Accurate identification of cell types from single-cell RNA sequencing data remains challenging due to high dimensionality, sparsity, and the limited availability of expert annotations. We propose a dynamic supervised prelabel diffusion framework that leverages a small set of verified cell-type labels to guide clustering through iterative representation learning. The framework couples a purity-controlled diffusion mechanism with a supervised contrastive objective, forming a self-reinforcing loop in which improved cell representations enable more accurate and adaptive prelabel propagation, which in turn enriches the supervisory signal for subsequent training. An adaptive Leiden clustering strategy automatically matches the target number of cell types, eliminating the need for manual resolution tuning. Experiments on five benchmark datasets show that the proposed method consistently outperforms both unsupervised and semi-supervised baselines in clustering accuracy, normalized mutual information, and adjusted Rand index, while achieving substantially lower computational cost. These results demonstrate the effectiveness of dynamic prelabel diffusion as a principled semi-supervised strategy for single-cell clustering under limited annotation budgets.
The MAP kinase cascade of the epidermal growth factor receptor network is analyzed to further develop and demonstrate the application of Petri net models for the analysis of cellular systems. The mitogen-activated protein kinase/extracellular-regulated kinase (ERK) (subreaction H) model is then coupled to the model previously developed for the GTP:Ras complex (subreaction A) to demonstrate the computational scalability and analytical utility of the Shannon-like partition information entropy metric in the analysis of ever larger biochemical signaling networks. Unique minimal cycles (UMCs) have been shown to provide a natural partitioning of biochemical reaction networks. The complete listing of UMCs for two coupled subnetworks is found to contain all the UMCs of the individual subnetworks plus additional UMCs introduced by the cross-coupling of the two subnetworks by the shared species, in this case, ERK-PP and GTP:Ras, leading to a significant increase in the number and complexity of UMCs available for the control of the coupled subnetworks. The partition information entropy metric is used to rank order the UMCs to provide insight into the most critical pathways, thereby identifying which loops provide the most dynamic controllability over the cell's fate. Analysis of UMCs with some feed-backward transitions is found to be at least as important to the functioning of the network as UMCs with feed-forward only transitions.
Recently, link prediction (LP) based on graph neural networks (GNNs) methods has achieved notable successes in biological networks (BNs), since it can reveal the organizational principles, functional mechanisms, and dynamic properties of biological systems. However, these LPs still face some significant challenges that need to be addressed in BNs. The first is the complex and heterogeneous characteristics of BNs. Moreover, BN structures often have the dynamic addition and removal of nodes and edges over time or across physiological states. Afterward, there are asymmetries and hierarchical modularity in the structures of BNs. Finally, high computational complexity has resulted from the above challenges in BNs. Therefore, this article proposes a novel GNN-based LP model via local clustering and subgraphs, termed LCS in short, to address these issues in BNs. LCS introduces a subgraph-based GNN approach to effectively address the heterogeneous characteristics inherent in complex BNs, along with the consequent challenges of asymmetry and hierarchical modularity. Furthermore, LCS designs a dynamic local subgraph extraction (SE) mechanism based on heat kernel diffusion and the Chopper pruning algorithm. This mechanism leverages the effective local clustering properties of heat diffusion and uses Chopper to achieve linear-time SE, mitigating subgraph size explosion and enhancing LP efficiency. Additionally, by imposing diversity regularization constraints, the method reduces computational complexity and improves generalization performance. Experimental results on four BN benchmarks demonstrate that LCS achieves significant improvements over existing state-of-the-art LP methods. The implementation of LCS is publicly available at https://github.com/XL0104/LCS-Model.git.
Predicting drug-target interactions is critical for drug discovery, yet many deep learning methods overlook atom-residue-level relationships. We propose Protein Heterogeneous Graph learning for Drug-Target Interaction prediction (PHGDTI), a multimodal framework that integrates sequence and structural cues for binding prediction. Drug and protein sequences are embedded with Mol2Vec and Tasks Assessing Protein Embeddings (TAPE) and refined by a self-attention module. In parallel, a drug-protein graph encoder models three complementary graphs: a drug atom graph, a protein residue graph, and a heterogeneous atom-residue graph. Graph attention layers propagate intra- and intermolecular information, and SAGPooling yields compact structural representations. Fusing these structural and sequence features enables accurate affinity estimation. Experiments on the Davis kinase dataset and GalaxyDB dataset show PHGDTI surpasses competitive baselines, and ablation results highlight the benefit of heterogeneous graph modeling.
