TransCell: In silico characterization of genomic landscape and cellular responses from gene expressions through a two-step deep transfer learning

Abstract

Gene expression profiling of new or modified cell lines becomes routine today; however, obtaining comprehensive molecular characterization and cellular responses for a variety of cell lines, including those derived from underrepresented groups, is not trivial when resources are minimal. Using gene expression to predict other measurements has been actively explored; however, systematic investigation of its predictive power in various measurements has not been well studied. We evaluate commonly used machine learning methods and present TransCell, a two-step deep transfer learning framework that utilizes the knowledge derived from pan-cancer tumor samples to predict molecular features and responses. Among these models, TransCell has the best performance in predicting metabolite, gene effect score (or genetic dependency), and drug sensitivity, and has comparable performance in predicting mutation, copy number variation, and protein expression. Notably, TransCell improved the performance by over 50% in drug sensitivity prediction and achieved a correlation of 0.7 in gene effect score prediction. Furthermore, predicted drug sensitivities revealed potential repurposing candidates for new 100 pediatric cancer cell lines, and predicted gene effect scores reflected BRAF resistance in melanoma cell lines. Together, we investigate the predictive power of gene expression in six molecular measurement types and develop a web portal (http://apps.octad.org/transcell/) that enables the prediction of 352,000 genomic and cellular response features solely from gene expression profiles. Key PointsO_LIWe provide a systematic investigation on evaluating the predictive power of gene expression in six molecular measurement types including protein expression, copy number variation, mutation, metabolite, gene effect score, and drug sensitivity. C_LIO_LITransCell took advantage of the transfer learning technique, showing how to learn knowledge from the source tumors, and transfer learned weight initializations to the downstream tasks in cell lines. C_LIO_LICompared to the baseline methods, TransCell outperformed in metabolite, gene effect score, and drug sensitivity predictions. C_LIO_LITwo cases studies demonstrate that TransCell could identify new repurposing candidates for pediatric cancer cell lines as well as capture the differences of genetic dependencies in melanoma resistant cell lines. C_LI