Abstract Personalized oncology aims to match effective treatments for individual cancer patients based on the biology of an individual tumor. This approach is not possible for metabolic therapies due to our inability to quantify metabolic activity in patient tumors, even when patients are infused with 13C-glucose. We overcame this challenge by generating training data with a prior knowledge of 13C-glucose infused patient data and implementing a machine learning model to predict fluxes. We trained a convolutional neural network (CNN) to predict flux ratios. To validate our model, we compared predicted flux ratios to the experimental 13C-glucose-infused mouse models to predict treatment responses for: (1) pharmacologic inhibition of de novo GMP synthesis and (2) environmental serine restriction. Our GMP synthesis model predicted elevated GMP synthesis in GBM compared to cortex in a mouse model and discriminated between sources of GMP synthesis in patients. These findings allow prediction of potential responders to mycophenolate mofetil (MMF), which inhibits de novo but not salvage GMP synthesis. Furthermore, we previously showed that circulating serine uptake is higher in GBM than normal cortex, and a serine- and glycine-restricted diet slows tumor growth in mice. While de novo serine synthesis, circulating and microenvironmental-derived serine could account for serine sources in GBM, our model can distinguish between serine sources in patients, hence potentially predict response to dietary serine depletion. Thus, we have shown that our model may identify patients who will benefit from MMF treatment or a serine- and glycine-restricted diet, potentially enabling administration of highly effective personalized metabolic treatments for GBM patients.
Read full abstract