To protect the images and provide a more secure cipher image, DNA encoding is crucial in image encryption. Applying a single, easily detectable coding rule to the image during DNA encoding has no impact on the encryption model's security level. Therefore, using various coding rules while applying encryption to the image, dynamic DNA-coding techniques have emerged to strengthen and improve the encryption of the image and its security. This study integrates a dynamic DNA-coding method with an encryption model. The model is applied to gray-scale images, where using a predetermined coding rule, every two bits are DNA-encoded in the image. The proposed model generates the key by sending the image and its metadata to hash functions. Following that, the hyperchaotic system constructs three chaotic sequences using the key, and the Lorenz–Liu chaotic system generates a sequence of coding rules. Then, the image is passed to Arnold Transform, where the resulted image is diffused by applying five chaotic maps. Last, using the coding rules, it is DNA-encoded, provided with the chaotic sequences to DNA, and DNA-decoded. Twelve metrics were used to assess the proposed model on ten widely used images. Results show a promising improvement in performance, since it enhanced the security of the model.

Post-translational glycosylation and glycation are common types of protein post-translational modifications (PTMs) in which glycan binds to protein enzymatically or nonenzymatically, respectively. They are associated with various diseases such as coronavirus, Alzheimer’s, cancer, and diabetes diseases. Identifying glycosylation and glycation sites is significant to understanding their biological mechanisms. However, utilizing experimental laboratory tools to identify PTM sites is time-consuming and costly. In contrast, computational methods based on machine learning are becoming increasingly essential for PTM site prediction due to their higher performance and lower cost. In recent years, advances in Transformer-based Language Models based on deep learning have been transferred from Natural Language Processing (NLP) into the proteomics field by developing language models for protein sequence representation known as Protein Language Models (PLMs). In this work, we proposed a novel method, PTG-PLM, for improving the performance of PTM glycosylation and glycation site prediction. PTG-PLM is based on convolutional neural networks (CNNs) and embedding extracted from six recent PLMs including ProtBert-BFD, ProtBert, ProtAlbert, ProtXlnet, ESM-1b, and TAPE. The model is trained and evaluated on two public datasets for glycosylation and glycation site prediction. The results show that PTG-PLM based on ESM-1b and ProtBert-BFD has better performance than PTG-PLM based on the other PLMs. Comparison results with the existing tools and representative supervised learning methods show that PTG-PLM surpasses the other models for glycosylation and glycation site prediction. The outstanding performance results of PTG-PLM indicate that it can be used to predict the sites of the other types of PTMs.

Purpose: Autism spectrum disorder(ASD) is a disease associated with the neurodevelopment of the brain. The autism spectrum can be observed in early childhood, where the symptoms of the disease usually appear in children within the first year of their life. Currently, ASD can only be diagnosed based on the apparent symptoms due to the lack of information on genes related to the disease. Therefore, in this paper, we need to predict the largest number of disease-causing genes for a better diagnosis. Methods: A hybrid stacking ensemble model with Synthetic Minority Oversampling TEchnique (Stack-SMOTE) is proposed to predict the genes associated with ASD. The proposed model uses the gene ontology database to measure the similarities between the genes using a hybrid gene similarity function(HGS). HGS is effective in measuring the similarity as it combines the features of information gain-based methods and graph-based methods. The proposed model solves the imbalanced ASD dataset problem using the Synthetic Minority Oversampling Technique (SMOTE), which generates synthetic data rather than duplicates the data to reduce the overfitting. Sequentially, a gradient boosting-based random forest classifier (GBBRF) is introduced as a new combination technique to enhance the prediction of ASD genes. Moreover, the GBBRF classifier combined with random forest(RF), k-nearest neighbor, support vector machine(SVM), and logistic regression(LR) to form the proposed Stacking-SMOTE model to optimize the prediction of ASD genes. Results: The proposed Stacking-SMOTE model is evaluated using the Simons Foundation Autism Research Initiative (SFARI) gene database and a set of candidates ASD genes.The results of the proposed model-based SMOTE outperform other reported undersampling and oversampling techniques. Sequentially, the results of GBBRF achieve higher accuracy than using the basic classifiers. Moreover, the experimental results show that the proposed Stacking-SMOTE model outperforms the existing ASD prediction models with approximately 95.5% accuracy. Conclusion: The proposed Stacking-SMOTE model demonstrates that SMOTE is effective in handling the autism imbalanced data. Sequentially, the integration between the gradient boosting and random forest classifier (GBBRF) support to build a robust stacking ensemble model(Stacking-SMOTE).

In the software development process, more than one developer may work on developing the same program and bugs in the program may be fixed by a different developer; therefore, understanding the source code is an important issue. Pseudocode plays an important role in solving this problem, as it helps the developer to understand the source code. Recently, transformer-based pre-trained models achieved remarkable results in machine translation, which is similar to pseudocode generation. In this paper, we propose a novel automatic pseudocode generation from the source code based on a pre-trained Bidirectional and Auto-Regressive Transformer (BART) model. We fine-tuned two pre-trained BART models (i.e., large and base) using a dataset containing source code and its equivalent pseudocode. In addition, two benchmark datasets (i.e., Django and SPoC) were used to evaluate the proposed model. The proposed model based on the BART large model outperforms other state-of-the-art models in terms of BLEU measurement by 15% and 27% for Django and SPoC datasets, respectively.

N-linked glycosylation is one of the most common protein post-translation modifications (PTMs) in humans where the Asparagine (N) amino acid of the protein is attached to the glycan. It is involved in most biological processes and associated with various human diseases as diabetes, cancer, coronavirus, influenza, and Alzheimer’s. Accordingly, identifying N-linked glycosylation sites will be beneficial to understanding the system and mechanism of glycosylation. Due to the experimental challenges of glycosylation site identification, machine learning becomes very important to predict the glycosylation sites. This paper proposes a novel N-linked glycosylation predictor based on bagging positive-unlabeled (PU) learning and stacking ensemble machine learning (PUStackNGly). In the proposed PUStackNGly, comprehensive sequence and structural-based features are extracted using different feature extraction descriptors. Then, ensemble-based feature selection is employed to select the most significant and stable features. The ensemble bagging PU learning selects the reliable negative samples from the unlabeled samples using four supervised learning methods (support vector machines, random forest, logistic regression, and XGBoost). Then, stacking ensemble learning is applied using four base classifiers: logistic regression, artificial neural networks, random forest, and support vector machine. The experiments results show that PUStackNGly has a promising predicting performance compared to supervised learning methods. Furthermore, the proposed PUStackNgly outperforms the existing N-linked glycosylation prediction tools on an independent dataset with 95.11% accuracy, 100% recall 80.7% precision, 89.32% F1 score, 96.93% AUC, and 0.87 MCC.