Prediction Of LncRNA-protein Interactions Based On Graph Autoencoders

Noncoding RNA(nc RNA)is a kind of RNA that do not obey the biological central dogma.Due to the fact that nc RNAs cannot code for proteins to realize gene expression directly,it had been considered to be as the dark matter in the past few decades.Since1990 s,benefitting from the fast development of modern biomedicine,more and more new nc RNAs have been found,which caused researchers’ great enthusiasm to study the various molecular functions of nc RNAs.Long noncoding RNA(lnc RNA)is a kind of noncoding RNA with a length of more than 200 nucleotide units.Numerous researches have proved that although lnc RNAs cannot be directly translated into proteins,lnc RNAs still play an important role in human growth process.Furthermore,one of the most important ways for lnc RNA to perform its biological function is interacting with relevant proteins,and then regulate the gene expression.The research of non-coding RNA and protein interactions is crucial for many important cellular processes.The traditional methods used to detect lnc RNA-protein interactions are often realized through large-scale biological experiments.However,these methods often require a lot of time and material costs.In recent years,with the rapid development of computer technology,many calculation models based on statistical science,machine learning have been proposed to predict potential LPI.However,their comprehensive performance such as accuracy and efficiency still need to be improved.Based on this basis,this thesis proposes a novel deep learning method named as combined graph auto-encoders(LPICGAE)to predict potential lnc RNA-protein interactions.In a short word,at first,LPICGAE apply variational graph auto-encoder to learn the low dimensional representations from the high-dimensional features of lnc RNAs and proteins.Then the graph auto-encoder is used to reconstruct the adjacency matrix for inferring potential interactions between lnc RNAs and proteins.Finally,LPICGAE minimize the loss of the two processes alternately to gain the final predicted interaction matrix.This thesis compares the prediction result of LPICGAE with six state-of-the-art lnc RNA-protein interaction prediction methods under the same dataset.The result in 5-fold cross-validation experiment illustrates that LPICGAE achieves a AUC of 0.974 and a ACC of 0.985,which is better than all the comparison methods.To verify the robustness of LPICGAE,this thesis also tests LPICGAE on an external validation dataset.In can be concluded from the results that LPICGAE can help researchers to gain more potential relationships between lnc RNAs and proteins effectively.