TransG4: an interpretable deep-learning approach for sequence-based G-quadruplex prediction.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• G-quadruplexes (G4) are crucial non-canonical nucleic acid structures in guanine-rich regions.
• G4s regulate vital cellular processes including gene expression, DNA replication, and telomere maintenance.
• Dysregulation of G4s is linked to cancer and other human diseases, driving research into their formation.

Purpose of the Study:

• To address the challenge of accurately predicting G-quadruplex formation from nucleotide sequences.
• To develop an interpretable and generalizable computational framework for G4 propensity prediction.
• To improve upon existing G4 prediction methods that often lack interpretability or struggle with long-range dependencies.

Main Methods:

• Introduced TransG4, a novel neural network architecture integrating Convolutional Neural Networks (CNN), Transformers, and Bidirectional Gated Recurrent Units (BiGRUs).
• Evaluated TransG4's performance on G4-seq and rG4-seq datasets for DNA and RNA G4 prediction.
• Utilized attention-based mechanisms for model interpretability, identifying biologically relevant motifs.

Main Results:

• TransG4 demonstrated strong predictive performance on both DNA and RNA G4 datasets.
• The model accurately predicted DNA mismatch scores.
• TransG4 consistently outperformed existing methods in RNA RSR-ratio prediction, showing superior accuracy.

Conclusions:

• TransG4 provides an accurate, interpretable, and generalizable framework for sequence-based G4 propensity prediction.
• The model's ability to capture biologically meaningful motifs advances the understanding of G4 structures.
• This work represents a significant contribution to computational genomics and G4 research, with implications for disease studies.