XSXiumin Shi

Xiumin Shi

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Xiumin Shi has a total of 31 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Diff-SelfA-PepGSI: A Diffusion-Based Framework for De Novo Generation, Screening, and Morphological Identification of Self-Assembled Peptide Sequences

    by Weizhong Sun, Yuning Ma, Xiumin Shi, Mingshan Wei, Ao He, Weizhi Wang

    Peptide self-assembly is a ubiquitous molecular phenomenon in which short amino acid sequences spontaneously organize into diverse nanostructures. These assemblies exhibit distinct morphologies closely linked to their biological functions and potential applications, making them attractive candidates for biomaterials in drug delivery, regenerative medicine, and diagnostics. However, designing and predicting peptide sequences with specific self-assembled morphologies remains highly challenging due to the complex interplay of noncovalent interactions, including hydrogen bonding, hydrophobic effects, and π–π stacking. Although deep learning has enabled accurate prediction of peptide self-assembly propensity, rational de novo design of peptides with targeted assembly behavior remains largely unexplored. In this study, we propose Diff-SelfA-PepGSI, a novel pipeline for the morphology-specific design of self-assembling peptides. This three-stage framework comprises: (i) generation, in which a diffusion-based generative model produces candidate self-assembling peptide sequences; (ii) screening, in which pretrained protein language models extract deep semantic features and a deep discriminative network performs sequence optimization and property screening; and (iii) identification, in which a dual-branch structure extracts features and a dedicated classifier identifies the specific nanostructures formed. By integrating sequence generation, predictive screening, and morphology identification, this workflow establishes a proof-of-concept de novo design pipeline that links peptide sequences directly to their target morphologies. Diff-SelfA-PepGSI demonstrates superior performance compared to state-of-the-art models in both screening and identification tasks. Experimental validation preliminarily demonstrates the feasibility of the de novo design pipeline, successfully generating self-assembling peptides such as LAPFLA. Under the reported experimental conditions, 4 of 11 generated candidates (36.4%) exhibited the target morphology predicted by the model. Within the challenging context of de novo morphology-specific peptide design, this success rate supports the potential of Diff-SelfA-PepGSI as an early stage self-assembling peptide generation platform.

  • Prot-ΔΔG: Prediction of protein–protein binding affinity changes upon mutations with pre-training strategies

    by Han Zhou, Yuxiang Wang, Xiumin Shi, Yongfeng Ma

    Protein-protein interactions are essential for diverse biological activities, but amino acid mutations can disrupt these interactions, leading to dysfunction and disease. Mutation impacts can be quantified via change in protein-protein binding affinity (ΔΔG) before and after mutation. Accurate prediction of ΔΔG is critical for understanding disease mechanisms, guiding drug discovery, and advancing protein engineering. Existing computational methods often exhibit reduced accuracy in the absence of high-resolution protein structural data, failing to fully capture sequence-embedded patterns and evolutionary information. To address this limitation, we introduce Prot-ΔΔG, a purely sequence-based deep learning framework that integrates large-scale pre-trained protein language models with a BiGRU-DBRNN encoder. By leveraging solely on wild-type and mutant amino acid sequences, Prot-ΔΔG effectively captures evolutionary and context-dependent patterns without relying on structural inputs. Comprehensive experiments demonstrate that Prot-ΔΔG achieves competitive performance across single-point, mixed, and multi-point mutation prediction scenarios, with the largest improvement observed in protein-level blind testing. This sequence-based approach eliminates the dependency on protein structural information, thereby broadening its applicability, especially in cases where protein structures are unavailable or unreliable.


  • Deep3D-DTA: A Tri-Modal Deep Learning Framework for Binding Affinity Prediction Leveraging 3D Structural Representations of Drugs and Targets

    by Han Zhou, Xiumin Shi, Lu Wang

    Predicting drug-target affinity (DTA) plays a pivotal role in drug discovery and repurposing. While existing computational approaches predominantly rely on 1D sequences or 2D structural data, they often fail to fully capture the intricate nature of molecular interactions. To address this limitation, we propose Deep3D-DTA, a novel tri-modal deep learning framework that integrates 1D sequence semantics, 2D graph topology, and 3D spatial geometry complementary representations for both drugs and target proteins. The proposed architecture offers three key advancements: First, it employs a pre-trained protein language model to encode amino acid sequences, effectively capturing long-range sequential dependencies. Second, it constructs precise 3D structural representations by computing interatomic distances and bond angles, enabling accurate modeling of the spatial conformations of both proteins and compounds. Third, it leverages a hybrid feature extraction module that combines graph neural networks with multi-head attention mechanisms to learn hierarchical structural patterns. Extensive experiments on three widely used benchmark datasets (Davis, KIBA, and Metz) demonstrate that Deep3D-DTA significantly outperforms state-of-the-art methods in DTA prediction. These results highlight its potential as a robust and reliable computational tool for accelerating drug discovery and reducing development costs through more accurate affinity prediction.

