Long video delivery over wireless links whose SNR (Signal-to-Noise Ratio) falls below 5 dB often hits a cliff: conventional codecs combined with FEC remain decodable, but the decoded frames collapse into blocky mosaics that are semantically unusable. This paper presents AdaGenCom (Adaptive Generative Communication), a multi-modal semantic communication framework that preserves meaning when pixel recovery becomes infeasible. AdaGenCom decomposes a source video into (i) structural semantics as sparse visual anchors (motion-aware keyframes and optional flow cues), (ii) textual semantics as compact scene descriptions, and (iii) conditional semantics as lightweight channel-state metadata. Each stream is transmitted with learned joint source-channel coding to enable graceful degradation. At the receiver, a pre-trained video diffusion model (LTX-Video) reconstructs the full sequence while a Channel-Aware Parameter Mapping (CAPM) mechanism adapts diffusion sampling—denoising steps, guidance scale, and sampling stochasticity—to the estimated channel quality without retraining. On 400+ frame sequences from UVG, AdaGenCom maintains CLIP similarity above 0.96 at SNRs where H.264/H.265 outputs become unrecognizable (CLIP below 0.80). These results indicate that generative priors with channel-aware inference provide a practical path for long-video delivery under extreme wireless conditions.

MOTIVATION: The goal of molecular representation learning is to automate the extraction of molecular features, a critical task in cheminformatics and drug discovery. While pretraining models using multiple views like SMILES, 2D graphs, and 3D conformations have advanced the field, integrating them effectively to produce superior representations remains a challenge. RESULTS: To bridge this gap, we propose a novel multi-view molecular pretraining method termed MMPCS, which explicitly factorizes representations into consistency and specific information. Our approach utilizes the Graph Isomorphism Network and the RoBERTa model to encode 2D molecular topological graphs and SMILES sequences, respectively. Each resulting molecular embedding is decomposed into a shared consistency component and a view-specific remainder. An autoencoder then aligns the consistency information across views. The combined consistency and view-specific representations serve as input for downstream tasks, enabling precise and task-aware predictions. When benchmarked against 16 state-of-the-art molecular pretraining methods, MMPCS achieved the highest average performance across both classification and regression tasks for molecular property prediction. It also delivered outstanding results in predicting drug-target binding affinity and cancer drug response, demonstrating its robustness and broad applicability. Additionally, a case study on the SARS-CoV-2 Omicron variant highlights the potential of MMPCS in facilitating drug repurposing efforts. AVAILABILITY AND IMPLEMENTATION: The source code and datasets supporting this study are publicly available at GitHub (https://github.com/xmubiocode/MMPCS) and Zenodo (https://doi.org/10.5281/zenodo.18182748).

In recent years, denoising diffusion models have exhibited exceptional performance in molecular generation tasks. However, existing approaches still face two major challenges. First, most diffusion-based methods rely on discretization operations and assume a Gaussian distribution, which often results in ambiguous or meaningless intermediate values when generating atom and bond types. Additionally, discretization introduces information loss, further compromising the quality of generated molecules. Second, current methods fail to incorporate constraints on bond angles and atomic valency, potentially leading to molecules that lack geometric consistency and chemical validity. To address these limitations, this paper introduces GeoSDE, a geometric diffusion model based on stochastic differential equations. Unlike conventional approaches, GeoSDE employs a continuous diffusion model that directly generates molecules in continuous space, effectively eliminating discretization-related issues and enhancing the stability and chemical validity of the generated structures. Moreover, GeoSDE explicitly incorporates bond angle and atomic valency constraints during training, providing richer molecular representations and ensuring that the generated molecules adhere more closely to real-world chemical principles and spatial structures. Extensive experiments on the GEOM-QM9 and GEOM-Drugs datasets demonstrate that GeoSDE surpasses existing state-of-the-art models across multiple metrics, particularly in terms of stability, validity, and uniqueness, highlighting its effectiveness in molecular generation tasks.

