Hongyuan Zhou 周宏远

Hongyuan is a master’s student in the Xu Group at South China University of Technology, joining the 2026 cohort. His background is in biotechnology, with additional training in economics. He is interested in AI for biological research and in testing model predictions against experimental evidence and known biological constraints.

His experience spans protein expression and purification, computational protein analysis, causal inference and AI agent development. Across these areas, he asks how to assess whether a prediction or a data-driven conclusion is sufficiently supported before using it to guide the next decision.

Synthetic biology and experimental work

As a Student Leader of the silver-medal-winning SCUT-China team at iGEM 2025, Hongyuan coordinated computational and experimental work on human alpha-lactalbumin production for a space-nutrition concept. His hands-on work included strain engineering, protein expression and purification. The project explored computational designs for altered digestibility; improved digestion and uptake were not established by functional experiments.

Evaluating protein language models

In hLA-AI-redesign, he evaluated ESM2 predictions for human alpha-lactalbumin using known disulfide-bond constraints as a reference. This computational case study examines where sequence-based predictions may need additional structural evidence before informing protein design.

Causal data analysis

Hongyuan worked on data cleaning, panel-data construction and instrumental-variable analysis for a study of administrative shareholding and CBA team performance. He is the second author of Does Administrative Shareholding Affect Athletic Performance? Evidence from the CBA, listed in his supplied materials as under review at the Journal of Sports Economics in 2026.

Evidence-grounded AI agents

He independently developed Finer, a multi-agent research workflow that connects extracted investment-research claims to their source evidence. His work includes an annotation platform, preference-data pipelines, an initial DPO-LoRA fine-tuning experiment and verifiable reward rules for evidence grounding, calibration and abstention. The reported model evaluation is an initial small-sample study.