Cancer treatments often yield different results from patient to patient due to the genomic heterogeneity of tumors. Accurately predicting a patient's response to an anticancer drug is challenging, especially when using traditional machine learning models trained on cell lines and applied to patient data. These models struggle with domain shift (out-of-distribution data due to differences between cell line and patient data), loss of tissue specificity, and data imbalance across domains. To address these issues, we developed the Tissue-Aware and Prototype Learning for Drug Response Prediction (TAPL-DRP) model. This model predicts anticancer drug responses using a two-stage process. At the stage of tissue-aware cross-domain feature extraction, we integrate a Variational Autoencoder (VAE) and a Generative Adversarial Network (GAN) to extract features from both cell line and patient data. This process incorporates tissue prototypes, InfoNCE loss, and class-balance loss to ensure the features are not only domain-invariant but also biologically meaningful and tissue-specific. At the drug response prediction stage, the model combines these extracted features with molecular drug graph features. It then uses the tissue prototypes to guide the training of the classifier, which further improves the prediction accuracy. We tested TAPL-DRP on the TCGA clinical dataset and the PDTC in vitro dataset. The results show that our model significantly outperforms other methods in key metrics like AUC, AUPRC, ACC, and MCC. This demonstrates its effectiveness in handling domain shift and maintaining tissue specificity. Further analysis confirmed that the tissue prototypes, mutual information loss, and class-balance strategies are all crucial components of the model. In summary, TAPL-DRP offers an effective and precise solution for predicting anticancer drug responses in personalized medicine.The source code is available at https://github.com/weiba/TAPL-DRP.

Accurate prediction of drug responses is critical for advancing personalized cancer therapies. Although current graph neural network (GNN)-based approaches predominantly focus on pairwise interactions between cell lines and drugs, they often neglect the potential of higher-order interactions. In this study, we present HRLCDR, a novel computational framework that utilizes Hypergraph Representation Learning to predict Cancer Drug Responses. HRLCDR begins by constructing hypergraphs for both cell lines and drugs and then processes through low-pass and high-pass hypergraph convolutions, allowing the model to extract both common and different features from the complex higher-order interactions between cell lines and drugs. After that, HRLCDR constructs a heterogeneous graph using known cell line responses to drugs. Parallel heterogeneous graph convolution operations are then employed to extract primary interaction features between cell lines and drugs from these associations. Finally, HRLCDR integrates the features learned from both the hypergraphs and the heterogeneous graph, predicting drug response via Classifiers. We evaluated HRLCDR's performance on two major cancer drug response datasets: the Cancer Drug Sensitivity Data (GDSC) and the Cancer Cell Line Encyclopedia (CCLE). The results demonstrate that HRLCDR outperforms current state-of-the-art methods, underscoring its potential to enhance the accuracy and reliability of cancer drug response predictions.

Anticancer drug response prediction is crucial in developing personalized treatment plans for cancer patients. However, High-quality patient anticancer drug response data are scarce and cell line data and patient data have different distributions, models trained solely on cell line data perform poorly. Some existing methods predict anticancer drug response by transferring knowledge from the cell line domain to the patient domain using transfer learning. However, the robustness of these classifiers is affected by anomalies in the cell line data, and they do not utilize the knowledge in the unlabeled target domain data. To this end, we proposed a model called DAPL to predict patient responses to anticancer drugs. The model extracts domain-invariant features from cell lines and patients by constructing multiple VAEs and extracts drug features using GNNs. These features are then combined for prototypical learning to train a classifier, resulting in better predictions of patient anticancer drug response. We used the cell line datasets CCLE and GDSC as source domains and the patient datasets TCGA and PDTC as target domains and conducted experiments. The results indicate that DAPL shows excellent performance in predicting patient anticancer drug response compared to other state-of-the-art methods.