  • MuloAD: A Multiomics Integration Model Utilizing Graph Convolutional Networks for Alzheimer's Disease Diagnosis and Biomarker Identification

    by Han Zhou, Yixuan Zhu, Xiumin Shi, Lu Wang

    Early and accurate diagnosis of Alzheimer's disease ( AD ) remains challenging due to the limitations of single biomarker approaches and the complexity of disease pathogenesis. In this study, we present MuloAD, a graph convolutional neural network‐based framework that integrates multiomics data for enhanced AD diagnosis and biomarker identification. The MuloAD leverages GraphSAGE to extract features from DNA methylation, mRNA expression and microRNA expression data independently, followed by a view correlation discovery network that captures cross‐omics relationships in a higher dimensional label space. Using 350 samples from the ROSMAP cohort, MuloAD demonstrates superior classification performance compared to existing multiomics methods, delivering robust accuracy across different omics combinations. Feature‐importance analysis identifies several key molecular biomarkers that exhibit significant differential expression between patients with AD and healthy controls. The results of ablation studies confirm the complementary value of multiomics integration compared to single‐omics approaches. These findings highlight the potential of deep learning‐driven multiomics integration in the diagnosis of neurodegenerative diseases and offer novel molecular targets for AD detection and therapeutic development.


  • DeepMoDRP: A Multi‐Omics‐Based Deep Learning Framework for Drug Response Prediction in Brain Cancer

    by Yuxuan Li, Xiumin Shi, Lu Wang, Lianzhong Zhang

    Considering the limited efficacy of existing pharmacotherapies for brain tumors, the development of accurate predictive models is essential for advancing neuro‐oncology treatment strategies. In this article, we introduce a drug response prediction model, DeepMoDRP, specifically designed for brain cancer. This model integrates genomic, transcriptomic, and epigenomic data from various brain tumor cell lines, including low‐grade glioma, glioblastoma multiforme, and diffuse large B‐cell lymphoma. To address the high‐dimensional complexity inherent in gene expression and copy number variations within cell line data, we have integrated sparse autoencoders (AEs) and denoising AEs to reduce noise and redundancy. Meanwhile, one‐dimensional convolutional neural networks are utilized to process the low‐dimensional mutation and DNA methylation data. Additionally, a multiscale graph neural network is implemented to handle the drug‐related data. Finally, fully connected networks are employed to generate predictions of drug responses. A series of experiments were conducted utilizing a brain tumor dataset that was extracted and curated from public databases. The experimental results demonstrate that the proposed DeepMoDRP outperforms the performance of state‐of‐the‐art pan‐cancer baseline models in predicting drug responses for brain tumors. The downstream analysis indicates that the DeepMoDRP holds significant promise for the treatment of brain tumors.


  • DeepMCL-DTI: predicting drug-target interactions using multi-channel deep learning with attention mechanism

    by Han Zhou, Yijie Guo, Xiumin Shi, Yuxuan Li, Lu Wang

    Drug-target interaction (DTI) prediction is crucial for drug discovery. Deep learning has been extensively utilized to reduce costs and expedite this process. However, most existing methods employ a two-channel architecture that separately constructs feature extraction networks for the drug and the target. These approaches fail to fully leverage the original input data and are unable to completely learn the features from it. In this study, we propose DeepMCL-DTI, an attention-based multi-channel deep learning model with four feature extraction channels: Graph Sample and Aggregate and convolutional neural network for drug features, and ProtBert and bidirectional convolutional long short-term memory for protein features. An interact-attention module models drug-target interactions across both spatial and channel dimensions. Extensive experiments conducted on the DrugBank and Davis datasets demonstrate that DeepMCL-DTI outperforms state-of-the-art methods. A case study on the angiotensin-converting enzyme 2 receptor further confirms its effectiveness as a pre-screening tool for drug discovery.


