DeepCSO: a deep-learning network approach to predicting Cysteine S-sulphenylation sites

Abstract

Cysteine S-sulphenylation (CSO), as a novel post-translational modification (PTM), has emerged as a potential mechanism to regulate protein functions and affect signal networks. Because of its functional significance, several prediction approaches have been developed. Nevertheless, they are based on a limited dataset from Homo sapiens and there is a lack of prediction tools for the CSO sites of other species. Recently, this modification has been investigated at the proteomics scale for a few species and the number of identified CSO sites has significantly increased. Thus, it is essential to explore the characteristics of this modification across different species and construct prediction models with better performances based on the enlarged dataset. In this study, we constructed a few classifiers and fond that the long short-term memory model with the word-embedding encoding approach, dubbed LSTMWE, performs favorably to the traditional machine-learning models and other deep-learning models across different species, in terms of cross-validation and independent test. The area under the ROC curve values for LSTMWE ranged from 0.82 to 0.85 for different organisms, which is superior to the reposted CSO predictors. Moreover, we developed the general model based on the integrated data from different species and it showed great universality and effectiveness. We provided the on-line prediction service called DeepCSO that included both species-specific and general models, which is accessible through http://www.bioinfogo.org/DeepCSO.

DeepKcrot: A Deep-Learning Architecture for General and Species-Specific Lysine Crotonylation Site Prediction

“…In this study, we compiled a benchmark data set of known Kcrot sites and evaluated the performance of different machine-learning approaches, including deep-learning algorithms. We found that the DLbased classifier CNNWE had the best performance compared with the traditional ML model and the LSTMWE model that showed superior to CNNWE for the prediction of cysteine sulphenylation sites, even for the limited training dataset [24]. It suggests that CNN and LSTM may have distinct characteristics that are feasible to extract different PTM features.…”

“…1) The input layer. Each peptide segment is converted into an integer vector with the NUM encoding approach, where each type of amino acid residues was mapped to a different integer [24].…”

“…They are generally based on different ML algorithms combined with various pre-defined features encoded from peptide sequences [25]. The RF algorithm is widely applied to the PTM prediction as it is robust and insensitive to data imbalance [24]. We ever compared the effect of the imbalanced dataset on the potential overfitting of the classifiers and found that the RF model constructed using an imbalanced training dataset had a similar performance to that built using a balanced training dataset [24].…”

“…The RF algorithm is widely applied to the PTM prediction as it is robust and insensitive to data imbalance [24]. We ever compared the effect of the imbalanced dataset on the potential overfitting of the classifiers and found that the RF model constructed using an imbalanced training dataset had a similar performance to that built using a balanced training dataset [24]. According to our previous study on the papaya proteome, we constructed RF-based predictors with the EGAAC encoding scheme, dubbed RFEGAAC that showed the best performance in the RF-based models [18].…”

“…We ever applied the CNN models for the prediction of the papaya Kcrot sites and the CNN model with the word embedding encoding approach compared favourably to other CNN-based approaches [18]. Additionally, we developed here the classifier based on long short-term memory (LSTM) with word embedding named LSTMWE, which was previously constructed to predict Cysteine S-Sulphenylation Sites and had better performance than CNNWE [24].…”