MSADN: A Method for Predicting ncRNA Family Based on Deep Learning with Chemical Properties

ncRNAs are vital biomolecules critically involved in chromatin architectural and gene regulation and signal transduction. The prediction of complex and diverse ncRNAs is needed to understand their functions, but accurate prediction of ncRNA families remains difficult. Current computational approaches are contraindicated to use without sequence or structural information, often limiting accuracy, and are costly and time consuming. To address this, we propose MSADN, a novel computational method that combines encoding scheme (NCP-ND) to capture the chemical properties and positional density of nucleotides within ncRNA sequences along with a deep learning framework incorporating multi scale attention (attention mechanism that assigns attention weight to the sequence regions and multi scale information via convolutional layers with different kernel size), bidirectional LSTM (type of recurrent neural network for processing sequence data in both forward and backward direction to capture contextual dependencies) and MSADN achieved state-of-the-art performance on public datasets nRC and NCY, with up to 6% and 2% improvement in precision, recall, f-score, MCC respectively over previous methods. Moreover, its framework can be improved in other aspects by incorporating profound NLP technologies, like pre-trained language models, to discover intricate pattern in ncRNA sequences. These advances can enhance prediction accuracy, uncover new insights into ncRNA mechanism and ultimately aid in biomarker discovery and RNA targeted therapeutics.