Kaihui Cheng | 成凯回



Google Scholar GitHub

Portrait of Kaihui Cheng

Biography

I am a Ph.D. student at Fudan University, advised by Prof. Siyu Zhu. My research focuses on generative models for protein dynamics and 3D generation. I study how these models can capture protein conformational changes over time, drawing on molecular dynamics simulations and physical properties.

Research Interests

  • Generative Models for Protein Dynamics
  • 3D Generation

Honors

  • NeurIPS 2026 Top Reviewer

Projects & Publications

Framework for history-dependent bias in protein dynamics emulation
Protein Dynamics

Learning Implicit Bias in Generative Spaces for Accelerating Protein Dynamics Emulation

Authors: Kaihui Cheng, Zhiqiang Cai, Wenkai Xiang, Zhihang Hu, Siyu Zhu†, Tzuhsiung Yang†, Yuan Qi†

Venue: NeurIPS 2026

Description: A history-dependent bias guides a frozen protein dynamics emulator toward underexplored conformations. Paired with score-based refinement, it accelerates exploration while preserving structural validity during long-horizon rollouts.

arXiv
Protein Dynamics
Protein Dynamics

AlphaFolding: 4D Diffusion for Dynamic Protein Structure Prediction with Reference and Motion Guidance

Authors: Kaihui Cheng, Ce Liu, Qingkun Su, Jun Wang, Liwei Zhang, Yining Tang, Yao Yao, Siyu Zhu†, Yuan Qi†

Venue: AAAI 2025

Description: AlphaFolding jointly models protein backbones and side chains across multiple time steps. Trained on molecular dynamics trajectories, it combines reference-structure guidance with motion alignment to improve structural consistency and temporal coherence.

arXiv AAAI
Protein Dynamics Dataset
Dataset

Dynamic PDB: A New Dataset and a SE(3) Model Extension by Integrating Dynamic Behaviors and Physical Properties in Protein Structures

Authors: Ce Liu, Jun Wang, Zhiqiang Cai, Yingxu Wang, Huizhen Kuang, Kaihui Cheng, Liwei Zhang, Qingkun Su, Yining Tang, Fenglei Cao, Limei Han, Siyu Zhu, Yuan Qi

Status: Preprint, 2024

Description: Dynamic PDB contains approximately 12,600 proteins, each with a one-microsecond all-atom molecular dynamics simulation. It includes atomic velocities and forces, potential and kinetic energies, and temperature recorded at one-picosecond intervals. An accompanying SE(3) diffusion model incorporates these physical properties for protein trajectory prediction.

arXiv Dataset
Text-Video Retrieval
Text-Video Retrieval

Text-Video Retrieval Re-ranking via Multi-grained Cross Attention and Frozen Image Encoders

Authors: Zuozhuo Dai, Kaihui Cheng, Fangtao Shao, Zilong Dong, Siyu Zhu†

Venue: Pattern Recognition 2025

Description: CrossTVR refines text-video retrieval results with multi-grained cross attention. Frame- and video-level token selectors identify salient visual information, while a frozen vision backbone reduces training overhead. The approach improves retrieval performance and scales to larger pretrained vision models.

Paper Preprint