Learning distributed representations of RNA and protein sequences and its application for predicting lncRNA-protein interactions

The long noncoding RNAs (lncRNAs) are ubiquitous in organisms and play crucial role in a variety of biological processes and complex diseases. Emerging evidences suggest that lncRNAs interact with corresponding proteins to perform their regulatory functions. Therefore, identifying interacting lncRNA-protein pairs is the first step in understanding the function and mechanism of lncRNA. Since it is time-consuming and expensive to determine lncRNA-protein interactions by high-throughput experiments, more robust and accurate computational methods need to be developed. In this study, we developed a new sequence distributed representation learning based method for potential lncRNA-Protein Interactions Prediction, named LPI-Pred, which is inspired by the similarity between natural language and biological sequences. More specifically, lncRNA and protein sequences were divided into k-mer segmentation, which can be regard as “word” in natural language processing. Then, we trained out the RNA2vec and Pro2vec model using word2vec and human genome-wide lncRNA and protein sequences to mine distribution representation of RNA and protein. Then, the dimension of complex features is reduced by using feature selection based on Gini information impurity measure. Finally, these discriminative features are used to train a Random Forest classifier to predict lncRNA-protein interactions. Five-fold cross-validation was adopted to evaluate the performance of LPI-Pred on three benchmark datasets, including RPI369, RPI488 and RPI2241. The results demonstrate that LPI-Pred can be a useful tool to provide reliable guidance for biological research.