Abstract
AbstractWith recent advances in DNA sequencing technologies, researchers are able to acquire increasingly larger volumes of genomic datasets, enabling the training of powerful models for downstream genomic tasks. However, genome scale dataset often contain many missing values, decreasing the accuracy and power in drawing robust conclusions drawn in genomic analysis. Consequently, imputation of missing information by statistical and machine learning methods has become important. We show that the current state-of-the-art can be advanced significantly by applying a novel variation of the Transformer architecture, called Split-Transformer Impute (STI), coupled with improved preprocessing of data input into deep learning models. We performed extensive experiments to benchmark STI against existing methods using resequencing datasets from human 1000 Genomes Project and yeast genomes. Results establish superior performance of our new methods compared to competing genotype imputation methods in terms of accuracy and imputation quality score in the benchmark datasets.
Publisher
Cold Spring Harbor Laboratory
Cited by
2 articles.
订阅此论文施引文献
订阅此论文施引文献,注册后可以免费订阅5篇论文的施引文献,订阅后可以查看论文全部施引文献