I study how computational models represent, understand and design biological systems.我研究计算模型如何表征生物系统，如何理解它，又能否用它做设计。
Using machine learning to understand, predict and design biomolecules.用机器学习理解、预测和设计生物分子。
The common thread across my projects is not a particular model. It is an evaluation loop designed to make biological AI results harder to fool.贯穿这些项目的不是某一种模型，而是一套评估闭环，用来让生物 AI 结果更难自欺。
Define what the model should know — and what evidence would prove that it does not.先说清模型应该知道什么，也说清什么证据能证明它不知道。
Use family-held-out splits or external cohorts instead of convenient random partitions.用家族留出划分或外部队列，而不是图省事的随机划分。
Challenge apparent success with structured nulls, calibration checks and leakage audits.表面的成功要经得起结构化零假设、校准检查和泄漏审计。
Report what survives the controls, preserve limitations, and make failure reproducible.只报告经得住对照的结论，局限照写，失败也要能复现。
Explore the tree, or switch to the list for a quick, fully accessible overview — every node links to the work behind it.可以在树状图里点着看，也可以切到列表快速浏览。每个节点都连着它背后的工作。
I am a researcher working across AI for biology — protein language models, computational protein design, and trustworthy biomedical AI. My current focus is representation analysis under strict evaluation: structured nulls, family-held-out splits, and explicit claim boundaries.我做生物 AI 方向的研究，覆盖蛋白质语言模型、计算蛋白质设计和可信生物医学 AI。目前的重点是表征分析，评估必须足够严格：结构化零假设、家族留出划分，以及明确的论断边界。
I work with Python, PyTorch and reproducible Linux-based pipelines, and I write about evaluation design and interpretability in my research notes.工作环境是 Python、PyTorch 和可复现的 Linux 流程。评估设计和可解释性方面的思考，写在研究笔记里。
