Integrated Human Practices

Mentors Huo Yixin, Guo Shuyuan, Shao Bin, and Chen Zhenya provided initial direction for the project, emphasizing the importance of innovation and interdisciplinary collaboration. They initially identified proteins such as CRISPR-Cas13a as optimization cases and recommended focusing on the software and AI track.

Regarding direction selection, Dean Huo advised team members to consider innovation and application value when choosing topics, avoid blindly following popular fields, and focus on differentiated competition to enhance the project's competitiveness. Innovation is crucial for project success and future development, and the team should actively pursue innovation in protein design.

Dean Huo required the team to advance the project according to the plan, ensure timely completion of tasks in each phase, and adjust and optimize the project direction as needed. In terms of team member development and training, emphasis should be placed on fostering leadership, communication skills, and professional expertise, with the goal of promoting personal growth through project practice.

He also highlighted the promising prospects of the intersection of computer science and biology, encouraging members to accumulate experience through the project and enhance their competitiveness in the future job market.

Regarding team collaboration and communication, mentors clarified the division of labor and responsibilities among groups, requiring enhanced collaboration between group members to advance the project collectively. They also encouraged active communication among team members to share ideas and suggestions, and resolve problems in a timely manner.

Finally, mentors requested team members to regularly summarize work progress, promptly report problems and difficulties for adjustment and optimization, and acknowledged and encouraged the team members' performance, expressing hope that the team would achieve excellent results in competitions and bring honor to the university and the team.

Mastering the Overall Competition and Core Elements: The sharing systematically outlined the full process of the iGEM competition, emphasizing that the core competitiveness of a project lies in the deep integration of innovation, experimental rigor, and social value.

Optimizing Team Collaboration Models: Valuable advice was obtained on interdisciplinary team management, particularly regarding the importance of establishing an efficient collaborative mechanism between biological experiments and computational modeling.

Forward-looking Project Planning: It was clarified that Human Practices should run through the entire project rather than being added later, and that detailed documentation is critical for project demonstration and knowledge transfer.

Mentors pointed out that limiting the project to a single protein (e.g., Cas13a) might reduce initial difficulty but would greatly restrict the project's innovation and impact. They encouraged us to elevate the project's perspective, abstract the core technology, and build a universal AI-driven design platform independent of specific proteins. This aligns more with the positioning of a "universal software" and better leverages our team's strengths in computing and modeling.

Value of Hardware-Software Integration: The peer team demonstrated their mature automated hardware platform, which can automatically perform tests and upload results in real time, enabling seamless connection between "hardware execution" and "software analysis." They emphasized that this integration can greatly improve the efficiency and reproducibility of biological research.

Insights into iGEM Hardware Evaluation: They shared valuable competition experience: iGEM judges evaluate hardware projects not only based on technical maturity but also on the degree of integration with core biological problems and the ability to actually solve biological bottlenecks—rather than mere technical accumulation.

Importance of Project Implementation: Based on feedback from their participation last year, they pointed out that the project needs to demonstrate a clear and feasible implementation path, which is crucial for enhancing the project's credibility and impact.

Solidifying Basic Experimental Principles: When answering the high school students' questions about basic experiments such as PCR and primer design, we indirectly reinforced our team's understanding of core molecular biology principles.

Recognizing Experimental Complexity: The detailed challenges encountered by the high school team in experimental operations reminded us that wet experiments are full of uncertainties and technical barriers. AI model design must take into account feasibility and fault tolerance in real laboratory environments.

Clarifying Core HP Concepts: The lecture clarified that the core of Human Practices in iGEM is to "study how your work affects the world and how the world affects your work." Its key lies in fostering "ideological exchange" and "mutual shaping" between the project and society, rather than one-way education or promotion.

Mastering the HP Cycle Method: We learned the standard Human Practices Cycle—through continuous interaction, reflection, and integration, translating external feedback into specific actions for project improvement. This provided a methodological framework for us to systematically conduct HP work.

Distinguishing HP from Outreach: The lecture clarified a common misconception: not all external activities qualify as HP. Only interactions that actually influence the project's purpose, design, or implementation count as Integrated Human Practices. This helped us plan and document subsequent HP activities more accurately.

Understanding Global Governance Frameworks: Gained in-depth insights into cutting-edge trends and challenges in global biosecurity governance, and recognized the new biosecurity risks posed by AI in protein design (e.g., generation and abuse of harmful sequences).

Learning Domestic Regulatory Policies: Professor Xue Yang systematically interpreted domestic laws and regulations related to human genetic resources and biotechnology R&D, emphasizing the importance of conducting compliance assessments in the early stages of the project.