Patients with the same type of cancer often respond differently to identical drug treatments due to unique genomic traits. Accurately predicting a patient's response to drug is crucial in guiding treatment decisions, alleviating patient suffering, and improving cancer prognosis. Current computational methods utilize deep learning models trained on extensive drug screening data to predict anti-cancer drug responses based on features of cell lines and drugs. However, the interaction between cell lines and drugs is a complex biological process involving interactions across various levels, from internal cellular and drug structures to the external interactions among different molecules.To address this complexity, we propose a novel Hierarchical graph representation Learning with Multi-Granularity features (HLMG) algorithm for predicting anti-cancer drug responses. The HLMG algorithm combines features at two granularities: the overall gene expression and pathway substructures of cell lines, and the overall molecular fingerprints and substructures of drugs. Subsequently, it constructs a heterogeneous graph including cell lines, drugs, known cell line-drug responses, and the associations between similar cell lines and similar drugs. Through a graph convolutional network model, the HLMG learns the final cell line and drug representations by aggregating features of their multi-level neighbor in the heterogeneous graph. The multi-level neighbors consist of the node self, directly related drugs/cell lines, and indirectly related similar drugs/cell lines. Finally, a linear correlation coefficient decoder is employed to reconstruct the cell line-drug correlation matrix to predict anti-cancer drug responses. Our model was tested on the Genomics of Drug Sensitivity in Cancer (GDSC) and the Cancer Cell Line Encyclopedia (CCLE) databases. Results indicate that HLMG outperforms other state-of-the-art methods in accurately predicting anti-cancer drug responses.

Using deep learning models on spatial transcriptomics data to identify the spatial domain is crucial for uncovering the spatial distribution of cells and gene expression patterns within tissues, essential for understanding complex biological processes and disease mechanisms. Existing methods for spatial domain partitioning often rely on predefined adjacency relationships at a single scale, overlooking the hierarchical structure and functional characteristics of biological tissues. In this paper, we propose SpaNFM, a novel method that leverages sparse attention-based hierarchical node representation and multi-view contrastive learning for spatial domain identification in spatial transcriptomics data. The SpaNFM first treats each spot as a node and constructs two views using different data augmentation techniques based on tissue image information, gene expression profiles, and spatial coordinates of cells. Subsequently, SpaNFM utilizes a sparse attention-based hierarchical node fusion module to generate coarse-grained node representations. This fine-to-coarse hierarchical structure integrates complementary information from multi-granularity node features and reduces model complexity due to the decreased node size. The model parameters are updated using gene expression reconstruction loss and contrastive loss on the coarse-grained node representations from the two views. Finally, the learned node features are subjected to downstream clustering using the Leiden algorithm. We tested SpaNFM on the human dorsolateral prefrontal cortex dataset. The results demonstrate that SpaNFM outperforms other state-of-the-art methods in most cases. The data and code are available at: https://github.com/weiba/SpaNFM

The correct prediction of disease-associated miRNAs plays an essential role in disease prevention and treatment. Current computational methods to predict disease-associated miRNAs construct different miRNA views and disease views based on various miRNA properties and disease properties and then integrate the multiviews to predict the relationship between miRNAs and diseases. However, most existing methods ignore the information interaction among the views and the consistency of miRNA features (disease features) across multiple views. This study proposes a computational method based on multiple hypergraph contrastive learning (MHCLMDA) to predict miRNA-disease associations. MHCLMDA first constructs multiple miRNA hypergraphs and disease hypergraphs based on various miRNA similarities and disease similarities and performs hypergraph convolution on each hypergraph to capture higher order interactions between nodes, followed by hypergraph contrastive learning to learn the consistent miRNA feature representation and disease feature representation under different views. Then, a variational auto-encoder is employed to extract the miRNA and disease features in known miRNA-disease association relationships. Finally, MHCLMDA fuses the miRNA and disease features from different views to predict miRNA-disease associations. The parameters of the model are optimized in an end-to-end way. We applied MHCLMDA to the prediction of human miRNA-disease association. The experimental results show that our method performs better than several other state-of-the-art methods in terms of the area under the receiver operating characteristic curve and the area under the precision-recall curve.

