Explainable AI for molecular toxicity prediction (GNN + transformer architecture)

An explainable AI framework for predicting molecular toxicity, combining graph neural networks (GNNs) and transformer-based architectures with built-in interpretability. It matters because current toxicity screening is either slow and expensive (traditional in vitro/in vivo assays) or fast but opaque (existing deep learning models), leaving scientists, safety teams, and regulators unable to interrogate why a prediction was made. Roughly 30% of drug candidates fail in pre-clinical stages due to toxicity issues, a bottleneck with direct cost and safety consequences across pharmaceuticals, chemicals, and industrial products.

The framework integrates knowledge graphs (ToxKG, PubChem, ChEMBL) with molecular graph representations for GNN-based structure encoding, and transformer models operating on SMILES and clinical/genomic data for broader toxicity signal capture. Explainability is embedded throughout using SHAP, GNNExplainer, and attention visualisation, so outputs are interrogable rather than black-box, benchmarked across Tox21, ToxCast, and MoleculeNet.

This is a PhD research programme beginning October 2026 at Brunel University of London (Computer Science, AI), supervised by Dr Matloob Khushi and Dr Alessandro Pandini, running through September 2029. The methodology and phased research plan are fully defined; model development has not yet begun. We're seeking sponsored research or funding partners, and collaborators able to provide real-world chemical, toxicological, or clinical data for validation as the models are developed, with open-source tools and joint publication as key outputs.