http://theses.ncl.ac.uk/jspui/handle/10443/5223 2026-05-19T15:40:06Z http://theses.ncl.ac.uk/jspui/handle/10443/6774 Title: Enhanced predictive models for macular hole surgery outcomes Authors: Kucukgoz, Burak Abstract: This thesis presents research work conducted in the field of retinal image analysis. More specifi cally, the work is directed at the employment of deep learning (DL) based image informatics for the analysis of diverse real world phenomena where features of interest are very difficult to distinguish. The evaluation of idiopathic full-thickness macular holes (MHs) holds critical clinical importance as MHs represent one of the strongest predictors of surgical success, influ encing both anatomical closure and functional visual recovery– a key motivation for developing robust deep learning frameworks to quantify their characteristics and predict postoperative out comes. In this context, three distinct parts to retinal image analysis are proposed. The first part addresses critical research questions on the quantitative assessment of MH, the role of DL in postoperative visual acuity (VA) prediction, the integration of automated optical coherence tomography (OCT) analysis for clinical decision-making, and the potential of DL models to improve diagnostic accuracy and support clinical practices. Hence, this part presents a compre hensive image informatics framework to create a high-quality spectral-domain OCT (SD-OCT) image dataset, providing a robust DL-based predictive model of VA in patients following surgery with MH and presenting an automated solution for non-standardised SD-OCT datasets. The imaging data undergoes preprocessing, quality assurance, and anomaly detection procedures. Seven state-of-the-art DL predictive models are then designed, implemented, trained, and tested with multiple two-dimensional (2D) input channels on the SD-OCT dataset. The models are quantitatively compared using four evaluation metrics. The method concludes the impact of the following surgery by predicting VA. Overall, the obtained results confirm that the fully automated approach with input from seven central SD-OCT images from each patient may robustly predict VA measurements using a high-quality SD-OCT image dataset. Following this, three-dimensional (3D) convolutional neural networks are integrated to train the model. 3D networks generally outperformed the 2D networks in some evaluation metrics; however, it came with the sacrifice of significantly more computational complexity. The second part identifies key research questions related to common sources of uncertainty in OCT images and proposes an effective method for representing and quantifying this uncertainty in DL-based predictive models. Furthermore, the study compares the proposed UQ method with existing approaches. In this context, the study highlights the significance of uncertainty, especially in dealing with the SD-OCT images. Predicting postoperative VA through DL models is crucial for decision-making and patient advisement, though their black-box behaviour is opaque to users and uncertainty associated with their predictions is not typically stated, leading to a lack of trust among clinicians and patients. To meet this need, an uncertainty-aware regression model is introduced for predicting postoperative VA using 3D SD-OCT images. The model not only x predicts VA post-surgery but also quantifies the associated uncertainty, enhancing reliability and trustworthiness. Qualitative evaluation shows that the proposed model outperforms commonly used methods in terms of prediction accuracy and reliability, demonstrating robust performance on out-of-sample data, including low-quality images and previously unseen instances. This makes the model a promising tool for clinical settings, improving the reliability of DL models in predicting VA. The third is the segmentation of the retinal external limiting membrane layer, where any disruptions in this layer are associated with worse visual outcomes in patients with idiopathic full-thickness MHs. Precise image-wise binary annotations are used to segment the retinal external limiting membrane (ELM) layer. Finally, qualitative and quantitative results are systematically compared with seven state-of-the-art DL-based segmentation methods to identify the ELM layer with an automated system. Additionally, it examines the feasibility of integrating automated ELM layer segmentation into clinical workflows while incorporating the latest advancements in DL-based ELM detection. The results confirm the efficacy of DL in retinal image analysis, providing a foundation for future enhancements in clinical applications. Future work will explore enhancing the models’ performance and efficiency, and extending the approach to other retinal conditions. Keywords: Image Analysis, Machine Learning, Deep Learning, Visual Acuity Measurement, Optical Coherence Tomography Description: Ph. D. Thesis. 