CORE CONCEPT
BNU-China focuses on "continuous feedback and iterative optimization" as the core, driving the deep integration and positive cycle between science, education, and industry in the iGEM project. Through systematic dialogues with experts from diverse backgrounds in research, education, and industry, we continuously incorporate external insights into the project design and iteration process, ensuring that the project evolves in terms of scientific rigor, social relevance, and technical feasibility.
EXPERIMENT
During the project development process, we conducted multiple rounds of interviews with 12 experts from different fields. Their feedback was integrated into every stage of the project, forming a continuous DBTL optimization cycle. In the Design stage, expert input reshaped our technical framework and validated key hypotheses. In the Build stage, we addressed expression, assembly, and purification bottlenecks. In the Test stage, we established a rigorous verification system. In the Learn stage, we clarified the green transformation direction and societal impact of our technology.
Design: Shaping the Core Concept
Design: Shaping the Core Concept
At the beginning of the project design, we faced dual challenges: validating our core hypotheses and optimizing the system architecture. To ensure the scientific rigor and feasibility of our technical roadmap, we proactively engaged in in-depth conversations with experts across multiple disciplines. Their insights provided a solid theoretical foundation for the project.
Professor Chunfu Xu (Modeling and Prediction)
Mr. Chunfu Xu
Click here to learn more
Assistant Professor, TIMBR, Assistant Investigator, NIBS, Beijing, China
Main takeaways
- Developing deep-learning-based protein design approaches;
- Designing functional protein devices for basic research and disease therapeutics and diagnosis;
- Designing novel enzymes to address environmental and energy crises.
His laboratory will develop new computational protein design methods and explore the applications of de novo designed proteins in numerous research areas. The primary research directions in his lab include:
Why We Reached Out:
By chance, during a lecture by Professor Yan Ning, we raised questions about applying AI-based protein structure prediction to support structural biology research, hoping to integrate cutting-edge computational methods into our project design.
What we learned:
Professor Yan emphasized that although AI tools like AlphaFold provide powerful predictive capabilities, they will not replace experimental structural biology in the foreseeable future. She clarified that while AI models excel at rapidly proposing structural hypotheses, they often lack mechanistic and dynamic insights and therefore should not be trusted uncritically.
How we integrated this:
We adopted a combined computational-experimental strategy, integrating AI predictions into three key aspects of our initial design: prediction of the capsid structure, prediction of the SlPPK structure, and molecular docking of protein-protein interactions.
Professor Jiming Zheng (Ferritin System)
Mr. Jimin Zheng
Click here to learn more
Professor of Chemistry, Beijing Normal University
Main takeaways
A professor and doctoral supervisor at the College of Chemistry, Beijing Normal University,he possesses a profound academic background and extensive research experience. Professor Zheng's main research focuses on natural resource products and food biotechnology, with long-term dedication to interdisciplinary studies in chemistry and life sciences. Currently, he offers several core courses for undergraduate and graduate students, including "Chemistry and Life," "Natural Resource Products," and "Food Biotechnology," emphasizing the integration of theoretical knowledge with practical applications to cultivate students' innovative thinking and practical skills.
Why we sought help:
Beyond validating our current technical route, we also wanted to explore alternative encapsulation systems with greater potential. We turned to Professor Zheng, an expert in metalloproteins and biomineralization, to discuss ferritin-based inorganic crystallization strategies.
What we learned:
He raised key technical questions, such as whether ferritin (with an ~8 nm inner cavity) could accommodate zinc oxide crystals and how to chemically verify their presence. He advised us to first procure commercial ferritin to bypass expression challenges and to use molecular modeling to predict the effects of targeting peptides on structure.
How we integrated this:
Based on his advice, we developed an evaluation plan for the ferritin system, including procuring commercial materials, simulating peptide–protein interactions, and designing chemical validation experiments. These steps informed our early technical route design before entering the build stage.
