A directional multi-LSTM framework integrated BERT for S-sulfhydration sites prediction

Abstract

Protein S-sulfhydration is an important post-translational modification that regulates signaling pathways in animal cells by influencing protein activity and function. It also plays a crucial role in regulating plant metabolism and morphogenesis. Therefore, the identification of S-sulfhydration sites is crucial for cellular biology research. In this study, we propose a deep learning framework with directional multi-LSTM (Long Short-Term Memory) for predicting protein S-sulfhydration sites. In this study, we propose a deep learning framework utilizing a directional multi-LSTM (Long Short-Term Memory) network to predict protein S-sulfhydration sites. Initially, protein sequence data is preprocessed via an improved BERT strategy to extract high-dimensional sequence features. Hypothesizing that S-sulfhydration modification exhibits directionality, we partition sequences around cysteine residues and extract features using directional multi-LSTM, simulating the enzymatic reaction conditions. Subsequently, a convolutional neural network (CNN) is employed to capture deep local information features. On an independent test set, the accuracy, sensitivity, specificity, Matthews correlation coefficient, area under the curve, and precision are 76.76%, 85.45%, 67.21%, 53.77%, 76.33% and 74.11% respectively. The results demonstrate that the multi-directional LSTM deep learning framework is an effective tool for predicting protein S-sulfhydration. The source code is available on the website https://github.com/endeavor-zzc/Multi-LSTM.