Disease gene recognition and editing optimization through knowledge learned from domain feature spaces

Publication Type:
Issue Date:
Full metadata record
This thesis presents computational methods used for the recognition of disease genes and for the optimal design of disease gene CRISPR/Cas9 editing systems. The key innovation in these computational methods is the feature space and characteristics captured from the biology domain knowledge through machine learning algorithms. The disease-gene association prediction problems are studied in Chapters 3-5. Disease gene recognition is a hot topic in various fields, especially in biology, medicine and pharmacology. Non-coding genes, a type of genes without protein products, have been proved to play important roles in disease development. Particularly, the two kinds of non-coding gene products such as microRNA (miRNA) and long non-coding RNA (lncRNA) have caught much attention as they are abundantly expressed in various tissues and frequently interact with other biomolecules, e.g. DNA, RNA and protein. The disease-ncRNA relationships remain largely unknown. Computational methods can immensely help replenish this kind of knowledge. To overcome existing computational methods’ limitations such as significantly relying on network structures and similarity measurements, or lacking reliable negative samples, this thesis presents two novel methods. One is the precomputed kernel matrix support vector machine (SVM) method to predict disease related miRNAs in Chapter 3. The precomputed kernel matrix was built by integrating several kinds of similarities computed with effective characteristics for miRNAs and diseases. The reliable negative samples were collected through analyzing the published array and sequencing data. This binary classification method accurately predicts disease-miRNA associations, which outperforms those state-of-the-art methods. In Chapter 4, the predicted novel disease-miRNA associations were combined with known relationships of diseases, miRNAs and genes to reconstruct a disease-gene-miRNA (DGR) tripartite network. Reliable multi-disease associated co-functional miRNA pairs were extracted from this DGR for cross-disease analysis by defining the co-function score. This not only proves the proposed method’s effectiveness but also contributes to the study of multi-purpose miRNA therapeutics. Another is the bagging SVM-based positive-unlabeled learning method for disease-lncRNA prioritizing that is described in Chapter 5. It creatively characterized a disease with its related genes’ chromosome distribution and pathway enrichment properties. The disease-lncRNA pairs were represented as novel feature vectors to train the bagging SVM for predicting disease-lncRNA associations. This novel representation contributes to the superior performance of the proposed method in disease-lncRNA prediction even when a given disease has no currently recognized lncRNA genes. After confirming the relationships between genes and diseases, one of the most difficult tasks is to investigate the molecular mechanism and treatment of the diseases considering their related genes. The CRISPR/Cas9 system is a promising gene editing tool for operating the genes to achieve the goals of disease-gene function clarification and genetic disease curing. Designing an optimal CRISPR/Cas9 system can not only improve its editing efficiency but also reduce its side effect, i.e. off-target editing. Furthermore, the off-target site detection problem involves genome-wide sequence observing which makes it a more challenging job. The CRISPR/Cas9 system on-target cutting efficiency prediction and off-target site detection questions are discussed in Chapters 6 and 7 respectively. To accurately measure the CRISPR/Cas9 system’s cutting efficiency, the profiled Markov properties and some cutting position related features were merged into the feature space for representing the single-guide RNAs (sgRNAs). These features were learned by a two-step averaging method where an XGBoost’s predictions and an SVM’s predictions were averaged as the final results. Later performance evaluations and comparisons demonstrate that this method can predict a sgRNA’s cutting efficiency with consistently good performance no matter it is expressed from a U6 promoter in cells or from a T7 promoter in vitro. In the off-target site detection, a sample was defined as an on-target-off-target site sequence pair to turn this problem into a classification issue. Each sample was numerically depicted with the nucleotide composition change features and the mismatch distribution properties. An ensemble classifier was constructed to distinguish real off-target sites and no-editing sites of a given sgRNA. Its excellent performance was confirmed with different test scenarios and case studies.
Please use this identifier to cite or link to this item: