Advances in spatial transcriptomics have promoted spatial multi-omics technologies, and the integration of multi-omics data is a key strategy to analyze complex biological systems. Current methods incorporate spatial context yet often fail to fully leverage local spatial structures during feature fusion, limiting their ability to capture complex spatial patterns. Moreover, in discrete tissue regions, unsupervised integration faces interference from false negative samples, compromising integration robustness and the granularity of spatial domain identification. Here, we introduce SMGC, a spatial domain analysis method that integrates spatial multi-omics data using multi-view autoencoders and granular-ball contrastive learning. SMGC proposes a strategy to generate granular balls from high-level features, which defines each ball as a cluster to naturally partition the sample set into multiple local clusters, thereby enabling subsequent feature fusion to fully exploit the underlying local topological structures. To mitigate the interference from false negative samples, our method achieves feature alignment and contrastive learning at the local structural level by constructing intra-view and cross-view associative relationships based on the granular balls of each view. Extensive experiments demonstrate that SMGC achieves competitive and consistent performance for spatial domain identification on simulated and real-world datasets, outperforming representative baseline methods in terms of the Adjusted Rand Index and Normalized Mutual Information. The source code is available at https://github.com/smile1112223333/SMGC.

Molecular property prediction is essential in drug discovery for early-stage compound evaluation. Recently, contrastive learning has demonstrated significant potential under limited labeled data by constructing augmented views. However, current augmentation strategies often disrupt molecular semantics and ignore chemical priors, limiting representation quality. Moreover, molecular data is inherently multimodal, including graphs, fingerprints, and sequences, yet how to effectively integrate their complementary information remains challenging. Therefore, we propose MPMFMol, a unified framework that integrates multitask self-supervised pretraining with multimodal fine-tuning for molecular property prediction. During pretraining, we construct heterogeneous augmented views based on molecular fragments to preserve original molecular semantics, enabling the graph encoder to capture fragment-level information. Meanwhile, fingerprint features are integrated into a multitask learning objective, reducing reliance on negative sampling and enhancing the encoder's representation capability. During fine-tuning, we further incorporate functional group and SMILES sequence information and design a stage-aware modality fusion strategy. Specifically, pretrained graph features are injected into the initial representation of functional groups to guide feature extraction and then fused with SMILES features to enable deep cross-modal interaction and enhance downstream predictive performance. Experimental results on six classification and three regression data sets demonstrate that MPMFMol outperforms state-of-the-art baselines.

2'-O-methylation (2OM) of ribose is a widespread RNA modification that significantly impacts RNA stability, structure, and function. Accurately predicting 2OM sites is crucial for understanding RNA's biological functions and related pathologies. Traditional detection methods pose challenges such as resource intensiveness, potential RNA sample damage, and high costs. However, recent advancements in machine learning, particularly deep learning techniques, offer rapid and cost-effective prediction solutions. In this study, we introduce DeepR2OM, a novel method integrating feature selection and deep learning for 2OM sites prediction. DeepR2OM encodes sequences using eight RNA descriptors, employs feature selection algorithms to reduce dimensions, and then utilizes a deep learning network for training. After evaluating various deep learning architectures, we selected Convolutional Neural Network (CNN), Multi-Head Self-Attention mechanism, and Deep Neural Network (DNN) as our final prediction models. Experimental results demonstrate DeepR2OM's effectiveness, achieving 87.1% accuracy (ACC), 85.5% recall rate (Recall), 87.9% precision (PRE), and a Matthews correlation coefficient (MCC) of 75.7% on an independent test set. This tool serves as a valuable resource for exploring the functional and bioinformatic aspects of 2OM sites.

