The type of primary tumor responsible for brain metastases is crucial for clinical treatment. However, current approaches predominantly focus on issues such as lesion segmentation in brain MRI imaging. While some studies have attempted to predict primary tumor types using statistical methods, none have established a direct predictive relationship between brain imaging and primary tumor types. With the rapid advancement of large models, adapting pre-trained models for various tasks has become feasible. In this paper, we propose a primary tumor type prediction method called clinicViT, based on a pre-trained large model. This method integrates MRI imaging features with clinical information features. By fine-tuning a pretrained encoder model with the addition of a fusion module and classifier, the clinicViT model directly predicts primary tumor types. Furthermore, utilizing the imaging and clinical features extracted by clinicViT, survival analysis for different primary tumor types is achieved. Extensive experimental results demonstrate that clinicViT outperforms state-of-the-art methods.

Protein metabolite interactions (PMIs) play a crucial role in cellular homeostasis, supporting drug development and revealing biological processes. Traditional detection methods are constrained by resource limitations, making large-scale PMI identification challenging. Existing machine learning approaches for PMI prediction often fail to capture correlation features between proteins and metabolites and struggle with crossspecies and batch effect problems. We propose GCN Embedding-Multihead Focusing Attention (GE-MFAT), a novel method that combines heterogeneous graph neural networks with a focusing attention mechanism to extract complex PMI relationships. Our model employs transfer learning to address cross-species and batch effect challenges. Experimental results across three datasets demonstrate that GE-MFAT significantly outperforms state-of-the-art methods across all evaluation metrics.

Function annotation of proteins is fundamental for revealing the nature of life phenomena and understanding the mechanisms of disease. Several computational methods have been developed to predict protein functions. However, existing methods ignore the prior knowledge among Gene ontology (GO) terms and simply regard the prediction problem as an independent multilabel classification task. To mitigate these issues, we propose a novel predictor based on graph representation learning and iterative optimization, named GDTGO, to explore the potential protein function on the GO. GDTGO employs a graph convolutional network to learn semantically rich and topologically aware knowledge from GO terms, which serves as a strong prior to guide prediction. Furthermore, we model the function prediction task as an iterative optimization problem for set prediction. A DETR-based decoder dynamically refines predictions through a series of layers, where each layer provides corrective feedback to progressively enhance the final output. Experimental results show that GDTGO achieves state-of-the-art performance on the PDB dataset.

Non-coding RNAs (ncRNAs) play crucial roles in drug resistance and sensitivity, making them important biomarkers and therapeutic targets. However, predicting ncRNA-drug associations is challenging due to issues such as dataset imbalance and sparsity, limiting the identification of robust biomarkers. Existing models often fall short in capturing local and global sequence information, limiting the reliability of predictions. This study introduces DMGAT (diffusion map and heterogeneous graph attention network), a novel deep learning model designed to predict ncRNA-drug associations. DMGAT integrates diffusion maps for sequence embedding, graph convolutional networks for feature extraction, and GAT for heterogeneous information fusion. To address dataset imbalance, the model incorporates sensitivity associations and employs a random forest classifier to select reliable negative samples. DMGAT embeds ncRNA sequences and drug SMILES using the word2vec technique, capturing local and global sequence information. The model constructs a heterogeneous network by combining sequence similarity and Gaussian Interaction Profile kernel similarity, providing a comprehensive representation of ncRNA-drug interactions. Evaluated through five-fold cross-validation on a curated dataset from NoncoRNA and ncDR, DMGAT outperforms seven state-of-the-art methods, achieving the highest area under the receiver operating characteristic curve (0.8964), area under the precision-recall curve (0.8984), recall (0.9576), and F1-score (0.8285). The raw data are released to Zenodo with identifier 13929676. The source code of DMGAT is available at https://github.com/liutingyu0616/DMGAT/tree/main.

