Chi-Wei Chen

Accurate SUMOylation site prediction is crucial for deciphering gene regulation and disease mechanisms. However, distinguishing SUMO1 and SUMO2 modifications remains a major challenge due to their structural similarities. Conventional prediction models often struggle to differentiate between these paralogues, limiting their applicability in biological research. To address this, we introduce SUMO-LMNet, a deep learning-based framework for the precise prediction of SUMO1 and SUMO2 sites. Unlike previous models, SUMO-LMNet integrates a lossless mapping strategy and deep learning architectures to enhance both prediction accuracy and interpretability. Our model extracts high-dimensional features from sequences and transforms them into two-dimensional feature maps, enabling convolutional neural networks (CNNs) to effectively capture both local and global dependencies within the data. By leveraging a Lossless Mapping Network (LM-Net), this approach preserves the original feature space, ensuring that feature integrity is retained without loss of spatial information. While Grad-CAM highlights key features in individual predictions, it lacks consistency across samples and does not provide a dataset-wide evaluation of feature importance. To address this, we introduce Combined Heatmap Feature Analysis (CHFA), which systematically aggregates feature importance across multiple samples, providing a more reliable and interpretable dataset-wide assessment. Experimental results reveal distinct feature dependencies between SUMO1 and SUMO2, underscoring the necessity of paralogue-specific predictive models. Through a systematic comparison of multiple neural network architectures, we demonstrate that our model achieves over 80 % accuracy in distinguishing SUMO1 and SUMO2 modification sites. By prioritizing candidate sites for further study, our model aids experimental design and accelerates the discovery of biologically relevant SUMOylation targets. SUMO-LMNet is publicly available at https://predictor.isu.edu.tw/sumo-lmnet.

The impact of COVID-19 on the human body after infection requires further investigation. The protease carried by the virus may interact with cellular proteins, thereby affecting the regulatory pathways. Machine learning can provide predictive information to aid the research design in such studies. This study examined the effectiveness of different encoding methods for training models using traditional machine learning and deep learning methods. Using the advantages of deep learning methods for high-dimensional data analysis, amino acid properties were extracted and classified. Viral polyprotein data were used as the training set and human protein substrates were collected from the literature as the independent test set. In the one-dimensional convolutional neural network model, the accuracy of the independent test set was the highest, at 0.832.

Bridge failures inevitably cause severe social consequences in terms of the economic losses or even human fatalities. Scour is one of the leading causes for bridge failures. Due to its complexity and multidisciplinary nature, the mechanism of scour has not been completely understood yet. Inspired by the rapid development of Machine Learning (ML) techniques, this paper aims to construct a novel data-driven extreme gradient boosting (XGBoost) algorithm based model to predict the scour risk around bridge piers. Data used in the present study is provided by the French National Railway Company (SNCF). The performance of XGBoost-based model is compared with three commonly used algorithms: support vector machine, random forest and multilayer perceptron. Results show that the XGBoost classifier achieves high accuracy (0.959/0.938), precision (0.970/0.961), recall (0.974/0.956) and low false positive rate (0.085/0.114) for training and test set respectively. Moreover, the classifier obtains an area under the ROC curve (AUC) score equal to 0.974 for test set (a perfect classifier has an AUC equal to 1). This paper presents a cutting-edge application of XGBoost algorithm in the maintenance of railway bridges. The proposed methodology may allow engineers to determine the scour risk of bridge piers in an accurate and rapid way.

Protein phosphorylation is one of the most important post-translational modifications, and many biological processes are related to phosphorylation, such as DNA repair, transcriptional regulation and signal transduction and, therefore, abnormal regulation of phosphorylation usually causes diseases. If we can accurately predict human phosphorylation sites, this could help to solve human diseases. Therefore, we developed a kinase-specific phosphorylation prediction system, GasPhos, and proposed a new feature selection approach, called Gas, based on the ant colony system and a genetic algorithm and used performance evaluation strategies focused on different kinases to choose the best learning model. Gas uses the mean decrease Gini index (MDGI) as a heuristic value for path selection and adopts binary transformation strategies and new state transition rules. GasPhos can predict phosphorylation sites for six kinases and showed better performance than other phosphorylation prediction tools. The disease-related phosphorylated proteins that were predicted with GasPhos are also discussed. Finally, Gas can be applied to other issues that require feature selection, which could help to improve prediction performance.

Mutation of a single amino acid residue may change protein structure which affect protein function and diseases. Increasing protein stability or maintaining its stability while changing protein properties is often a goal in protein engineering, drug design or industrial optimization. A variety of methods and features have been proposed to predict the stability of protein mutations, the conflicting prediction results from different tools could cause confusion to users. Therefore, this study integrates structure and sequence-based prediction tools with machine learning and adds information of protein sequences. The best model is selected through feature selection methods to improve accuracy and reduce the time complexity of the training model. The PCC (Pearson correlation coefficient) of the integrated model can be increased from 0.669 to 0.702. Not only successfully integrates predictors, but also improves the accuracy of integration tools.

Thermostability is a protein property that impacts many types of studies, including protein activity enhancement, protein structure determination, and drug development. However, most computational tools designed to predict protein thermostability require tertiary structure data as input. The few tools that are dependent only on the primary structure of a protein to predict its thermostability have one or more of the following problems: a slow execution speed, an inability to make large-scale mutation predictions, and the absence of temperature and pH as input parameters. Therefore, we developed a computational tool, named KStable, that is sequence-based, computationally rapid, and includes temperature and pH values to predict changes in the thermostability of a protein upon the introduction of a mutation at a single site. KStable was trained using basis features and minimal redundancy–maximal relevance (mRMR) features, and 58 classifiers were subsequently tested. To find the representative features, a regular-mRMR method was developed. When KStable was evaluated with an independent test set, it achieved an accuracy of 0.708.

Protein quaternary structure complex is also known as a multimer, which plays an important role in a cell. The dimer structure of transcription factors is involved in gene regulation, but the trimer structure of virus-infection-associated glycoproteins is related to the human immunodeficiency virus. The classification of the protein quaternary structure complex for the post-genome era of proteomics research will be of great help. Classification systems among protein quaternary structures have not been widely developed. Therefore, we designed the architecture of a two-layer machine learning technique in this study, and developed the classification system PClass. The protein quaternary structure of the complex is divided into five categories, namely, monomer, dimer, trimer, tetramer, and other subunit classes. In the framework of the bootstrap method with a support vector machine, we propose a new model selection method. Each type of complex is classified based on sequences, entropy, and accessible surface area, thereby generating a plurality of feature modules. Subsequently, the optimal model of effectiveness is selected as each kind of complex feature module. In this stage, the optimal performance can reach as high as 70% of Matthews correlation coefficient (MCC). The second layer of construction combines the first-layer module to integrate mechanisms and the use of six machine learning methods to improve the prediction performance. This system can be improved over 10% in MCC. Finally, we analyzed the performance of our classification system using transcription factors in dimer structure and virus-infection-associated glycoprotein in trimer structure. PClass is available via a web interface at http://predictor.nchu.edu.tw/PClass/.