TransCell: In Silico Characterization of Genomic Landscape and Cellular Responses by Deep Transfer Learning.

Area of Science:

• Genomics
• Computational Biology
• Machine Learning

Background:

• Comprehensive molecular characterization of cell lines is resource-intensive, especially for underrepresented groups.
• Predictive power of gene expression for various molecular measurements requires systematic investigation.

Purpose of the Study:

• To develop and evaluate a deep transfer learning framework (TransCell) for predicting molecular features and cellular responses from gene expression.
• To systematically assess the predictive power of gene expression across diverse molecular measurements.

Main Methods:

• Implemented a two-step deep transfer learning framework (TransCell).
• Utilized knowledge from pan-cancer tumor samples to train the model.
• Evaluated machine learning methods for predicting metabolite, gene effect score, drug sensitivity, mutation, copy number variation, and protein expression.

Main Results:

• TransCell demonstrated superior performance in predicting metabolite, gene effect score, and drug sensitivity.
• Achieved over 50% improvement in drug sensitivity prediction and a 0.7 correlation for gene effect score.
• Identified potential drug repurposing candidates for pediatric cancer cell lines and revealed BRAF resistance mechanisms in melanoma.

Conclusions:

• Gene expression is a powerful predictor for multiple molecular and cellular response types.
• TransCell provides a valuable tool for comprehensive cell line characterization with minimal resources.
• A web portal is available for predicting 352,000 genomic and cellular response features from gene expression profiles.