Piwi-interacting RNAs (piRNAs) play a crucial role in various biological processes and are implicated in disease. Consequently, there is an escalating demand for computational tools to predict piRNA-disease interactions. Although there have been computational methods proposed for the detection of piRNA-disease associations, the problem of imbalanced and sparse dataset has brought great challenges to capture the complex relationships between piRNAs and diseases. In response to this necessity, we have developed a novel computational architecture, denoted as PUTransGCN, which uses heterogeneous graph convolutional networks to uncover potential piRNA-disease associations. Additionally, the attention mechanism was used to adjust the weight parameters of aggregation heterogeneous node features automatically. For tackling the imbalanced dataset problem, the combined positive unlabelled learning (PUL) method comprising PU bagging, two-step and spy technique was applied to select reliable negative associations. The features of piRNAs and diseases were derived from three distinct biological sources by PUTransGCN, including information on piRNA sequences, semantic terms related to diseases and the existing network of piRNA-disease associations. In the experiment, PUTransGCN performs in 5-fold cross-validation with an AUC of 0.93 and 0.95 on two datasets, respectively, which outperforms the other six state-of-the-art models. We compared three different PUL methods, and the results of the ablation experiment indicate that the combined PUL method yields the best results. The PUTransGCN could serve as a valuable piRNA-disease prediction tool for upcoming studies in the biomedical field. The code for PUTransGCN is available at https://github.com/chenqiuhao/PUTransGCN.

Spatial transcriptomics unveils the complex dynamics of cell regulation and transcriptomes, but it is typically cost-prohibitive. Predicting spatial gene expression from histological images via artificial intelligence offers a more affordable option, yet existing methods fall short in extracting deep-level information from pathological images. In this paper, we present THItoGene, a hybrid neural network that utilizes dynamic convolutional and capsule networks to adaptively sense potential molecular signals in histological images for exploring the relationship between high-resolution pathology image phenotypes and regulation of gene expression. A comprehensive benchmark evaluation using datasets from human breast cancer and cutaneous squamous cell carcinoma has demonstrated the superior performance of THItoGene in spatial gene expression prediction. Moreover, THItoGene has demonstrated its capacity to decipher both the spatial context and enrichment signals within specific tissue regions. THItoGene can be freely accessed at https://github.com/yrjia1015/THItoGene.

The prediction of molecular interactions is vital for drug discovery. Existing methods often focus on individual prediction tasks and overlook the relationships between them. Additionally, certain tasks encounter limitations due to insufficient data availability, resulting in limited performance. To overcome these limitations, we propose KGE-UNIT, a unified framework that combines knowledge graph embedding (KGE) and multi-task learning, for simultaneous prediction of drug-target interactions (DTIs) and drug-drug interactions (DDIs) and enhancing the performance of each task, even when data availability is limited. Via KGE, we extract heterogeneous features from the drug knowledge graph to enhance the structural features of drug and protein nodes, thereby improving the quality of features. Additionally, employing multi-task learning, we introduce an innovative predictor that comprises the task-aware Convolutional Neural Network-based (CNN-based) encoder and the task-aware attention decoder which can fuse better multimodal features, capture the contextual interactions of molecular tasks and enhance task awareness, leading to improved performance. Experiments on two imbalanced datasets for DTIs and DDIs demonstrate the superiority of KGE-UNIT, achieving high area under the receiver operating characteristics curves (AUROCs) (0.942, 0.987) and area under the precision-recall curve ( AUPRs) (0.930, 0.980) for DTIs and high AUROCs (0.975, 0.989) and AUPRs (0.966, 0.988) for DDIs. Notably, on the LUO dataset where the data were more limited, KGE-UNIT exhibited a more pronounced improvement, with increases of 4.32$\%$ in AUROC and 3.56$\%$ in AUPR for DTIs and 6.56$\%$ in AUROC and 8.17$\%$ in AUPR for DDIs. The scalability of KGE-UNIT is demonstrated through its extension to protein-protein interactions prediction, ablation studies and case studies further validate its effectiveness.

MOTIVATION: Exploring human-virus protein-protein interactions (PPIs) is crucial for unraveling the underlying pathogenic mechanisms of viruses. Limitations in the coverage and scalability of high-throughput approaches have impeded the identification of certain key interactions. Current popular computational methods adopt a two-stream pipeline to identify PPIs, which can only achieve relation modeling of protein pairs at the classification phase. However, the fitting capacity of the classifier is insufficient to comprehensively mine the complex interaction patterns between protein pairs. RESULTS: In this study, we propose a pioneering single-stream framework HBFormer that combines hybrid attention mechanism and multimodal feature fusion strategy for identifying human-virus PPIs. The Transformer architecture based on hybrid attention can bridge the bidirectional information flows between human protein and viral protein, thus unifying joint feature learning and relation modeling of protein pairs. The experimental results demonstrate that HBFormer not only achieves superior performance on multiple human-virus PPI datasets but also outperforms 5 other state-of-the-art human-virus PPI identification methods. Moreover, ablation studies and scalability experiments further validate the effectiveness of our single-stream framework. AVAILABILITY AND IMPLEMENTATION: Codes and datasets are available at https://github.com/RmQ5v/HBFormer.