Clustering individual cells or spots based on their gene expression profiles in a spatial context is a powerful approach to uncovering the underlying biological diversity and relationships among cells. The intricate information within spatial transcriptomics data demands sophisticated algorithms that effectively integrate gene expression, cell position, and tissue image data for accurate cell or spot clustering. This work proposes a Multi-View Comparative Learning method for clustering Spatial Transcriptomics data (MVCLST). MVCLST first builds on two data views using gene expression profiles, cell space coordinates, and image features. Then it employs four different encoders to capture the common and private features of the two views. The model employs a contrastive learning loss to encourage effective interaction between the two views and ensure feature consistency. The shared and private features from both views are fused using corresponding decoders. Finally, the model employs the Leiden algorithm for downstream clustering of the learned features. We test the MVCLST method on a human dorsolateral prefrontal cortex dataset. The results show that MVCLST outperforms other state-of-the-art methods in most cases. Additionally, the clusters identified by MVCLST align closely with manual annotations and established neuroscience definitions.

BACKGROUND: Correctly identifying the driver genes that promote cell growth can significantly assist drug design, cancer diagnosis and treatment. The recent large-scale cancer genomics projects have revealed multi-omics data from thousands of cancer patients, which requires to design effective models to unlock the hidden knowledge within the valuable data and discover cancer drivers contributing to tumorigenesis. RESULTS: In this work, we propose a graph convolution network-based method called MRNGCN that integrates multiple gene relationship networks to identify cancer driver genes. First, we constructed three gene relationship networks, including the gene-gene, gene-outlying gene and gene-miRNA networks. Then, genes learnt feature presentations from the three networks through three sharing-parameter heterogeneous graph convolution network (HGCN) models with the self-attention mechanism. After that, these gene features pass a convolution layer to generate fused features. Finally, we utilized the fused features and the original feature to optimize the model by minimizing the node and link prediction losses. Meanwhile, we combined the fused features, the original features and the three features learned from every network through a logistic regression model to predict cancer driver genes. CONCLUSIONS: We applied the MRNGCN to predict pan-cancer and cancer type-specific driver genes. Experimental results show that our model performs well in terms of the area under the ROC curve (AUC) and the area under the precision-recall curve (AUPRC) compared to state-of-the-art methods. Ablation experimental results show that our model successfully improved the cancer driver identification by integrating multiple gene relationship networks.

MiRNAs are reported to be linked to the pathogenesis of human complex diseases. Disease-related miRNAs may serve as novel bio-marks and drug targets. This work focuses on designing a multi-relational Graph Convolutional Network model to predict miRNA-disease associations (HGCNMDA) from a Heterogeneous network. HGCNMDA introduces a gene layer to construct a miRNA-gene-disease heterogeneous network. We refine the features of nodes into initial and inductive features so that the direct and indirect associations between diseases and miRNA can be considered simultaneously. Then HGCNMDA learns feature embeddings for miRNAs and disease through a multi-relational graph convolutional network model that can assign appropriate weights to different types of edges in the heterogeneous network. Finally, the miRNA-disease associations were decoded by the inner product between miRNA and disease feature embeddings. We apply our model to predict human miRNA-disease associations. The HGCNMDA is superior to the other state-of-the-art models in identifying missing miRNA-disease associations and also performs well on recommending related miRNAs/diseases to new diseases/ miRNAs. The codes are available at https://github.com/weiba/HGCNMDA.

