Deep Learning Models for Profiling of Kinase Inhibitors

Abstract: With the advent of fluorescence microscopy and image analysis, quantitative information from images can be extracted and changes in cell morphology can be studied. Microscopy-based morphological profiling assays with multiplexed fluorescent dyes, like Cell Painting, can be used for this purpose. It has been shown that morphological profiles can be used to train AI models to classify images into different biological mechanisms. Hence, the goal of this project was to study the possibilities for Deep Learning models and Convolutional Neural Networks to distinguish between different classes of kinase inhibitors based on their morphological profiles. Three different Convolutional Neural Network architectures were used: ResNet50, MobileNetV2, and VGG16. They were trained with two different inputs and two different optimisers: Adam and SGD. Also, a comparison between the performances with and without Transfer Learning through ImageNet weights was executed. The results indicate that MobileNetV2 with Adam as an optimiser performed the best, with a micro average of 0.93 and higher ROC areas compared to the other models. The study also highlighted the importance of utilizing Transfer Learning.