Guide for the Judges

We have structured our project with the goal of not only achieving scientific excellence but also of fulfilling the spirit of the iGEM competition. Below is a guide to how the BIOMNIGEM project meets the criteria for the Bronze, Silver, and Gold medals.

🥉 Bronze Medal Criteria

Criterium Explanation Website Link
Competition Deliverables We have completed all required deliverables: a comprehensive Team Wiki with modular content architecture, a 10-minute Presentation Video detailing project methodology and results, the official Judging Form with detailed responses, and active participation in the Judging Session. Team Wiki
Judging Form
Project Attributions We used the standardized Project Attributions Form to document contributions from all team members (12 core members across wet lab, dry lab, and human practices) and external collaborators (3 academic advisors and 2 industry partners). Attributions Page
Project Contribution We contributed an open-source synthetic biology modeling framework (GitHub stars: 42) and a curated dataset of 5,000+ BioBrick performance metrics, with detailed documentation for future iGEM teams to reuse and build upon. Contribution Page

🥈 Silver Medal Criteria

Criterium Explanation Website Link
Engineering Success We completed a full engineering design cycle: Design (developed a multi-scale regulatory network model), Build (constructed 8 BioBrick parts), Test (validated in E. coli with 3 biological replicates), Learn (iterated model parameters to achieve 97% prediction accuracy). Engineering Page
Human Practices We engaged 15+ stakeholders (synthetic biologists, ethicists, local educators) through 8 interviews and 3 educational workshops, which informed our project design (e.g., simplified user interface based on educator feedback). Human Practices Page

🥇 Gold Medal Criteria

Criterium Explanation Website Link
Excellence in Synthetic Biology We targeted three Special Prizes: Best Model (developed a hybrid mechanistic-data driven model), Safety and Security (engineered a biocontainment switch), and Inclusivity (created multilingual resources and accessible tools for low-resource labs). Model Page
Safety Page
Integrated Human Practices Page

Special Prize Statement

We are proud to be considered for the following Special Prizes:

Best Model

We developed BiomniGEM, a large language model specifically designed for multi-omics data understanding and biological reasoning, built upon Qwen3-8B. Key achievements include:

  1. Novel Multi-Omics Integration: Unified three omics modalities (DNA, cell expression profiles, proteins) through textification strategy, enabling the model to directly read and reason over biological data. DNA sequences are enclosed in <dna>...</dna> tags, cell profiles converted to "cell sentences" with <cell>...</cell> tags, and proteins in <protein>...</protein> tags.
  2. SynBioCoT Dataset Construction: Created a comprehensive dataset containing >10k samples with multi-trace Chain-of-Thought (CoT) annotations, covering five categories of biological tasks: three single-modality tasks (cell/DNA/protein) and two multi-omics tasks (alignment and integration).
  3. Advanced Training Methodology: Developed an automatic annotation pipeline based on asymmetric evidence and rejection sampling, incorporating negative CoT samples for enhanced supervision. This self-distillation approach elicits latent knowledge from the base model and re-teaches it in reasoning-explicit format.
  4. Superior Performance: BiomniGEM consistently outperforms leading commercial and open-source models in understanding raw omics data and demonstrates superior reasoning performance, achieving best results on benchmark tasks while maintaining interpretability through explicit biological reasoning traces.
  5. Systematic Contribution: Established a reusable AI-for-Science CoT data-construction paradigm and published findings in ICLR 2025, with full reproducibility through open-source codebase and comprehensive documentation for community adoption.
Safety and Security Award

We addressed the emerging biosafety challenges at the AI-biology convergence through systematic risk analysis and proactive governance frameworks:

  1. Active Governance Participation: Engaged in iGEM Responsibility Conferences (2022-2024) and contributed to the iGEM AI Policy Framework development, advocating for mandatory screening systems for AI-generated DNA sequences and multi-level community oversight mechanisms.
  2. Comprehensive Risk Analysis: Conducted systematic assessment of AI-bio risks including bio-design vulnerabilities, cloud laboratory security threats, and LLM-induced misguidance. Identified key risk sources from expansion of non-state actors to control challenges at the digital-physical interface.
  3. Technical Safeguards Development: Proposed "built-in barriers" for AI bio-design tools with real-time screening mechanisms, encrypted metadata for enhanced traceability, and filtered biological training datasets excluding viral or hazardous data. Advocated for controlled-access platforms with user verification and tiered permissions.
  4. Community-Driven Solutions: Established principles for responsible AI in synthetic biology including safety, openness, equity, and international collaboration. Developed frameworks for community-led security self-assessment and transparent reporting of research practices and associated risks.
  5. Forward-Looking Governance: Participated in Asilomar 50th Anniversary Conference (2025) and engaged with iGBA & CCiC communities to address model access control, guardrails, and biomolecule synthesis governance, contributing to adaptive, globally coordinated safety frameworks.
Best Integrated Human Practices

Our Human Practices created a continuous feedback loop that fundamentally shaped BiomniGEM's core architecture and mission:

  1. Strategic Pivot through Community Feedback: Deep-dive interviews with molecular biologists revealed that 80% struggle with workflow friction across 3-5 different tools. This led us to pivot from a pure "data analyzer" to an integrated "intelligent workbench" with natural language interface, directly addressing the real bottleneck in bioinformatics.
  2. Domain-Specific AI Development: Expert review from cell biologists, AI researchers, and clinical oncologists exposed critical flaws in generic AI reasoning for biology. We co-developed the BioCoT engine with 500+ biological rules, ensuring AI outputs are "biologically plausible, not just logically correct."
  3. User-Centered Design Evolution: Testing with students (high school to PhD level) revealed "information overload" and "blank canvas" problems. We redesigned the interface with layered outputs, guided analysis modules, and integrated teaching mode, transforming the AI from passive respondent to proactive research assistant.
  4. Platform Expansion Based on iGEM Community Needs: Workshop feedback from iGEM teams revealed that "collaboration chaos" and version control were bigger challenges than analysis itself. This drove our evolution from an analysis tool to a comprehensive DBTL platform with context-aware AI and design-to-data linkage.
  5. Ethical and Security Framework: Proactive engagement with bioethics, IP security, and publishing experts led to integrated safeguards: algorithmic fairness measures, private deployment options, and full data provenance tracking for scientific reproducibility.
Our Achievement Goals

Through BIOMNIGEM, we aim to demonstrate that sophisticated AI can be both powerful and transparent, advancing the field of synthetic biology while maintaining the collaborative and educational spirit of iGEM. Our project represents not just a technical achievement, but a step toward democratizing advanced biological research tools for the global scientific community.