Leyi Wei

JBHI Journal 2026 Journal Article

LLM-Enhanced Knowledge Distillation for Sequence-Based Protein-Ligand Interaction Prediction

• Wenyu Xi
• Ruheng Wang
• Xiucai Ye
• Tetsuya Sakurai
• Leyi Wei

Accurate prediction of protein-ligand interactions is essential for drug discovery, supporting critical stages from lead optimization to therapeutic development. Many existing methods depend on high-resolution protein-ligand complex structures, which limits scalability and reduces robustness in structure-limited settings. To address these challenges, we introduce Multi-Combinatorial Knowledge Distillation (MCKD), a sequence-based framework that predicts protein-ligand interactions without requiring explicit three-dimensional structures at inference time. MCKD represents proteins and ligands as two-dimensional molecular graphs derived from their sequences and physicochemical properties, enabling effective learning from readily available inputs. To incorporate structural knowledge beyond sequence information, MCKD employs a hybrid distillation strategy that combines cross-modal distillation from a structure-based teacher with self-distillation to improve representation consistency across layers. To model protein-ligand interactions explicitly, MCKD integrates a bilinear attention network that captures residue-atom level associations and supports both binding affinity regression and binary interaction classification. Evaluations on multiple public benchmark datasets show that MCKD consistently outperforms existing sequence-based methods and achieves performance comparable to structure-based approaches. The model also generalizes well to unseen proteins and novel ligand scaffolds, while providing interpretable insights into key molecular interaction regions. These results suggest that MCKD offers a scalable and effective solution for protein-ligand interaction prediction, particularly for structure-free and data-limited drug discovery applications.

JBHI Journal 2026 Journal Article

PGST: A prototype-guided parameter-efficient network for spatial transcriptomics prediction

• Yuan He
• Kaimiao Hu
• Changming Sun
• Leyi Wei
• Ran Su

Spatial transcriptomics (ST) aims to decode spatially resolved gene expression patterns while preserving tissue morphology. Current methods tend to use lower-cost deep learning approaches for gene expression prediction, yet face severe challenges. First, existing methods fail to give sufficient consideration to the spatial specificity of positional encoding inherent in ST; second, they neglect to leverage spatially coherent co-expression patterns across different domains; third, their reliance on linearly weighted aggregation induces vulnerability to noise and distribution shifts; and finally, these architectures exhibit limited parameter efficiency. To address these issues, we introduce prototype-guided network for spatial transcriptomics (PGST), which includes four parts: (1) oriented signal propagation through polar embedding strategy for spatial transcriptomics (PEST); (2) prototype-guided aggregation for global co-feature preservation; (3) global consistency enforcement via shared decoder with reconstruction loss; and (4) lightweight architectural design. Our framework integrates contrastive learning with graph neural networks to balance local-global spatial dependencies and cross-modal consistency. Experimental results on multiple datasets from ST demonstrate the superior performance of our PGST model than existing methods. Our source code is available at: https://github.com/RanSuLab/PGST https://github.com/RanSuLab/PGST.

AAAI Conference 2026 Conference Paper

TRACE: Transformation-Aware Graph Refinement for Reaction Condition Prediction

• Yujie Chen
• Tengfei Ma
• Yuansheng Liu
• Leyi Wei
• Shu Wu
• Dongsheng Cao
• Yiping Liu
• Xiangxiang Zeng

Identifying suitable reaction conditions is critical for chemical synthesis, as they directly affect yield, selectivity, and transformation feasibility. While recent methods have shown promising results, most approaches either encode reactants and products independently or rely on rule-based reaction graphs, both of which constrain the ability of the model to capture condition-relevant structural transformations. In this work, we propose TRACE, a transformation-aware graph refinement framework for reaction condition prediction. TRACE constructs atom-level joint graphs that integrate both reactant and product structures to represent condition-relevant transformations. A structure-aware encoder enriches atom features with local chemical context, followed by a dynamic interaction refinement module that adaptively infers task-specific edges. To further guide the model toward condition-relevant patterns, a mechanism regularized graph encoder incorporates reaction center information, enabling more accurate modeling of transformation mechanisms. Experiments on benchmark datasets show that TRACE achieves state-of-the-art performance across multiple condition types. The integration of transformation-aware refinement leads to improvements in prediction accuracy and generalization, while maintaining robust performance in challenging and realistic synthesis planning scenarios.

