DeepCKI, a Bioinformatics Model for Predicting Cell-Cytokine Interactions Based on Variational Graph Auto-Encoder / 中国生物化学与分子生物学报

Cytokines are a class of signaling molecules that are synthesized and secreted by immune cells and certain non-immune cells, and regulate cell growth, differentiation, and immune response by binding to corresponding receptors in the immune system. Most of the current studies focus on investigating the intercellular communication network using experimental methods to detect the interaction between cytokines and receptors, but there are short comings such as long experimental cycles, high equipment requirements, and high costs. Therefore, it is necessary to accelerate the systematic study of cell-cytokine interactions (CKIs) through computational methods. In this paper, we propose DeepCKI, a deep learning model based on variational graph auto-encoder (VGAE) to predict cell-cytokine interactions, which can effectively fuse protein interactions and different types of protein features, and fully exploit the effective information in network topology and node properties to achieve efficient prediction of cell-cytokine interactions. Compared with variational auto-encoder and deep neural network methods, DeepCKI with graph structure designs exhibits optimal prediction performance. The AUC values of the DeepCKI model for four different types of cell-cytokine interactions are higher than 0. 8, and the model has certain robustness and effectiveness. Among the top 100 cell-cytokine interactions scored for prediction, 36 pairs have been validated by the latest published literature, indicating that the model can discover new cell-cytokine interactions.

Study on the inverse problem of diffuse optical tomography based on improved stacked auto-encoder / 生物医学工程学杂志

The inverse problem of diffuse optical tomography (DOT) is ill-posed. Traditional method cannot achieve high imaging accuracy and the calculation process is time-consuming, which restricts the clinical application of DOT. Therefore, a method based on stacked auto-encoder (SAE) was proposed and used for the DOT inverse problem. Firstly, a traditional SAE method is used to solved the inverse problem. Then, the output structure of SAE neural network is improved to a single output SAE, which reduce the burden on the neural network. Finally, the improved SAE method is used to compare with traditional SAE method and traditional levenberg-marquardt (LM) iterative method. The result shows that the average time to solve the inverse problem of the method proposed in this paper is only 1.67% of the LM method. The mean square error (MSE) value is 46.21% lower than the traditional iterative method, 61.53% lower than the traditional SAE method, and the image correlation coefficient(ICC) value is 4.03% higher than the traditional iterative method, 18.7% higher than the traditional SAE method and has good noise immunity under 3% noise conditions. The research results in this article prove that the improved SAE method has higher image quality and noise resistance than the traditional SAE method, and at the same time has a faster calculation speed than the traditional iterative method, which is conducive to the application of neural networks in DOT inverse problem calculation.

Psychosis speech recognition algorithm based on deep embedded sparse stacked autoencoder and manifold ensemble / 生物医学工程学杂志

Speech feature learning is the core and key of speech recognition method for mental illness. Deep feature learning can automatically extract speech features, but it is limited by the problem of small samples. Traditional feature extraction (original features) can avoid the impact of small samples, but it relies heavily on experience and is poorly adaptive. To solve this problem, this paper proposes a deep embedded hybrid feature sparse stack autoencoder manifold ensemble algorithm. Firstly, based on the prior knowledge, the psychotic speech features are extracted, and the original features are constructed. Secondly, the original features are embedded in the sparse stack autoencoder (deep network), and the output of the hidden layer is filtered to enhance the complementarity between the deep features and the original features. Third, the L1 regularization feature selection mechanism is designed to compress the dimensions of the mixed feature set composed of deep features and original features. Finally, a weighted local preserving projection algorithm and an ensemble learning mechanism are designed, and a manifold projection classifier ensemble model is constructed, which further improves the classification stability of feature fusion under small samples. In addition, this paper designs a medium-to-large-scale psychotic speech collection program for the first time, collects and constructs a large-scale Chinese psychotic speech database for the verification of psychotic speech recognition algorithms. The experimental results show that the main innovation of the algorithm is effective, and the classification accuracy is better than other representative algorithms, and the maximum improvement is 3.3%. In conclusion, this paper proposes a new method of psychotic speech recognition based on embedded mixed sparse stack autoencoder and manifold ensemble, which effectively improves the recognition rate of psychotic speech.

Application of neural network autoencoder algorithm in the cancer informatics research / 生物工程学报

Cancers have been widely recognized as highly heterogeneous diseases, and early diagnosis and prognosis of cancer types have become the focus of cancer research. In the era of big data, efficient mining of massive biomedical data has become a grand challenge for bioinformatics research. As a typical neural network model, the autoencoder is able to efficiently learn the features of input data by unsupervised training method and further help integrate and mine the biological data. In this article, the primary structure and workflow of the autoencoder model are introduced, followed by summarizing the advances of the autoencoder model in cancer informatics using various types of biomedical data. Finally, the challenges and perspectives of the autoencoder model are discussed.

