Details
Original language | English |
---|---|
Qualification | Doctor rerum naturalium |
Awarding Institution | |
Supervised by |
|
Date of Award | 22 Mar 2023 |
Place of Publication | Hannover |
Publication status | Published - 2023 |
Abstract
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
Hannover, 2023. 139 p.
Research output: Thesis › Doctoral thesis
}
TY - BOOK
T1 - Classifying distinct data types: textual streams protein sequences and genomic variants
AU - Melidis P, Damianos
N1 - Doctoral thesis
PY - 2023
Y1 - 2023
N2 - Artificial Intelligence (AI) is an interdisciplinary field combining different research areas with the end goal to automate processes in the everyday life and industry. The fundamental components of AI models are an “intelligent” model and a functional component defined by the end-application. That is, an intelligent model can be a statistical model that can recognize patterns in data instances to distinguish differences in between these instances. For example, if the AI is applied in car manufacturing, based on an image of a part of a car, the model can categorize if the car part is in the front, middle or rear compartment of the car, as a human brain would do. For the same example application, the statistical model informs a mechanical arm, the functional component, for the current car compartment and the arm in turn assembles this compartment, of the car, based on predefined instructions, likely as a human hand would follow human brain neural signals. A crucial step of AI applications is the classification of input instances by the intelligent model. The classification step in the intelligent model pipeline allows the subsequent steps to act in similar fashion for instances belonging to the same category. We define as classification the module of the intelligent model, which categorizes the input instances based on predefined human-expert or data-driven produced patterns of the instances. Irrespectively of the method to find patterns in data, classification is composed of four distinct steps: (i) input representation, (ii) model building (iii) model prediction and (iv) model assessment. Based on these classification steps, we argue that applying classification on distinct data types holds different challenges. In this thesis, I focus on challenges for three distinct classification scenarios: (i) Textual Streams: how to advance the model building step, commonly used for static distribution of data, to classify textual posts with transient data distribution? (ii) Protein Prediction: which biologically meaningful information can be used in the input representation step to overcome the limited training data challenge? (iii) Human Variant Pathogenicity Prediction: how to develop a classification system for functional impact of human variants, by providing standardized and well accepted evidence for the classification outcome and thus enabling the model assessment step? To answer these research questions, I present my contributions in classifying these different types of data: temporalMNB: I adapt the sequential prediction with expert advice paradigm to optimally aggregate complementary distributions to enhance a Naive Bayes model to adapt on drifting distribution of the characteristics of the textual posts. dom2vec: our proposal to learn embedding vectors for the protein domains using self-supervision. Based on the high performance achieved by the dom2vec embeddings in quantitative intrinsic assessment on the captured biological information, I provide example evidence for an analogy between the local linguistic features in natural languages and the domain structure and function information in domain architectures. Last, I describe GenOtoScope bioinformatics software tool to automate standardized evidence-based criteria for pathogenicity impact of variants associated with hearing loss. Finally, to increase the practical use of our last contribution, I develop easy-to-use software interfaces to be used, in research settings, by clinical diagnostics personnel.
AB - Artificial Intelligence (AI) is an interdisciplinary field combining different research areas with the end goal to automate processes in the everyday life and industry. The fundamental components of AI models are an “intelligent” model and a functional component defined by the end-application. That is, an intelligent model can be a statistical model that can recognize patterns in data instances to distinguish differences in between these instances. For example, if the AI is applied in car manufacturing, based on an image of a part of a car, the model can categorize if the car part is in the front, middle or rear compartment of the car, as a human brain would do. For the same example application, the statistical model informs a mechanical arm, the functional component, for the current car compartment and the arm in turn assembles this compartment, of the car, based on predefined instructions, likely as a human hand would follow human brain neural signals. A crucial step of AI applications is the classification of input instances by the intelligent model. The classification step in the intelligent model pipeline allows the subsequent steps to act in similar fashion for instances belonging to the same category. We define as classification the module of the intelligent model, which categorizes the input instances based on predefined human-expert or data-driven produced patterns of the instances. Irrespectively of the method to find patterns in data, classification is composed of four distinct steps: (i) input representation, (ii) model building (iii) model prediction and (iv) model assessment. Based on these classification steps, we argue that applying classification on distinct data types holds different challenges. In this thesis, I focus on challenges for three distinct classification scenarios: (i) Textual Streams: how to advance the model building step, commonly used for static distribution of data, to classify textual posts with transient data distribution? (ii) Protein Prediction: which biologically meaningful information can be used in the input representation step to overcome the limited training data challenge? (iii) Human Variant Pathogenicity Prediction: how to develop a classification system for functional impact of human variants, by providing standardized and well accepted evidence for the classification outcome and thus enabling the model assessment step? To answer these research questions, I present my contributions in classifying these different types of data: temporalMNB: I adapt the sequential prediction with expert advice paradigm to optimally aggregate complementary distributions to enhance a Naive Bayes model to adapt on drifting distribution of the characteristics of the textual posts. dom2vec: our proposal to learn embedding vectors for the protein domains using self-supervision. Based on the high performance achieved by the dom2vec embeddings in quantitative intrinsic assessment on the captured biological information, I provide example evidence for an analogy between the local linguistic features in natural languages and the domain structure and function information in domain architectures. Last, I describe GenOtoScope bioinformatics software tool to automate standardized evidence-based criteria for pathogenicity impact of variants associated with hearing loss. Finally, to increase the practical use of our last contribution, I develop easy-to-use software interfaces to be used, in research settings, by clinical diagnostics personnel.
U2 - 10.15488/13514
DO - 10.15488/13514
M3 - Doctoral thesis
CY - Hannover
ER -