Microbes play a crucial role in human health and disease. Figuring out the relationship between microbes and diseases leads to significant potential applications in disease treatments. It is an urgent need to devise robust and effective computational methods for identifying disease-related microbes. This work proposes a Multi-View Feature Aggregation (MVFA) scheme that integrates the linear and nonlinear features to identify disease-related microbes. We introduce a non-negative matrix tri-factorization (NMTF) model to extract linear features for diseases and microbes. Then we learn another type of linear feature by utilizing a bi-random walk model. The nonlinear feature is obtained by inputting the two kinds of linear features into a capsule neural network. These three types of features describe the associations between diseases and microbes from different views. Finally, considering the complementary of these features, we leverage a logistic regression model to combine the NMTF model predictions, bi-random walk model predictions, and the capsule neural network predictions to obtain the final microbe-disease pair scores. We apply our method to predict human microbe-disease associations on two datasets. Experimental results show that our multi-view model outperforms the state-of-the-art models in recovering missing microbe-disease associations and predicting associations for new microbes. The ablation study shows that aggregating multi-view linear and nonlinear features can improve the prediction performance. Case studies on two diseases, i.e. Type 1 diabetes and Liver cirrhosis, further validate our method effectiveness.

The correct identification of the driver genes that lead to cancer development is essential for understanding the mechanisms of cancer and developing drugs to treat it. Currently, most computational methods for identifying cancer driver genes are based on a cohort of patients. However, due to the heterogeneity of cancers, patients diagnosed with the same cancers may have different genomic characteristics and present varied clinical symptoms. It requires devising effective methods to identify personalized cancer driver genes in an individual. This work developed a novel method to predict personalized cancer driver genes of a single sample based on graph convolution networks, namely pDriverGCN. pDriverGCN constructed a mutant gene-sample heterogeneous network according to the known driver genes of samples. Then it employed two separate graph convolution network models to learn feature representations for genes and samples by gathering the features of themselves and their neighbors. Finally, pDriverGCN used the feature representations to reconstruct the association matrix between genes and samples through a linear correlation coefficient decoder. We apply our model to identify personalized driver genes of samples on the TCGA datasets. The experimental results show that our model outperforms state-of-the-art methods being evaluated at both population and individual levels.

Cancer is thought to be caused by the accumulation of driver genetic mutations. Therefore, identifying cancer driver genes plays a crucial role in understanding the molecular mechanism of cancer and developing precision therapies and biomarkers. In this work, we propose a Multi-Task learning method, called MTGCN, based on the Graph Convolutional Network to identify cancer driver genes. First, we augment gene features by introducing their features on the protein-protein interaction (PPI) network. After that, the multi-task learning framework propagates and aggregates nodes and graph features from input to next layer to learn node embedding features, simultaneously optimizing the node prediction task and the link prediction task. Finally, we use a Bayesian task weight learner to balance the two tasks automatically. The outputs of MTGCN assign each gene a probability of being a cancer driver gene. Our method and the other four existing methods are applied to predict cancer drivers for pan-cancer and some single cancer types. The experimental results show that our model shows outstanding performance compared with the state-of-the-art methods in terms of the area under the Receiver Operating Characteristic (ROC) curves and the area under the precision-recall curves. The MTGCN is freely available via https://github.com/weiba/MTGCN.

