Identification of novel lncRNAs potentially associated with diabetic retinopathy based on machine learning prediction.

objectiveDiabetic retinopathy (DR), a neurovascular complication, stands as a major contributor to global vision impairment, substantially affecting patients' quality of life. Increasing evidence suggests that lncRNAs fundamentally contribute to the development of DR; however, their precise molecular functions remain largely unclear. This study aimed to uncover lncRNAs potentially implicated in DR pathogenesis and to lay the groundwork for identifying new therapeutic targets.

methodsMicroarray profiling was performed using vitreous samples from patients with PDR and idiopathic macular hole (IMH). Differentially expressed non-coding RNA transcripts were screened using the criteria of a single isoform, an absolute fold change of ≥1.5, and P < 0.05. Least absolute shrinkage and selection operator (LASSO) regression was then applied to prioritize candidate lncRNAs. The expression of the selected lncRNAs was validated by quantitative real-time PCR (qRT-PCR) in human retinal microvascular endothelial cells (HRMECs) exposed to normal- or high-glucose conditions. For the consistently validated lncRNAs, an exploratory random forest analysis with nested leave-one-out cross-validation was performed, and SHAP values were calculated to assess their relative contributions to model output. In addition, cis- and trans-associated analyses were conducted to explore potential lncRNA-protein-coding gene relationships.

resultsEighty-six lncRNA transcripts were initially identified based on the established criteria. Subsequent LASSO regression analysis refined the list to eight lncRNAs that exhibited significant expression differences between the PDR and IMH groups: RP11-133K1.5, RP11-483C6.1, RP11-91G21.1, AP000487.4, RP11-715F3.2, LOC101927131, RP5-956O18.3, and ATP2B2-IT2. Validation via qRT-PCR in high-glucose-treated HRMECs confirmed that four lncRNAs exhibited expression patterns consistent with the microarray findings-RP11-483C6.1, RP11-91G21.1, and LOC101927131 were downregulated, while ATP2B2-IT2 was upregulated. Among them, exploratory SHAP analysis highlighted ATP2B2-IT2 as a potential DR-associated candidate lncRNA with the highest relative contribution to the model output, warranting further functional investigation. In addition, integrated cis- and trans-associated analyses identified several potential relationships between the candidate lncRNAs and protein-coding genes, providing preliminary insights into their possible regulatory context; however, these associations require further experimental validation.

conclusionVitreous microarray profiling combined with LASSO selection and cellular validation identified several candidate lncRNAs potentially associated with PDR. ATP2B2-IT2 emerged as a prioritized candidate in the exploratory feature-contribution analysis. Given the small discovery cohort and lack of external validation, these findings should be considered preliminary and require confirmation in larger independent cohorts and further functional studies.