Repository logo

Integrating biclustering techniques with de novo gene regulatory network discovery using RNA-seq from skeletal tissues



Journal Title

Journal ISSN

Volume Title






Degree Level



In order to improve upon stem cell therapy for osteoarthritis, it is necessary to understand the molecular and cellular processes behind bone development and the differences from cartilage formation. To further elucidate these processes would provide a means to analyze the relatedness of bone and cartilage tissue by determining genes that are expressed and regulated for stem cells to differentiate into skeletal tissues. It would also contribute to the classification of differences in normal skeletogenesis and degenerative conditions involving these tissues. The three predominant skeletal tissues of interest are bone, immature cartilage and mature cartilage. Analysis of the transcriptome of these skeletal tissues using RNA-seq technology was performed using differential expression, clustering and biclustering algorithms, to detect similarly expressed genes, which provides evidence for genes potentially interacting together to produce a particular phenotype. Identifying key regulators in the gene regulatory networks (GRNs) driving cartilage and bone development and the differences in the GRNs they drive will facilitate a means to make comparisons between the tissues at the transcriptomic level. Due to a small number of available samples for gene expression data in bone, immature and mature cartilage, it is necessary to determine how the number of samples influences the ability to make accurate GRN predictions. Machine learning techniques for GRN prediction that can incorporate multiple data types have not been well evaluated for complex organisms, nor has RNA-seq data been used often for evaluating these methods. Therefore, techniques identified to work well with microarray data were applied to RNA-seq data from mouse embryonic stem cells, where more samples are available for evaluation compared to the skeletal tissue RNA-seq samples. The RNA-seq data was combined with ChIP-seq data to determine if the machine learning methods outperform simple, correlation-based methods that have been evaluated using RNA-seq data alone. Two of the best performing GRN prediction algorithms from previous large-scale evaluations, which are incapable of incorporating data beyond expression data, were used as a baseline to determine if the addition of multiple data types could help reduce the number of gene expression samples. It was also necessary to identify a biclustering algorithm that could identify potentially biologically relevant modules. Publicly available ChIP-seq and RNA-seq samples from embryonic stem cells were used to measure the performance and consistency of each method, as there was a well-established network in mouse embryonic stem cells to compare results. The methods were then compared to cMonkey2, a biclustering method used in conjunction with ChIP-seq for two important transcription factors in the embryonic stem cell network. This was done to determine if any of these GRN prediction methods could potentially use the small number of skeletal tissue samples available to determine transcription factors orchestrating the expression of other genes driving cartilage and bone formation. Using the embryonic stem cell RNA-seq samples, it was found that sample size, if above 10, does not have a significant impact on the number of true positives in the top predicted interactions. Random forest methods outperform correlation-based methods when using RNA-seq, with area under ROC (AUROC) for evaluation, but the number of true positive interactions predicted when compared to a literature network were similar when using a strict cut-off. Using a limited set of ChIP-seq data was found to not improve the confidence in the transcription factor interactions and had no obvious affect on biclustering results. Correlation-based methods are likely the safest option when based on consistency of the results over multiple runs, but there is still the challenge of determining an appropriate cut-off to the predictions. To predict the skeletal tissue GRNs, cMonkey was used as an initial feature selection method to identify important genes in skeletal tissues and compared with other biclustering methods that do not use ChIP-seq. The predicted skeletal tissue GRNs will be utilized in future analyses of skeletal tissues, focussing on the evolutionary relationship between the GRNs driving skeletal tissue development.



Gene Regulatory Networks, Biclustering, Skeletal Tissues



Master of Science (M.Sc.)


Computer Science


Computer Science


Part Of