MetalPrognosis-WebServer

Protein-metal ion interactions are pivotal in the mechanisms underlying the development of various diseases. Changes in amino acids can result in missense mutations within the metal-binding sites of proteins, disrupting their interaction with metal ions. This, in turn, affects the protein's function and can lead to the onset of serious human diseases. Accurately identifying disease-associated mutation sites within protein metal-binding sites is of great importance. It aids researchers in comprehending protein functionality and designing innovative drugs. Many existing methods that are based on deep learning often suffer from limited accuracy, primarily due to manual feature extraction and a lack of structural information.

In this study, we present MetalPrognosis, an alignment-free approach designed to precisely predict disease-associated mutations within the metal-binding sites of metalloproteins. MetalPrognosis leverages sliding window sequences as its input and extracts semantic information from pretrained protein language models. This information is then integrated into a convolutional neural network to extract highly abstract features. Our experimental results demonstrate the outstanding predictive performance of MetalPrognosis when compared to state-of-the-art approaches, such as MCCNN and PolyPhen-2, across independent test sets for various metalloproteins. Additionally, an ablation experiment underscores the efficiency of our model architecture.