Antibody design is critically important in biomedical and therapeutic contexts but remains extremely challenging due to the complexity of antibody sequence–structure relationships and stringent antigen specificity requirements. Traditional computational approaches rely on multi-stage pipelines and often overlook full-atom details (e.g., side-chain conformations) as well as fine-grained geometric features, resulting in limited effectiveness. To overcome these limitations, we propose Dynamic Geometric Equivariant Network (DGENet), an end-to-end full-atom antibody design model that integrates a geometric-kinematic equivariant dynamic optimization module (GK-EDO) with an full-atom E(3)-equivariant message-passing architecture. This framework enables iterative optimization of antibody structures under explicit geometric and kinematic constraints, generating complete antibody structures (including backbone and side chains) and simultaneously jointly optimizing the sequences and 3D structures of the complementarity-determining regions (CDRs). DGENet also introduces a novel virtual anchor docking mechanism that employs an adaptive PNet-Kabsch module to explicitly guide antibody–antigen binding and achieve precise bound conformations. Evaluations on multiple benchmark datasets demonstrate that DGENet exhibits outstanding performance in antibody structure and sequence generation as well as in designing high-affinity antibodies, underscoring its reliability as an advanced antibody design model.

Protein subcellular localization prediction is essential for understanding protein function and cellular organization. However, existing methods exhibit two major limitations: (1) they overlook the critical role of evolutionarily conserved protein domains, which are fundamental functional and structural units that significantly influence functions and subcellular localization, and (2) they rarely learn residue order and backbone coordinates simultaneously, neglecting the complementary information inherent in multi-modal representations. In this paper, we propose a novel Domain-Aware Multi-View Contrastive Representation Learning for Protein Subcellular Localization prediction, named DMVCL. Firstly, it devises domain-sequence/structure attention modules, which identify functionally significant regions in protein structures/sequences that critically determine subcellular localization. Secondly, it introduces a multi-view contrastive learning framework that unites inter-view and intra-view objectives. Inter-view contrastive learning aligns protein sequences with their corresponding structures by maximizing mutual information, thereby capturing the consistency of protein residue order and backbone coordinates. Intra-view contrastive learning enhances the representation discriminability of each modality by explicitly separating proteins with no common location and attracting those with any shared localization. Extensive experiments demonstrate that DMVCL significantly outperforms existing baselines. Ablation studies and visualizations further highlight the contributions of domain-sequence/structure attention and multi-view contrastive learning in achieving superior predictive performance.

Polyps are abnormally growing tissues on the intestinal mucosal surface, often serving as precursors to cellular carcinogenesis. Accurate polyp segmentation is crucial for the diagnosis and treatment of early-stage colorectal cancer. Despite significant advancements in this field enabled by deep learning techniques, complex backgrounds in the gastrointestinal tract-such as luminal folds, air bubbles, and gastric acid secretions-still pose major challenges for existing models. These complexities hinder the models' ability to accurately perceive polyp boundaries and effectively extract contextual information. Due to the prevalent neglect of frequency-domain features in existing segmentation methods, this paper proposes SwinV2-SF-a dual-branch hybrid model integrating the spatial-frequency dualdomain network SFUNet with Swin Transformer V2. It extracts frequency-domain features through Residual Fast Fourier Transform blocks (RFFT), combines spatial and frequency features via a dynamically weighted Spatial-Frequency Attention Assembly (SFAA), and introduces the AutoDFLoss function, to resolve collectively boundary ambiguity and class imbalance. We evaluated the segmentation performance of SwinV2-SF on benchmark datasets including Kvasir-SEG and ETIS-Larib. Experimental results demonstrate that our model achieves competitive performance in terms of mDice and mIoU metrics, notably outperforming state-of-the-art models on the Kvasir-SEG dataset. The code is located at https://github.com/LeedJohn/SwinV2-SF.

