Cancer classification with a network of chemical oscillators

Abstract

We discuss chemical information processing considering dataset classifiers formed with a network of interacting droplets. Our arguments are based on computer simulations of droplets in which a photosensitive variant of the Belousov–Zhabotinsky (BZ) reaction proceeds. By applying optical control we can adjust the time evolution of individual droplets and prepare the network to perform a specific computational task. We demonstrate that chemical classifiers made of droplets can be designed in computer simulations based on evolutionary algorithms. The mutual information between the dataset and the observed time evolution of droplets in the network is taken as the fitness function in the optimization process. We show that a classifier of the Wisconsin Breast Cancer Dataset made of a relatively small number of droplets can distinguish between malignant and benign forms of cancer with an accuracy exceeding 97%. The reliability of the optimized chemical classifiers of this dataset as a function of optimization time, number of droplets involved in data processing and the method of extracting the output information is discussed.