MRes - From Sequence to Structure: Building an AI Model for G4 DNA Prediction

G-quadruplexes (G4s) are non-canonical nucleic acid structures formed by guanine-rich DNA sequences and stabilized by Hoogsteen hydrogen bonds and monovalent cations. These structures are biologically significant, particularly in telomeres and promoter regions, where they regulate transcription, replication, and genomic stability. 

While recent breakthroughs such as AlphaFold3 have revolutionized protein and RNA structure prediction, they currently fail to reliably generate correct G-quadruplex structures from sequence alone. These AI models are largely trained on protein-centric datasets and lack sufficient architectural constraints or topological formalism necessary for capturing G4 diversity. Despite the presence of many experimentally determined G4 structures in the Protein Data Bank (PDB), there is no dedicated AI pipeline that maps a single-stranded DNA sequence to a valid 3-loop G4 fold. 

This project addresses the critical gap in predictive capabilities for non-canonical DNA folding—specifically in translating sequence to realistic three-loop G4 structures. Bridging this gap requires a model that can integrate biological sequence input with structural constraints defined by G4 topological formalism.

Objectives of the research:

To develop a hybrid AI pipeline that predicts three-loop G4 structures from sequence by extending AlphaFold’s sequence-to-structure functionality with a specialized deep learning model trained on known G4 topologies.

Methods to be used:

Structural Bioinformatics: Extraction and annotation of known G-quadruplex (G4) structures from the Protein Data Bank (PDB) using topological classification frameworks, particularly the formalism described by Silva et al.

Machine Learning / AI: Design and training of a neural network—likely a transformer or graph neural network (GNN)—to learn sequence-to-topology relationships based on curated G4 datasets. Training will use PyTorch or TensorFlow.

Deep Learning Integration: Modification of AlphaFold3 components (if source code access is available) to integrate the trained G4 prediction model into a custom structure generation pipeline.

Scripting and Automation: Development of Python-based scripts for sequence parsing, structure annotation, and prediction pipeline automation.

Skills required of applicant:

An interest in learning the skills mentioned in the methods section. A background in Chemistry, Pharmacy, Biomedical Sciences, Engineering, Data Science or related sciences.