The study introduces an advanced deep learning framework developed at HU that leverages graph neural networks (GNNs) to identify subtle biological signals of lung cancer from blood plasma samples. The model, called Metabolite Graph Neural Network (M-GNN), was designed by HU faculty and Ph.D. students in collaboration with BioMark Diagnostics, Inc., St. Boniface Hospital Research Centre, and Asper Clinical Research Centre. Researchers at HU collected additional data from public repositories and were responsible for designing and implementing the model architecture.

AI models are as powerful as the data and knowledge we provide them with, said Dr. Vaida. These models don’t inherently understand what each metabolite does, what normal levels look like, which pathways they act on, or what diseases they’re associated with. It’s our job as researchers to guide them by embedding accurate information about metabolic processes so that they can learn the complex interactions within pathways and ultimately detect subtle signals of disease, like early-stage lung cancer.

Unlike conventional AI models that treat data as independent features, the M-GNN framework models the relationships between biological elements as a connected graph. This allows it to learn the structure of metabolic activity more accurately and detect early signs of disease with high precision. When tested on blood plasma samples provided by BioMark Diagnostics, the model achieved a 96% AUC, effectively distinguishing cancer from non-cancer samples.

Source: https://www.harrisburgu.edu/news/2025-06-24-research-publication-ai-lung-cancer-early-detection/