Acquiring Cutting-edge Mitigation Strategies: The lecture introduced technical and policy tools for addressing AI biosecurity risks, such as "unlearning" technology, digital watermarking, and autonomous AI agent monitoring, providing concrete ideas for building a responsible technology platform.

Understanding Tool Integration Paths: The team is committed to integrating existing protein design tools to build an all-in-one software platform, significantly improving the efficiency of the design process—representing a practical and efficient technical path in the field.

Clarifying Our Innovative Positioning: Through comparison, we clearly identified that PROTEUS's core advantage lies in exploring innovative protein design paradigms (i.e., using language models for targeted modification) rather than tool integration.

Identifying Key Validation Bottlenecks: The exchange revealed a common challenge faced by innovative design paradigms—lack of efficient and reliable validation methods, which directly affects the credibility of design results.

Recognizing the Core Role of Data Closed Loops: Industry experts unanimously emphasized that the implementation of AI in life sciences relies on a "data-driven" closed loop—wet experiment validation data must be fed back to the model to achieve precise design and continuous optimization.

Addressing Our Own Data Bottlenecks: We realized that our team's current wet experiment capabilities are the main bottleneck in building this closed loop, making it difficult to generate the high-throughput data required for model iteration.

Recognizing AI's Dual-Edged Sword Effect: The meeting clearly stated that unconstrained powerful AI generation capabilities may be abused to design harmful biological agents, posing serious biosecurity risks.

Learning Cutting-edge Mitigation Technologies: The meeting systematically introduced three collaborative mitigation strategies: "unlearning" technology (to remove dangerous knowledge at the source), "digital watermarking" technology (for output tracking), and "autonomous AI agents" (for real-time risk monitoring).

Nanjing University Team: Develops language model-based gene mining tools, aiming to break free from sequence homology limitations and integrate high-dimensional information for function determination.

Xi'an Jiaotong-Liverpool University Team: Focuses on building multi-modal large models for biology, using language models as backbone networks to uniformly process multi-modal data.

Jilin University Team: Specializes in antimicrobial peptide databases and design, using large models to screen antimicrobial peptides with specific functions from massive data.

Tongji University Team: Focuses on enzyme activity prediction and modification, using pre-trained models to obtain sequence embeddings, then training simple regression models to quickly predict enzyme property parameters.

Identifying Common Challenges: All teams face the core bottleneck of "experimental validation."

Clarifying Our Own Positioning: We gained a clearer understanding of the unique value of PROTEUS as a "universal protein optimization platform"—not limited to specific proteins or functions, but providing a universal AI-driven design solution.

Integration of AI and Physical Models: Proposed a "hierarchical coupling" paradigm—AI for rapid initial screening and physical models (e.g., molecular dynamics, free energy perturbation) for detailed evaluation.

Improvement of Evaluation Systems: Pointed out that evaluating protein variants requires introducing more physicochemical properties (e.g., entropic effects, interface flexibility) rather than just macroscopic properties.

Computational Validation: Recommended embedding computational validation modules (e.g., rapid energy optimization, short-term MD simulations) into the platform to ensure the physical rationality of results.

Importance of Data: Emphasized that high-quality data is the foundation of the model, reminding us to invest significant effort in this area.

Obtaining Comprehensive Technical Roadmap Confirmation: The mentor group fully affirmed the core technical roadmap, including fine-tuning of protein language models based on ESM-2, AlphaFold structural validation, and the dual closed loop of dry and wet experiments.

Clarifying the Platform-oriented Development Path: Mentors emphasized that we should always adhere to the positioning of a "universal AI-driven protein optimization platform."

Resolving Key Bottleneck Issues: For technical bottlenecks mentioned in the report, the mentor group provided specific resource allocation suggestions and technical solutions to help the team overcome obstacles.

Integration of Language Models and Biological Sequences: We gained a deeper understanding of the similarities between language model architectures and biological sequences, enabling effective processing by Transformer-based models.

Application of Pre-training and Fine-tuning Strategies: Learning universal features through large-scale pre-training and then fine-tuning for specific targets can significantly improve model performance.

Importance of Data Processing: High-quality data processing and parameter adjustment are key to improving model performance.

Critical Role of Experimental Validation: Experimental validation is the gold standard for evaluating model performance.

Future Directions: We will explore developing the platform into an NLP-enabled project, allowing users to describe functional requirements in natural language.

