Einladung zum Statusvortrag im Promotionsver­fahren von Neringa Jurenaite

Abstract: 
The field of biological representation learning, in particular, application of biomedical language models on a variety of downstream tasks, has ballooned in the last couple years, facilitated by the exponential growth in the amount of heterogeneous multi-omics data. This has become of particular interest in clinical application, such as in understanding the underlying genomic changes in cancer, however this research shift has lacked proper study of model efficiency. Therefore, my status talk delves into the understanding and representation of these data structures for meaningful clinical insight extraction. Furthermore, I will discuss the trade-off between model performance and hardware efficiency of foundation models trained on human DNA sequences. Lastly, a comparative analysis is conducted of data loading strategies, architectural design, hyper-parameter tuning and parallel efficiency of these models with the overarching goal of furthering clinical applicability of these approaches.