Abstract
G-quadruplexes (G4) are prevalent non-B DNA structures playing crucial biological roles in cells. Although experimental technologies for G4 identification in vitro and in vivo are advancing, computational prediction methods are increasingly preferred for their efficiency, convenience, and cost effectiveness. However, existing tools primarily perform in vitro G4 predictions that lack cell-specific information and are often non-genome-wide, with high-performance genome-wide in vivo cell-specific G4 prediction models still lacking. Here, we present G4Beacon2, a genome-wide cell-specific G4 prediction model based on multilevel ensemble learning. G4Beacon2 leverages DNABERT2 to capture semantic information from DNA sequences, normalizes chromatin accessibility data using Z-score, and establishes a three-level ensemble framework to achieve accurate in vivo G4 predictions across the genome. Across the evaluated intra-cell-line and cross-cell-line tests, G4Beacon2 shows consistently favorable performance. Exploratory analyses further indicate that mouse-derived training information may contribute to human prediction under the currently available data sets. Moreover, the fusion model, integrating high-quality multisource data, improves the stability and performance of human cell-specific G4 prediction. In summary, G4Beacon2 presents a novel solution for human genome-wide in vivo G4 prediction by integrating multisource data, offering a user-friendly and advanced prediction tool for researchers.