"We have managed to develop a procedure that is able to evaluate the effect of selected drugs on a culture of cancer cells from microscopic images using artificial intelligence tools. The probability of correct classification into one of the three groups in our method exceeds 98 percent, thus outperforming previous methods. This paves the way for much simpler high-throughput cell screening, which will contribute to faster development of new anticancer drugs," says Prof. Jan Kybic, head of the Algorithms for Biomedical Imaging group at FEE CTU.
"Artificial intelligence and its practical applications find extensive use in medicine. Today it is already routinely used in radiological diagnostics, histopathology or endoscopic examinations. A new application that we have developed in cooperation with Prof. Kybic's team at the FEE CTU in Prague is the use of AI in the field of drug development. For a long time, we have believed that from the reaction of cells to a potential drug over time, we can estimate the mechanism of its effect, which is, however, not evaluable by the human eye and brain," explains Associate Professor MUDr. Marián Hajdúch, Ph. D., Director of the Institute of Molecular and Translational Medicine (IMTM) at the Faculty of Medicine of Palacký University and Medical Director of the National Cancer Research Institute (NCRI).
The two research teams and the unique research infrastructures CZ-OPENSCREEN and EATRIS-CZ have therefore joined forces to analyse the image of cells exposed to drugs using artificial intelligence tools. "The results exceeded our expectations, we managed to distinguish the effect of even chemically and mechanistically very related substances. We will be pleased to continue this project and our successful cooperation with CTU in the future and will be happy to extend our cooperation to other AI teams," adds Associate Professor Marián Hajdúch.
RCI supercomputer speeds up image analysis and neural network training
Computer scientists at FEE CTU use images from phase contrast microscopy to analyse cell images, which can be obtained more easily and quickly than conventional fluorescence images. In addition, phase contrast microscopy images do not damage cells and can thus be used to image live cell cultures. The facts of the research are summarised in an article* published in the December issue of the scientific journal Computers in Biology and Medicine.
The researchers use the extensive chemical banks and instrumentation at the UMTM at the Faculty of Medicine, which can be used to automatically perform a large number of experiments. Time-lapse images that capture the reaction of cell cultures exposed to drugs are then sent to computer scientists for processing. At FEE CTU, they undergo analysis using artificial intelligence tools.
"We use a convolutional neural network that has already been trained directly on the input images from phase contrast microscopy. Evaluating the effect of a single contrast agent then takes on the order of a few seconds per image, assuming that the neural network is already trained, which takes several days," explains Prof. Kybic about the data evaluation process. When processing the data, the scientists use the RCI supercomputer at the CTU premises, whose robust computing power significantly speeds up the work.
The most challenging part, however, is to program the network and design the appropriate algorithms, which is the result of many months of work, mainly involving Denis Baručič and Sumit Kaushik, a PhD student and postdoctoral fellow from Prof. Kybic's team.
The research will expand from units to hundreds of chemicals
Prof. Jan Kybic points out that high-throughput cell screening under standardized conditions is only one of many steps in the drug discovery process. Scientists from the Algorithms for Biomedical Imaging group at the Faculty of Electrical Engineering of the Czech Technical University and the Institute of Molecular and Translational Medicine (IMTM) of the Faculty of Medicine of Palacký University in Olomouc have so far succeeded in analysing the effect of a few chemicals in this way; in the next phases of the research, the size of the dataset will need to be expanded to hundreds. The researchers will also take into account different mechanisms of action and possibly multiple cell line types and other factors.