JBHI Journal 2025 Journal Article

A Multi-Objective Comprehensive Framework for Predicting Protein-Peptide Interactions and Binding Residues

• Ruheng Wang
• Xuetong Yang
• Chao Pang
• Leyi Wei

Identifying protein–peptide interaction pairs and their corresponding binding residues is crucial and can greatly facilitate peptide therapeutic design as well as improve our understanding of protein function mechanisms. Recently, several computational approaches have been proposed to solve the protein-peptide interaction prediction problem. However, most existing prediction methods cannot simultaneously predict protein–peptide interaction pairs and their binding residues directly from the protein and peptide sequences. Here, we developed a Comprehensive Protein-Peptide Interaction prediction Framework (CPPIF), to predict both binary protein-peptide interaction and their binding residues. We also constructed a benchmark dataset containing more than 8900 protein-peptide interacting pairs with non-covalent interactions and their corresponding binding residues to systematically evaluate the performances of existing models. Comprehensive evaluation on the benchmark datasets demonstrated that CPPIF can successfully predict the non-covalent protein-peptide interactions that cannot be effectively captured by previous prediction methods. Moreover, CPPIF outperformed other state-of-the-art methods in predicting binding residues in the peptides and achieved good performance in the identification of important binding residues in the proteins.

JBHI Journal 2025 Journal Article

BFGTP: A BERT-Guided Two-Stage Molecular Representation Learning Framework for Toxicity Prediction

• Kaimiao Hu
• Yuan He
• Jianguo Wei
• Changming Sun
• Jie Geng
• Leyi Wei
• Ran Su

Accurate prediction of molecular toxicity is vital for drug development. Most mainstream methods rely on fingerprints or graph-based feature extraction, the emergence of large language models (LLMs) offers new prospects for molecular representation learning in toxicity prediction. Although several studies attempt to leverage LLMs to integrate molecular sequence data for pretraining molecular representations, certain limitations remain. Current LLM-based approaches usually utilize solely on class embedding features, overlooking the rich information in sequence embedding. Moreover, integrating pre-trained molecular representations with multi-modal molecular data may further enhance performance in toxicity prediction. To address these challenges, we propose BFGTP, a BERT-guided two-stage molecular representation learning framework for toxicity prediction. Firstly, we design independent encoders for molecular descriptions of three modalities, where the fingerprint encoder with dual level attention mechanisms effectively integrates multi-category fingerprints. Then, the two-stage guide strategy is introduced to fully utilize the prior knowledge of LLMs, employing contrastive learning to align and fuse the tri-modal representations and knowledge distillation to align predicted value distributions. BFGTP ultimately combines fingerprint and graph representations to predict molecular toxicity. Experiments on seven toxicity datasets show that BFGTP outperforms baselines, achieving the highest AUC on five datasets and the best average performance across five evaluation metrics. Ablation studies, t-SNE visualization and case study confirm the effectiveness of BFGTP's components and its ability to capture meaningful molecular representations.

JBHI Journal 2025 Journal Article

ERSR: An Ellipse-constrained pseudo-label refinement and symmetric regularization framework for semi-supervised fetal head segmentation in ultrasound images

• Linkuan Zhou
• Zhexin Chen
• Yufei Shen
• Junlin Xu
• Ping Xuan
• Yixin Zhu
• Yuqi Fang
• Cong Cong

Automated segmentation of the fetal head in ultrasound images is critical for prenatal monitoring. How-ever, achieving robust segmentation remains challenging due to the poor quality of ultrasound images and the lack of annotated data. Semi-supervised methods alleviate the lack of annotated data but struggle with the unique characteristics of fetal head ultrasound images, making it challenging to generate reliable pseudo-labels and enforce effective consistency regularization constraints. To address this issue, we propose a novel semi-supervised framework, ERSR, for fetal head ultrasound segmentation. Our framework consists of the dual-scoring adaptive filtering strategy, the ellipse-constrained pseudo-label refinement, and the symmetry-based multiple consistency regularization. The dual-scoring adaptive filtering strategy uses boundary consistency and contour regularity criteria to evaluate and filter teacher outputs. The ellipse-constrained pseudo-label refinement refines these filtered outputs by fitting leastsquares ellipses, which strengthens pixels near the center of the fitted ellipse and suppresses noise simultaneously. The symmetry-based multiple consistency regularization enforces multi-level consistency across perturbed images, symmetric regions, and between original predictions and pseudo-labels, enabling the model to capture robust and stable shape representations. Our method achieves stateof-the-art performance on two benchmarks. On the HC18 dataset, it reaches Dice scores of 92. 05% and 95. 36% with 10% and 20% labeled data, respectively. On the PSFH dataset, the scores are 91. 68% and 93. 70% under the same settings.

