engleski

Speed and accuracy benchmarks of large-scale microbial gene function prediction with supervised machine learning

Machine learning approaches for microbial gene function prediction (MGFP) from genome context data are mostly unsupervised [1] and rely on pairwise distances between individual examples arranged into “functional interaction networks” [2]. When supervised approaches were used, most of them typically predicted a limited set of functions and/or used a single-label approach to classification [3, 4], constructing a separate classifier for each function and ignoring the relationships between the functions. Multilabel approaches may perform better, especially those that can exploit the relations between functions readily available in gene function ontologies [5]. Our aim is to compare predictive accuracy and computational efficiency of single vs. multi-label approaches on supervised MGFP. High accuracy is a prerequisite for applying the classifier in real- life tasks, where confidence in predicted functions is of key importance for prioritizing downstream experimental work. Many such predictions have indeed been validated in biological experiments [6, 7]. A lower demand for computational time is of importance when the number of considered functions is high.

gene function prediction, supervised machine learning, bencmark

nije evidentirano