Antibody affinity maturation
抗体亲和力成熟
Optimize an existing binder with small, interpretable interface changes—preserving the binding mode while improving affinity, breadth and developability.
Interface mutationΔΔGWet-lab feedback
AI for biology
From plausible proteins to experimentally useful ones.
从“看起来合理”，走向“真实可验证”。聚焦蛋白质表示、抗体亲和力成熟和实验引导的模型纠偏。
01
Current research
Working direction
Model priors meet experimental evidence.
Working thesis“A protein can be structurally plausible and still fail to bind.”
Iterative research loop
01Priorlanguage + structure
→02Proposemutations + paths
→03Evaluatephysics + uncertainty
→04Correctwet-lab evidence
Language × structure priors
序列—结构先验
Treat protein language and structure models as compatibility priors, then identify where naturalness scores stop being reliable proxies for real binding.
PLMSE(3)Structure compatibility
Epistasis-guided evolution
上位性引导的进化
Use sparse single- and double-mutant evidence to reason about higher-order combinations, uncertainty and the routes that directed evolution should test next.
Fitness landscapeFew-shotActive learning
?
The question underneath the methods
How can sparse experimental evidence correct a strong—but imperfect—model prior?
如何让少量真实实验数据，纠正强大但并不完美的模型先验？
02
Research notes
Questions in progress
Notes from the edge
of the current work.
A growing space for reading, model critique and questions that shape the next experiment.
Structural compatibility is not affinity
结构自洽为什么不等于真实结合？
From two mutations to many
稀疏实验数据如何约束高阶突变预测？
Putting physics into a few-shot model
物理规则应该进入数据、损失，还是推理过程？