Prediction of drug-induced cell viability by SAE-XGBoost algorithm based on LINCS-L1000 perturbation signal / 生物工程学报

Different cell lines have different perturbation signals in response to specific compounds, and it is important to predict cell viability based on these perturbation signals and to uncover the drug sensitivity hidden underneath the phenotype. We developed an SAE-XGBoost cell viability prediction algorithm based on the LINCS-L1000 perturbation signal. By matching and screening three major dataset, LINCS-L1000, CTRP and Achilles, a stacked autoencoder deep neural network was used to extract the gene information. These information were combined with the RW-XGBoost algorithm to predict the cell viability under drug induction, and then to complete drug sensitivity inference on the NCI60 and CCLE datasets. The model achieved good results compared to other methods with a Pearson correlation coefficient of 0.85. It was further validated on an independent dataset, corresponding to a Pearson correlation coefficient of 0.68. The results indicate that the proposed method can help discover novel and effective anti-cancer drugs for precision medicine.

Single-modal neuroimaging computer aided diagnosis for schizophrenia based on ensemble learning using privileged information / 生物医学工程学杂志

Neuroimaging technologies have been applied to the diagnosis of schizophrenia. In order to improve the performance of the single-modal neuroimaging-based computer-aided diagnosis (CAD) for schizophrenia, an ensemble learning algorithm based on learning using privileged information (LUPI) was proposed in this work. Specifically, the extreme learning machine based auto-encoder (ELM-AE) was first adopted to learn new feature representation for the single-modal neuroimaging data. Random project algorithm was then performed on the learned high-dimensional features to generate several new feature subspaces. After that, multiple feature pairs were built among these subspaces to work as source domain and target domain, respectively, which were used to train multiple support vector machine plus (SVM+) classifier. Finally, a strong classifier is learned by combining these SVM+ classifiers for classification. The proposed algorithm was evaluated on a public schizophrenia neuroimaging dataset, including the data of structural magnetic resonance imaging (sMRI) and functional MRI (fMRI). The results showed that the proposed algorithm achieved the best diagnosis performance. In particular, the classification accuracy, sensitivity and specificity of the proposed algorithm were 72.12% ± 8.20%, 73.50% ± 15.44% and 70.93% ± 12.93%, respectively, on the sMRI data, and it also achieved the classification accuracy of 72.33% ± 8.95%, sensitivity of 68.50% ± 16.58% and specificity of 75.73% ± 16.10% on the fMRI data. The proposed algorithm overcomes the problem that the traditional LUPI methods need the additional privileged information modality as source domain. It can be directly applied to the single-modal data for classification, and also can improve the classification performance. Therefore, it suggests that the proposed algorithm will have wider applications.

The research of sleep staging based on single-lead electrocardiogram and deep neural network

The polysomnogram (PSG) analysis is considered the golden standard for sleep staging under the clinical environment. The electroencephalogram (EEG) signal is the most important signal for classification of sleep stages. However, in-vivo signal recording and analysis of EEG signal presents us with a few technical challenges. Electrocardiogram signals on the other hand, are easier to record, and can provide an attractive alternative for home sleep monitoring. In this paper we describe a method based on deep neural network (DNN), which can be used for the classification of the sleep stages into Wake (W), rapid-eye-movement (REM) and non-rapid-eye-movement (NREM) sleep stage. We apply the sleep stage stacked autoencoder to constitute a 4-layer DNN model. In order to test the accuracy of our method, eighteen PSGs from the MIT-BIH Polysomnographic Database were used. A total of 11 features were extracted from each electrocardiogram recording The experimental design employs cross-validation across subjects, ensuring the independence of the training and the test data. We obtained an accuracy of 77% and a Cohen's kappa coefficient of about 0.56 for the classification of Wake, REM and NREM.

VASC: Dimension Reduction and Visualization of Single-cell RNA-seq Data by Deep Variational Autoencoder / 基因组蛋白质组与生物信息学报·英文版

Single-cell RNA sequencing (scRNA-seq) is a powerful technique to analyze the transcriptomic heterogeneities at the single cell level. It is an important step for studying cell sub-populations and lineages, with an effective low-dimensional representation and visualization of the original scRNA-Seq data. At the single cell level, the transcriptional fluctuations are much larger than the average of a cell population, and the low amount of RNA transcripts will increase the rate of technical dropout events. Therefore, scRNA-seq data are much noisier than traditional bulk RNA-seq data. In this study, we proposed the deep variational autoencoder for scRNA-seq data (VASC), a deep multi-layer generative model, for the unsupervised dimension reduction and visualization of scRNA-seq data. VASC can explicitly model the dropout events and find the nonlinear hierarchical feature representations of the original data. Tested on over 20 datasets, VASC shows superior performances in most cases and exhibits broader dataset compatibility compared to four state-of-the-art dimension reduction and visualization methods. In addition, VASC provides better representations for very rare cell populations in the 2D visualization. As a case study, VASC successfully re-establishes the cell dynamics in pre-implantation embryos and identifies several candidate marker genes associated with early embryo development. Moreover, VASC also performs well on a 10× Genomics dataset with more cells and higher dropout rate.