MOTIVATION: Due to cancer heterogeneity, the therapeutic effect may not be the same when a cohort of patients of the same cancer type receive the same treatment. The anticancer drug response prediction may help develop personalized therapy regimens to increase survival and reduce patients' expenses. Recently, graph neural network-based methods have aroused widespread interest and achieved impressive results on the drug response prediction task. However, most of them apply graph convolution to process cell line-drug bipartite graphs while ignoring the intrinsic differences between cell lines and drug nodes. Moreover, most of these methods aggregate node-wise neighbor features but fail to consider the element-wise interaction between cell lines and drugs. RESULTS: This work proposes a neighborhood interaction (NI)-based heterogeneous graph convolution network method, namely NIHGCN, for anticancer drug response prediction in an end-to-end way. Firstly, it constructs a heterogeneous network consisting of drugs, cell lines and the known drug response information. Cell line gene expression and drug molecular fingerprints are linearly transformed and input as node attributes into an interaction model. The interaction module consists of a parallel graph convolution network layer and a NI layer, which aggregates node-level features from their neighbors through graph convolution operation and considers the element-level of interactions with their neighbors in the NI layer. Finally, the drug response predictions are made by calculating the linear correlation coefficients of feature representations of cell lines and drugs. We have conducted extensive experiments to assess the effectiveness of our model on Cancer Drug Sensitivity Data (GDSC) and Cancer Cell Line Encyclopedia (CCLE) datasets. It has achieved the best performance compared with the state-of-the-art algorithms, especially in predicting drug responses for new cell lines, new drugs and targeted drugs. Furthermore, our model that was well trained on the GDSC dataset can be successfully applied to predict samples of PDX and TCGA, which verified the transferability of our model from cell line in vitro to the datasets in vivo. AVAILABILITY AND IMPLEMENTATION: The source code can be obtained from https://github.com/weiba/NIHGCN. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Different cancer patients may respond differently to cancer treatment due to the heterogeneity of cancer. It is an urgent task to develop an efficient computational method to identify drug responses in different cell lines, which guides us to design personalized therapy for an individual patient. Hence, we propose an end-to-end algorithm, namely MOFGCN, to predict drug response in cell lines based on Multi-Omics Fusion and Graph Convolution Network. MOFGCN first fuses multiple omics data to calculate the cell line similarity and then constructs a heterogeneous network by combining the cell line similarity, drug similarity, and the known cell line-drug associations. Secondly, it learns the latent features for cancer cell lines and drugs by performing graph convolution operations on the heterogeneous network. Finally, MOFGCN applies the linear correlation coefficient to reconstruct the cancer cell line-drug correlation matrix to predict drug sensitivity. To our knowledge, this is the first attempt to combine graph convolutional neural network and linear correlation coefficient for this significant task. We performed extensive evaluation experiments on the Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE) databases to validate MOFGCN's performance. The experimental results show that MOFGCN is superior to the state-of-the-art algorithms in predicting missing drug responses. It also leads to higher performance in predicting drug responses for new cell lines, new drugs, and targeted drugs.

Microbes play a crucial role in human health and disease. Understanding the relationship between microbes and diseases is conducive to the treatment and diagnosis of diseases. Recently, many computational methods have been proposed to predict disease-microbe associations. However, most of the existing methods only consider a single model and explore the disease-microbe associations from a single view. To improve the prediction accuracy, we propose a novel multi-view approach that fuses the linear and nonlinear features to predict new potential associations between diseases and microbes. We first design a non-negative matrix tri-factorization method to extract the linear features of diseases and microbes. We input the linear features from the non-negative matrix tri-factorization model and bi-random walk model into a capsule neural network to obtain the diseases and microbes' nonlinear features. Finally, we leverage a logistic regression model to combine the non-negative matrix tri-factorization model predictions, bi-random walk model predictions and the capsule neural network predictions to obtain the final association scores between microbes and diseases. We apply our method to predict human microbedisease associations. Experimental results show that our fusion model outperforms the non-negative matrix tri-factorization model, bi-random walk model and other existing models.

Protein function prediction is necessary for understanding life and is valuable for application on drug design and health care. It is still a big challenge to predict protein function correctly by integrating multi-biological information. In this work, we propose a novel non-Negative Matrix Factorization (NMF) based method regularized by PPI network and GO similarity network, namely PONMF, for protein function prediction, which decomposes the known GO-protein association matrix into two low-rank matrixes for proteins and GO terms. In the course of factorization, PONMF incorporates the label matrix factorization term, an additional network regularization term and a GO similarity regularization term into the objective function. Finally, the potential protein functions are predicted by referring to the product of the two low-rank matrixes. PONMF not only successfully integrates diverse biological information to predict protein functions, but also naturally partitions proteins into different modules and infers functions from the proteins in the same modules. Our methods as well as the other two state-of-the-art methods (UBiRW and NMFGO) are applied to predict functions for protein of S. cerevisiae and H. sapiens. The prediction results show that PONMF outperforms the other two existing methods.