G4mismatch: Deep neural networks to predict G-quadruplex propensity based on G4-seq data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• G-quadruplexes are crucial non-B-DNA structures implicated in various molecular and disease phenotypes.
• Genome-wide measurement of G-quadruplex formation is essential but experimentally challenging and time-consuming.
• Existing computational methods for G-quadruplex prediction are limited by small datasets or rule-based approaches.

Purpose of the Study:

• To develop a novel, accurate, and efficient algorithm for predicting G-quadruplex propensity genome-wide.
• To address the limitations of current computational methods for G-quadruplex prediction.
• To enable high-throughput analysis of G-quadruplex formation across genomic sequences.

Main Methods:

• Development of G4mismatch, a convolutional neural network-based algorithm.
• Training the model on a large dataset of nearly 400 million human genomic loci from a G4-seq experiment.
• Validation on held-out chromosomes and independent datasets from various animal species.

Main Results:

• G4mismatch achieved a Pearson correlation over 0.8 when predicting mismatch scores on held-out human chromosome sequences.
• The algorithm demonstrated high accuracy (Pearson correlations > 0.7) on independent datasets from different species, even when trained on human data.
• G4mismatch outperformed existing methods in detecting G-quadruplexes genome-wide using predicted mismatch scores.
• The model's learned principles provided unique visualization for deducing G-quadruplex formation mechanisms.

Conclusions:

• G4mismatch represents a significant advancement in the computational prediction of G-quadruplex propensity.
• The algorithm enables accurate and efficient genome-wide G-quadruplex prediction across diverse species.
• G4mismatch facilitates deeper understanding of G-quadruplex formation mechanisms and their biological implications.