Drug repositioning, which involves identifying new therapeutic indications for approved drugs, is pivotal in accelerating drug discovery. Recently, to mitigate the effect of label sparsity on inferring potential drug-disease associations (DDAs), graph contrastive learning (GCL) has emerged as a promising paradigm to supplement high-quality self-supervised signals through designing auxiliary tasks, then transfer shareable knowledge to main task, i.e. DDA prediction. However, existing approaches still encounter two limitations. The first is how to generate augmented views for fully capturing higher-order interaction semantics. The second is the optimization imbalance issue between auxiliary and main tasks. In this paper, we propose a novel heterogeneous Graph Contrastive learning method with Gradient Balance for DDA prediction, namely GCGB. To handle the first challenge, a fusion view is introduced to integrate both semantic views (drug and disease similarity networks) and interaction view (heterogeneous biomedical network). Next, inter-view contrastive learning auxiliary tasks are designed to contrast the fusion view with semantic and interaction views, respectively. For the second challenge, we adaptively adjust the gradient of GCL auxiliary tasks from the perspective of gradient direction and magnitude for better guiding parameter update toward main task. Extensive experiments conducted on three benchmarks under 10-fold cross-validation demonstrate the model effectiveness.

Drug repositioning, which identifies new therapeutic potential of approved drugs, is instrumental in accelerating drug discovery. Recently, to alleviate the effect of data sparsity on predicting possible drug-disease associations (DDAs), graph contrastive learning (GCL) has emerged as a promising paradigm for learning discriminative representations of drugs and diseases through distilling informative self-supervised signals. However, existing GCL-based methods devised for DDA prediction still encounter two limitations. Firstly, the crucial heterogeneous property, which allows for capturing nuanced interaction semantics between biological entities, is overlooked. The second is how to perform contrastive view augmentation without relying on stochastic perturbation. In this study, we propose a novel multi-view contrastive learning approach for DDA prediction, namely MICLE. To handle the first issue, protein-related bipartite graphs are integrated with the original DDA network in advance, thereby composing a heterogeneous biological network (HBN). Besides, heterogeneous graph neural network is applied to mine the rich connectivity patterns implicit in the above HBN. For the second limitation, we design the complementary inter-view and intra-view contrastive learning tasks. Specifically, the former ensures that the mutual information between paired nodes across views is maximized, the latter enhances the agreement between each node and its first-order neighbors on similarity networks. Extensive experiments conducted on three benchmarks under 10-fold cross-validation demonstrate the model effectiveness.

Drug repositioning, which identifies new therapeutic potential of approved drugs, is pivotal in accelerating drug discovery. Recently, growing efforts are devoted to applying graph neural networks (GNNs) for effectively modeling drug-disease associations (DDAs). However, current GNN-based methods are generally designed for unsigned graphs and fail to gain complementary insights provided by negative links. Despite the proposal of sign-aware GNNs in general fields, there exist two intractable challenges when indiscriminately deploying prior solutions into drug repositioning. (i) How to explicitly connect the nodes within the same set (disease-disease and drug-drug)? (ii) How to design the contrastive learning objective for signed graphs? To this end, we propose a novel sign-aware graph contrastive learning approach, namely SIGDR, which takes both the positive and negative links from signed biological networks into consideration to identify underlying DDAs. To handle the first challenge, we measure the drug and disease similarity and form signed unipartite graphs according to similarity scores. For the second challenge, a signed bipartite graph is then constructed from the annotated DDA dataset. Through dividing above obtained signed graphs into positive and negative subgraphs respectively, we devise the inter-view contrastive learning auxiliary task to enhance the consistency of node representations derived from partitioned subgraphs with the same link type. Extensive experiments conducted on three benchmarks under 10-fold cross-validation demonstrate the model effectiveness.

Survival analysis is critical to cancer prognosis estimation. High-throughput technologies facilitate the increase in the dimension of genic features, but the number of clinical samples in cohorts is relatively small due to various reasons, including difficulties in participant recruitment and high data-generation costs. Transcriptome is one of the most abundantly available OMIC (referring to the high-throughput data, including genomic, transcriptomic, proteomic and epigenomic) data types. This study introduced a multitask graph attention network (GAT) framework DQSurv for the survival analysis task. We first used a large dataset of healthy tissue samples to pretrain the GAT-based HealthModel for the quantitative measurement of the gene regulatory relations. The multitask survival analysis framework DQSurv used the idea of transfer learning to initiate the GAT model with the pretrained HealthModel and further fine-tuned this model using two tasks i.e. the main task of survival analysis and the auxiliary task of gene expression prediction. This refined GAT was denoted as DiseaseModel. We fused the original transcriptomic features with the difference vector between the latent features encoded by the HealthModel and DiseaseModel for the final task of survival analysis. The proposed DQSurv model stably outperformed the existing models for the survival analysis of 10 benchmark cancer types and an independent dataset. The ablation study also supported the necessity of the main modules. We released the codes and the pretrained HealthModel to facilitate the feature encodings and survival analysis of transcriptome-based future studies, especially on small datasets. The model and the code are available at http://www.healthinformaticslab.org/supp/.