Professor Youjun Wang(Feasibility and Risk Assessment)
Mr. Youjun Wang
Click here to learn more
Professor of Chemistry, Beijing Normal University
Main takeaways
The mechanisms of calcium signal transduction and their physiological as well as pathological roles. In recent years, the main focus has been on calcium signaling pathways during the interaction of organelles (endoplasmic reticulum-plasma membrane), dedicated to revealing the activation and regulatory mechanisms of store-operated calcium entry (SOCE) mediated by STIM-Orai proteins.
Why we sought help:
Our project relies on the fundamental assumption that polyphosphate can enter VLPs to provide energy for the ATP regeneration system. Before experimental work began, we consulted Professor Wang to evaluate the feasibility and risks.
What we learned:
He emphasized the importance of verifying substrate entry and characterizing the internal environment of VLPs. He also highlighted a hidden risk: acidification induced by polyphosphate might inhibit enzymes, and recommended starting with empty VLPs and progressing incrementally.
How we integrated this:
This directly defined our initial research pipeline. We prioritized experiments to detect VLP permeability and designed a protocol to monitor and control pH levels. We strictly followed the sequence: empty VLP → substrate entry → functional assembly.
Professor Zhanxin Wang(System Simplification)
Mr. Zhanxin Wang
Click here to learn more
Professor of Life Sciences, Beijing Normal University
Main takeaways
Using approaches in structural biology, biochemistry, and cell biology, the research focuses on the molecular mechanisms by which important protein complexes participate in epigenetic regulation, as well as the relationship between their abnormalities and human diseases.
Why we sought help:
We faced issues with system complexity and enzyme leakage. We consulted Professor Wang for advice on whether multi-layer assembly was necessary or if a simpler strategy could improve stability.
What we learned:
He pointed out that open systems are prone to enzyme leakage and recommended closed systems to enhance stability. He encouraged us to reconsider the necessity of using both CP and SP simultaneously, suggested His-tag/Ni-column enzyme immobilization, and emphasized structural validation using EM and centrifugation.
How we integrated it:
We reassessed the rationality of our complex design, strengthened experimental evidence, optimized the assembly plan, and introduced His-tag immobilization strategies. This resulted in a more streamlined and efficient system while retaining functionality.
Build: Troubleshooting and Optimization
Build Phase: Troubleshooting and Optimization
As the project entered the experimental construction phase, protein soluble expression and efficient assembly became critical bottlenecks restricting progress. By consulting experts in protein folding and nano-assembly, we obtained a series of targeted solutions that successfully transformed our theoretical designs into practical, operable experimental protocols.
Professor Sen Li(Protein Expression and Purification)
Mr. Sen Li
Click here to learn more
Chair, Department of Biochemistry and Biotechnology, School of Life Sciences, Beijing Normal University
Main takeaways
With extensive work experience and research achievements in the field of biochemistry and molecular biology, the research has innovatively proposed the concept of single-chain antibody-type molecular chaperones, screened and identified single-chain antibodies against several major disease-related proteins, developed a set of single-chain antibody-oligopeptide chaperone systems effectively aiding protein refolding, and systematically studied the effects of mixed crowded systems on protein folding and aggregation.
Why we sought help:
In the protein expression and purification phase, the formation of inclusion bodies was our primary challenge. We sought practical solutions from protein experts, such as Professor Li Sen.
What we learned:
He provided a comprehensive toolkit: adjusting IPTG concentration, using chaperone strains, and optimizing refolding conditions during dialysis (e.g., using DTT/β-mercaptoethanol for slow reduction of urea).
How we integrated it:
We systematically applied his suggestions in the lab: adjusting IPTG levels, switching to chaperone-assisted strains, using slow urea removal during 4 °C dialysis with reducing agents—greatly improving soluble, functional protein yield.
