engleski

Selection of the most informative genomic regions for the determination of melanoma cell-of-origin using machine learning methods

Although being the least common form of skin cancer melanoma is by far its deadliest form characterized by high invasiveness and a large metastatic potential. The metastatic potential of melanoma proves to be a problem in a fraction of patients who have no obvious site of primary tumor which is the strongest predictor of tumor behavior. Novel approaches based on machine learning methods have been successful in identifying different cancer types cells-of-origin solely by using cancer passenger mutation and somatic cell epigenomic regional profiles. The aim of this research was to investigate whether melanoma cell-oforigin can be determined with a fraction of genomic regions used in previous studies. We investigated whether the use of principal component analysis can reduce the number of regions that machine learning models need to use to successfully predict melanoma cell-of- origin. Here we show that with even 10% of the profile size used in previous research is enough to predict melanoma cell-of-origin with high accuracy. Moreover, we also found that the best regions have a larger proportional fraction of transcribed sequence and have relatively few of discovered known mutations which are connected with melanoma development. Those findings reveal a potential path to new research into the diagnostic potential of machine learning methods not only for melanoma but for other cancer types.

melanoma ; machine learning ; epigenetics ; bioinformatics ; PCAWG

nije evidentirano