| Abstract |
Since the beginning of humanity, the mankind struggles with diseases whose mechanisms are hard to be deciphered. Diseases such as cancer and dementia are the open wounds of the civilization of our time, but today, humanity has an advantage that never had before: the aid of smart machines. Machine learning (ML), an artificial intelligence based field, aims at the creation of systems capable of learning and excellence in a given task. Scientists that work within the field of bioinformatics, turn this task into biomarker discovery, patient classification or time-to event predictions. Besides the advances in the field of ML, there are also advances in the field of Biology and Medicine at the same time. The problem is that sometimes, the advances of these two fields are not synchronized. One of the biggest breakthroughs of our era, is the single-cell technologies, that can measure multi-parameters of one cell at a time, for up to thousands of cells from a given sample, in just one run. The information provided by a single-cell experiment is of great interest and can empower our knowledge about many abnormal conditions in the human body and beyond, but ML lacks the needed algorithms and approaches that could truly understand and model single-cell data. There are some approaches that have emerged in the past few years, each having its advantages and disadvantages in comparison to others, but the community of scientists that try to create and advance them is in its infancy and there is a need for new proposals for sure.
In this thesis, a new approach for single-cell classification tasks is added to the list: an assumption free approach that compares the matrices themselves through the dissimilarities of their distributions. Two algorithms are proposed, a KNN like one that uses Maximun Mean Discrepancy (MMD) as its distance metric, and Kernel Based Custom SVMs (KBCsvm), which can be thought as the advancement of the former. KBCsvm is a supervised approach that combines both the generative and discriminative natures of ML to create a hybrid model, that uses MMD, kernels and Support Vector Machines (SVMs) in its core. The validation of these methods and their comparison with other approaches took place with the use of 8 datasets, and showed that the algorithms perform equally or, in some cases, even better than some existing solutions. In the cases where other algorithms may outperform them, MMD-KNN and KBCsvm may offer better interpretability, assumption-free calculations and lesser human intervention, plus the much needed ability to be extended and adapted to different environments and tasks such as regression and biomarker discovery, following the steps of the state of the art algorithms emerged in the past five years.
|
Search