Professor Xiao Zhao & Kekman Cheng (Assembly and Encapsulation)
Mr. Xiao Zhao
Click here to learn more
National Center for Nanoscience and Technology, Chinese Academy of Sciences
Main takeaways
- Modification and application of biogenic nanoparticles
- Construction of nano-artificial hybrid biological systems and their applications in tumor immunodiagnosis and therapy
- Precise protein assembly at the nanoscale and its spatial biological effects
The research primarily focuses on the development and application of synthetic biology-based biogenic nanocarriers, modification and application of biogenic nanoparticles, precise protein assembly at the nanoscale, and their spatial biological effects.
Mr. Keman Cheng
Click here to learn more
National Center for Nanoscience and Technology, Chinese Academy of Sciences
Main takeaways
He is now engaged in teaching and research in the field of nanoscience and technology at the University of Chinese Academy of Sciences. His primary recruitment disciplines include Nanoscience and Technology,Nanoscience and Technology, and Nanoscience and Technology. His research focuses on nanobiology, chemical biology, and nanomedicine. He has accumulated in-depth research expertise in the interaction mechanisms between nanomaterials and biological systems, biological effects at the nanoscale, and the application of nanotechnology in medical diagnosis and therapy. Actively participating in domestic and international academic exchanges, Keman Chen is dedicated to advancing the integration of nanoscience and biomedicine. He has achieved multiple research outcomes in related fields and has accumulated extensive experience in education and scientific research.
Why we sought help:
Efficient enzyme encapsulation is a core step in the “Build” stage and is crucial for maintaining system activity and stability. To address challenges related to loading efficiency and structural assembly, we consulted Professors Xiao Zhao and Keman Cheng.
What we learned:
They advised us to carefully consider the VLP structure to avoid spike regions blocking SpyTag sites, which would prevent SpyCatcher from binding. They also suggested using an excess of internal enzymes to increase loading efficiency and performing ratio modeling to improve encapsulation uniformity and yield. Additionally, they emphasized the importance of precise buffer control during assembly to maintain structural stability.
How we integrated it:
Based on their suggestions, we adjusted our assembly strategy by repositioning SpyTag to reduce spatial conflicts, optimizing the internal enzyme-to-capsid protein ratio, and refining buffer conditions. These changes significantly improved encapsulation efficiency and particle uniformity.
Professor Li Tong (Purification Strategy)
Ms. Li Tong
Click here to learn more
Senior Engineer at the School of Life Sciences, Beijing Normal University.
Main takeaways
Long engaged in undergraduate and graduate teaching, co-authored textbooks such as Fundamental Biochemistry and Biochemistry Experiments, and responsible for the development and reform of the national first-class undergraduate course "Molecular Biology Experiments".
Why we sought help:
In the later stages of the project, purification became a key factor affecting product quality and subsequent analyses, particularly the challenge of removing empty capsids without disrupting enzyme encapsulation. Therefore, we consulted Professor Tong.
What we learned:
Professor Tong emphasized leveraging the hydrophobic–hydrophilic properties of the capsid structure and the His-tag to design an effective separation strategy. She suggested exploring multiple methods, including centrifugation, ammonium sulfate precipitation (with caution to avoid disrupting the encapsulation), and IP-based solid-phase separation. She also raised an important question: Is it necessary to completely remove empty capsids? If their impact is minimal, retaining a portion could simplify the process, and their effect on the effective enzyme concentration could be quantified to support decision-making.
How we integrated this:
We conducted parallel testing of multiple purification methods, evaluating their effects on VLP integrity and activity. Additionally, we began quantifying the impact of empty capsids on overall catalytic performance, enabling us to develop a more flexible and efficient purification strategy.
Test: Rigorous Validation and Characterization
Test: Rigorous Validation and Characterization
Establishing a reliable validation system is the core process for evaluating project results. Faced with challenges in characterizing a complex system, we consulted experts in biophysics and systems engineering to build a multi-layer, cross-scale testing scheme, ensuring both academic rigor and industrial application value.
