AWTI-Net Enables Accurate and Interpretable Functional Assessment of Disease-Associated LncRNA Mutations

Long non-coding RNA (lncRNA) variants contribute to human disease, yet systematic evaluation of their functional impact is limited by the lack of high-quality benchmark datasets for computational tool development. To address this, we curated three comprehensive datasets, lncRNAVar-Literature, lncRNAVar-GWAS, and Mpravardb, by integrating disease-associated variants from published studies and genome-wide association analyses with putatively benign variants derived from the 1000 Genomes Project. Using these datasets, we characterized structural, miRNA-targeting, and evolutionary features distinguishing functional from non-functional lncRNA variants. Based on these insights, we developed an adaptive wavelet-Transformer interpretable network (AWTI-Net), an integrative and interpretable framework combining multi-scale wavelet-based sequence encoding, LSTM-driven temporal modeling, a frequency-aware memory Transformer, and an adaptive decision-making mechanism to capture hierarchical lncRNA variant features. AWTI-Net consistently achieved superior performance across accuracy, specificity, sensitivity, and F1 score on both benchmark datasets. Ablation analyses demonstrated complementary contributions of individual modules and synergistic effects of multidimensional features, while entropy-guided path selection enabled biologically interpretable predictions. Across independent evaluations, AWTI-Net outperformed widely used noncoding variant predictors-including CADD, Eigen, GWAVA, DANN, and FATHMM-achieving AUCs of 0.9948 and 0.9852 on lncRNAVar-GWAS and lncRNAVar-Literature, respectively. AWTI-Net was evaluated against 13 mainstream machine learning models and 15 state-of-the-art deep learning frameworks across three independent benchmark datasets, consistently demonstrating superior and more balanced performance across most evaluation metrics compared with competing methods. Importantly, it effectively prioritized hepatocellular carcinoma-associated lncRNA variants, highlighting its potential as a robust and interpretable tool for genetic studies of disease-associated noncoding variation.