Learning the Fragment Size Distribution in Liquid Biopsy Sequencing

Abstract

Cancer as a disease affects thousands of patients every year. Earlier, cancer was analyzed through tissue biopsies derived during surgery. However, due to improved sequencing methods, liquid biopsies have become more common, as these are convenient and provide the opportunity to monitor cancer evolution in real time. The project aim was to employ computational methods to analyze fragment length distributions from liquid biopsies by finding characteristics related to cancer and then filter appropriately. We carried out the project by utilizing a combination of machine learning and statistical learning. The machine learning models were performed for two labels, where one was based on purity and one was generated through a minimalistic cell death model. We found characteristic information linked to cancer by evaluating the models based on feature importance. The project resulted in one label sufficient enough for usage, which led to several models outperforming relevant baselines. As the models were somewhat flawed due to comprised results and insufficient data, no filtering could be made with the guarantee of only removing healthy data. However, we still managed to find characteristic features because of synergistic results across the models.