Deep Learning Framework for G-Quadruplex Stability Prediction

Guanine-quadruplexes (G4s) are non-canonical higher-order structures formed by guanine rich nucleic acid sequences. G4s naturally occur throughout human chromosomes, where they mediate a host of critical biological functions, such as gene expression, telomere maintenance and regulating RNA processing. Studies in the past decade have revealed that G4s are involved in the pathogenesis of cancer. The stability of G4 structures regulates oncogene expression through transcription factor binding, inhibiting telomerase regulating cellular division and causing genomic damage via stalling replication forks. Thus, understanding the driving factors of G4 stability is crucial in the development of anticancer therapeutic drugs. To investigate G4 stability, we measured melting temperatures while examining two parameters: the primary nucleotide sequence responsible for G4 formation and the ionic environment that supports and stabilizes the G4 structure. We developed a deep learning model that combines recurrent neural network (RNN) to capture sequence patterns with a convolutional neural network (CNN) to detect spatial motifs in G4-forming sequences. Our highly accurate model was validated by predicting the melting temperatures of previously unreported G4 sequences, which were experimentally determined through UV melting analysis in a laboratory setting. Our model aims to provide insights into the sequence-structure relationships governing G4 stability, facilitate experimental validation and accelerates the discovery of G4 structures with customised thermal properties for therapeutic applications.