ncRNA-protein interactions generated by negative sample selection

The original datasets are NPInter4158 [1], NPInter10412 [2], RPI7317 [3], RPI2241 [4], and RPI369 [4]. Only positive samples of them were used in our work.We used a different strategy to select more reliable negative samples rather than randomly pairing, which was originally introduced by Zhang et al. in the LPI-CNNCP [5] study.First, we calculated the Smith-Waterman similarity between each pair of proteins. Next, wecalculated interaction scores between each pair of protein and RNA based on the known interaction pairs and protein similarities. Then,we sorted the interaction scores of all pairs in an ascending order. Finally, negative samples were selected sequentially from the head of the sorted list with the same number as positives.[1]H. Zhang, Z. Ming, C. Fan, Q. Zhao, and H. Liu, “A path-based computational model for long non-coding RNA-protein interaction prediction,” Genomics, vol. 112, no. 2, pp. 1754–1760, Mar. 2020, doi: 10.1016/j.ygeno.2019.09.018.[2] J. Yuan, W. Wu, C. Xie, G. Zhao, Y. Zhao, and R. Chen, “NPInter v2.0: an updated database of ncRNA interactions,” Nucl. Acids Res., vol. 42, no. D1, pp. D104–D108, Jan. 2014, doi: 10.1093/nar/gkt1057.[3]X.-N. Fan and S.-W. Zhang, “LPI-BLS: Predicting lncRNA–protein interactions with a broad learning system-based stacked ensemble classifier,” Neurocomputing, vol. 370, pp. 88–93, Dec. 2019, doi: 10.1016/j.neucom.2019.08.084.[4]U. K. Muppirala, V. G. Honavar, and D. Dobbs, “Predicting RNA-Protein Interactions Using Only Sequence Information,” BMC Bioinformatics, vol. 12, no. 1, p. 489, Dec. 2011, doi: 10.1186/1471-2105-12-489.[5] S.-W. Zhang, X.-X. Zhang, X.-N. Fan, and W.-N. Li, “LPI-CNNCP: Prediction of lncRNA-protein interactions by using convolutional neural network with the copy-padding trick,” Analytical Biochemistry, vol. 601, p. 113767, Jul. 2020, doi: 10.1016/j.ab.2020.113767.