Integrative interpretable learning reveals shared patterns of epitranscriptomic regulation across multiple cancer types

Cancer is a complex set of diseases caused by the dysregulation of cell proliferation, differentiation, and apoptosis, ultimately leading to malignant tumor development and metastasis. Recent studies have revealed that dysregulation of N4-acetylcytidine (ac⁴C) acetylation is associated with enhanced metastatic potential and tumor progression in various cancer types. However, it remains unclear whether diverse cancer types share common epitranscriptomic regulatory patterns or engage in interconnected networks. In this study, leveraging an extensive collection of 88 ac⁴C epitranscriptome datasets profiled across multiple cancer types and normal conditions, we developed the first ac⁴C pan-cancer model based on an ensemble strategy that combines classical machine learning algorithms and a deep learning-based transformer architecture. Our interpretable model uncovered, for the first time, the shared epitranscriptomic patterns of dysregulated ac⁴C across diverse cancers, particularly associated with low GC-content region of 3’UTR and internal 3’UTR splicing. Furthermore, we discovered a set of candidate ac⁴C-mediated genes that may function in cancers through epitranscriptomic regulation, including three novel ac⁴C-mediated candidates (<i>SMARCD1</i>, <i>SENP5</i>, <i>RNF207</i>) that exhibit consistently dysregulated ac⁴C level, expression patterns, and are associated with poor clinical prognosis across cancers. Taken together, our findings highlight the importance of a comprehensive characterization of ac⁴C epitranscriptome in pan-cancer landscape, with potential implications for developing RNA modification-based biomarkers and therapeutic strategies.