Neuropeptides are endogenous signaling molecules that regulate diverse physiological and cognitive processes. However, reliable identification from primary sequence remains challenging due to high sequence diversity, weak motif conservation, and the limited experimental annotations. Solving this challenge is crucial for elucidating the molecular structure of neural communication and accelerating the development of neuropeptide-based therapies and peptide drugs. Existing computational approaches for neuropeptide identification range from traditional machine-learning models relying on handcrafted features to deep-learning architectures that learn sequence representations. However, both types of methods struggle with the high heterogeneity and weak motif conservation of neuropeptides, resulting in limited generalization and highlighting the need for more robust predictive frameworks. To address these limitations, we propose a unified multitask neuropeptide identification framework that integrates ESM-derived protein representations, a BiLSTM encoder, and multihead self-attention to capture local and long-range sequence dependencies jointly. Within this framework, the model further leverages attention-based pooling, auxiliary knowledge distillation, and contrastive representation learning to enhance generalization and ultimately improve the accuracy and robustness of neuropeptide identification. On the independent test set, our proposed multitask learning method (NeuroPred-MTCL) demonstrates strong generalization performance, achieving an accuracy of 93.6% and an AUROC of 0.977. It further maintains a balanced trade-off between precision (92.9%) and recall (94.4%), yielding an F1-score of 0.936 and an MCC of 0.872. These results highlight the method's ability to effectively capture discriminative sequence characteristics and substantially enhance the reliability of neuropeptide identification. These results establish NeuroPred-MTCL as a robust and generalizable approach that meaningfully advances the computational identification of neuropeptides.

Accurate prediction of drug-target affinity (DTA) is essential for accelerating drug discovery. Although pretrained protein language models have achieved significant progress, existing methods predominantly focus on bottom-up sequence patterns and lack explicit constraints from high-level biological functions. We propose GoMA-DTA, a framework integrating gene ontology (GO) functional annotations with protein semantic features. GoMA-DTA introduces a channelwise gating mechanism that uses functional semantics as anchors to dynamically recalibrate ESM-2embeddings, achieving adaptive semantic filtering. For drugs, the model integrates Molformer-based semantic and TransConv-derived structural features. These dual-modality drug representations interact with calibrated protein features through a parallel synergistic architecture of cross-attention and Mamba modules, ensuring precise cross-modal alignment and efficient long-range dependency modeling. Evaluations on PDBBind, BindingDB, and ChEMBL benchmarks demonstrate that GoMA-DTA significantly outperforms state-of-the-art models across various evaluation scenarios. Its superior screening power is further validated on CASF-2016. Moreover, virtual screening of 200 000compounds against the SARS-CoV-2Spike protein, supported by experimental evidence (ZINC2111387), underscores its practical utility as a robust and biologically reliable tool. The datasets and codes are publicly available at https://github.com/xa-123955/GoMA-DTA.

With the rapid advance of spatial multi-omics technologies, it has become possible to simultaneously profile transcripts, proteins and chromatin states at their native spatial coordinates, thereby uncovering molecular architecture that transcends any single-omics perspective. However, the resulting data matrices are often highly sparse and suffer from unstable dimensionality. Graph-based neural methods capture only local neighborhood information, whereas conventional Transformers, although capable of modelling long-range dependencies, incur prohibitive computational costs on such data. To overcome these limitations, we propose TLAGC—a Taylor-Linear-Attention-Guided Graph Convolutional framework that couples a Taylor-expanded linear attention (TLA) mechanism with graph convolutional networks. By eliminating the soft-max operation and linking the LocalGCN via residual connections, TLA preserves local structural information while enabling the integration of global and local contexts, thereby alleviating ineffective information propagation between spatially distant yet transcriptionally similar regions. Theoretical analysis confirms that TLA indeed reduces computational complexity, and extensive experiments on multiple spatial multi-omics benchmarks demonstrate that TLAGC consistently outperforms state-of-the-art baselines in delineating spatial domains.