2025-01-01T00:00:00Z http://theses.ncl.ac.uk/jspui/handle/10443/6762 Title: Advancing scientific knowledge representation : standardisation and integration in tolerogenic therapies Authors: Sahar, Ayesha Abstract: In this thesis, we use data integration and analysis methods and examine the impact of data standardisation to enhance our understanding of tolerogenic dendritic cell (tolDC) therapies. Standardisation and structuring of the data are extremely valuable for it to be useful and accessible. Emerging biological fields face unique difficulties, including limited data availability, a lack of standardisation and challenges in knowledge management from different studies due to varied methodologies. These issues demand the development and application of specialised techniques and strategies tailored to their specific data handling and management needs. This thesis focuses on one such emerging field, “tolerogenic Dendritic Cell Therapy”, which has demonstrated significant potential. Like all biomedical experiments, developing these therapies involves several crucial steps that must be well-documented for comparison and replication purposes. Reporting frameworks, like Minimum Information Models can aid in standardising these descriptions; Minimum Information about Tolerogenic AntigenPresenting cells (MITAP) was created in 2016 in this field for this purpose. We evaluate MITAP’s impact on the field of tolDC therapies by analysing a selection of literature. We found that MITAP is utilised in a minority of relevant papers (14%), but where it is applied, there is slightly more metadata available. This suggests that while MITAP has had some success, further efforts are needed for standardised reporting to become widespread in the discipline. In order to further aid the comparison, re-purposing and re-use of data about tolDC therapies, we built a method to identify and integrate the most significant information related to tolerogenic dendritic cell therapies into a knowledge graph structure. A key aspect of the knowledge graph is ensuring that the merged data is relevant to the field. We employ knowledge extraction techniques to identify and collect relevant information from research articles, integrating this with publicly available datasets to enrich the knowledge base. We successfully embedded this data into a comprehensive knowledge graph comprising 120k entities extracted from full-text articles and additional integration of 92k relationships from other relevant databases. The use of knowledge extraction techniques from research articles ensured the relevance of the integrated data to the field. It also allowed us to gain more insights from publications with unpublished experimental data, as shown in the example queries. This knowledge graph can act as a base for the generation of further hypotheses as well as a database for the storage and retrieval of relevant information about tolDC therapies. Having built the knowledge graph our focus shifts to considering queries about the tolDC therapies that give us a better understanding of the degree of standardisation, about the underlying biology and the social environment in the field. We formulated diverse queries encompassing heterogeneity concerns. The results demonstrated the effectiveness of tolKG in promptly addressing these queries, a task that would either necessitate specialised expertise or significant manual scrutiny if pursued conventionally. Through the utilisation of tolKG, we streamline tasks such as comparison and analysis and even facilitate the generation of novel hypotheses. In summary, we found that a knowledge graph is an effective way to integrate data. Moreover, the addition of data from the literature makes it more meaningful, especially for emerging fields where there is a lack of experimental data sharing. Text mining from literature enables the extraction of more relationships that are specific to a field. As a result, it can help to perform an effective analysis and comparison of the tolDC therapy field. Together, this work helps establish the groundwork for applying data science methods in tolDC therapies making several kinds of comparisons possible which are not possible without it. The methodologies employed are specifically tailored to the data sources of tolDC therapies. Nonetheless, these strategies are not restricted to this particular domain; they primarily depend on the input data sources, which makes them usable in other areas of biology as well. Description: PhD Thesis 2025-01-01T00:00:00Z http://theses.ncl.ac.uk/jspui/handle/10443/6731 Title: Automated design, build, test, learn workflows to engineer synthetic genetic networks Authors: Vidal Peña, Gonzalo Andrés Abstract: Synthetic biology is an interdisciplinary field that pursues the engineering of biological systems. The design, build, test, learn (DBTL) cycle is at the core of engineering disciplines and is iterated until a desired goal is achieved. Synthetic biology is still defining abstractions, standards and developing a software ecosystem to iterate the DBTL cycle. The aim is to work in a similar way as other engineering disciplines, making designs with a computational aided design (CAD) tool that can simulate the expected behaviour of the designed biological system, and that can communicate to build tools to create a physical implementation of the biological system. After the biological system is built, it is tested by taking measurements of its behaviour. The test has to be automated, calibrated and standardised to get high quantity and quality data that can inform the learn stage properly. Given the diversity of synthetic biology and its applications the DBTL cycle could have different needs when the researcher needs to engineer a genetic network, a metabolic pathways, a strain or a protein, among others. The focus of this work is in creating DBTL cycle workflows for engineering synthetic genetic network dynamics, because it allows to control the logic of a system and how that logic state is reached and maintained over time with direct applications in biochemical production, drug dosage, and the study of pattern formation and developmental biology. Existing tools for engineering genetic network dynamics do not cover the whole DBTL cycle and lack connections, leaving several gaps. Most tools do not use standardised inputs and outputs hindering the connectivity between tools and slowing the research process. To iterate faster through the DBTL cycle it has to be closed and automated by leveraging software tools and liquid handling robots. Software tools have to be compatible with standards to make them useful and accessible for the community, promoting the use of best practices. The workflow has to be flexible to accommodate different needs and resources, to be used for researchers without a wetlab, with non-automated wetlab and with lab automation. Here I have created a set of software tools tackling different DBTL cycle stages that are modular and leverage standards to connect and automate the DBTL cycle for genetic network engineering. The workflows developed in this work provides novel teaching and research tools available for different needs. Description: Ph. D. Thesis. 