Convolutional Neural Networks for Cellular Drug Response Prediction Using Immunofluorescence Images of Intracellular Actin Filament Networks

Abstract

Actin cytoskeleton has been identified as a potential therapeutic target for cancer. Therefore, to identify cell responses to such chemical agents has been an essential part in the past studies, which is often measured visually. This kind of visual recognition task currently is performed by human experts, which poses a great challenge since the features can hardly be detected using only human eyes. This article presents the application of convolutional neural networks (CNNs) in classifying human breast epithelial cells based on different dosages of drug exposure. MCF-10A cell line was chosen for the experiments and was treated with 90 nM and 400 nM cytochalasin D. The CNNs were evaluated on a large immunofluorescence images of intracellular actin filament networks captured after the exposure of different drug concentrations. During the image pre-processing, we implemented image enhancement and data augmentation approaches. Two well-known CNNs, VGG-16 and ResNet-50, were trained with or without transfer learning. The study revealed that the CNN performed better in the classification task compared to human experts. In conclusion, ResN et-50 with transfer learning achieved the best performance.