Protein-RNA interactions (RPIs) stand for the central process in post-transcriptional regulation and have catalyzed a fast proliferation of computational approaches in recent years. Adopting a task-oriented classification method, RPIs calculation prediction schemes proposed over the period 2010-2025 fall into five primary categories: RNA-binding protein (RBP) classification, RPIs prediction, binding site and binding profile modeling on RNA, residue-level RNA-binding interface prediction on proteins, and quantitative estimation of binding affinity and mutation effects. This study reviews the methodological evolution from conventional machine learning to deep learning, graph neural networks and large-scale pre-trained language models, and compares their differences in data preparation, evaluation protocols and generalization behavior. Particular emphasis is placed on recent advances in structure-aware and condition-aware models, as well as learning in low-data regimes. Finally, the study outlines practical recommendations for field-wide benchmarking and looks ahead to the integration with spatial omics and the development of dynamic, generative landscapes of RPIs to better empower biomedical research.

Genetic perturbation response prediction plays a critical role in virtual cell research, yet the performance of current deep learning models still leaves room for improvement. In this study, we present a prior-guided response inference model (PRIM) that leverages a valuable priori knowledge of gene expression in control cells to model the effects of perturbations at the cellular level. This allows PRIM to predict the amount of change in gene expression after perturbation to better approximate the real situation. Compared with existing deep learning approaches, PRIM achieves superior performance across multiple data sets and notable advantages in predicting combinatorial perturbation responses. Moreover, it is more lightweight than current deep learning models and enables faster forward inference. Importantly, PRIM effectively captures nonadditive genetic interactions and shows the potential to uncover associations between combinatorial perturbations and new biologically meaningful phenotypes. These findings provide new insights into the application of deep learning for predicting cellular responses to genetic perturbations.

Drug-target interaction (DTI) prediction is vital for computer-aided drug discovery, yet current methods struggle with cross-domain generalization and multimodal fusion. To address these challenges, we propose MMTF-DTI, a multimodal framework that establishes deep interaction between drug and target representations via multimodal feature extraction and a Transformer-based fusion mechanism. Specifically, we design a dual-path encoder: Transformers decode semantic information from drug SMILES and protein sequences, while graph neural networks capture structural features of molecules and proteins. We then introduce a dynamic multimodal fusion module with learnable attention weights for adaptive feature interaction, enhancing compound-target representation learning. To further improve generalization, a domain adversarial neural network is incorporated to mitigate cross-domain distribution shifts. Comprehensive evaluations on benchmark datasets show that MMTF-DTI achieves state-of-the-art performance, with average gains of 1.7% AUROC and 2.1% AUPR over baselines. In cross-domain adaptation, the model also exhibits superior transferability, highlighting its potential for large-scale drug discovery where reliable cross-domain prediction is crucial.

Protein phosphorylation is a key post-translational modification involved in signal transduction, disease progression and therapeutic regulation. Reliable phosphorylation-site prediction remains challenging because existing methods often have limited feature integration, weak generalization and reduced performance on low-sample residue types. Here, we propose AdaptIPs, a dual-channel deep-learning framework for phosphorylation-site prediction. AdaptIPs integrates residue-level sequence representations with pretrained ESMC embeddings through element-wise multiplicative fusion, and uses Bahdanau attention, mirror loss and adaptive early stopping to improve feature learning and training stability. To address the limited number of tyrosine phosphorylation sites, we further use transfer learning from serine/threonine site prediction to tyrosine-site prediction. On independent test sets, AdaptIPs achieved an accuracy of 84.48% and an AUC of 91.83% for serine/threonine sites, and an accuracy of 92.86% and an AUC of 94.33% for tyrosine sites. Comparative experiments, ablation analysis and fusion-strategy evaluation confirmed the contribution of the dual-channel architecture and multiplicative fusion. AdaptIPs provides a reproducible framework for phosphorylation-site prediction and may support broader few-shot biological sequence prediction tasks. The source code is available at https://github.com/scarlet211/AdaptIPs.

