Predicting drug response through tumor deconvolution by cancer cell lines

Area of Science:

• Computational biology
• Genomics
• Pharmacogenomics

Background:

• Limited patient drug response data hinders pharmacogenomics research.
• Bridging the gap between in vitro (cell line) and in vivo (patient) data is challenging.

Purpose of the Study:

• To develop a deep learning model (Scaden-CA) for deconvoluting tumor data into cancer cell line proportions.
• To create a drug response prediction method using deconvoluted data and cell line sensitivity information.

Main Methods:

• Trained a deep learning model, Scaden-CA, for tumor deconvolution.
• Validated the model using Cancer Cell Line Encyclopedia (CCLE) bulk RNA data.
• Applied the model to The Cancer Genome Atlas (TCGA) dataset for drug response prediction.

Main Results:

• Scaden-CA demonstrated high performance with concordance correlation coefficients >0.9 and a deconvolution rate >70% in validation.
• The model successfully deconvoluted TCGA tumor data into cancer cell line proportions.
• Associations between predicted cell viability and genomic features were examined.

Conclusions:

• Scaden-CA effectively deconvolutes tumor data, enabling robust drug response prediction.
• The findings support the potential for drug repurposing by identifying mechanisms linked to predicted cell viability.
• This work advances pharmacogenomics by integrating in vitro and in vivo data.