Xiao Kang

Xiao Kang

Auto-generated profile, based on publicly available information

Xiao Kang

Ningbo Artificial Intelligence Institute, Shanghai Jiao Tong University, Ningbo, China

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
4 September 2025 in IEEE Journal of Biomedical and Health Informatics
Verified
Journal Article
Closed access
LZLicai Zhang
Licai Zhang
XKXiao Kang
Xiao Kang
XYXinxing Yang
Xinxing Yang
LWLin Wang
Lin Wang
YGYang Genke
Yang Genke
JCJian Chu
Jian Chu
The personalization of cancer treatment through drug combinations is critical for improving healthcare outcomes, increasing effectiveness, and reducing side effects. Computational methods have become increasingly important to prioritize synergistic drug pairs because of the vast search space of possible chemicals. However, existing approaches typically rely solely on global molecular structures, neglecting information exchange between different modality representations and interactions between molecular and fine-grained fragments, leading to limited understanding of drug synergy mechanisms for personalized treatment. To address these limitations, we propose MOSAIC (Multi-granularity crOSs-modAl method for synergIstic drug combinations prediCtion), an AI-driven multi-granularity cross-modal method for personalized synergistic drug combination prediction that considers both molecular and fragment-level features. MOSAIC employs a dual-layer representation system, decomposing molecules into chemically meaningful fragments using the BRICS algorithm, facilitating information exchange between graph and SMILES representations through a bidirectional cross-attention mechanism, and ensuring semantic consistency of different modal representations of the same molecular fragment through a contrastive learning framework. Additionally, we designed a bilinear attention network to capture interactions between fragments of different drugs and dynamically integrate multi-granularity feature relationships through a multi-head attention mechanism. Through extensive experiments on multiple real-world datasets, MOSAIC demonstrates superior performance over state-of-the-art methods. Literature validation confirms its predicted novel drug combinations align with existing clinical evidence, while visualization analyses elucidate its capability to pinpoint key molecular fragments critical for drug synergy, providing valuable insights for personalized treatment planning and remote patient monitoring.

24 November 2025 in IEEE Transactions on Computational Biology and Bioinformatics
Verified
Journal Article
Closed access
XKXiao Kang
Xiao Kang
LZLicai Zhang
Licai Zhang
XYXinxing Yang
Xinxing Yang
LWLin Wang
Lin Wang
YGYang Genke
Yang Genke
JCJian Chu
Jian Chu
Computational drug repositioning technology can identify potential uses for existing drugs and reduce the time and cost required in the drug development process. How to find appropriate representations of drugs and diseases to predict the associations between the two is the main objective of such tasks. With the emergence of graph neural networks in recent years, researchers learned drugs and diseases via graphs in a bid to improve the prediction accuracy. However, there are three key problems that have not been adequately studied: 1) They usually place drug disease association graphs, drug-drug similarity graphs, and disease-disease similarity graphs under the same se mantic space for learning, which lose the higher-order features of the different graphs. 2) They assign equal weight to each neighbor node when aggregating based on the drug disease association graph, but the effect and mechanism of a drug in treating different diseases are not consistent. 3) They adopt residual connections to enhance the role of the root node without considering that this operation amplifies the effect of anomalous features. In view of this, we first propose a graph attention network-based spatial decomposition method for drug repositioning. Specifically, we reduce the dimensions of the feature space by spatial decomposition and initialize the drug and diseaseem bedding in the drug-similarity subspace, disease-similarity subspace, anddrug-disease association subspace, respectively. The representations of drugs and diseases are jointly captured in the corresponding subspace based on similarity and therapeutic associations. Moreover, the extent of the associations is measured through the graph attention mechanism to explore higher-order neighborhood relationships between drugs and diseases. Finally, we introduce a targeted residual connection for personalized propagation of node features. Experiments on four benchmark datasets show that our proposed architecture outperforms current state-of-the-art approaches. The code of this paper can be accessed at https://github.com/shaoqiezi/GATSD.

10 June 2026 in IEEE Journal of Biomedical and Health Informatics
Verified
Journal Article
Closed access
XKXiao Kang
Xiao Kang
LZLicai Zhang
Licai Zhang
ZWZekun Wang
Zekun Wang
XYXinxing Yang
Xinxing Yang
LWLin Wang
Lin Wang
JCJian Chu
Jian Chu
Predicting drug-target binding affinity is a crucial step in drug discovery, with the aim of estimating the strength of interactions between unknown drug-target pairs. In recent years, with the advancement of graph neural networks and self-supervised learning, numerous novel approaches have been proposed by researchers. However, these methods still exhibit the following drawbacks: 1) They only consider drug-target interaction information when learning relationship features, resulting in low information density and an inability to obtain high-quality representations. 2) The integration of cross-scale features is rigid, with features across different scales showing no correlation, failing to reflect the consistency of the same entity's features across different perspectives. 3) They employ outdated Graph Contrastive Learning (GCL) strategies, which have been proven to yield highly biased estimates. Therefore, to address these challenges, we propose a method that Enhances cross-scale feature Mutual information via Heterogeneous Graph Contrastive Learning for predicting Drug-Target binding Affinity (EMHGCL-DTA). The proposed model reconstructs the relationship network on the macro scale, incorporating similarity information while preserving drug-target interaction data. Furthermore, to establish a connection between macro-scale and micro-scale features, our similarity information is derived directly from the molecular structures of drugs and targets. In addition, we introduce a pertinent unbiased feature augmentation strategy to enhance the representations of drugs and targets. The novelty and effectiveness of the EMHGCL-DTA model are validated on two datasets, where it outperforms state-of-the-art methods.

1 June 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
ZWZekun Wang
Zekun Wang
GPGuojin Pei
Guojin Pei
XZXintao Zheng
Xintao Zheng
YGYang Genke
Yang Genke
JCJian Chu
Jian Chu
XYXinxing Yang
Xinxing Yang
Scene text detection in arbitrary shapes remains a critical challenge in computer vision due to complex backgrounds, irregular geometric layouts, and extreme scale variations. Existing segmentation-based methods typically employ feature pyramids for multi-scale representation but suffer from long information paths between low-level details and high-level semantics, while attention mechanisms often introduce significant parameter overhead. In this paper, we propose CSPA-Net, which incorporates three main components: First, we introduce a Cross-Scale Morphology-Semantic Alignment (CMSA) mechanism that establishes bidirectional feature pathways, combining top-down semantic propagation with bottom-up morphological enhancement to effectively shorten the fusion path and preserve fine-grained boundary information. Second, we adopt a Parameter-Free 3-D Attention (PFA-3D) module following the formulation of SimAM [28], which derives attention weights from neuron energy optimization. This yields a lightweight attention mechanism that captures spatial, channel, and depth correlations without introducing any learnable parameters. Third, CSPA-Net achieves a favorable efficiency-accuracy trade-off: it performs competitively with Transformer-based methods in accuracy while maintaining CNN-like inference speed. Integrated with differentiable binarization [23], our method is extensively evaluated on five benchmarks: ICDAR2015 [6], CTW1500 [7], Total-Text [8], ICDAR2017 [61], and SynthText [55]. Experimental results demonstrate that CSPA-Net achieves new state-of-the-art F-measures of 88.5%, 86.1%, 86.7%, and 74.8% on these datasets, respectively. Comprehensive ablation studies validate the contribution of each component, and complexity analysis shows that CSPA-Net runs at 85 FPS with only 12.3M parameters, significantly outperforming recent Transformer-based methods [72, 75] in efficiency. Our code will be made publicly available.