The University of Edinburgh
Network-based embedding methods for multi-omics data analysis
Abstract
dc:description.abstractThe development of high-throughput technologies has resulted in a significant increase in data, opening up new opportunities to study and better understand how biological systems dynamically interact. Network analysis, which is based on graph theory, can offer a framework for combining these high-throughput multi-omics datasets and examining interactions inside biological systems. Even though it is beneficial to analyze vast volumes of data within the networks, employing traditional statistical approaches on these large networks can be challenging. Network embedding methodologies can be used to effectively handle the complexity of large biological networks. Network embedding approaches can represent node connectivity in large-scale networks as low dimensional vectors, which minimizes the downstream analysis complexity. These lower-dimensional vectors can then be used in machine learning prediction tasks with a wide range of applications in computational biology and bioinformatics. In recent years, there has been considerable interest in the application of network embedding methods in biological studies. Following a comprehensive review of the uses of network embedding techniques for link prediction tasks in biological networks, this thesis aims to (1) investigate different approaches to the creation of embedding features from biological networks by taking into account the sparsity and incompleteness of biological networks and the trade-off between local and global structure properties in network embedding methods; (2) assess the use of these embeddings in supervised learning link prediction tasks, taking into account the imbalances in class distributions observed in biological networks; (3) investigate approaches to integrate network embedding approaches in order to obtain more comprehensive vector representations from networks. To achieve embeddings from the networks, two novel pipelines have been designed in the context of this thesis; Individual Network Embedding (INE) and Heterogeneous Network Embedding (HeNE). INE applies network embedding methods on each input network individually and then integrates the embeddings achieved from these different networks. In contrast, HeNE initially integrates networks and applies network embedding methods on an integrated network. The embeddings created by these methods have been used in supervised learning models to predict drug-target interaction (DTI), protein-disease interaction (PDI), and miRNA-disease interaction (MDI). The results show that in DTI and PDI, performance is slightly higher when the INE pipeline is applied. The average AUROC curves in DTI are 0.92 and 0.91 in INE and HeNE respectively. INE and HeNE are 0.94 and 0.93 in PDI. In contrast, HeNE is more advantageous in MDI where networks are sparse and network integration results in improved connectivity following the implementation of the HeNE pipeline. The average AUROC curves in MDI are 0.77 and 0.87 in INE and HeNE respectively. The latter part of this thesis focuses on developing ensemble learning methods. Ensemble learning is a framework to integrate several embedding techniques for use in prediction tasks. This framework aims to assess whether combined vector representations from multiple embedding methods offer complementary information with regard to network features and thus, better performance on prediction tasks. In this framework, different network embedding methods can be compared and the effect of sampling methods, feature selection methods, cross-validation type, and cross-validation parameters can be tested on the integrated embeddings. Here, the integration of embeddings is compared to different embedding methods individually in the prediction of DTIs, PDIs, and MDIs. The results show that embedding integration slightly outperforms some of the embedding methods individually, likely due to their ability to obtain more comprehensive information from networks and thus higher prediction performance in link prediction. Furthermore, the results show that eliminating some of the embedding methods from ensemble learning might change the prediction performance. Although variations in the results can be observed, the changes in performance caused by the elimination of these network embeddings are not significant. The studies conducted in this thesis contribute to examining the application of network embedding methods on biological networks to reduce the analysis complexity of large networks. The results show that the performance following the INE and HeNE pipelines is dependent on the task and networks that are used. In conclusion, in future studies, to find the approach with the highest performance both pipelines should be considered and tested. This thesis additionally concludes that the integration of embedding methods in link prediction tasks may result in the accumulation of more comprehensive network features, which improves link prediction performance.
Degree
thesis:*- Grantor dc:publisher
- The University of Edinburgh
- Year dc:date.issued
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Parvizi, Poorya
- Advisors dc:contributor.advisor
-
- Luz Filho, Saturnino
- Theodoratou, Evropi
Subjects
dc:subject × 9Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Identifier URI
- http://dx.doi.org/10.7488/era/3989