AutoML approaches to the identification of novel biomarkers associated with thalassemia

Abstract

Thalassemias are a group of genetic blood disorders in which abnormal hemoglobin production occurs. Currently, there are obstacles in its diagnostic methods and approaches. In addition, its treatment represents a significantly high cost. This work proposes the use of machine learning techniques, and prior knowledge of known genes associated with thalassemia, to find novel biomarkers associated with the disease. This may eventually help in detection efforts. Also, we propose to evaluate automated machine learning (AutoML) approaches as an alternative to using traditional algorithms. The AutoML tools we decided to use were Auto-Sklearn and Tree-based Pipeline Optimization Tool (TPOT). In this way, we synthesize the experience of using these tools and compared their performance against a Support Vector Machine based Model. This was done through a comparison of performance metrics. Finally, we found that TPOT offers certain ease of use, such as the option to export the best pipeline found, as well as an improvement in performance compared to other methods. This opens the possibility to test new configurations of the tool, as well as other AutoML tools.