Single-cell classification based on single-omics data is often constrained by the one-sidedness of the data. With the advancement of single-cell sequencing technology, it has become possible to classify single cells using multi-omics data. However, integration and classification of multi-omics data are still challenging. In this study, we propose a model named MoSCHG. In this model, we first construct a heterogeneous bipartite graph based on the data of each omics, where the two types of nodes represent cells and their features (e.g., genes, chromatin) respectively, and the edge weights represent the relationship between cells and features; then GCN is applied with a residual mechanism to learn node embeddings in each bipartite graph and cell embeddings from different graphs are aligned based on supervised contrastive learning; finally, the aligned multi-omics cell embeddings are concatenated and the classification task is completed. Experimental results on three real datasets show that the proposed MoSCHG model outperforms the current state-of-the-art algorithms in classification performance, and through ablation studies, we validate the effectiveness of each module in MoSCHG.

Traditional drug-target and drug-disease associations prediction tasks have been performed independently, without fully exploiting the relationships between drugs and various other entities, leading to inaccurate predictions. With the emergence of large-scale heterogeneous biological networks, multi-task learning can effectively enhance the accuracy of association prediction based on the associations between entities. In this study, we propose a multi-task learning framework named DTD-MTL to predict drug-target and drug-disease associations simultaneously. Firstly, it utilizes a multi-layer relational graph convolutional network (RGCN) to learn the features of each node in the drug-target-disease network. Subsequently, it obtains the initial feature of an edge by concatenating the features of the two nodes on the same edge. To coordinate different prediction tasks, drug features are shared among different tasks. Afterwards, the autoencoder (AE) is used to extract features from different types of edges. In order to make the learned edge features more suitable for different prediction tasks, the distance covariance (DC) is utilized to eliminate the specificity between different types of edges, thereby leveraging the relationships between different tasks more effectively. Finally, the drug-target and drug-disease associations predictions are achieved based on the edge features extracted by the AE. Experimental results on a widely-used dataset show that DTD-MTL outperforms the state-of-the-art methods in the prediction task of drug-target and drug-disease associations.

Drug-target binding affinity (DTA) is a critical criterion for drug screening. Accurate affinity prediction will significantly cut the cost of new drug development and accelerate the drug discovery process. However, most existing approaches frequently utilize sequence or structure information without incorporating any additional information. At the same time, they encode drugs and targets separately, ignoring the important existing drug-target relationships. In this study, we propose a novel DTA prediction approach, named HSGCL-DTA, which is based on hybrid-scale graph contrastive learning. To completely capture the global information and discriminative properties of the heterogeneous graphs, HSGCL-DTA divides the drug-target affinity graph into two subgraphs with stronger and weaker affinities respectively, and the node embeddings of the two subgraphs are obtained based on node-graph level contrastive learning. Afterwards, graph convolutional network (GCN) is used to encode the molecular graph of drugs and targets, and the node embeddings in the strong affinity subgraph are fused with molecule graph embeddings to fully utilize the distinct information in two different views. Another node-node level contrastive learning is performed between the affinity graph and molecular graphs, thereby filtering out task-independent noise that only appears in one graph. The final drug-target embeddings are put into a multilayer perceptron (MLP) for affinity prediction. Experiments on two widely-used datasets have shown that HSGCL-DTA achieves better prediction performance and generalization than the state-of-the-art DTA prediction methods.

Unsupervised clustering of single-cell RNA sequencing (scRNA-seq) data is essential because it allows us to identify putative cell types. With the rapid growth of scRNA-seq data, it is difficult for traditional learning-based single-cell analysis methods to efficiently and continuously deal with them due to catastrophic forgetting. Inspired by how the human brain learns and remembers, we propose a novel unsupervised continual learning method for single-cell analysis, namely Continual Unsupervised Memory Replay (CUMR). We first employ a novelty detection algorithm to learn and assign pseudo-labels to unlabeled scRNA-seq data. Then, we apply memory-replay-based continual learning networks to achieve unsupervised continual learning of single-cell analysis. Experiments on real datasets show that CUMR has superior performance over other state-of-the-art continual learning methods in single-cell analysis tasks such as cell typing.

