Accurate prediction of candidate lncRNAs associated with DNA damage response based on gene expression patterns from graph neural networks.

DNA damage response (DDR) is essential for maintaining genome stability and preventing tumorigenesis. While protein-coding DDR genes have been extensively investigated, long non-coding RNAs (lncRNAs) remain relatively understudied despite the growing evidence of their involvement in DDR. Particularly, it is rather challenging to systematically identify DDR-associated lncRNAs through experimental approaches, which are often time-consuming, labor-intensive, and expensive. Moreover, lncRNAs lack translational open reading frames often targeted by experimental methods. In this study, we have developed a new machine learning approach, GlncDDR, which utilizes graph-based node embedding of gene expression features and supervised learning algorithms to predict candidate lncRNAs associated with DDR. GlncDDR models achieved robust predictive performance with ROC-AUC reaching ∼0.93 on test data. We used the models to predict 1232 candidate lncRNAs, including several known DDR regulators such as JADRR, PINCR, TP53TG1, HOTAIR, MALAT1, ENRICD, and DINOL. Interestingly, 212 of the candidates were found to be located near known DDR genes in the genome, supporting the potential functions of these lncRNAs in DDR. The results demonstrate the effectiveness of predicting DDR-associated lncRNAs based on cancer transcriptomic data and provide valuable targets for exploring the non-coding regulatory landscape of genome stability and cancer drug discovery. The source code and datasets used in the study are available at https://github.com/BioDataLearning/GlncDDR.git.