A PiRNA-disease association model incorporating sequence multi-source information with graph convolutional networks

Abstract

There is growing evidence that PIWI-interacting RNA (piRNA) is widely involved in the proliferation, invasion, and metastasis of malignant tumors, playing an important regulatory role in numerous human physiological and pathological processes. Disease-associated piRNAs are expected to be biomarkers and novel therapeutic targets for early diagnosis and prognosis of malignant tumors. However, most previous computational models did not fully focus on the rich representation ability of multiple sources of information in piRNA sequences, which affected their performance in predicting piRNA-disease associations (PDAs). In this work, we propose a model, iSG-PDA, which combines the multi-source information of piRNA sequences with graph convolutional neural networks to predict potential PDAs. More specifically, we first fuse multi-source information including piRNA sequences and disease semantics to enhance the expressiveness of data, then deeply mine the advanced hidden features of PDA using graph convolutional networks, and finally exploit random forest to accurately determine the associations between piRNAs and diseases. In the golden standard dataset, the proposed model realized a prediction accuracy of 91.96% at the AUC of 0.9184. In ablation experiments and comparisons with other different models, iSG-PDA exhibits strong competitiveness. Moreover, the results of the case study indicate that 17 of the top 20 PDAs in the proposed model predictive score were confirmed. These preliminary results reveal that iSG-PDA is an effective computational method for predicting PDAs and can provide reliable disease candidate piRNAs for biological experiments.