Drug-target interaction (DTI) prediction plays an important role in drug repositioning, drug discovery and drug design. However, due to the large size of the chemical and genomic spaces and the complex interactions between drugs and targets, experimental identification of DTIs is costly and time-consuming. In recent years, the emerging graph neural network (GNN) has been applied to DTI prediction because DTIs can be represented effectively using graphs. However, some of these methods are only based on homogeneous graphs, and some consist of two decoupled steps that cannot be trained jointly. To further explore GNN-based DTI prediction by integrating heterogeneous graph information, this study regards DTI prediction as a link prediction problem and proposes an end-to-end model based on HETerogeneous graph with Attention mechanism (DTI-HETA). In this model, a heterogeneous graph is first constructed based on the drug-drug and target-target similarity matrices and the DTI matrix. Then, the graph convolutional neural network is utilized to obtain the embedded representation of the drugs and targets. To highlight the contribution of different neighborhood nodes to the central node in aggregating the graph convolution information, a graph attention mechanism is introduced into the node embedding process. Afterward, an inner product decoder is applied to predict DTIs. To evaluate the performance of DTI-HETA, experiments are conducted on two datasets. The experimental results show that our model is superior to the state-of-the-art methods. Also, the identification of novel DTIs indicates that DTI-HETA can serve as a powerful tool for integrating heterogeneous graph information to predict DTIs.

Drug-target interaction prediction is an important research field in computer-aided drug discovery. The data involved in drug-target interaction prediction are characterized by noise, high dimensionality, and sparseness, which leads to poor prediction performance of traditional machine learning methods. Matrix factorization methods are often used to predict unknown or missing data, and can deal with data with the above characteristics. Therefore, a drug-target interaction prediction model based on non-negative and self-representative matrix factorization is proposed in this study. The proposed model performs matrix factorization based on the topological structure of the drug-target interaction data, and focuses on capturing the internal structural information of the drug-target data for representation learning. At the same time, it introduces nonnegative and non-trivial solution constraints to optimize the representation learning results, and integrates the graphs regularization method to optimize the low-dimensional key latent factor matrix, and finally realizes the prediction of drug-target interactions. Experimental results show that the model effectively mines the structural information of drug-target interactions, and is superior to other benchmark methods in the prediction performance.

Unsupervised ensemble learning aims to estimate ground-truth labels via integrating noisy and unreliable labeling results from multiple annotators. Although many techniques have been proposed to deal with this challenging task, there still exists some "tough" instances with noisy labels that are misclassified after the integration, which significantly affect the classification performance. This paper introduces a novel approach to improve the label accuracy based on unsupervised ensemble learning. First, we apply the expectation maximization (EM) algorithm to aggregate labels for all the instances. Then we identify instances that are most likely to be "tough" through a two-stage filtering method. Finally, an ensemble of AdaBoost-based classification models is trained on the high-quality dataset, and predicts new labels for these "tough" instances. The results of empirical investigation on binary classification task show that: (1) our approach can identify "tough" instances from the input dataset effectively; (2) our approach achieves a better performance on improving the accuracy of labels produced by unsupervised ensemble algorithms.

Drug-target interactions (DTIs) identification is an important issue of drug research, and many methods proposed to predict potential DTIs based on machine learning treat it as a binary classification problem. However, the number of known interacting drug-target pairs (positive samples) is far less than that of non-interacting pairs (negative samples). Most methods do not utilize these large numbers of negative samples sufficiently, which limits their prediction performance. To address this problem, we proposed a stacking framework named NegStacking. First, it uses sampling to obtain multiple completely different negative sample sets. Then, each weak learner is trained with a different negative sample set and the same positive sample set, and the logistic regression (LR) is used as a meta-learner to adaptively combine these weak learners. Moreover, in the training process, feature subspacing and hyperparameter perturbation are applied to increase ensemble diversity. Finally, the trained model could be used to predict new samples. We compared NegStacking with other methods, and the experimental results show that our model is superior. NegStacking can improve the performance of predictive DTIs, and it has broad application prospects for improving the drug discovery process. The source code and datasets are available at https://github.com/Open-ss/NegStacking.