2025-01-01T00:00:00Z http://theses.ncl.ac.uk/jspui/handle/10443/6729 Title: Computational Approaches to Drug Repurposing Through Probabilistic Functional Integration of Disease-Gene networks and Graph Neural Networks Authors: Alsobhe, Aoesha Abstract: Drug discovery is a time-consuming, costly, high-risk, and complex process. An alternative to traditional drug development is drug repurposing, which aims to find new uses for existing drugs. This approach significantly reduces time and cost, as much of the safety evaluation has already been completed. Computational approaches to drug repurposing help generate hypotheses about potential drug-disease indications, which can later be validated experimentally in the lab. Network integration is a common computational technique in drug repurposing applications. These approaches combine multiple diverse data sources into a single heterogeneous biomedical integrated network. Such networks combine various types of biological data, including drugs, diseases, genes, and proteins, into a unified framework where biomedical entities are represented as nodes and their interactions as edges. Integrating diverse data sources is essential to gain a comprehensive picture of interconnected biological entities, which can then be mined to infer new hypotheses about drug repurposing opportunities. The quality of these integrated networks is highly dependent on the experimental data they include. However, biomedical data is often noisy and incomplete, leading to a high rate of false results in existing networks. Therefore, there is an important need for methods to reduce noise during network integration. One proposed technique to produce accurate integrated networks is Probabilistic Functional Integrated Networks (PFINs), which assess data quality and generate confidence scores to filter out low-quality data before mining these networks for drug repurposing opportunities. Disease-Gene Association (DGA) networks, where nodes represent diseases and genes and edges represent their associations, are the major building blocks for most biomedical integrated networks used in drug repurposing applications. Unfortunately, many available DGA networks contain a high rate of false results due to the quality of the biomedical data, which faces numerous challenges, including incorrect entries, missing values, inconsistencies, duplication, and various forms of bias. For instance, high-throughput experimental studies, which are commonly used to generate biological data, often produce incomplete and noisy data containing both false positives and false negatives. Although methods exist to score the confidence of DGAs, they are often unreliable. Many of these scoring approaches rely on heuristic strategies that do not assess data quality prior to integration. For example, they often overlook the impact of duplicated data, which can artificially inflate confidence scores and distort the strength of associations. To address this gap, we investigated the applicability of PFINs to DGA networks by researching and developing novel strategies to build and evaluate DGA PFINs. These accurate integrated DGA networks can be employed in various computational drug repurposing applications, including deep learning techniques. Deep learning has become the leading technique in most in silico applications for drug repurposing. Among deep learning methods, Graph Neural Networks (GNNs) have gained considerable attention due to their ability to learn complex relationships between drugs and related biological entities from heterogeneous biomedical integrated networks. Existing GNN applications in drug repurposing often overlook important aspects of data quality, such as noise and incompleteness. Given that the performance of GNNs is highly dependent on the quality of the integrated networks used for training, incorporating PFINs with GNNs could enhance their performance by reducing noise during network integration. To address these issues, we investigated the impact of incorporating the PFINs approach within GNNs on their performance. The constructed DGA PFIN was integrated with an existing network and used to train GNN models. Another factor impacting the performance of GNNs, beyond data quality, is the lack of diverse data types in the integrated networks. Most existing GNN approaches are trained on networks with a limited number of node and edge types, often ignoring node features in the training process. We explored the impact of adding various types of nodes and edges to the integrated networks on GNN performance, as well as incorporating node features in the training process. The results showed that the performance of GNN models improved by incorporating these additional types of nodes and edges into the training networks. Furthermore, the proposed GNN models demonstrated significant enhancement by incorporating node features. Finally, the proposed GNN models were employed to predict drug-disease indications, and these predictions were validated and supported by the literature. Description: PhD Thesis 2025-01-01T00:00:00Z