Protein phosphorylation is one of the most critical post-translational modifications of proteins in eukaryotes, which is essential for a variety of biological processes. Plenty of attempts have been made to improve the performance of computational predictors for phosphorylation site prediction. However, most of them are based on extra domain knowledge or feature selection. In this article, we present a novel deep learning-based predictor, named TransPhos, which is constructed using a transformer encoder and densely connected convolutional neural network blocks, for predicting phosphorylation sites. Data experiments are conducted on the datasets of PPA (version 3.0) and Phospho. ELM. The experimental results show that our TransPhos performs better than several deep learning models, including Convolutional Neural Networks (CNN), Long-term and short-term memory networks (LSTM), Recurrent neural networks (RNN) and Fully connected neural networks (FCNN), and some state-of-the-art deep learning-based prediction tools, including GPS2.1, NetPhos, PPRED, Musite, PhosphoSVM, SKIPHOS, and DeepPhos. Our model achieves a good performance on the training datasets of Serine (S), Threonine (T), and Tyrosine (Y), with AUC values of 0.8579, 0.8335, and 0.6953 using 10-fold cross-validation tests, respectively, and demonstrates that the presented TransPhos tool considerably outperforms competing predictors in general protein phosphorylation site prediction.

Identifying compound-protein (drug-target, DTI) interactions (CPI) accurately is a key step in drug discovery. Including virtual screening and drug reuse, it can significantly reduce the time it takes to identify drug candidates and provide patients with timely and effective treatment. Recently, more and more researchers have developed CPI’s deep learning model, including feature representation of a 2D molecular graph of a compound using a graph convolutional neural network, but this method loses much important information about the compound. In this paper, we propose a novel three-channel deep learning framework, named SSGraphCPI, for CPI prediction, which is composed of recurrent neural networks with an attentional mechanism and graph convolutional neural network. In our model, the characteristics of compounds are extracted from 1D SMILES string and 2D molecular graph. Using both the 1D SMILES string sequence and the 2D molecular graph can provide both sequential and structural features for CPI predictions. Additionally, we select the 1D CNN module to learn the hidden data patterns in the sequence to mine deeper information. Our model is much more suitable for collecting more effective information of compounds. Experimental results show that our method achieves significant performances with RMSE (Root Mean Square Error) = 2.24 and $R^2$ (degree of linear fitting of the model) = 0.039 on the GPCR (G Protein-Coupled Receptors) dataset, and with RMSE = 2.64 and $R^2$ = 0.018 on the GPCR dataset RMSE, which preforms better than some classical deep learning models, including RNN/GCNN-CNN, GCNNet and GATNet.

Cancer prognosis is an essential goal for early diagnosis, biomarker selection, and medical therapy. In the past decade, deep learning has successfully solved a variety of biomedical problems. However, due to the high dimensional limitation of human cancer transcriptome data and the small number of training samples, there is still no mature deep learning-based survival analysis model that can completely solve problems in the training process like overfitting and accurate prognosis. Given these problems, we introduced a novel framework called SAVAE-Cox for survival analysis of high-dimensional transcriptome data. This model adopts a novel attention mechanism and takes full advantage of the adversarial transfer learning strategy. We trained the model on 16 types of TCGA cancer RNA-seq data sets. Experiments show that our module outperformed state-of-the-art survival analysis models such as the Cox proportional hazard model (Cox-ph), Cox-lasso, Cox-ridge, Cox-nnet, and VAECox on the concordance index. In addition, we carry out some feature analysis experiments. Based on the experimental results, we concluded that our model is helpful for revealing cancer-related genes and biological functions.

There is a strong need to eliminate batch-specific differences when integrating single-cell RNA-sequencing (scRNA-seq) datasets generated under different experimental conditions for downstream task analysis. Existing batch correction methods usually transform different batches of cells into one preselected “anchor” batch or a low-dimensional embedding space, and cannot take full advantage of useful information from multiple sources. We present a novel framework, called IMGG, i.e., integrating multiple single-cell datasets through connected graphs and generative adversarial networks (GAN) to eliminate nonbiological differences between different batches. Compared with current methods, IMGG shows excellent performance on a variety of evaluation metrics, and the IMGG-corrected gene expression data incorporate features from multiple batches, allowing for downstream tasks such as differential gene expression analysis.

Single-cell RNA sequencing (RNA-seq) has been demonstrated to be a proven method for quantifying gene-expression heterogeneity and providing insight into the transcriptome at the single-cell level. When combining multiple single-cell transcriptome datasets for analysis, it is common to first correct the batch effect. Most of the state-of-the-art processing methods are unsupervised, i.e., they do not utilize single-cell cluster labeling information, which could improve the performance of batch correction methods, especially in the case of multiple cell types. To better utilize known labels for complex dataset scenarios, we propose a novel deep learning model named IMAAE (i.e., integrating multiple single-cell datasets via an adversarial autoencoder) to correct the batch effects. After conducting experiments with various dataset scenarios, the results show that IMAAE outperforms existing methods for both qualitative measures and quantitative evaluation. In addition, IMAAE is able to retain both corrected dimension reduction data and corrected gene expression data. These features make it a potential new option for large-scale single-cell gene expression data analysis.

Predicting protein function from sequences through machine learning can improve the understanding of novel proteins and biological mechanisms. Existing methods mainly rely on one-dimensional convolution or natural language processing (NLP) techniques to extract features from sequences, but they suffer from limited predictive performance. To address this challenge, we propose MulAxialGO, a new method that leverages multi-modal feature fusion to improve prediction accuracy. MulAxialGO integrates the prior features of a large-scale pre-trained protein language model and the posterior features of dynamic embedding coding and sequence homology. In addition, MulAxialGO employs a comprehensive image feature encoder to extract features from sequences, providing a novel perspective for protein function prediction. MulAxialGO is tested on two benchmark datasets and achieves state-of-the-art results. On the 2016 dataset, MulAxialGO significantly outperforms DeepGOPlus, improving molecular function by 4.5 points, biological process by 2.4 points and cellular component by 1.6 points for the AUPR metric. Similarly, on the NetGO dataset, MulAxialGO outperforms the state-of-the-art NetGO2.0, improving Fmax by 1.1 points for biological process and 2.3 points for cellular component.