Data biases in genomics

Machine learning (ML) is developing into an inherent part of genomic research due to the ever-increasing amounts of genomic data. However, data-driven algorithms are strongly dependent on good quality and representative data, which can be problematic in genomics due to various reasons. One of these reasons is data biases-flawed or incomplete data often containing systematic errors that compromise its representativeness. In this review, we examine different categories of data biases in genomics and translate them into the framework of general ML. We give examples of different types of biases present in widely used databases such as NCBI ClinVar and gnomAD and illustrate how data biases can influence model performance in assorted studies.