JBHI Journal 2025 Journal Article

GCNLA: Inferring Cell-Cell Interactions From Spatial Transcriptomics With Long Short-Term Memory and Graph Convolutional Networks

• Chao Yang
• Xiuhao Fu
• Zhenjie Luo
• Leyi Wei
• Jingbing Li
• Feifei Cui
• Quan Zou
• Qingchen Zhang

Spatial transcriptomics analysis methods offer an opportunity to investigate highly diverse biological tissues. Cell-cell communication is fundamental for maintaining physiological homeostasis in organisms and coordinating complex biological processes. Identifying cell-cell interactions is critical for understanding cellular activities. The interaction of a cell with other cells depends on several factors, and most of the existing methods that consider only gene expression information of neighbouring cells and spatial location information are somewhat limited. In this paper, we propose a network architecture based on graph convolution network and long short-term memory attention module-GCNLA, which contains graph convolution layer, long short-term memory network, attention module, and residual connections. GCNLA not only learns the spatial structure of cells but also captures interaction information between distal cells, the attention module further extracting and enhancing features related to cell-cell interactions. Finally, the inner product decoding calculates the cosine similarity, which is used to infer cell-cell interactions. In addition, GCNLA is capable of reconstructing the complete cell-cell interaction network. The experimental results on seqFISH and MERFISH demonstrate that the GCNLA network structure has better robustness and noise immunity. The potential features learned by GCNLA enable other downstream analyses, including single-cell resolution cell clustering based on spatial information resolving cell heterogeneity.

JBHI Journal 2025 Journal Article

Guest Editorial: Large Language Models With Applications in Bioinformatics and Biomedicine

• Quan Zou
• Limin Jiang
• Leyi Wei

JBHI Journal 2025 Journal Article

HRProtoKD: A hierarchical and relational prototype based knowledge distillation framework for few-shot cancer molecular subtyping

• Fei Guo
• Rui Shi
• Jia Zhou
• Junlin Xu
• Hui Cui
• Ping Xuan
• Xikang Feng
• Leyi Wei

Accurate molecular subtyping of cancers is critical for drug discovery and disease treatment but re-mains challenging due to the scarcity of labeled data and the intrinsic heterogeneity of cancer biology. While current methods address this issue via few-shot learning (FSL), they often overlook the hierarchical relation-ships among cancer subtypes and the transfer of relational knowledge between models, both of which are crucial for improving generalization under limited supervision. In this work, we propose HRProtoKD, a novel hierarchical and relational prototype-based knowledge distillation framework designed for few-shot cancer subtype classification. HRProtoKD first employs hierarchical inter-class prototype clustering to capture the underlying class relationships. A prototype-wise contrastive loss is then introduced to enhance intra-class compactness and interclass separability. Furthermore, relational prototype knowledge distillation is applied to transfer structural knowledge from a teacher model to a student model. In addition, we construct three benchmark datasets for few-shot cancer molecular subtyping. Extensive experiments conducted on these datasets demonstrate that HRProtoKD consistently outperforms state-of-the-art meta-learning baselines under both 1-shot and 5-shot learning settings, thereby advancing precision oncology and supporting data-driven approaches for drug discovery and therapeutic development.

JBHI Journal 2025 Journal Article

Multi-Modal Deep Representation Learning Accurately Identifies and Interprets Drug-Target Interactions

• Jiayue Hu
• Yuhang Liu
• Xiangxiang Zeng
• Quan Zou
• Ran Su
• Leyi Wei

Deep learning offers efficient solutions for drug-target interaction prediction, but current methods often fail to capture the full complexity of multi-modal data (i. e. , sequence, graphs, and three-dimensional structures), limiting both performance and generalization. Here, we present UnitedDTA, a novel explainable deep learning framework capable of integrating multi-modal biomolecule data to improve the binding affinity prediction, especially for novel (unseen) drugs and targets. UnitedDTA enables automatic learning unified discriminative representations from multi-modality data via contrastive learning and cross-attention mechanisms for cross-modality alignment and integration. Comparative results on multiple benchmark datasets show that UnitedDTA significantly outperforms the state-of-the-art drug-target affinity prediction methods and exhibits better generalization ability in predicting unseen drug-target pairs. More importantly, unlike most “black-box” deep learning methods, our well-established model offers better interpretability which enables us to directly infer the important substructures of the drug-target complexes that influence the binding activity, thus providing the insights in unveiling the binding preferences. Moreover, by extending UnitedDTA to other downstream tasks (e. g. , molecular property prediction), we showcase the proposed multi-modal representation learning is capable of capturing the latent molecular representations that are closely associated with the molecular property, demonstrating the broad application potential for advancing the drug discovery process.