Single-cell RNA sequencing (scRNA-seq) provides a novel perspective to explore cellular biology at the single-cell resolution. Single-cell clustering is a crucial step to reveal cell types and the corresponding biological functions. However, when dealing with the high dimensionality and complexity of scRNA-seq data, existing deep models fail to comprehensively capture the intrinsic attribute information and structural relationships within the data. In this study, we propose a novel single-cell deep clustering model named scDFVA. The proposed scDFVA consists of a variational graph attention autoencoder (AE), a zero-inflated negative binomial (ZINB) based AE, and a self-supervised clustering. To better simulate sparse and zero-inflated scRNA-seq data, we incorporate the ZINB model into the AE. The variational graph attention AE is introduced to learn the cell structure information. scDFVA achieves representation learning within a joint framework comprising a ZINB-based AE and a variational graph attention AE, effectively fusing gene expression and cell structure information. Furthermore, scDFVA performs self-supervised clustering training on the latent fusion representations of cells to achieve mutual supervision between representation learning and clustering. Experiments indicated that scDFVA outperformed several other competing methods, demonstrating that our method is beneficial in single-cell clustering.
Spatial transcriptomics (ST) often requires aligning multiple tissue slices to reconstruct three-dimensional biological structures, a task hindered by complex deformations and structural heterogeneity. We propose GMSA, a synergistic alignment framework that integrates graph matching with point cloud registration. Unlike existing tools, GMSA first identifies high-confidence correspondences through subgraph matching based on gene expression and spatial topology, which provides a robust initialization for subsequent rigid Iterative Closest Point (ICP) or Nonrigid Iterative Closest Point (NICP) registration. Benchmark results on dorsolateral prefrontal cortex (DLPFC), spatially-resolved transcript amplicon readout mapping (STARmap), and multiplexed error-robust fluorescence in situ hybridization (MERFISH) datasets demonstrate that GMSA consistently outperforms state-of-the-art methods in alignment accuracy. Notably, GMSA’s nonrigid strategy successfully resolves complex structural distortions in MERFISH data where traditional methods fail, while maintaining stable gene expression distributions across aligned slices. This framework provides a flexible and precise solution for multimodal spatial transcriptomics integration.
Single-cell RNA sequencing (scRNA-seq) techniques for measuring gene expression in individual cells have developed rapidly. Recently, the identification of cell types in scRNA-seq analysis has been accomplished using deep learning. Most methods utilize a dataset containing cell-type labels to train the model and then apply this model to other datasets. However, the integration of multiple datasets leads to unexpected batch effects caused by differences in laboratories, experimenters, and sequencing techniques. As the batch effect interrupts the biological signal of interest, an effective batch correction method is essential. In this article, we present scUDAS, a cell-type prediction model for scRNA-seq that utilizes unsupervised domain adaptation and semi-supervised learning (SSL) to reduce the differences in distributions between datasets. First, we pretrain the proposed model based on the source dataset, which contained cell-type information. Subsequently, scUDAS is trained on the target dataset by leveraging adversarial training to align the distribution of the target dataset with that of the source dataset. Finally, scUDAS was retrained to improve its performance through SSL by leveraging both the source and target datasets with consistency regularization. scUDAS outperformed the other deep learning-based batch correction models by appropriately removing the batch effect. scUDAS is publicly available at https://github.com/cbi-bioinfo/scUDAS.
The prediction of protein-protein interaction (PPI) can be insightful for exploring the molecular mechanisms of cellular functions. Constructing the negative datasets of PPI is related to the assessment of the prediction accuracy and evaluation of the prediction performance. Aiming at the problem of unstable prediction accuracy in the current method of building negative sets using random sampling, we proposed a method of constructing negative sets based on a conditional generative adversarial network (CGAN), named PPIGAN. This method generates negative samples through a generative network, and the PPI prediction model uses these generated negative samples along with positive samples to learn interaction features. Simultaneously, the generator and the prediction model continuously compete against each other during the learning process, which enhances the model's generalization ability and prediction accuracy. Experimental results show that the accuracy of our proposed method reaches 94.68% and 98.22% in 5-fold cross-validation on yeast and human datasets, respectively. These results either surpass or closely approach the performance of advanced PPI prediction models such as PIPR, convolutional neural network, DeepTrio, and DeepFE, indicating that the method proposed in this article provides an effective solution for the work related to PPI prediction.
