MLPPF: Multi-Label Prediction of piRNA Functions Based on Pretrained k-mer, Positional Embedding and an Improved TextRNN Model

Abstract

PIWI-interacting RNAs (piRNAs) are a kind of important small non-coding RNAs and play a vital role in maintaining the stability of genome. Previous studies have revealed that piRNAs not only silence transposons, but also mediate the degradation of a large number of mRNAs and lncRNAs. Existing computational models only focus on mRNA-related piRNAs and rarely concentrate on lncRNA-related piRNAs. In this study, we propose a novel method, MLPPF, which is designed for multi-label prediction of piRNA functions based on pretrained k-mer, positional embedding and an improved TextRNN model. First, a benchmark dataset, which contains two types of functional labels, namely mRNA-related and lncRNA-related piRNAs, was constructed by processing piRNA-function-annotated data and sequence data. Moreover, pretrained k-mer embedding fused with positional embedding was applied to get the sequence representation with biological significance. Finally, an improved textRNN model with Bi-GRU and an attention mechanism was employed for implementing the piRNA functional label prediction task. Experiments substantiate that our model can effectively identify the piRNA functional labels, reveal the key factors of its subsequences and be helpful for in-depth investigations into piRNA functions.