EAAI Journal 2025 Journal Article

PKDF-Net: Anticancer peptide prediction via a prior-knowledge-aware dual-path feature-entangled network

• Qiangguo Jin
• Ankang Wu
• Leyi Wei
• Hui Cui
• Ping Xuan
• Xikang Feng
• Ran Su

Anticancer peptides (ACPs) have emerged as promising therapeutic agents due to their ability to selectively target cancer cells. In recent years, various computational methods have been developed to identify ACPs. However, these approaches often struggle to effectively handle the complexity and variability inherent in peptide sequences. To address these challenges, we propose PKDF-Net, a prior-knowledge-aware dual-path feature-entangled network, designed to provide accurate ACP prediction. PKDF-Net consists of three key modules: the prior-knowledge-aware residue interaction representation (PKRI), the Mamba-enhanced amino acid residue representation (MER), and the dual-path feature-entangled prediction (DPFE). The PKRI module integrates prior biological knowledge, the MER module enhances sparse word embeddings within the sequence state space, and the DPFE module entangles dual-path features for precise prediction. Extensive experiments on benchmark datasets demonstrate that PKDF-Net outperforms state-of-the-art models, offering superior performance in ACP prediction. This study presents a promising tool to accelerate ACP development and provides valuable insights for the broader field of biological sequence analysis.

AIIM Journal 2017 Journal Article

A novel hierarchical selective ensemble classifier with bioinformatics application

• Leyi Wei
• Shixiang Wan
• Jiasheng Guo
• Kelvin KL Wong

Selective ensemble learning is a technique that selects a subset of diverse and accurate basic models in order to generate stronger generalization ability. In this paper, we proposed a novel learning algorithm that is based on parallel optimization and hierarchical selection (PTHS). Our novel feature selection method is based on maximize the sum of relevance and distance (MSRD) for solving the problem of high dimensionality. Specifically, we have a PTHS algorithm that employs parallel optimization and candidate model pruning based on k-means and a hierarchical selection framework. We combine the prediction result of each basic model by majority voting, which employs the divide-and-conquer strategy to save computing time. In addition, the PT algorithm is capable to transform a multi-class problem into a binary classification problem, and thereby allowing our ensemble model to address multi-class problems. Empirical study shows that MSRD is efficient in solving the high dimensionality problem, and PTHS exhibits better performance than the other existing classification algorithms. Most importantly, our classifier achieved high-level performance on several bioinformatics problems (e. g. tRNA identification, and protein-protein interaction prediction, etc.), demonstrating efficiency and robustness.

AIIM Journal 2017 Journal Article

Improved prediction of protein–protein interactions using novel negative samples, features, and an ensemble classifier

• Leyi Wei
• Pengwei Xing
• Jiancang Zeng
• JinXiu Chen
• Ran Su
• Fei Guo

Computational methods are employed in bioinformatics to predict protein–protein interactions (PPIs). PPIs and protein–protein non-interactions (PPNIs) display different levels of development, and the number of PPIs is considerably greater than that of PPNIs. This significant difference in the number of PPIs and PPNIs increases the cost of constructing a balanced dataset. PPIs can be classified as either physical or genetic. However, ready-made PPNI databases were proven only to have no physical interactions and were not proven to have no genetic interactions. Hence, ready-made PPNI databases contain false negative non-interactions. In this study, two PPNI datasets were artificially generated from a PPI database. In contrast to various traditional PPI feature extraction methods based on sequential information, two types of novel feature extraction methods were proposed. One is based on secondary structure information, and the other is based on the physicochemical properties of proteins. The experimental results of the RandomPairs dataset validate the efficiency and effectiveness of the proposed prediction model. These results reveal the potential of constructing a PPI negative dataset to reduce false negatives. Related datasets, tools, and source codes are accessible at http: //lab. malab. cn/soft/PPIPre/PPIPre. html.