Predicting ligand-receptor interactions based on LSTM network with the attention mechanism and its application on cell-cell communication inference

Abstract Cell-cell communication (CCC) inference provides many important information for deciphering the mechanisms of various life processes. With the development of single-cell RNA sequencing (scRNA-seq), multiple computational tools have been developed to predict CCC based on scRNA-seq data. These tools significantly improved the prediction of CCC which is mediated by ligand-receptor interactions (LRIs). In this study, we propose an LRI-mediated CCC inference method called CellAL. CellAL mainly consists of two procedures: LRI prediction by feature extraction via BioTriangle, feature selection through XGBoost, and classification using LSTM network with the attention mechanism, along with CCC inference based on LRI filtering, CCC inference and CCC network visualization. The LRI prediction performance of CellAL significantly surpassed four classical methods (i.e., ACT-SVM, decision tree, PPISP-XGBoost, and PIPR) on four LRI datasets. The identified LRIs by CellAL were validated by counting the overlapping LRIs between CellAL and four other CCC tools (i.e., CellChat, Connectome, CytoTalk, and NATMI). After that, CellAL was applied to infer CCC within melanoma and the predictions were further validated by comparing with three other representative CCC tools (i.e., CellPhoneDB, iTALK, and CellChat). The results demonstrated that caner-associated fibroblasts could have signaling with melanoma cells, which was the same as those predicted by iTALK and CellChat. Finally, the CCC network was visualized using network view and heatmap view. CellAL provides a reference for CCC prediction from single-cell resolution.