What is it about?
Glioblastoma is the most aggressive and common brain cancer. Fortunately we have learnt a lot about this cancer in last decades. We know that tumor cells migrate, proliferate, die and consume oxygen and nutrients, and we also know that, under hypoxic conditions, these cells migrate towards more oxygenated regions, close to the blood vessels and then start to fast proliferate near them. This mechanism, known as the go-or-grow behaviour, is indeed one of the reasons of the tumor fast propagation and its invassive capacity. But, how is exactly this transition? Is it sharp? Smooth? Is it reversible or irreversible? In this work we define a method able to unveil this transition, something that enables a better tumor evolution prediction. The method is based on the hybridization of Neural Network technologies with the known physics of the problem. The predictive power of Neural Newtorks is concentrated in the unknown biophysics of the problem, whereas the known phenomena are introduced using mathematical equations. The resulting Netowrk (Physically-Guided Neural Network) looks for a solution that fits the data while satisfying the physics of the problem, enhancing the predictive and explanatory capacity. The presented method demonstrates to be able to unravel this switch between proliferation and migration for different metabolic behaviours and to predict the cell culture evolution using computers for different external stimuli.
Photo by National Cancer Institute on Unsplash
Why is it important?
Although a patient is diagnosed with glioblastoma, each tumor is different and therefore may be addressed differently. This include the use of different drugs, surgery, radiotherapy or chemotherapy treatments. If a patient with glioblastoma came to the hospital. and a biopsy were removed and cultured in a device that subject the tumoral tissue to variable stimuli, the proposed methodology could unravel the specific response of this particular treatment to hypoxia. That is, we would be able to unravel the specific metabolic processes that govern that specific tumor. It is possible to use that knowledge to simulate its response to a certain drug or therapy and thus decide which treatment would be the best one for that particular patient. In that case, we would be taking the first steps towards personalized medicine… and perhaps towards a brighter future for glioblastoma patients.
Read the Original
This page is a summary of: Understanding glioblastoma invasion using physically-guided neural networks with internal variables, PLoS Computational Biology, April 2022, PLOS, DOI: 10.1371/journal.pcbi.1010019.
You can read the full text:
Dataset for the training procedure
The folder contains the data used to train PGNNIVs to unravel the go or grow behaviour of glioblastoma under 4 different parametric models. These data consist on the solutions (in time and space) of eleven simulations with different oxygen boundary conditions. Each folder is named as DATA_"ModelName", where "ModelName" can be: "Sigmoid", "ReLU", "MichaelisMenten" or "Heaviside". Inside each folder, the multidimensional arrays for input and output data for the network training can be found. These arrays have dimension [nExp,TimeStep,x,field], where: * nExp = 11 and corresponds to the number of different configurations or experiments simulated. * TimeStep = 1000 and correspond to the different temporal frames where the solution is given. * x = 51 and corresponds to the different spatial points where the solution is given. * field = 2 and correspond to the different solution fields (1: cells, 2: oxygen).
The following have contributed to this page