TransCell

TransCell utilizes deep transfer learning to predict molecular features and cellular responses from gene expression data, enabling predictions for measurements including drug sensitivity and gene effect scores.

Key Features:

• Two-step deep transfer learning framework: TransCell leverages knowledge derived from pan-cancer tumor samples to predict molecular features and cellular responses.
• Multi-measurement prediction: TransCell predicts six molecular measurement types including metabolite levels, gene effect scores (genetic dependency), drug sensitivity, mutation, copy number variation, and protein expression.
• High-performance drug sensitivity prediction: TransCell improved performance by over 50% in drug sensitivity prediction compared to other evaluated machine learning methods.
• Gene effect score prediction: TransCell achieved a correlation of 0.7 in gene effect score prediction.
• Large-scale feature prediction: TransCell enables prediction of 352,000 genomic and cellular response features from gene expression profiles.

Scientific Applications:

• Pediatric cancer drug repurposing: Predicted drug sensitivities revealed potential repurposing candidates for 100 new pediatric cancer cell lines.
• BRAF resistance characterization: Predicted gene effect scores reflected BRAF resistance in melanoma cell lines.
• Underrepresented cell line characterization: TransCell enables comprehensive molecular characterization and cellular response prediction for cell lines derived from underrepresented groups where resources are minimal.

Methodology:

TransCell employs a two-step deep transfer learning framework that utilizes knowledge derived from pan-cancer tumor samples. The framework was evaluated against commonly used machine learning methods for systematic investigation of gene expression predictive power across various measurements.