Protein function prediction is one of the most important biological problems in the field of bioinformatics. The functions of proteins are generally described by a series of Gene Ontology (GO) terms that have hierarchical relationships. Two factors hinder the effective prediction of protein functions using current methods: 1) they cannot well model and learn the topological semantic similarity between residues and GO terms, resulting in a huge semantic gap; 2) they predict the functions of proteins by calculating the semantic similarity between protein-level embeddings and GO terms, which does not effectively learn the protein-function relationship. To address the above issues, we propose the Topological-aware Residue-Gene Ontology Attention Network (TRGOA) for protein function prediction. First, a topological-aware attention module is designed to leverage attention scores within this joint semantic space allowing for modeling the fine-grained semantic similarity between residues and GO terms, thereby narrowing the semantic gap. Second, a multi-head aggregator is proposed, which adeptly captures the functions relevant fine-grained semantic similarity and filters out function-irrelevant components, which effectively reveal protein-function relationships, thereby enhancing generality and robustness. Finally, TRGOA has demonstrated promising outcomes, revealing our model can understand the protein-function relationship in deep insights.

The majority of spatial transcriptomics datasets are characterized by low resolution, wherein each spot generally encompasses multiple cells. This limitation poses challenges for exploring biological insights at the cellular level. Consequently, the development and application of robust deconvolution methods for spatial transcriptomics data are imperative to address this challenge. Addressing the limitations of previous deconvolution methods—such as the lack of consideration cell type labels from single-cell sequencing data and the inability to adaptively capture local relationship among points—we propose a novel spatial transcriptomics data deconvolution model based on label-guided Multi-Head Dynamic Graph Attention Networks with Optimal Transport(MHDGATOT). Our approach leverages an advanced multi-head dynamic graph attention network to adaptively capture inter-data relationships and generate effective low-dimensional embeddings. Subsequently, we employ optimal transport based on fused gromov-wasserstein to derive the transport matrix between spatial transcriptomics data and single-cell sequencing data, facilitating the accurate deconvolution of spatial transcriptomics datasets. Experimental validation substantiates the effectiveness of our model.

Enzymatic Reaction feasibility prediction is used to determine whether the reaction generated by computational methods can actually occur, which can effectively reduce the complexity of synthetic pathway design. Existing methods often use SMILES or molecular fingerprints to represent molecules, resulting in a lack of molecular structural information. Although some GNNs have been leveraged to solve this problem, the complexity of the intra and inter-substructures interactions of molecules in microorganisms makes it difficult for traditional GNNs to accurately model it. To address these problems, we propose a subgraph-based self-supervised learning framework to predict the feasibility of enzymatic reactions. Specifically, we first propose a subgraph-based two-branch graph neural network. This network leverages the atom graph and substructure graph of a molecule to thoroughly capture its structural and semantic information. Besides, a subgraph interaction module is designed to facilitate the full integration of features. Subsequently, we propose a domain knowledge-guided self-supervised learning task, utilizing molecular fingerprints and substructures to capture the consistency between them effectively. The experimental results show that the proposed method outperforms existing state-ofthe-art methods significantly on all datasets.

Genome-scale metabolic models (GEMs) are powerful tools for predicting cellular metabolic and physiological states. However, there are still missing reactions in GEMs due to incomplete knowledge. Recent gaps filling methods suggest directly predicting missing responses without relying on phenotypic data. However, they do not differentiate between substrates and products when constructing the prediction models, which affects the predictive performance of the models. In this paper, we propose a hyperedge prediction model that distinguishes substrates and products based on dual-scale fused hypergraph convolution, DSHCNet, for inferring the missing reactions to effectively fill gaps in the GEM. First, we model each hyperedge as a heterogeneous complete graph and then decompose it into three subgraphs at both homogeneous and heterogeneous scales. Then we design two graph convolution-based models to, respectively, extract features of the vertices in two scales, which are then fused via the attention mechanism. Finally, the features of all vertices are further pooled to generate the representative feature of the hyperedge. The strategy of graph decomposition in DSHCNet enables the vertices to engage in message passing independently at both scales, thereby enhancing the capability of information propagation and making the obtained product and substrate features more distinguishable. The experimental results show that the average recovery rate of missing reactions obtained by DSHCNet is at least 11.7% higher than that of the state-of-the-art methods, and that the gap-filled GEMs based on our DSHCNet model achieve the best prediction performance, demonstrating the superiority of our method.