  • AbEgDiffuser: Antibody Sequence-Structure Codesign with Equivariant Graph Neural Networks and Diffusion Models

    by Yibo Zhu, Xiumin Shi, Jingjuan Zhang, Weizhong Sun, Lu Wang

    Antibodies are crucial immune proteins with high antigen specificity. Conventional antibody engineering is time-consuming and inefficient, whereas deep learning-driven specific antibody design offers an innovative avenue for drug discovery. In this work, we introduce AbEgDiffuser, a deep generative framework that enables the codesign of antibody sequences and structures conditioned on target antigens. Our model integrates diffusion models with equivariant graph neural networks and further incorporates evolutionary constraints. During forward diffusion, amino acid sequences, Cα atom coordinates, and residue orientations are progressively corrupted toward a prior distribution. In reverse, a bilevel equivariant graph neural network captures both residue- and atom-level interactions to reconstruct functional antibodies. To enforce evolutionary plausibility, we encode noisy sequences with the pretrained protein language model ESM-2. Extensive experiments on de novo antibody design and optimization tasks demonstrate that the model generates antibodies with accurate sequences and structures, as well as high binding affinity, outperforming existing design methods.


  • MDL-HTI: A Multimodal Deep Learning Approach for Predicting Herb-Target Interactions

    by Lianzhong Zhang, Xiumin Shi, Xiaohong Deng

    Purpose Traditional Chinese medicine (TCM) has garnered increasing attention from the global medical community due to its unique therapeutic principles and extensive medicinal resources. Understanding herb-target interactions (HTIs) is crucial for elucidating the pharmacological mechanisms that link herbal medicines to biological targets, offering valuable insights into the precise effects of herbal therapeutics. However, current methods exhibit limited effectiveness and fail to fully leverage the biological information associated with herbs and targets. Methods We propose MDL-HTI, a novel framework that integrates heterogeneous graph learning with multimodal biological data. The architecture employs a heterogeneous graph learning network based on the multi-view heterogeneous relation embedding (MV-HRE) algorithm to extract structural patterns from subgraphs, meta-paths, and communities, alongside a biological multimodal information network that encodes herbal ingredients, target pathways, and ligand properties into unified vectors. A relational prediction network with self-attention dynamically fuses features from both components to identify potential HTIs. Results MDL-HTI demonstrates superior performance compared to state-of-the-art baselines. Furthermore, case study validation confirms that our model can serve as an effective tool for identifying potential HTIs. Conclusion This work establishes a novel computational paradigm for TCM pharmacology by integrating topological learning with multimodal biological encoding. MDL-HTI provides a robust platform for elucidating TCM mechanisms and accelerating the discovery of multi-target herbs. The framework has potential applications in precision and personalized medicine, and its predictive capability may significantly reduce experimental costs while improving therapeutic outcomes for complex conditions. Graphical Abstract

  • TrambaHLApan: A Transformer and Mamba-based Neoantigen Prediction Method Considering both Antigen Presentation and Immunogenicity

    by Yibo Zhu, Xiumin Shi, Lu Wang, Jingjuan Zhang

    Neoantigens, tumor-specific peptides with immunogenic potential, represent pivotal targets for cancer immunotherapy. Existing methods prioritize HLA-peptide binding but often fail to adequately address immunogenicity, limiting their clinical utility. This study introduces TrambaHLApan, a novel neoantigen prediction framework that integrates Transformer and Mamba architectures to concurrently predict antigen presentation likelihood (TrambaHLApan-EL) and immunogenic potential (TrambaHLApan-IM). A Transformer-based encoding module is employed to generate unique representations for HLA molecules and peptides. Subsequently, a hybrid fusion module, which combines merged-attention mechanisms with Mamba-based sequential modeling, is deployed to evaluate interaction patterns. TrambaHLApan-IM incorporates antigen presentation scores derived from TrambaHLApan-EL to explicitly model the interplay between antigen presentation and immunogenic potential, thereby enhancing the identification of neoantigens with high confidence. Experimental results on independent datasets demonstrate that TrambaHLApan outperforms state-of-the-art methods, establishing it as a reliable tool for advancing personalized cancer immunotherapies. Graphical Abstract