Cell type annotation in single-cell RNA sequencing is a critical bottleneck, with existing automated methods facing limitations in accuracy, interpretability, and generalization to novel cell types. Although Large Language Models (LLMs) have recently shown potential in single-cell annotation, they are prone to inherent "hallucinations". Furthermore, a critical challenge is utilizing imperfect and potentially noisy external knowledge bases in a principled and robust manner to effectively constrain and enhance the LLM's reasoning capabilities. To address this, we propose scHilda, a novel framework designed to tackle this challenge. It deeply integrates an external Knowledge Graph into the LLM's reasoning process and employs a hierarchical arbitration annotation strategy. This strategy first identifies major cell lineages with the support of a global knowledge base and then dynamically retrieves focused subgraph domain information related to that lineage to precisely resolve cell subtypes. This dynamic knowledge-enhanced reasoning mechanism effectively constrains the LLM's decision space, reduces the risk of hallucination, and mitigates potential misguidance from knowledge base deficiencies. Tests on multiple benchmark datasets show that scHilda outperforms existing methods, achieving state-of-the-art (SOTA) performance. Notably, scHilda demonstrates exceptional robustness when handling complex mixed samples and enables lower-cost lightweight LLMs to achieve annotation performance close to that of top-tier models. Furthermore, rigorous statistical evaluations, alongside detailed interpretability case studies and query complexity analyses, validate the framework's efficiency and transparent decision-making. By deeply integrating the reasoning power of LLMs with structured biological knowledge, scHilda not only improves the accuracy and interpretability of cell annotation but also provides a new paradigm for building the next generation of trustworthy biological AI systems.

Leveraging the wealth of biomedical data available, we can derive insights into the relationships between biological entities from various angles. This underscores the complexity and significance of developing a dynamic approach for integrating data from multiple sources, a critical endeavor in drug recommendation. In this study, we introduce an innovative deep learning approach termed “Multi-View Fusion Strategy with Self-Attention” (FusionMVSA), designed to predict associations between drugs and diseases. To effectively amalgamate data from diverse sources and extract representative features, we have developed a feature extraction mechanism that capitalizes on similarities. This mechanism computes self-attention across multiple perspectives using shared group parameters, thereby highlighting common characteristics. Simultaneously, we utilize biomedical similarities among multi-source data as guiding factors for calculating similarity, enabling the capture of more nuanced features. Subsequently, we integrate these features through a feature fusion process, where known associations between drugs and diseases act as guiding terms. This strategy allows us to uncover the complementary aspects of different viewpoints. Ultimately, we predict potential drug-disease associations using a multi-layer perceptron neural network. Our methodology has undergone rigorous testing through various cross-validation experiments and case studies. We are confident that FusionMVSA will prove to be a valuable tool in drug recommendation, offering new avenues for exploration and discovery in the quest to combat diseases.

The rapid development of spatial multi-omics technologies has enabled the simultaneous acquisition of transcriptomic, proteomic, and epigenomic information from the same tissue section. However, substantial differences in distributional properties, data dimensionality, and noise levels across modalities, together with the inherent sparsity and incompleteness of spatial information, pose major challenges for data integration and modeling. In recent years, deep learning-based spatial multi-omics integration algorithms have emerged rapidly, offering new approaches for constructing unified latent representations and achieving cross-modal fusion. In this review, we systematically summarize existing spatial multi-omics integration methods for the first time, categorizing and comparing them from two perspectives. We not only systematically surveyed the datasets employed by these methods, but also highlighted the key downstream analytical tasks they support, and further summarized the major challenges currently faced in spatial multi-omics integration research. Furthermore, we compare the strengths and limitations of different approaches to assist researchers in selecting appropriate methods more efficiently, thereby advancing the application of spatial multi-omics in uncovering multilayer regulatory mechanisms of tissue microenvironments and disease processes.

