Junwen Liao (廖俊文)

Ph.D. candidate, Qiuzhen College, Tsinghua University

headshot.jpg

Qiuzhen College, Tsinghua University

Advised by Prof. Chenglong Bao

liaojw24@mails.tsinghua.edu.cn

I am a third-year Ph.D. candidate in mathematics at Qiuzhen College, Tsinghua University, advised by Prof. Chenglong Bao. My current research focuses on combining AlphaFold with cryo-EM observations for protein structure prediction.

More broadly, I am interested in AI for science, computational imaging, and data-driven methods for molecular and biological systems. Before joining Tsinghua University, I received my B.S. in Statistics from the School of Mathematics, Sichuan University.

My recent work develops methods that connect deep protein-structure priors with limited or imperfect experimental observations, aiming to improve atomic model refinement when reconstructed maps are weak, anisotropic, or incomplete.

Research interests: AI for science, protein structure prediction, computational imaging.

news

Oct 21, 2026 Selected for the 2026 CAST Young Science and Technology Talent Cultivation Program (Doctoral Student Special Plan).
Oct 18, 2026 Presented our work on CoCoFold at ORSC 2026 in Suzhou, China. Talk details.

Preprints

  1. LTS Preprints
    cocofold2-flow.png
    CoCoFold2: scalable latent refinement of diffusion-based protein structure predictions from limited-particle cryo-EM data
    Junwen Liao, Mingxu Hu, and Chenglong Bao
    2026
    LangTaoSha Preprint Server. CoCoFold2 refines protein structure predictions using limited-particle cryo-EM observations and a frozen Protenix-v1 diffusion prior. It optimizes target-specific latent perturbations and supports component-parallel refinement across multiple GPUs.
    DOI Code

Selected Publications

  1. Commun. Chem.
    pipeline.png
    Fine-tuning AlphaFold with limited cryo-EM observations
    Junwen Liao, Dihan Zheng, Hui Zhang, and 3 more authors
    Communications Chemistry, 2026
    CoCoFold adapts AlphaFold using limited cryo-EM particles to refine atomic models when reconstructed maps are weak, anisotropic, or incomplete. It connects predicted structures to experimental particles through differentiable projection and particle-level supervision while preserving the AlphaFold structural prior.
    DOI Code

Beyond Research

Beyond Research, I enjoy music and anime.