Professor Yong Tao(Value Validation)
Mr. Yong Tao
Click here to learn more
Univercity of Chinese Academy of Sciences
Main takeaways
Research focuses on extracellular ATP regeneration systems using PPK2 for ATP regeneration. A recently published paper, "Cell-free regeneration of ATP based on polyphosphate kinase 2 facilitates cytidine 5'-monophosphate production" (ScienceDirect, 2023), discusses an extracellular ATP regeneration system using PPK2. The cell-free system employs polyphosphate kinase 2 (PPK2) to regenerate ATP, thereby promoting the production of cytidine 5'-monophosphate (5'-CMP).
Why we sought help:
We wanted to identify the key success evidence for our project and ensure that our validation methods met the highest academic and industrial standards, clearly demonstrating the performance of our system.
What we learned:
Professor Tao emphasized the need for "physical evidence" of encapsulation and functionality, rather than relying solely on activity measurements. He suggested conducting direct comparative experiments between our system and conventional ATP regeneration methods, while also cautioning us about potential issues related to enzyme activity mismatch and waste accumulation.
How we integrated it:
We expanded our characterization plan to include DLS and size-exclusion chromatography, and we designed controlled experiments to quantitatively compare our system with traditional one-pot methods. These steps strengthened the scientific rigor and persuasiveness of our validation process.
Professor Kunhui Liu(Theoretical and Analytical Framework)
Mr. Kunhui Liu
Click here to learn more
Associate Professor, College of Chemistry, Beijing Normal University.
Main takeaways
Research focuses on laser spectroscopy, bio-organic photochemistry, surface plasmon nanophotonics, photodegradation and photocatalysis, and singlet oxygen radical reaction kinetics.
Why we sought help:
When analyzing the multi-enzyme catalytic processes within the confined space of VLPs, we encountered significant challenges in understanding the complex reaction kinetics and substrate transport across the capsid boundary. To establish a robust analytical and modeling framework, we consulted Professor Kunhui Liu, an expert in theoretical analysis and systems modeling.
What we learned:
Professor Liu pointed out the inherent complexity of describing self-assembly dynamics and confined multi-enzyme reaction kinetics, emphasizing the necessity of using computational simulations to capture the spatiotemporal behavior of multi-enzyme systems. He also highlighted the critical importance of precise temperature and pH control in experiments to ensure data reliability and reproducibility.
How we integrated it:
During the testing and analysis stage, we adopted his suggestions by implementing strict environmental control mechanisms and introducing kinetic modeling and simulation tools. These measures allowed us to better interpret the system’s internal reaction dynamics and provided a stronger theoretical basis for our experimental design.
Learning: Defining Purpose and Impact
Learning: Defining Purpose and Impact
At the project maturation stage, we focused on the practical pathways for technology transfer and its broader social responsibilities. Through in-depth discussions with experts in green biomanufacturing and biosafety, we reexamined the application scenarios and potential impacts of our project, establishing a development direction that balances innovation with sustainability.
Professor Shuwen Liu (Application Scope Definition and Green Chemistry)
Mr. Shuwen Liu
Click here to learn more
Univercity of Chinese Academy of Sciences
Main takeaways
Guided by industrial technical indicators and production costs, research is conducted on the biosynthesis and green manufacturing of vitamins, amino acids, and material monomers. The main focus is on key core technologies for microbial metabolic engineering breeding assisted by synthetic biology, chemical biology, and systems biology, as well as industrial integration and application technologies integrating fermentation engineering and biochemical engineering. Independently developed technologies for producing vitamin B5 via fermentation and pentamethylenediamine via whole-cell catalysis have been transferred to enterprises, leading to the construction of the world's first industrial production lines for both.
Why we sought help:
As the project entered a more mature stage, we aimed to extend our technology beyond laboratory research into real industrial applications while highlighting its unique advantages. To achieve this, we consulted Professor Shuwen Liu, an expert in green biomanufacturing.