Cyclic peptides represent a rapidly growing class of therapeutics, yet their development is often hindered by the challenge of predicting cell membrane permeability, a critical determinant of drug efficacy. Existing computational methods often struggle to integrate the diverse structural information inherent in these complex molecules, resulting in suboptimal predictive accuracy. Here, we introduce MCPerm, a multi-modal deep learning framework that synergistically integrates 1D SMILES, 2D topological, and 3D geometric information through a novel modality share and contrastive learning strategy to accurately predict cyclic peptide permeability. MCPerm fine-tunes a pretrained peptide language model for SMILES encoding and uses a parameter-sharing graph transformer for structural representation, while a dual contrastive learning mechanism enforces representational consistency both within and between modalities. On the benchmark PAMPA dataset, MCPerm achieves state-of-the-art performance, significantly outperforming leading methods. We further demonstrate its robustness and competitive transferability across three independent assays (Caco-2, MDCK, and RRCK). Our work presents a robust in silico framework that holds potential to accelerate the rational design and discovery of cell-permeable cyclic peptide drugs. Furthermore, to move beyond predictive accuracy, we introduced an attention-based visualization analysis. The results demonstrate that our model is not a “black box”; it has learned key chemical principles governing cyclic peptide permeability.

Signal peptides are short amino acid sequences located at the N-terminus of proteins. They guide newly synthesized proteins to their correct cellular destinations, playing a crucial role in protein localization and transport. Traditional experimental methods for identifying signal peptides are typically time-consuming, costly, and labor-intensive, driving rapid development of computational alternatives. Over the past two decades, researchers have proposed various computational approaches, with prediction accuracy continuously improving through evolution from early statistical and rule-based algorithms to deep learning. In this review, we systematically summarize these computational approaches, emphasizing methodological evolution and framework design. We compile representative computational methods, comparing their prediction outcomes and identifying existing limitations. Finally, we discuss current challenges and emerging opportunities, aiming to advance the development of computational frameworks characterized by unified evaluation, biologically grounded interpretation, and generative modeling.

Accurate prediction of B-cell epitopes plays a key role in facilitating advancements in vaccines, therapeutics, and diagnostics. In contrast to labor-intensive experimental approaches, computational strategies provide a more economical and efficient means of identifying potential epitopes. Existing methods are often limited by their reliance on experimentally resolved protein structures or by the use of lower-accuracy predicted structures. Sequence-based approaches, while fast, largely fail to capture the 3D spatial context essential for conformational epitopes. With the breakthroughs achieved by AlphaFold3 in predicting protein structures, we present MsgaBpred, the model to apply AlphaFold3-derived structures to B-cell epitope identification. Given only a protein sequence, our model employs a multi-scale graph convolutional network and additive attention to capture complex structural dependencies without relying on experimentally determined structures. The multi-scale design allows for effective modeling of both local and global contexts by aggregating information across different neighborhood ranges. Additionally, we leverage ESM-C, a more expressive protein language model than ESM-2, to enhance feature representation for B-cell epitope prediction. Extensive evaluations across multiple benchmark datasets demonstrate that MsgaBpred achieves competitive and robust performance; notably, it yields a statistically significant improvement in AUC compared to existing state-of-the-art methods. Moreover, the modular and scalable architecture of MsgaBpred holds promise for broader applications, including the structural analysis of other biomolecular entities such as nucleic acids and carbohydrates.