Drug-target interaction (DTI) prediction plays an important role in drug repositioning, drug discovery, and drug design. In recent years, some DTI prediction methods based on machine learning have been proposed. They usually extract features from chemical genomics data. However, these methods are easy to extract redundant information that is not fully related with the prediction task and ignore the latent relationship between drug and target. This paper presents a new DTI prediction model named DTIGCCN. The model uses a spectral-based graph convolutional network (GCN) to extract features from drug and target expression profiles respectively, and a convolutional neural network (CNN) to extract latent associations between drug and target. Finally, the extracted features are concatenated together and fed into an effective classifier for prediction. The advantage of DTIGCCN is that the extracted features are more refined and targeted and the correlation between drug and target is fully applied to the prediction. Experimental results show that our model is superior to the conventional DTI prediction methods based on feature extraction and provides a new idea and method for DTI prediction.

MOTIVATION: With the rapid development of high-throughput technologies, parallel acquisition of large-scale drug-informatics data provides significant opportunities to improve pharmaceutical research and development. One important application is the purpose prediction of small-molecule compounds with the objective of specifying the therapeutic properties of extensive purpose-unknown compounds and repurposing the novel therapeutic properties of FDA-approved drugs. Such a problem is extremely challenging because compound attributes include heterogeneous data with various feature patterns, such as drug fingerprints, drug physicochemical properties and drug perturbation gene expressions. Moreover, there is a complex non-linear dependency among heterogeneous data. In this study, we propose a novel domain-adversarial multi-task framework for integrating shared knowledge from multiple domains. The framework first uses an adversarial strategy to learn target representations and then models non-linear dependency among several domains. RESULTS: Experiments on two real-world datasets illustrate that our approach achieves an obvious improvement over competitive baselines. The novel therapeutic properties of purpose-unknown compounds that we predicted have been widely reported or brought to clinics. Furthermore, our framework can integrate various attributes beyond the three domains examined herein and can be applied in industry for screening significant numbers of small-molecule drug candidates. AVAILABILITY AND IMPLEMENTATION: The source code and datasets are available at https://github.com/JohnnyY8/DAMT-Model. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Drug-target interaction identification is of highly importance in drug research and development. The traditional experimental paradigm is costly, while the previous in silico prediction paradigm remains a challenge because of diversified data production platforms and data scarcity. In this paper, we modeled drug-target interaction prediction as a binary classification task based on transcriptome data of drug stimulation and gene knockout from LINCS project and developed a framework with a deep-learning-based model to predict potential interactions. The evaluation results showed that not only did our framework fit data with better accuracy than other classical methods, but predicted more credible drug-target interactions. What's more, the prediction has high percentage of overlap interactions across other platforms.

Most of the existing researches simply convert associations of nodes within the snapshot of the evolutionary social network to the weight of edges. However, because of the obvious Matthew effect existing in the interactions of nodes in the real social network, the association strength matrices extracted directly by snapshots are extremely uneven. This paper introduces a new evolutionary social network model. Firstly, we generate probabilistic snapshots of the evolutionary social network data. Afterwards, we use the probabilistic factor model to detect the variation points brought by network events. According to experimental results, our proposed probabilistic snapshot model of evolutionary social network is effective for network events detection.

A hyperclique pattern [H. Xiong et al. (2003)] is a new type of association pattern that contains items which are highly affiliated with each other. More specifically, the presence of an item in one transaction strongly implies the presence of every other item that belongs to the same hyperclique pattern. We present a new algorithm for mining maximal hyperclique patterns, which are desirable for pattern-based clustering methods [H. Xiong et al. (2004)]. This algorithm exploits key advantages of both the depth first search (DFS) strategy and the breadth first search (BFS) strategy. Indeed, we adapt the equivalence pruning method, one of the most efficient pruning methods of the DFS strategy, into the process of the BFS strategy. As demonstrated by our experimental results, the performance of our algorithm can be orders of magnitude faster than standard maximal frequent pattern mining algorithms, particularly at low levels of support.