Recently, deep cell clustering, which employs deep neural networks to learn cell representation for clustering purposes, has attracted increasing research interests. Traditional deep cell clustering models for single-cell RNA sequencing data rely only on the cell's internal features for learning the representation and suffer from the insufficient problem of representation learning. In this article, we introduce a deep structural enhanced network for cell clustering, namely, Deep Structure-Enhanced Cell Clustering (scDSEC). The scDSEC model uses the internal features of the cells as a foundation and enhances them by incorporating the external structural semantics of the cells. An integrated reinforcement enhancement strategy is designed, in which a complete cell representation, captured by fusing cell internal information and external information, and an enhanced cell internal representation, captured with the help of complete cell representation, are learned in a layer-by-layer reinforcement manner. Experimental results show that the scDSEC model outperforms various existing mainstream deep cell clustering algorithms in terms of performance.
Drug-drug interaction (DDI) prediction remains a significant challenge due to the complexity of biological systems and the growing demand for precise predictions. Recent advances in deep learning have been successfully applied to DDI prediction. However, the asymmetrical nature of DDIs is always neglected, which can lead to some information loss during the feature learning process. To address the issue, a novel DDI prediction method based on knowledge graph and generative adversarial network, KGGAN-DDI, is proposed to predict potential DDIs. First, a knowledge graph embedding module is designed to capture and encode asymmetric associations between drug pairs, which can enhance the feature representation and contextual relevance of drug interactions. Then, a dual-generator GAN is adopted to produce realistic samples and improve the prediction accuracy further. Moreover, a least squares loss function is utilized to mitigate the vanishing gradient problem, thereby providing smoother gradients and enhancing the efficiency and stability of the optimization process. Extensive experiments demonstrate that the proposed KGGAN-DDI performs better than state-of-the-art methods. Case study further shows the effectiveness of KGGAN-DDI.
Intratumor heterogeneity (ITH) impacts cancer progression, and its characterization is crucial. Clustering algorithms applied to the variant allele frequency (VAF) of mutations can facilitate the exploratory analysis of ITH. This study comparatively evaluated six clustering algorithms to characterize ITH by clustering mutations based on their VAFs. We utilized data from The Cancer Genome Atlas to analyze three cancer types by examining the distribution of clusters in the results from various methods and four internal validation metrics. The results indicated that the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Variational Bayesian Gaussian Mixture Model methods identified an insufficient number of clusters in most tumor samples. The Hierarchical DBSCAN (HDBSCAN) and Ordering Points to Identify the Clustering Structure (OPTICS) algorithms exhibited greater variability in the number of clusters, while Affinity Propagation (AP) showed controlled behavior, and Mean-Shift demonstrated greater consistency. The Mean-Shift and AP methods were consistently superior in the validation metrics, in contrast to HDBSCAN and OPTICS, which had inferior performance. We conclude that Mean-Shift and AP are promising and accessible alternatives for the initial exploratory analysis of ITH by VAFs. A computational pipeline is provided on the Google Colab platform to facilitate future studies.
The spread of a virus can be modeled as the diffusion of virion populations along the edges of a host contact network. Most prior research assumes that the virus maintains the same genome throughout the diffusion process, and, consequently, the genome itself is of no modeling interest. Eletreby et al. do consider a model of multiple variants in which a virus can mutate within a host into another variant before transmission across an edge. Still, they fail to model the genome or any established fitness models. We incorporate three biological notions, fitness landscapes, viral quasispecies, and genome structure, to more accurately model the diffusion of a virus with mutations. We investigate established fitness landscape models and various simulated contact networks. We simulate the diffusion process across the contact network, incorporating co-occurring evolutionary and epidemiological processes. In our proof-of-concept simulations, while determining which variant should infect an exposed individual, we consider the fitness values of the variants. ViraFit tunes the ruggedness of the fitness landscape by varying several attributes: genome length, network size, mutation rate, infection probability, and infection time. Our simulation results demonstrate the importance of including the fitness value of each variant and how the fitter ones persist over time.
A novel algorithm is proposed to increase the effectiveness of model selection and parameter estimation for the fractional susceptible-infected-recovered model. It combines reinforcement learning (RL) and approximate Bayesian computation sequential Monte Carlo (ABC-SMC) instead of ABC to improve the process of model selection and parameter estimation, where RL is used for model selection and ABC-SMC is exploited for parameter estimation. Numerical simulations illustrate that the combined algorithm (RL-ABC-SMC) significantly outperforms the ABC-SMC algorithm in terms of model selection. Finally, we consider the application of the proposed methodology.