Distant metastasis is the leading cause of cancer death. Identifying the tendency of a given cancer to metastasize could be conducive to cancer diagnosis and therapeutic schedules. In cancer studies, mRNA gene expression data have been widely used to predict cancer metastasis due to the ease with which they can be obtained. Moreover, mRNA gene expression data represent cancer progression directly and in detail. In these studies, feature extraction followed by a prediction model has been a commonly used solution to predict pan-cancer prognosis and tumor stage. Limitations of these studies include a lack of comprehensive feature extraction, relatively low prediction accuracy of cancer outcomes and a lack of precise pan-cancer metastasis site prediction.To address the questions mentioned above, we designed an innovative pipeline to determine the heterogeneity of pan-cancer distant metastatic sites using mRNA gene expression data. We used a directed relational graph convolutional network (DRGCN) for feature extraction and a multikernel support vector machine (SVM) for pan-cancer distant metastasis site prediction. DR-GCN successfully excavated hidden features from relational networks and effectively extracted features from gene-cancer relations, cancer-disease relations and gene-gene relational networks. DR-GCN was demonstrably able to deal with complex prior knowledge-based feature extraction tasks. A dynamic weight multikernel SVM was then applied to predict pan-cancer distant metastasis sites. By this method, the AUROC (0.7542) of the multikernel SVM outperformed that of the single kernel SVM (polykernel: 0.7346, RBF kernel: 0.72, linear kernel: 0. 725S). We last applied our pipeline to an extremely unbalanced small sample dataset and obtained a higher AUPRC (0.2606) than other semisupervised learning methods (Laplacian SVM: 0. 1S, TSVM: 0.21, SSL-EM: 0. 1S, RRLSL: 0.22) while predicting TCGA glioblastoma (GBM) patient prognosis.

MOTIVATION: The functional changes of the genes, RNAs and proteins will eventually be reflected in the metabolic level. Increasing number of researchers have researched mechanism, biomarkers and targeted drugs by metabolites. However, compared with our knowledge about genes, RNAs, and proteins, we still know few about diseases-related metabolites. All the few existed methods for identifying diseases-related metabolites ignore the chemical structure of metabolites, fail to recognize the association pattern between metabolites and diseases, and fail to apply to isolated diseases and metabolites. RESULTS: In this study, we present a graph deep learning based method, named Deep-DRM, for identifying diseases-related metabolites. First, chemical structures of metabolites were used to calculate similarities of metabolites. The similarities of diseases were obtained based on their functional gene network and semantic associations. Therefore, both metabolites and diseases network could be built. Next, Graph Convolutional Network (GCN) was applied to encode the features of metabolites and diseases, respectively. Then, the dimension of these features was reduced by Principal components analysis (PCA) with retainment 99% information. Finally, Deep neural network was built for identifying true metabolite-disease pairs (MDPs) based on these features. The 10-cross validations on three testing setups showed outstanding AUC (0.952) and AUPR (0.939) of Deep-DRM compared with previous methods and similar approaches. Ten of top 15 predicted associations between diseases and metabolites got support by other studies, which suggests that Deep-DRM is an efficient method to identify MDPs. CONTACT: [email protected]. AVAILABILITY AND IMPLEMENTATION: https://github.com/zty2009/GPDNN-for-Identify-ing-Disease-related-Metabolites.

Identification of new drug-target interactions (DTIs) is an important but a time-consuming and costly step in drug discovery. In recent years, to mitigate these drawbacks, researchers have sought to identify DTIs using computational approaches. However, most existing methods construct drug networks and target networks separately, and then predict novel DTIs based on known associations between the drugs and targets without accounting for associations between drug-protein pairs (DPPs). To incorporate the associations between DPPs into DTI modeling, we built a DPP network based on multiple drugs and proteins in which DPPs are the nodes and the associations between DPPs are the edges of the network. We then propose a novel learning-based framework, 'graph convolutional network (GCN)-DTI', for DTI identification. The model first uses a graph convolutional network to learn the features for each DPP. Second, using the feature representation as an input, it uses a deep neural network to predict the final label. The results of our analysis show that the proposed framework outperforms some state-of-the-art approaches by a large margin.

