Abstract: Activity cliffs (ACs) present a significant challenge in structure-activity relationship (SAR) studies, characterized by pairs of similar compounds exhibiting substantial differences in biological activity. This paper investigates the interactions between protein kinases and their inhibitors to predict ACs by utilizing advanced deep learning techniques. Matched Molecular Pairs (MMPs) and MMP-ACs are systematically defined based on a minimum 100-fold difference in potency. A deep autoencoder model is developed for feature extraction phase followed by classification phase using various algorithms, including Support Vector Machines (SVM) and neural networks. Our results demonstrate that deep learning approaches can effectively capture complex patterns in molecular data, leading to robust predictions of ACs. Across all classifiers, our experiments show a strong correlation between the structural properties of inhibitors and their activity profiles against specific protein targets.