Drug-target interaction (DTI) prediction is of great practical value for discovering, developing, and repurposing drugs, which has tremendous advantages to pharmaceutical industries and patients. However, the prediction of DTIs using wet-lab experimental methods is generally expensive and time-consuming. To date, numerous machine learning-based approaches show promising performance, which greatly improves DTI discovery efficiency, but there are still two challenges remained. One is how to represent the spatial structure features of drugs and targets appropriately, and the other is how to explicitly and effectively model and learn local interactions between drugs and targets, for better interpretation and prediction. In this work, we propose a novel framework that combines a transformer and graph attention convolutional network for DTI prediction (TransGAT-DTI), which not only precisely predicts the putative DTIs with satisfactory overall performance on three benchmark datasets, but also captures the pivotal sequence that contributes the most to the positive predictions. Experimental results demonstrate that the proposed TransGAT-DTI achieves the best performance on BindingDB, BioSNAP, and Human datasets. Importantly, the proposed approach provides a novel solution for discovering and developing novel drugs.
Principal component analysis (PCA) is one of the most frequently used dimensionality reduction methods for high-dimensional datasets, especially single-cell RNA sequencing (scRNA-seq). Despite its popularity, PCA faces challenges, particularly related to its performance degrading as the dataset size increases. Additionally, PCA is sensitive to outliers and assumes linearity. Random projection (RP) methods have emerged as a promising alternative to address several of PCA's limitations. In this study, we conduct a systematic and comprehensive evaluation of PCA and RP methods, including singular value decomposition (SVD) and randomized SVD approaches, against multiple RP methods including sparse random projection, Gaussian random projection, and we introduce a Matching Sparsity Random Projection algorithm that adaptively calibrates projection matrix density according to input data sparsity patterns, emphasizing both computational scalability and effectiveness in downstream analytical tasks. We evaluated these methods on multiple publicly available scRNA-seq datasets that include both labeled and unlabeled scenarios. Clustering performance is assessed using Hierarchical Clustering and Spherical K-Means algorithms, with labeled datasets evaluated through Hungarian algorithm accuracy and Mutual Information metrics. For unlabeled datasets, we used the Dunn Index and Gap Statistic to quantify cluster separation quality. Across both dataset types, the Within-Cluster Sum of Squares metric is used to assess variability. Moreover, locality preservation is examined, with RP methods, including our adaptive sparsity approach, outperforming PCA in several of the evaluated metrics. Our experimental results show that RP methods not only deliver substantial computational speed improvements over PCA but also rival, and in some cases, exceed PCA in preserving data variability and clustering quality. Through this comprehensive methodological comparison, our work provides critical guidance for selecting appropriate dimensionality reduction strategies that effectively balance computational demands, scalability requirements, and analytical quality in downstream analyses.
This study presents an improved mathematical model that incorporates negative interaction mechanisms to predict the dynamics of cell signaling pathways. By employing stochastic differential equations and the Euler-Maruyama method, we simulate the responses of proteins within the mitogen-activated protein kinase and oxytocin signaling pathways over time. Conventional signaling models that consider only positive interactions often lead to unrealistic signal over-amplification, the absence of oscillatory dynamics, and an inability to reproduce compensatory responses following targeted inhibition. To address these limitations, our model explicitly incorporates inhibitory interactions through a sign-changing characteristic function and bounds protein activity using a hyperbolic-tangent transfer function, ensuring biologically plausible saturation behavior. Our findings indicate that the inhibition of upstream proteins such as MEK1/2 leads to a rapid decrease in ERK1/2 activation while causing a compensatory increase in other proteins such as SOS, RAS, and RAF. Furthermore, we explore the synergistic effects of combination therapies, demonstrating that targeting multiple signaling pathways can enhance therapeutic efficacy. Through the application of the Bliss Independence Index, we assess the effectiveness of these therapeutic combinations. Additionally, we investigate the effects of abnormal activation increases caused by gain-of-function mutations on downstream proteins and the resulting changes in balance induced by negative interactions. Overall, our enhanced mathematical model serves as a valuable tool for simulating signaling dynamics with inhibitory crosstalk and for generating mechanistic hypotheses relevant to targeted and combination therapies.