Interactions between proteins and small molecule metabolites play vital roles in regulating protein functions and controlling various cellular processes. The activities of metabolic enzymes, transcription factors, transporters and membrane receptors can all be mediated through protein-metabolite interactions (PMIs). Compared with the rich knowledge of protein-protein interactions, little is known about PMIs. To the best of our knowledge, no existing database has been developed for collecting PMIs. The recent rapid development of large-scale mass spectrometry analysis of biomolecules has led to the discovery of large amounts of PMIs. Therefore, we developed the PMI-DB to provide a comprehensive and accurate resource of PMIs. A total of 49 785 entries were manually collected in the PMI-DB, corresponding to 23 small molecule metabolites, 9631 proteins and 4 species. Unlike other databases that only provide positive samples, the PMI-DB provides non-interaction between proteins and metabolites, which not only reduces the experimental cost for biological experimenters but also facilitates the construction of more accurate algorithms for researchers using machine learning. To show the convenience of the PMI-DB, we developed a deep learning-based method to predict PMIs in the PMI-DB and compared it with several methods. The experimental results show that the area under the curve and area under the precision-recall curve of our method are 0.88 and 0.95, respectively. Overall, the PMI-DB provides a user-friendly interface for browsing the biological functions of metabolites/proteins of interest, and experimental techniques for identifying PMIs in different species, which provides important support for furthering the understanding of cellular processes. The PMI-DB is freely accessible at http://easybioai.com/PMIDB.

MOTIVATION: Although long non-coding RNAs (lncRNAs) have limited capacity for encoding proteins, they have been verified as biomarkers in the occurrence and development of complex diseases. Recent wet-lab experiments have shown that lncRNAs function by regulating the expression of protein-coding genes (PCGs), which could also be the mechanism responsible for causing diseases. Currently, lncRNA-related biological data are increasing rapidly. Whereas, no computational methods have been designed for predicting the novel target genes of lncRNA. RESULTS: In this study, we present a graph convolutional network (GCN) based method, named DeepLGP, for prioritizing target PCGs of lncRNA. First, gene and lncRNA features were selected, these included their location in the genome, expression in 13 tissues and miRNA-mediated lncRNA-gene pairs. Next, GCN was applied to convolve a gene interaction network for encoding the features of genes and lncRNAs. Then, these features were used by the convolutional neural network for prioritizing target genes of lncRNAs. In 10-cross validations on two independent datasets, DeepLGP obtained high area under curves (0.90-0.98) and area under precision-recall curves (0.91-0.98). We found that lncRNA pairs with high similarity had more overlapped target genes. Further experiments showed that genes targeted by the same lncRNA sets had a strong likelihood of causing the same diseases, which could help in identifying disease-causing PCGs. AVAILABILITY AND IMPLEMENTATION: https://github.com/zty2009/LncRNA-target-gene. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: Millions of people are suffering from cancers, but accurate early diagnosis and effective treatment are still tough for all doctors. Common ways against cancer include surgical operation, radiotherapy and chemotherapy. However, they are all very harmful for patients. Recently, the anticancer peptides (ACPs) have been discovered to be a potential way to treat cancer. Since ACPs are natural biologics, they are safer than other methods. However, the experimental technology is an expensive way to find ACPs so we purpose a new machine learning method to identify the ACPs. RESULTS: Firstly, we extracted the feature of ACPs in two aspects: sequence and chemical characteristics of amino acids. For sequence, average 20 amino acids composition was extracted. For chemical characteristics, we classified amino acids into six groups based on the patterns of hydrophobic and hydrophilic residues. Then, deep belief network has been used to encode the features of ACPs. Finally, we purposed Random Relevance Vector Machines to identify the true ACPs. We call this method 'DRACP' and tested the performance of it on two independent datasets. Its AUC and AUPR are higher than 0.9 in both datasets. CONCLUSION: We developed a novel method named 'DRACP' and compared it with some traditional methods. The cross-validation results showed its effectiveness in identifying ACPs.

Cancer is the most threat to human's health and life. At present, people have developed several ways to against cancer, such as surgical operation, radiotherapy and chemotherapy. However, cancers still cause highly mortality rate. A main part of the reason is that the traditional methods bring treatment effect as well as the negative effect. Recently, the anticancer peptides (ACPs) have been discovered which can be a new way to treat cancer. Since ACPs are natural biologics, they are safer than other methods. However, the experimental technology is an expensive way to find ACPs so we purpose a new machine learning method to identify the ACPs which named Random Relevance Vector Machines (RRVMs). The cross validations experiments show the high accuracy and stability of this new method.