In the field of metabolic engineering, evaluating the feasibility of newly generated enzymatic reactions in retrobiosynthesis is a crucial process that helps biologists to efficiently screen out infeasible reactions. However, existing methods overlook the significance of sequence features in molecular SMILES and the ability of model to comprehensively mine and extract features needs to be strengthened. To address these issues, our work propose a novel attention-based multi-task learning (MTL) reaction feasibility checker, named AMTL-RFC, for enzymatic reaction feasibility classification. The model consists of two branches: a Transformer network and a 1-D convolutional neural network (CNN) that extracts SMILES sequence features and spatial structure features of the substrate and product in the reactant pair, respectively. Moreover, a multi-task learning strategy is employed to further enhance the model’s performance. Experimental results demonstrate that AMTL-RFC achieves a classification accuracy of 92.27% on the primary test set, which is highly competitive in the task of classifying the feasibility of enzyme reactions.

Gene Ontology (GO) is a framework that utilizes a series of GO terms in a Directed Acyclic Graph (DAG) to describe protein functions. Proteins are typically annotated with several or dozens of GO terms. However, existing methods often struggle to simultaneously annotate multiple relevant GO terms with hierarchical dependencies to proteins, as they solely rely on protein sequences or structures. To better utilize the hierarchical information of GO terms and improve protein function annotation performance, we propose the Protein Structure-Label Embedding Attention Network for Protein Function Annotation (SLPFA). SLPFA embeds proteins and GO terms into a joint latent space using attention mechanisms to bridge the semantic gap between them. Specifically, we employ a soft-mask GNN to learn the topological structure of proteins, allowing simultaneous focus on key nodes while remaining invariant to irrelevant parts. Additionally, we encode the ancestral information for each GO term in its embedding and utilize a learnable matrix to capture the hierarchical dependencies. Finally, SLPFA employs protein structure-label embedding attention to project the protein structure and label embedding together into a joint latent space. This enables the model to learn the high-level semantics of proteins and hierarchical GO terms, resulting in a reduced semantic gap between proteins and their functions. Experimental results demonstrate that SLPFA outperforms state-of-the-art deep learning-based methods on the PDB-cdhit dataset, which yields Fmax of 0.604, 0.478, 0.524 and the AUPRC of 0.630, 0.357, 0.452 for the MF, BP, CC ontology domains, respectively. Furthermore, when the training and testing proteins have less than 15% sequence identity, SLPFA also achieves competitive results in the MF, BP, and CC ontology domains.

BACKGROUND: Retrosynthesis prediction is the task of deducing reactants from reaction products, which is of great importance for designing the synthesis routes of the target products. The product molecules are generally represented with some descriptors such as simplified molecular input line entry specification (SMILES) or molecular fingerprints in order to build the prediction models. However, most of the existing models utilize only one molecular descriptor and simply consider the molecular descriptors in a whole rather than further mining multi-scale features, which cannot fully and finely utilizes molecules and molecular descriptors features. RESULTS: We propose a novel model to address the above concerns. Firstly, we build a new convolutional neural network (CNN) based feature extraction network to extract multi-scale features from the molecular descriptors by utilizing several filters with different sizes. Then, we utilize a two-branch feature extraction layer to fusion the multi-scale features of several molecular descriptors to perform the retrosynthesis prediction without expert knowledge. The comparing result with other models on the benchmark USPTO-50k chemical dataset shows that our model surpasses the state-of-the-art model by 7.4%, 10.8%, 11.7% and 12.2% in terms of the top-1, top-3, top-5 and top-10 accuracies. Since there is no related work in the field of bioretrosynthesis prediction due to the fact that compounds in metabolic reactions are much more difficult to be featured than those in chemical reactions, we further test the feasibility of our model in task of bioretrosynthesis prediction by using the well-known MetaNetX metabolic dataset, and achieve top-1, top-3, top-5 and top-10 accuracies of 45.2%, 67.0%, 73.6% and 82.2%, respectively. CONCLUSION: The comparison result on USPTO-50k indicates that our proposed model surpasses the existing state-of-the-art model. The evaluation result on MetaNetX dataset indicates that the models used for retrosynthesis prediction can also be used for bioretrosynthesis prediction.