Transfer-based adversarial attacks often transfer poorly across heterogeneous architectures because CNNs favor local textures while Vision Transformers rely on global shapes. Season decomposes each update into a low-frequency structural branch and a high-frequency textural branch, uses low-saliency guidance to steer high-frequency energy toward background regions, and applies an orthogonal projection so textural updates lie in the orthogonal complement of the structural direction. As a training-free plug-and-play wrapper, it improves transfer success on ImageNet against diverse CNN, ViT, and MLP targets.
@inproceedings{wang2026season,title={Season: Spectrum-Aware Orthogonal Gradient Refinement for Transfer-Based Adversarial Attacks},author={Wang, Tianyi and Gao, Zhenghao and Xu, Shengjie},booktitle={Proceedings of the IEEE International Conference on Multimedia and Expo (ICME)},year={2026},note={Accepted March 17, 2026. CCF-B. Spotlight presentation.},}
arXiv
End-to-End Reverse Screening Identifies Protein Targets of Small Molecules Using HelixFold3
Shengjie Xu, Xianbin Ye, Mengran Zhu, and 3 more authors
arXiv preprint arXiv:2601.13693, 2026
Accepted pending major revision at IEEE Journal of Biomedical and Health Informatics (JBHI)
We present an end-to-end reverse screening pipeline that leverages HelixFold3 to jointly fold proteins and dock small-molecule ligands, enabling large-scale identification of protein targets for small molecules without separate structure preparation. Validated across approximately 100 small molecules, our pipeline outperforms conventional reverse docking workflows in both hit rate and computational efficiency.
@article{xu2026helixfold3,title={End-to-End Reverse Screening Identifies Protein Targets of Small Molecules Using HelixFold3},author={Xu, Shengjie and Ye, Xianbin and Zhu, Mengran and Zhang, Xiaonan and Zhang, Shanzhuo and Fang, Xiaomin},journal={arXiv preprint arXiv:2601.13693},year={2026},note={Accepted pending major revision at IEEE Journal of Biomedical and Health Informatics (JBHI)},url={https://arxiv.org/abs/2601.13693},}
2025
Int. J. Mol. Sci.
Dumpling GNN: Hybrid GNN Enables Better ADC Payload Activity Prediction Based on the Chemical Structure
Shengjie Xu, Linyi Xie, Rui Dai, and 1 more author
International Journal of Molecular Sciences, May 2025
Antibody-drug conjugates (ADCs) represent a promising class of targeted cancer therapeutics that combine the specificity of monoclonal antibodies with the cytotoxicity of small-molecule drugs. The payload activity prediction is crucial for ADC design and optimization. In this study, we propose Dumpling GNN, a hybrid graph neural network architecture that leverages both molecular graph structure and chemical features to predict ADC payload activity. Our approach integrates multiple graph neural network layers with attention mechanisms to capture complex molecular interactions and structural patterns. Experimental results demonstrate that Dumpling GNN significantly outperforms traditional machine learning methods and existing graph neural network approaches in predicting ADC payload activity based on chemical structure. The model achieves improved accuracy in identifying potent payload candidates, which can accelerate the drug discovery process for ADCs.
@article{xu2025dumplinggnn,title={Dumpling GNN: Hybrid GNN Enables Better ADC Payload Activity Prediction Based on the Chemical Structure},author={Xu, Shengjie and Xie, Linyi and Dai, Rui and Lyu, Ziyi},journal={International Journal of Molecular Sciences},volume={26},number={10},pages={4859},year={2025},month=may,publisher={MDPI},doi={10.3390/ijms26104859},url={https://www.mdpi.com/1422-0067/26/10/4859},}
ICCV
FastJSMA: Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling
Zhenghao Gao, Shengjie Xu, Zijing Li, and 4 more authors
In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2025
FastJSMA is an efficient and interpretable adversarial attack that introduces a gradient decoupling mechanism to decompose Jacobian-based saliency into complementary suppression and excitation gradients, reducing the computational complexity of JSMA while maintaining high attack success rates on large-scale datasets.
@inproceedings{gao2025fastjsma,title={FastJSMA: Accelerating Jacobian-based Saliency Map Attacks through Gradient Decoupling},author={Gao, Zhenghao and Xu, Shengjie and Li, Zijing and Chen, Meixi and Yu, Chaojian and Shao, Yuanjie and Gao, Changxin},booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},year={2025},month=oct,pages={1506--1515},url={https://openaccess.thecvf.com/content/ICCV2025/html/Gao_FastJSMA_Accelerating_Jacobian-based_Saliency_Map_Attacks_through_Gradient_Decoupling_ICCV_2025_paper.html},}
2024
Watertox: The Art of Simplicity in Universal Attacks A Cross-Model Framework for Robust Adversarial Generation
Zhenghao Gao, Shengjie Xu, Meixi Chen, and 1 more author
@misc{gao2024watertoxartsimplicityuniversal,title={Watertox: The Art of Simplicity in Universal Attacks A Cross-Model Framework for Robust Adversarial Generation},author={Gao, Zhenghao and Xu, Shengjie and Chen, Meixi and Zhao, Fangyao},year={2024},eprint={2412.15924},archiveprefix={arXiv},primaryclass={cs.CV},url={https://arxiv.org/abs/2412.15924},}