identifying hits compared to random and greedy methods.
• Active learning strategies also showed enhanced response prediction performance for certain drugs and analysis runs.
• The study highlights the efficacy of intelligent experimental design in drug discovery.

Conclusions:

• Active learning strategies are superior to random and greedy selection for improving drug response prediction models and identifying effective treatments.
• These findings have implications for optimizing preclinical drug screening and designing personalized cancer therapies.
• Further research can explore advanced active learning techniques for more efficient drug discovery pipelines.