What we learned:
Professor Liu advised us to focus on the synthesis of high-value or toxic products in cell-free systems, emphasizing our advantages in enzyme stability, reusability, and green certification. He also reminded us of the potential environmental impact of phosphate waste streams and provided valuable industrial context to better define our application direction.
How we integrated it:
We shifted our project narrative to focus on the synthesis of high-value toxic products and established quantitative indicators for reusability and stability, highlighting the efficiency, economic, and environmental benefits of our system. This became the core value proposition for our technology’s application scenarios.
Professor Zhonglin Lu (Safety and Advancement)
Mr. Zhonglin Lu
Click here to learn more
Professor of Chemistry, Beijing Normal University
Main takeaways
- Chemical Biology: Design, synthesis, and property studies of organic small-molecule non-viral gene vectors and fluorescent probes; design and property studies of drug sustained-release systems; design, synthesis, and property studies of bioactive small-molecule fluorescent probes.
- Supramolecular Chemistry: Design and synthesis of molecular zippers and studies on peptide folding and assembly; design and synthesis of rigid macrocycles and their controllable self-assembly.
- Application Development: Industrial wastewater treatment and development of pharmaceutical intermediates.
Why we sought help:
We wanted to ensure that our system was not only technically feasible but also demonstrated clear advantages in biosafety and environmental compatibility over existing methods. Therefore, we consulted Professor Zhonglin Lu.
What we learned:
Professor Lu emphasized the need to assess the potential toxicity of polyphosphate and to demonstrate substantial improvements in both safety and performance. He also highlighted the importance of improving assembly uniformity, which is directly linked to the system’s stability and safety.
How we integrated this:
We incorporated biocompatibility assessments into our project considerations and applied his suggestions on improving assembly strategies. This ensured that our project was grounded in safety and responsibility from the very beginning.
iGEM Navigator LARGE MODEL
1. Introduction of iGEM Navigator large model
Throughout the project, we came to realize that, thousands of teams around the world have accumulated invaluable knowledge and experience in experimental design, human practices, education and outreach, and technology transfer. Yet these insights remain scattered across individual team wikis, lacking any structured curation or efficient retrieval mechanism. As a result, hard-won lessons are seldom reused or built upon over the past fifteen years of iGEM. To address this gap, the BNU-China tech team developed iGEM Navigator—an innovative knowledge-engine foundation-model powered by large language models (LLM) and retrieval-augmented generation (RAG). By systematically cleaning and indexing every iGEM wiki ever published, the platform combines semantic search with intelligent generation to act as a seasoned project consultant: it pinpoints critical information, distills core takeaways, and sparks creative directions for new teams.
More than a high-performance search tool, iGEM Navigator serves as a "navigator" for project innovation. It guides teams through the labyrinth of historical data, linking past knowledge to present designs and future applications, so that every new idea stands on the shoulders of collective wisdom—truly "letting the past inspire the future".
2. Target Problem
2.1 Fragmented iGEM wikis, no central index
Year-by-year team wikis sit on isolated URLs, lack uniform structure or tags, and force researchers to hop endlessly between pages without ever seeing the big picture.
2.2 Keyword search fails for cross-disciplinary, cross-track questions
Traditional keyword engines cannot bridge biology, hardware, policy and outreach in a single query, so iGEM-specific, multi-field information needs remain unsatisfied.
2.3 Knowledge is not reused
Teams repeatedly solve similar problems; continuity is broken and innovation chains never form. Every new project effectively "starts from zero" because past know-how is not systematically inherited or deepened.
2.4 No automated "expert consultant"
The community still relies on personal memory and scattered Google queries. There is no tool that, given a design brief, automatically retrieves analogous cases, synthesises findings and suggests novel directions.
iGEM Navigator was built to erase these pain-points. After crawling and structurally cleaning 10 years of award-winning & nominated iGEM wikis, we fine-tuned an LLM and layered it with a domain-specific knowledge graph. The once-siloed pages are now a live, query-ready knowledge network that teams can interrogate in natural language for instant, context-aware guidance.
