DeepSipred: A deep-learning-based approach on siRNA inhibition prediction

Abstract

Abstract Motivation: The use of exogenous small interfering RNAs (siRNAs) for gene silencing has become a widespread molecular tool for gene function study and new drug identification. Although the pathway of RNAi to mediate gene expression has been widely investigated, the selection of hyperfunctional siRNA with high inhibition remains challenging. Results: In this study, we build a deep-learning-based approach on siRNA inhibition prediction, named DeepSipred. It combines features from sequence context, thermodynamic property, and other expert knowledge together to predict the inhibition more accurately than existing methods. The sequence features from siRNA and local target mRNA are generated via one-hot encoding and pretrained RNA-FM encoding. The convolution layers with multiple kernels in DeepSipred can detect various decisive motifs, which will determine the actual inhibition of siRNA. The thermodynamic features are calculated from Gibbs Free Energy. In addition, the expert knowledge includes those design criteria from previous studies. Benchmarked on large available public datasets, the 10-fold cross-validation results indicate that our predictor achieving the state-of-the-art performance.