Item Details

title

Classification of Heterogeneous Datasets of Single Cell RNA Sequencing Experiments using Deep Learning

author

Bhandari, Sapan

abstract

With the increasing adoption of single-cell RNA sequencing for profiling gene expression at a high resolution, there is a need for accurate and efficient methods for the prediction of cell types. To date, unsupervised and supervised machine learning have been used to identify cell types in the high-dimensional, zero-inflated experimental data sets. This research focused on performing a comprehensive and reproducible evaluation of the performance of several supervised deep learning networks for cell-type prediction. Overall, 19 benchmark data sets, with diverse characteristics, have been collected, preprocessed and annotated. Among them, one benchmark was constructed by integrating heterogeneous experimental data. Nine deep learning networks were designed, implemented and validated using 19 benchmarks. Results demonstrate that the feed forward neural network architecture achieved the best performance in cross-validation and hold-out validation experiments. Feed forward neural network obtained an average accuracy of 88% for multi-class and average auc score of 0.884 for binary-class benchmarks. Further insights were derived about the impact of training size dependency and sensitivity to changes in batch sizes for the proposed feed forward neural network.

subject

Conv1D

Deep Learning

FFNN

Machine Learning

scRNA

XGBoost

contributor

Khuri, Natalia (committee chair)

Thomas, Stan (committee member)

Pauca, Victor Paul (committee member)

date

2021-06-03T08:36:07Z (accessioned)

2022-06-02T08:30:14Z (available)

2021 (issued)

degree

Computer Science (discipline)

embargo

2022-06-02 (terms)

identifier

http://hdl.handle.net/10339/98806 (uri)

language

en (iso)

publisher

Wake Forest University

type

Thesis