Human Leukocyte Antigen (HLA) is a type of molecule residing on the surfaces of most human cells and exerts an essential role in the immune system responding to the invasive items. The T cell antigen receptors may recognize the HLA-peptide complexes on the surfaces of cancer cells and destroy these cancer cells through toxic T lymphocytes. The computational determination of HLA-binding peptides will facilitate the rapid development of cancer immunotherapies. This study hypothesized that the natural language processing-encoded peptide features may be further enriched by another deep neural network. The hypothesis was tested with the Bi-directional Long Short-Term Memory-extracted features from the pretrained Protein Bidirectional Encoder Representations from Transformers-encoded features of the class I HLA (HLA-I)-binding peptides. The experimental data showed that our proposed HLAB feature engineering algorithm outperformed the existing ones in detecting the HLA-I-binding peptides. The extensive evaluation data show that the proposed HLAB algorithm outperforms all the seven existing studies on predicting the peptides binding to the HLA-A*01:01 allele in AUC and achieves the best average AUC values on the six out of the seven k-mers (k=8,9,...,14, respectively represent the prediction task of a polypeptide consisting of k amino acids) except for the 9-mer prediction tasks. The source code and the fine-tuned feature extraction models are available at http://www.healthinformaticslab.org/supp/resources.php.

MOTIVATION: DNA methylation is a biological process impacting the gene functions without changing the underlying DNA sequence. The DNA methylation machinery usually attaches methyl groups to some specific cytosine residues, which modify the chromatin architectures. Such modifications in the promoter regions will inactivate some tumor-suppressor genes. DNA methylation within the coding region may significantly reduce the transcription elongation efficiency. The gene function may be tuned through some cytosines are methylated. METHODS: This study hypothesizes that the overall methylation level across a gene may have a better association with the sample labels like diseases than the methylations of individual cytosines. The gene methylation level is formulated as a regression model using the methylation levels of all the cytosines within this gene. A comprehensive evaluation of various feature selection algorithms and classification algorithms is carried out between the gene-level and residue-level methylation levels. RESULTS: A comprehensive evaluation was conducted to compare the gene and cytosine methylation levels for their associations with the sample labels and classification performances. The unsupervised clustering was also improved using the gene methylation levels. Some genes demonstrated statistically significant associations with the class label, even when no residue-level methylation features have statistically significant associations with the class label. So in summary, the trained gene methylation levels improved various methylome-based machine learning models. Both methodology development of regression algorithms and experimental validation of the gene-level methylation biomarkers are worth of further investigations in the future studies. The source code, example data files and manual are available at http://www.healthinformaticslab.org/supp/.

The novel coronavirus (2019-nCoV) has recently caused a large-scale outbreak of viral pneumonia both in China and worldwide. In this study, we obtained the entire genome sequence of 777 new coronavirus strains as of 29 February 2020 from a public gene bank. Bioinformatics analysis of these strains indicated that the mutation rate of these new coronaviruses is not high at present, similar to the mutation rate of the severe acute respiratory syndrome (SARS) virus. The similarities of 2019-nCoV and SARS virus suggested that the S and ORF6 proteins shared a low similarity, while the E protein shared the higher similarity. The 2019-nCoV sequence has similar potential phosphorylation sites and glycosylation sites on the surface protein and the ORF1ab polyprotein as the SARS virus; however, there are differences in potential modification sites between the Chinese strain and some American strains. At the same time, we proposed two possible recombination sites for 2019-nCoV. Based on the results of the skyline, we speculate that the activity of the gene population of 2019-nCoV may be before the end of 2019. As the scope of the 2019-nCoV infection further expands, it may produce different adaptive evolutions due to different environments. Finally, evolutionary genetic analysis can be a useful resource for studying the spread and virulence of 2019-nCoV, which are essential aspects of preventive and precise medicine.

MOTIVATION: Deep neural network (DNN) algorithms were utilized in predicting various biomedical phenotypes recently, and demonstrated very good prediction performances without selecting features. This study proposed a hypothesis that the DNN models may be further improved by feature selection algorithms. RESULTS: A comprehensive comparative study was carried out by evaluating 11 feature selection algorithms on three conventional DNN algorithms, i.e. convolution neural network (CNN), deep belief network (DBN) and recurrent neural network (RNN), and three recent DNNs, i.e. MobilenetV2, ShufflenetV2 and Squeezenet. Five binary classification methylomic datasets were chosen to calculate the prediction performances of CNN/DBN/RNN models using feature selected by the 11 feature selection algorithms. Seventeen binary classification transcriptome and two multi-class transcriptome datasets were also utilized to evaluate how the hypothesis may generalize to different data types. The experimental data supported our hypothesis that feature selection algorithms may improve DNN models, and the DBN models using features selected by SVM-RFE usually achieved the best prediction accuracies on the five methylomic datasets. AVAILABILITY AND IMPLEMENTATION: All the algorithms were implemented and tested under the programming environment Python version 3.6.6. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

LIDAR (light detection and ranging) based real‐time 3D perception is crucial for applications such as autonomous driving. However, most of the convolutional neural network (CNN) based methods are time‐consuming and computation‐intensive. These drawbacks are mainly attributed to the highly variable density of LIDAR point cloud and the complexity of their pipelines. To find a balance between speed and accuracy for 3D object detection from LIDAR, authors propose RTL3D, a computationally efficient Real‐time LIDAR‐based 3D detector. In RTL3D, an effective voxel‐wise feature representation is utilised to organise unstructured point cloud. By employing a sparse feature learning network (SFLN) on voxelised 3D data, RTL3D exploits the sparsity of point cloud and down‐samples 3D data into 2D. Basing on the generated 2D feature map, an optimised dense detection network (DDN) is applied to regress the oriented bounding box without relying on any predefined anchor boxes. The authors also introduce an incremental data augmentation approach which greatly improves the performance of RTL3D. Empirical experiments on public KITTI benchmark demonstrate that RTL3D achieves a competitive performance with state‐of‐the‐art works on 3D detection task. Owning to the simplicity of its single‐stage and anchor‐free design, RTL3D has a real‐time inference speed of 40 FPS.