Multimodal Classification of Single-cell Transcriptional State with Deep Neural Networks
Location
Hager-Lubbers Exhibition Hall
Description
PURPOSE: Automatic curation and annotation of clinical samples is a growing challenge in modern healthcare. Once increasingly popular clinical assay is single-cell sequencing, a technology that measures the expression of all genes in individual cells from a patient tissue sample or blood draw. PROCEDURES: We are developing AI software that can annotate information about new single-cell sequencing samples including cell type, donor organ, and developmental time points. Our primary objective is to construct a deep neural network using the Keras machine learning library in R and train this network on data from the Human Cell Atlas CellxGene repository. OUTCOME: Our work has shown successful cell type classification on a large subset of this dataset using GVSU’s High-Performance Computing (HPC) cluster. To enhance model accuracy, we employ hyperparameter tuning. Our short-term goal is to train the model on over 28 million single-cell transcriptomes, spanning hundreds of cell types, developmental time points, and donor organs, which will uncover comprehensive insights into the underlying complexity of clinical samples. To do this, we are incorporating a multi-channel output architecture to allow for simultaneous classification of multiple output modalities in parallel from a single shared neural network. IMPACT: This research will enable better understanding of the etiology of complex human diseases and help personalize medical practice. Our long-term goal is to personalize treatment of rare diseases by providing detailed understanding of patient tissue differences compared to healthy human donors.
Multimodal Classification of Single-cell Transcriptional State with Deep Neural Networks
Hager-Lubbers Exhibition Hall
PURPOSE: Automatic curation and annotation of clinical samples is a growing challenge in modern healthcare. Once increasingly popular clinical assay is single-cell sequencing, a technology that measures the expression of all genes in individual cells from a patient tissue sample or blood draw. PROCEDURES: We are developing AI software that can annotate information about new single-cell sequencing samples including cell type, donor organ, and developmental time points. Our primary objective is to construct a deep neural network using the Keras machine learning library in R and train this network on data from the Human Cell Atlas CellxGene repository. OUTCOME: Our work has shown successful cell type classification on a large subset of this dataset using GVSU’s High-Performance Computing (HPC) cluster. To enhance model accuracy, we employ hyperparameter tuning. Our short-term goal is to train the model on over 28 million single-cell transcriptomes, spanning hundreds of cell types, developmental time points, and donor organs, which will uncover comprehensive insights into the underlying complexity of clinical samples. To do this, we are incorporating a multi-channel output architecture to allow for simultaneous classification of multiple output modalities in parallel from a single shared neural network. IMPACT: This research will enable better understanding of the etiology of complex human diseases and help personalize medical practice. Our long-term goal is to personalize treatment of rare diseases by providing detailed understanding of patient tissue differences compared to healthy human donors.