Using machine learning to parametrize large scale tumor simulations.

Abstract

While clinical imaging of tissues focuses on macroscopic tumors, many experiments investigate only small clusters of cells. We aim on providing a scale-bridging link by performing large scale tissue simulations. We employ highly parallelized code in an HPC setting to simulate mm-sized virtual tissues such as embryogenetic zebrafish tissue or breastcancer tumors with more than a million µm-resolved individual cells. We deploy Cells in Silico (CiS), which combines a cellular potts model with an agent based layer and is thus capable of accurately representing many physical and biological properties, such as individual cell shapes, cell division, cell motility etc. Using a model with such a strong representational capacity poses the task of adjusting a large number of parameters to reproduce experimental findings. Prior work has attempted to characterize the similarity between experimental and simulated data by extracting different features and using statistical tests to establish a distance measure. This work highlights how this difficult task can be circumvented by training neural networks to distinguish between experimental and simulated data while simultaneously optimizing the model parameters to maximize the error rate of the network.