Drug discovery is a complex and systematic process aimed at finding new treatment methods that can prevent or treat specific diseases. Accurately predicting the interaction and binding affinity between drugs and targets is one of the key steps in modern drug development. Although traditional experimental methods are accurate, they are difficult to meet the efficiency requirements of current drug development due to high costs, low throughput, and high failure rates. In contrast, computational prediction methods are gradually becoming an indispensable auxiliary tool that can not only significantly shorten the research and development cycle and reduce experimental costs, but also improve the success rate of candidate drug screening. This review focuses on the research of drug-target interaction (DTI) and drug-target binding affinity (DTA), and systematically reviews the relevant research progress. Distinct from existing reviews, we treat large pre-trained model-based approaches as an independent paradigm, rather than subsuming them under conventional sequence- or structure-based models. The article first outlines commonly used resources and methods from the perspective of data and representation, and the computational definition of drug target prediction problem was clarified. On this basis, we have summarized the development path of computational models, from early similarity and feature driven models, to matrix decomposition, network analysis, sequence and structure modeling, and then to the emergence of large-scale pre-trained models in recent years, forming a relatively complete technological evolution path. The article also summarizes the experience at the experimental level, such as the selection of evaluation indicators, handling of cold start scenarios, design of case studies, and analysis of model interpretability. Finally, we synthesize key challenges and identify several directions for future research.

Functional peptide discovery, particularly for blood-brain barrier-penetrating peptides (BBBPPs), is strictly limited by extreme data scarcity and the "black-box" nature of deep learning. Here, INB3P is presented as a physics-informed, multi-modal framework designed to address these challenges. Physicochemical-guided mutagenesis (PCGM), a novel augmentation strategy that enforces biochemical constraints to expand training diversity without violating the biological manifold. INB3P integrates PCGM with a bi-directional co-attention mechanism fusing sequence and structure, optimized via contrastive learning and a Stable-MCC loss. INB3P significantly outperforms state-of-the-art baselines on the same independent test set used in a prior study. Crucially, the model autonomously rediscovers known biophysical mechanisms-including amphipathic motifs and long-range contact stabilization-providing strong in silico validation of its learned representations. This work establishes a generalizable paradigm for learning from small, imbalanced biological datasets. To facilitate community adoption, a web server is provided at http://www.bioai-lab.com/INBP, featuring a standalone PCGM module, empowering researchers to apply physics-guided augmentation strategy to their own sparse datasets.

Single-cell multi-omics data reveal complex cellular states and deepen our understanding of tissue cell phenotypes and functions. However, data analysis remains challenging due to the discrete nature and high noise level of the data, as well as the lack of modality. Here, we propose scMultiNet, a multi-task deep adversarial neural network that can integrate different tasks to analyze single-cell multi-modal data. In particular, we achieve joint training of multi-modal integration and cross-modal prediction tasks by introducing a cross-modal bi-prediction module and a multi-head self-attention module. Data denoising is further enhanced by integrating an indicator matrix that constrains and precisely reconstructs the original expression values. Extensive simulations and real data experiments demonstrate that scMultiNet outperforms existing state-of-the-art methods in dimensionality reduction, visualization, clustering, batch elimination, data denoising, multi-modal integration, single-cell cross-modality translation, and in revealing cell type-specific biological insights. In addition, we demonstrate that scMultiNet can effectively transfer the complex relationships between modalities from one batch to another. In summary, scMultiNet stands as a comprehensive end-to-end framework, ideally suited for analyzing single-cell multi-omics data.

Drug-drug interaction (DDI) is a major safety concern in clinical practice, as unexpected interactions between drug mechanisms can cause severe adverse effects or therapeutic failure. While unimodal deep learning methods have shown potential in DDI prediction, their reliance on a single molecular representation limits the ability to capture the structural and functional complexity of drug molecules. To address this challenge, we propose MFDL-DDI, a multimodal deep learning framework that integrates directed message passing neural networks (D-MPNN), convolutional neural networks (CNN), and fully connected neural networks (FCNN) to jointly process molecular graphs, SMILES embeddings, and Morgan fingerprints. By leveraging an adaptive hybrid fusion, MFDL-DDI effectively captures complementary molecular features across chemical space. Experiments on benchmark datasets demonstrate that MFDL-DDI consistently outperforms state-of-the-art baselines under both warm-start and cold-start settings. These findings suggest that MFDL-DDI provides a robust and generalizable tool for DDI prediction, contributing to safer drug development and more reliable clinical decision-making.