Predicting Cancer Drug Response using a Recommender System.

Area of Science:

• Genomics
• Pharmacology
• Bioinformatics

Background:

• Precision medicine requires predicting patient-specific drug responses in cancer using molecular data like gene expression.
• High-dimensional data necessitates methods that learn interpretable models for drug response mechanisms and robust cross-dataset predictions.

Purpose of the Study:

• To propose a novel method, CaDRReS (Cancer Drug Response predictor), for predicting cancer drug responses.
• To develop a method that can accommodate high-dimensional gene expression data for interpretable and robust predictions.

Main Methods:

• CaDRReS utilizes principles from recommender systems to predict drug responses for new cell lines or patients.
• The method learns projections of drugs and cell lines into a latent pharmacogenomic space.
• Performance is evaluated using large public datasets (CCLE and GDSC) and compared against existing approaches.

Main Results:

• CaDRReS demonstrates consistently good models and robust predictions, even on unseen patient-derived cell-line datasets.
• The inferred pharmacogenomic spaces aid in understanding drug mechanisms.
• The method facilitates the identification of cellular subtypes and characterization of drug-pathway associations.

Conclusions:

• CaDRReS offers a powerful tool for predicting cancer drug responses and understanding underlying mechanisms.
• The method's robustness across datasets supports its utility in precision oncology.
• The inferred pharmacogenomic space provides valuable biological insights for drug discovery and personalized treatment.