BACKGROUND: Alzheimer's disease (AD) imposes a heavy burden on society and every family. Therefore, diagnosing AD in advance and discovering new drug targets are crucial, while these could be achieved by identifying AD-related proteins. The time-consuming and money-costing biological experiment makes researchers turn to develop more advanced algorithms to identify AD-related proteins. RESULTS: Firstly, we proposed a hypothesis "similar diseases share similar related proteins". Therefore, five similarity calculation methods are introduced to find out others diseases which are similar to AD. Then, these diseases' related proteins could be obtained by public data set. Finally, these proteins are features of each disease and could be used to map their similarity to AD. We developed a novel method 'LRRGD' which combines Logistic Regression (LR) and Gradient Descent (GD) and borrows the idea of Random Forest (RF). LR is introduced to regress features to similarities. Borrowing the idea of RF, hundreds of LR models have been built by randomly selecting 40 features (proteins) each time. Here, GD is introduced to find out the optimal result. To avoid the drawback of local optimal solution, a good initial value is selected by some known AD-related proteins. Finally, 376 proteins are found to be related to AD. CONCLUSION: Three hundred eight of three hundred seventy-six proteins are the novel proteins. Three case studies are done to prove our method's effectiveness. These 308 proteins could give researchers a basis to do biological experiments to help treatment and diagnostic AD.

BACKGROUND: Metabolites disrupted by abnormal state of human body are deemed as the effect of diseases. In comparison with the cause of diseases like genes, these markers are easier to be captured for the prevention and diagnosis of metabolic diseases. Currently, a large number of metabolic markers of diseases need to be explored, which drive us to do this work. METHODS: The existing metabolite-disease associations were extracted from Human Metabolome Database (HMDB) using a text mining tool NCBO annotator as priori knowledge. Next we calculated the similarity of a pair-wise metabolites based on the similarity of disease sets of them. Then, all the similarities of metabolite pairs were utilized for constructing a weighted metabolite association network (WMAN). Subsequently, the network was utilized for predicting novel metabolic markers of diseases using random walk. RESULTS: Totally, 604 metabolites and 228 diseases were extracted from HMDB. From 604 metabolites, 453 metabolites are selected to construct the WMAN, where each metabolite is deemed as a node, and the similarity of two metabolites as the weight of the edge linking them. The performance of the network is validated using the leave one out method. As a result, the high area under the receiver operating characteristic curve (AUC) (0.7048) is achieved. The further case studies for identifying novel metabolites of diabetes mellitus were validated in the recent studies. CONCLUSION: In this paper, we presented a novel method for prioritizing metabolite-disease pairs. The superior performance validates its reliability for exploring novel metabolic markers of diseases.

The collaborations of the diseases might be the key to understand the mechanism of the diseases since it is difficult to detect the role of complex genes and micro RNA in diseases. With the rapid development of technology, several metabolites of many kinds of diseases could be obtained by the advanced machines. Some diseases are related to several metabolites, and some metabolites have strong relationship with several diseases. Since there is certain relationship between different diseases, firstly we should find the similarity of different diseases. Then the similarity of different metabolites could be calculated by the relationship of diseases. Then a network of metabolites' similarity could be built. After building up the network, the diseases might not only relate to originally several metabolites, but more related metabolites to the disease could be found by the lines of the network. It can be used to explain the mechanism of the diseases more precisely. These metabolites could also be candidates to map the diseases in terms of the similarities. We can also sort these potentially relevant metabolites by similarity. It offers researchers a novel way to find out metabolites which is related to diseases.

More than 1/3 of human genes are regulated by microRNAs. The identification of microRNA (miRNA) is the precondition of discovering the regulatory mechanism of miRNA and developing the cure for genetic diseases. The traditional identification method is biological experiment, but it has the defects of long period, high cost, and missing the miRNAs that only exist in a specific period or low expression level. Therefore, to overcome these defects, machine learning method is applied to identify miRNAs. In this study, for identifying real and pseudo miRNAs and classifying different species, we extracted 98 dimensional features based on the primary and secondary structure, then we proposed the BP-Adaboost method to figure out the overfitting phenomenon of BP neural network by constructing multiple BP neural network classifiers and distributed weights to these classifiers. The novel method we proposed raised the accuracy and the stability. In this study, we verified the effectiveness and superiority over other methods by experiments.