Abstract
Background/Objectives: An important challenge in long non-coding RNA (lncRNA)–disease association prediction is prioritizing candidates when neither target node has association records in the training data, even though content features are available. This study aimed to develop and evaluate a content-based model for this both-cold setting. Methods: We developed CDELDA, a dual-encoder model combining lncRNA sequence, structure, and expression features with disease text, semantic similarity, and ontology-structure features. CDELDA was compared with five baseline methods on a normalized benchmark comprising 1662 lncRNAs, 220 diseases, and 9920 positive associations. Evaluation covered warm, lncRNA-cold, disease-cold, and both-cold settings, with association-derived inputs constructed from training labels. Disease-macro area under the receiver operating characteristic curve (AUROC) was the main comparison metric, complemented by average precision and top-ranked retrieval metrics. Feature ablation, disease-feature perturbation, sensitivity analyses, and literature-based case studies further characterized model performance. Results: Under both-cold evaluation, CDELDA achieved a mean disease-macro AUROC of 0.6617, compared with 0.6123 for SIMCLDA, the highest-scoring literature baseline on this metric. Relative performance varied across evaluation settings and metrics. Removing expression features produced the largest decrease in disease-macro AUROC among individual lncRNA feature removals. Reassigning disease features reduced performance, although ranking performance was partly retained. Literature review identified functional or observational support for 14 of 20 reviewed candidate pairs. Conclusions: These findings suggest that CDELDA can support content-based candidate prioritization under both-cold evaluation. The predictions provide candidates for biological follow-up, with disease-specific relevance requiring independent validation.