3. Data Collection
3.1 Data Sources
We systematically harvested, from the official iGEM website, every team wiki that won—or was nominated for—a special prize between 2014 and 2024. The dataset spans thousands of projects from dozens of countries and all major tracks, covering core synthetic-biology modules as well as education, human-practice and industry-translation content.
3.2 Content Extracted
Pages were segmented and structured to capture four key blocks:
(1) Design & Experimental Modules: protocols, part assembly, optimisation strategies, validation workflows.
(2) Human Practices: stakeholder interviews, policy analyses, social surveys, DBTL-cycle integration.
(3) Education & Communication: outreach materials, workshop designs, public-engagement metrics.
(4) References & Deliverables: team papers, technical reports, road-maps, retrospectives.
3.3 Data Cleaning and Structuring
The original Wiki pages varied significantly in structure, format, and content quality. To ensure efficient model usage, we performed multiple rounds of data cleaning and standardization, including:
(1) HTML Parsing and Text Extraction: Removing redundant formats, image tags, and hyperlink noise;
(2) Text Chunking: Dividing long texts into semantically coherent knowledge paragraphs for subsequent vector-based search;
(3) Module Labeling and Tagging System Construction: Based on the Wiki content structure and iGEM track characteristics, we categorized different types of knowledge with appropriate tags;
(4) De-duplication and Standardization: Eliminating duplicate content, standardizing terminology, and formatting.
In the end, we created a high-quality iGEM Wiki knowledge base covering the past 10 years, providing a solid data foundation for the large model's retrieval, understanding, and generation.
4. Technical Core of iGEM Navigator
To guarantee both pinpoint retrieval and high-quality generation across a sprawling knowledge space, iGEM Navigator follows a "RAG + fine-tuned LLM" pipeline that marries Retrieval-Augmented Generation with instruction tuning. This hybrid design lets the system remember more, understand deeper, answer accurately, and inspire creatively.
4.1 Core Technology 1: Retrieval-Augmented Generation (RAG)
Once relevant passages are retrieved, they are injected as context into the large language model. The model no longer relies on its parametric memory; instead, it grounds every answer in the authentic, curated iGEM knowledge base.
4.1.1 From Raw Data to Searchable Knowledge: Chunking & Indexing
(1) Hierarchical splitting: first by native Wiki headings (Overview / Design / Results / Human Practices / Education / Implementation), then by paragraph.
(2) Recommended chunk: 500–800 Chinese characters or 300–500 words per block, with 80–120-character overlap to preserve bilingual coherence.Every chunk is tagged with year, team, track, country/region, section, keywords, organisms, pathways, techniques.
(3) Noise & duplicate removal: navigation bars, footers, and templated menus are stripped; blocks with cosine similarity > 0.95 are clustered and deduplicated, keeping the earliest authoritative source.
(4) Vectorization: chunks are embedded with a bilingual Chinese-English text model; original text is preserved for citation and both vectors and raw text are stored in a vector DB (ANN/IVF-HNSW index).
4.1.2 Hybrid Retrieval
(1) Sparse + Dense: BM25 (keyword) and dense vector (semantic) run in parallel; scores fused as score = α·dense + (1–α)·BM25 with α≈0.6–0.8.
(2) Query Rewriting: user questions are synonym-expanded and entity-normalised(e.g., "cell-free ATP regeneration" ↔ "cell-free ATP regeneration / energy recycling"); explicit filters auto-appended (year:2015-2023, section:HP/Education).
(3) Deduplication & Diversity: Maximal Marginal Relevance (MMR) applied to Top-N candidates to return relevant yet mutually dissimilar chunks, preventing single-source domination.
(4) Re-ranking: a cross-encoder re-scores the shortlisted passages for fine-grained relevance; Top-k (typically 10) chunks fed to the generator.