Insects thrive in highly diverse environmental niches, exposing them to varying pathogens. This complex interaction has led to the development of a variety of defense mechanisms collectively termed as innate immunity. Innate immunity is the first line of defense against pathogens in most organisms. Despite the availability of genomic and protein data of insects, we do not understand how innate immunity has evolved. This study reports class-level analysis of innate immunity in Insecta spanning approximately 300 million years ago of evolution. We used the available data on 27 insect species of five predominant orders to track the evolutionary paths of innate immune proteins. We analyzed orthogroups and gene family dynamics to identify core conserved components and lineage-specific innovations in immune genes. Through orthogroup analysis, we find an asymmetrical or incoherent distribution of orthologs within the immune orthogroups among the orders that present intriguing events such as gene duplications, losses and functional diversification. For example, instances of missing orthologs in corresponding immune orthogroups were noted for specific orders, such as Defensin in Lepidoptera, transferrin in Hemiptera, and Cathepsin B in Hymenoptera. Lineage-specific conservation of antimicrobial peptides was observed in Lepidoptera and Hymenoptera. Interestingly, proteins like Peptidoglycan recognition protein (PGRP)-SA, Spatzle, and others have undergone major family expansion and contraction events in the insect species studied. By constructing phylogenies based on immune-related proteins and assessing the signatures of positive selection, we gained insights into diversification and adaptive evolution. Signatures of positive selection were only demonstrated by PGRP-SA and diptericin proteins, with a specific site under selection found in the latter. Our findings present a broad picture of how the five closely placed yet highly divergent orders of the class Insecta have maintained innate immunity in close reference to their ecological niches. These findings can have practical implications in strategizing pest management and insect conservation in the wild.
One way to understand biology is by finding genetic sequences that are related to each other. Often, a family of related sequences has position-varying probabilities of substitutions, insertions, and deletions: we can use these to find distantly related sequences. There are popular software tools to do this, which all have limitations. They either do not use all probability evidence (e.g., PSI-BLAST, MMseqs2) or have excessive complexity and minor biases (e.g., HMMER). This complexity inhibits fertile development of alternative tools.This study describes a simplest reasonable way to find related sequences, making full use of position-varying probabilities. The algorithms likely use the fewest operations that such algorithms possibly could, so they are fast and simple. This has been implemented in prototype software named DUMMER (Dumb Uncomplicated Match ModelER). Its sensitivity and specificity are competitive with HMMER. It finds evidence that the human genome has many more relics of some ancient transposons, including LF-SINE, which was co-opted for various functions in common ancestors of all land vertebrates.
Protein language models (PLMs) provide powerful sequence representations, yet their effectiveness for unsupervised viral clade assignment remains uncertain. In this study, we evaluated embeddings from ProtT5, ProtBert, CARP, and several ESM-2 variants on influenza A/H3N2 hemagglutinin sequences. Using dimensionality reduction (t-SNE, UMAP, PCA, MDS) and clustering with HDBSCAN, we compared PLM embeddings against baseline Hamming distance approaches. Our results show that t-SNE combined with PLM embeddings can recover clade structure, with ProtBert yielding the most stable performance and larger ESM-2 models occasionally achieving lower normalized variation of information scores but with greater variability. These findings suggest that while PLM embeddings capture clade-relevant signals, they also suffer from instability and the loss of site- or nucleotide-specific detail. Future improvements in pooling strategies may enhance their utility for viral surveillance.
The simple underlying pattern of presence-absence of a character within a species tree provides useful steps to trace complex evolutionary histories. Character-based models such as perfect transfer networks and its galled variant aim to leverage this information to predict horizontal gene transfers. Under the assumption that characters have a single origin, are rarely lost, and can be transferred horizontally, they remain an efficient inference method for almost tree-like scenarios. Nevertheless, they can sometimes predict overly complicated scenarios, and its simplest structural variants are too restrictive for practical uses. With the goal of extending this model to include loss events, we present a Sankoff-Rousseau-like algorithm that aims to recover the simplest possible scenarios that combine gene transfers and losses using solely the single character information already contained in a given species tree. We establish a link between the small parsimony problem and the inference of scenarios with a minimum number of losses and transfers, allowing losses and transfers to have a user-defined penalization for this end. We also explore the utility of our model for tracing possible highways of gene transfers by presenting a real case study on a dataset of bacterial species and Kyoto Encyclopedia of Genes and Genome functions as characters.
