AI Engineer SyntheticGestalt Kabushiki Kaisha, Japan
Abstract: Recently, deep learning models for single cell applications have been pre-trained on large-scale scRNA-seq data. These models have been utilized for dissecting disease phenotype, development trajectory, and cellular heterogeneity. However, there remain limitations regarding the model’s performance in identifying subpopulations and state-transitions. These limitations hinder our ability to trace cell fate decisions during finer differentiation processes and to comprehensively analyze disease-related cellular behavior within these subtypes. The gene expression data is represented as a bag in computer science. Using the data structure is suitable for analysis, whereas it has been challenging to directly incorporate with bag data in deep learning models. We addressed this issue by developing a new model which can process gene multiplicity and improved the performance of cell type classification. We applied the model to the classification of hepatocytes in liver, since it is crucial for their use in medical examinations. This model demonstrates improved prediction accuracy of the details of cell types compared to existing approaches.
Funding Source: Moonshot R&D Program promoted by Japan Science and Technology Agency: Project "Co-evolution of Human and AI-Robots to Expand Science Frontiers"