4.1.3 Context Assembly & Traceability Constraints
(1) Assembly order (highest → lowest priority):
1 Direct-evidence chunks (1–2 passages with strongest query match)
2 Background chunks (definitions / prerequisites, 1–3 passages)
3 Caveat / risk chunks (limitations, boundary conditions, 1–2 passages)
(2) Window budgeting: reserve 3–4 k tokens for knowledge inside the model's context window; automatically drop lower-weight background chunks if overflow occurs.
(3) Citation anchors: every appended chunk ends with a source tag, e.g. [Team: XYZ_2021 | Section: Education | Anchor: teacher-workshop]
(4) Generation rule: the model must insert the corresponding anchor immediately after any claim drawn from that chunk, ensuring full traceability.
4.1.4 Generation Control & Hallucination Mitigation
(1) Answer template (controlled generation):
Role statement: "You are iGEM Navigator. Answer solely from the provided context."
Structure: restate question → bullet-point evidence → concise conclusion → actionable recommendations → inline citations.
Tone: objective, neutral, no hype.
Safety rule: if evidence is insufficient, state the gap and suggest follow-up queries instead of inventing information.
(2) Post-hoc fact check:
Every key claim is matched back to the retrieved snippets; if no supporting sentence is found, the claim is down-weighted, rewritten, or flagged for additional retrieval.
(3) Out-of-domain guardrail:
Questions unrelated to iGEM (e.g., politics, medical diagnosis) trigger a polite refusal and redirect users to relevant iGEM topics.
4.1.5 Multilingual & Domain Adaptatio
(1) Cross-language retrieval: Queries are expanded in both Chinese and English, ensuring English wikis can answer Chinese questions and vice versa.
(2) Term dictionary:a curated lexicon maps genes, pathways, vectors, instruments, and company names to standard forms, boosting recall accuracy.
(3) Section-aware templates: for Human Practices, Education, and Implementation pages, additional templates trained on social-survey and education-design language emphasise "process/indicator" rather than "bench protocol" style answers.
4.1.6 Typical Prompt Snippets (Examples)
(1) What chassis organisms and related technologies have the BNU-China team used in recent years?
(2) Which teams in 2021 used bacteriophages or employed phage-related technologies? Please briefly describe their approaches.
(3) Could you please introduce the project topic and technical methods used by the BNU-China team in 2023?
4.1.7 Compliance & Ethics
(1) Copyright & Licence: Only publicly accessible iGEM Wiki pages are used; original URLs and team attributions are preserved.
(2) Citation Transparency: Source anchors are mandatory and one-clickable to the original wiki.
(3) Privacy: No personal-identifiable information is processed; passages containing sensitive personal data are excluded from retrieval.
(4) Content Boundaries: High-risk queries (medical, legal, etc.) are either declined or answered with an explicit risk disclaimer.
4.2 Core Technology 2: Large-Model Fine-Tuning
After building the knowledge base and the RAG retrieval pipeline, we carried out full fine-tuning of the base LLM, turning a general-purpose language model into a true "iGEM project innovation & management expert".
4.2.1 Model Selection & Ollama Deployment
Within the Ollama framework we benchmarked leading open-source models (LLaMA, Mistral, etc.) on:
(1) technical-QA accuracy
(2) fidelity to iGEM-wiki writing style
(3) adaptability to Human-Practice scenarios
(4) local inference speed vs. parameter count
The 7 B–13 B-class Qwen-plus (Tongyi Qianwen) delivered the best overall score. At ~7 B parameters it offers ample expressive and reasoning power while still fitting into a single A100 GPU at Beijing Normal University's School of Artificial Intelligence for efficient fine-tuning and local deployment.
