Deep neural architectures for highly imbalanced data in bioinformatics

Abstract

In the postgenome era, many problems in bioinfor-matics have arisen due to the generation of large amounts ofimbalanced data. In particular, the computational classificationof precursor microRNA (pre-miRNA) involves a high imbalancein the classes. For this task, a classifier is trained to identify RNAsequences having the highest chance of being miRNA precursors.The big issue is that well-known pre-miRNAs are usually just afew in comparison to the hundreds of thousands of candidatesequences in a genome, which results in highly imbalanceddata. This imbalance has a strong influence on most standardclassifiers and, if not properly addressed, the classifier is not ableto work properly in a real-life scenario. This work provides acomparative assessment of recent deep neural architectures fordealing with the large imbalanced data issue in the classificationof pre-miRNAs. We present and analyze recent architectures ina benchmark framework with genomes of animals and plants,with increasing imbalance ratios up to 1:2000. We also propose anew graphical way for comparing classifiers performance in thecontext of high-class imbalance. The comparative results obtainedshow that, at a very high imbalance, deep belief neural networkscan provide the best performance.