G4-Attention: an interpretable and lightweight deep learning framework for DNA G-quadruplex prediction

G-quadruplexes (G4s) are four-stranded non-canonical nucleic acid secondary structures, formed by the stacking arrangement of guanine tetramers. They are involved in a wide range of biological processes because of their exceptionally unique and distinct structural characteristics. After the completion of the human genome sequencing project, a lot of bio-informatics algorithms were introduced to predict the active G4 regions in vitro based on the canonical G4 sequence elements, G-richness and G-skewness, as well…

By the Truvace deskOctober 19, 20262 min read1 sourcePeer-reviewedTRV-2026-1927· cite & verify

In brief

Researchers describe G4-Attention, a deep learning framework that combines a lightweight convolutional network with Bi-LSTM and attention layers to predict which DNA sequences form G-quadruplex structures. It is presented as a response to earlier shallow CNN-based predictors whose accuracy was described as notably poor.

The work matters because G-quadruplexes are involved in a wide range of biological processes, and better computational mapping could support analysis of where they form. The abstract reports benchmark comparisons only, so real-world research impact is not demonstrated, and performance falls for evolutionarily distant prokaryotes.

Main points

The model combines a lightweight convolutional neural network with Bi-LSTM and attention layers.
It is reported to give the best results on G4 propensity and mismatch-score prediction against benchmark models in the literature.
It is said to predict G4 regions accurately in highly class-imbalanced datasets as well as balanced ones.
Trained on the human genome, it generalizes to non-human eukaryotic species but performs worse on evolutionarily distant prokaryotes.

The gain

A lightweight CNN, Bi-LSTM and attention model predicts G-quadruplex-forming sequences more accurately than earlier shallow CNN models, including on imbalanced data and cell-type-specific tasks.

The rundown

G-quadruplexes are four-stranded non-canonical nucleic acid structures formed by stacked guanine tetramers. After the human genome was sequenced, many bioinformatics algorithms were introduced to predict active G4 regions from canonical sequence elements, G-richness and G-skewness, and non-canonical features.

Sequencing techniques such as G4-seq and G4-ChIP-seq now map G4s in vitro and in vivo at a few hundred base resolution. Machine learning and deep learning approaches built on existing databases followed, but the abstract describes them as shallow CNN-based with poor accuracy.

The new model is also reported to improve cell-type-specific G4 prediction. Its cross-species results cover a range of eukaryotic genomes, with reduced performance for evolutionarily distant prokaryotes.