4.2.2 Fine-tuning Data Construction
Building on the preliminary work of converting RAG data into knowledge, we performed structured cleaning and annotation to create a task-specific training set that covers:
(1) Technical corpora: e.g., ATP regeneration systems, VLP technology, metabolic engineering, experimental design;
(2) Human-practice corpora: policy communication, regional alliances, educational initiatives, social surveys;
(3) Wiki-writing and case corpora: used to learn iGEM's distinctive expression formats and style.
During fine-tuning we adopted mainstream PEFT techniques such as LoRA (Low-Rank Adaptation) and Prefix-tuning, focusing the training on a small subset of the model's high-level parameters (e.g., low-rank updates to attention weights) without adjusting the full model weights.
Table 1. Performance Comparison of iGEM Navigator Model Before and After PEFT Fine-Tuning.
| Capability Dimension | Original Model | After PEFT Fine-Tuning(iGEM Navigator) |
|---|---|---|
| Wiki-Style Generation | Plain text | Clear structure, aligned with review logic |
| Scenario Adaptation | Generic and unfocused | Precisely addresses technical, educational, and HP (Human Practices) scenarios |
| Training Cost | High | Reduced, supports iterative refinement |
| Knowledge Updating | Requires full retraining | LoRA layers enable fast incremental fine-tuning and updates |
| iGEM Terminology Comprehension | Prone to confusion and over-generalization | Accurately identifies scenarios such as ATP regeneration, VLP, DBTL, etc. |
Click to view the demo video
DIALOGUE
During the implementation of Integrated Human Practice, we maintained an open and collaborative approach, actively engaging with universities and research institutions to gather valuable feedback. These collaborations helped us continuously optimize the project design, ensuring that the project is both scientifically sound and meets societal needs. We participated in the 12th China iGEMers Exchange Conference (CCIC) and exchanged ideas through letters with universities like Wuhan University(WHU-China) and Changchun Jingkai University Hospital (CJUH) - Jilin University (JLU), China to discuss project concepts. We also collaborated with China Agricultural University to discuss and refine the project design. Additionally, we co-hosted a Synthetic Biology Debate with Tsinghua University, guiding students to deeply consider the ethical and social impacts of the field. Furthermore, we worked with multiple universities to create the Crushing the Myths of Sythetic Biology, promoting the public's correct understanding of this cutting-edge discipline.
In terms of international cooperation, we had in-depth discussions with the Earth Charter International (ECI), exploring how to integrate the Sustainable Development Goals (SDGs) and synthetic biology into global education and action. Additionally, we communicated with the Hainan Provincial International Exchange Association, interviewing the association’s director to discuss opportunities for promoting green technology and educational projects in Southeast Asia.
Throughout the process, all our collaborations and interactions reflect the iHP (integrated Human Practice) concept, which involves continuously optimizing our project through feedback and practice to ensure it always aligns with societal and industrial needs. Each dialogue and collaboration has helped us adjust our direction, ensuring that the project brings real value to society and advances the application of synthetic biology. Ultimately, through these collaborations, we have effectively integrated education, technology, and industry, ensuring that our work not only has scientific significance but also generates sustainable social impact.
References
[1]Rice AJ, Sword TT, Chengan K, Mitchell DA, Mouncey NJ, Moore SJ, Bailey CB. Cell-free synthetic biology for natural product biosynthesis and discovery. Chem Soc Rev. 2025 May 6;54(9):4314-4352.
[2]Sun, C., Li, Z., Ning, X., Xu, W., & Li, Z. (2021). In vitro biosynthesis of ATP from adenosine and polyphosphate. Bioresources and Bioprocessing, 8, 117.
[3]Ryabova, L. A., Vinokurov, L. M., Shekhovtsova, E. A., Alakhov, Y. B., & Spirin, A. S. (1995). Acetyl phosphate as an energy source for bacterial cell-free translation systems. Analytical Biochemistry, 226, 184-186.
[4]Zubay, G. (1973). In vitro synthesis of protein in microbial systems. Annual Review of Genetics, 7, 267-287.