iDualG4: A Dual-Channel Deep Learning Framework for Predicting In Vivo G-Quadruplexes

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• G-quadruplexes (G4s) are crucial non-canonical DNA structures involved in genomic stability and gene regulation.
• While many potential G4-forming sequences (PQSs) exist, only a few form stable G4s in vivo due to chromatin complexity.
• Current deep learning methods for G4 prediction rely on expensive, cell-specific epigenetic data, limiting their application.

Purpose of the Study:

• To develop an interpretable deep learning framework, iDualG4, for predicting in vivo G-quadruplex formation.
• To eliminate the need for cell-specific epigenetic assays by inferring features from DNA sequence alone.
• To improve the accuracy and accessibility of G4 prediction for clinical and research applications.

Main Methods:

• iDualG4 utilizes a dual-channel deep learning architecture.
• It incorporates a pretrained Enformer module to infer epigenomic proxy features directly from DNA sequence.
• The framework integrates inferred epigenomic features with local sequence information for prediction.

Main Results:

• iDualG4 significantly outperforms existing methods in predicting G4 formation across multiple cell lines.
• The model demonstrates exceptional performance on imbalanced datasets, achieving an AUPR of 0.981 on K562 cells.
• DeepSHAP analysis confirms the interpretability of iDualG4, highlighting its predictive mechanisms.

Conclusions:

• iDualG4 provides a highly accurate and interpretable tool for predicting in vivo G4 formation without requiring additional sequencing data.
• The framework offers a novel computational approach to understand the interplay between DNA sequence and epigenetic factors in G4 formation.
• iDualG4 enhances the broader applicability of G4 prediction in diverse biological contexts.