Experts such as Researcher Yu Yang systematically elaborated on the new ethical and security challenges posed by the application of AI in synthetic biology.

Learned about governance frameworks such as the Tianjin Guidelines for Scientists' Biosecurity Code of Conduct.

In the youth roundtable, we exchanged ideas with representatives from other teams on project inspiration and ethical considerations. The public generally held an "open but prudent" attitude toward the technology.

Model Selection: Strongly recommended fine-tuning based on universal large models rather than developing small models to avoid falling into local optima.

Experimental Closed Loops: Re-emphasized that high-throughput experimental data is the prerequisite for building an effective reinforcement learning feedback loop.

Evaluation Dimensions: Proposed expanding the protein function evaluation system to include more dimensions such as subcellular localization and protein-protein interaction properties.

Future Directions: Pointed out that RNA and RNA-protein complex design will be the next disruptive direction.

Gaining Inspiring Public Perceptions: Students used metaphors such as "Life's Lego" to describe synthetic biology and "super navigation" to describe the role of AI in scientific research. These vivid analogies provided valuable inspiration for explaining complex technologies to the public.

Insights into the Technical Ethics of the Younger Generation: The students demonstrated admirable prudence, clearly expressing concerns about synthetic biology "opening Pandora's box" and warnings about AI leading to the degradation of human scientists' thinking.

Clarifying the Value and Direction of Science Popularization: The students' strong interest in solving real-world problems confirmed that linking technical value to concrete social needs is an effective way to stimulate public interest and understanding.

Learning Professional Multi-task Model Architectures: The XJTLU team focuses on antimicrobial peptide design, with their model adopting a refined "pre-training + multi-task sequential fine-tuning" strategy.

Learning from Comprehensive Evaluation Indicator Systems: The multi-dimensional evaluation system they built (covering activity, toxicity, and stability) aligns closely with our goal of building a more comprehensive protein optimization evaluation system.

Experiencing Best Practices in Open-Source and Collaboration: The team made all their models, code, and datasets fully public, setting an example for our team's open-source practices.

Clarifying the Feasibility of Pure Dry Experiment Projects: As a pure dry experiment team, their achievements enhanced our confidence in focusing on building powerful software tools with limited resources.

• Validation of Core Direction and Clarification of Capability Boundaries: Researcher Xu affirmed the value of the technical approach centered on ESM2 for protein optimization. He precisely indicated that protein language models are more suitable for the modification and optimization of natural proteins rather than purely de novo design. This provides a solid theoretical foundation for the platform positioning of PROTEUS.
• Identification of Data Bottlenecks and Optimization Priorities: It was explicitly pointed out that the limited fine-tuning dataset is a key factor affecting model performance. Furthermore, he emphasized that the strict separation of the development set and the test set to minimize sequence homology during model training is paramount for preventing overfitting and enhancing the model's generalization capability.
• Feasible Pathways for Establishing an Experimental Closed Loop: Addressing our bottleneck of low-throughput wet lab validation, he proposed highly constructive solutions: outsourcing gene synthesis to specialized companies and actively exploring the linkage of protein function with high-throughput screening methods (e.g., association with bacterial growth). This pointed out the direction for translating AI designs into experimental data on a larger scale.
• Enhancement of Project Narrative Depth: It was suggested that a clear scientific hypothesis should be formulated to explain why our fine-tuning strategy enhances the model's capability for specific tasks. This prompted deeper reflection and refinement of the project's underlying logical framework.

Stimulating Interest with Vivid Metaphors: We used simple metaphors such as "Life's Lego" to explain to children how synthetic biology designs and constructs biological systems, sparking their strong interest and imaginative ideas.

Collecting Pure Feedback: The children's imaginative questions and intuitive understandings (e.g., "Can this AI design glowing Pokémon?") made us realize the importance of maintaining pure curiosity and imagination in technology communication.

Understanding the Starting Point of Public Perception: Through conversations with parents, we learned that the public's understanding of AI and synthetic biology mostly comes from news and film/television works, which include both expectations and concerns.

Customized Science Popularization Content: We carefully prepared 3D models of animal cells, plant cells, E. coli, and DNA. Through interaction with the children, we transformed complex scientific concepts into intuitive and interesting experiences using stories and games.

Conveying Warmth and Care: Beyond science popularization, we focused on interacting with and accompanying the children. By giving carefully prepared science-themed gifts, we brought them knowledge and joy.

Practicing Inclusive Education: We recognized that the light of science should shine on every corner of society. This activity prompted us to reflect on the inclusiveness and accessibility of technological development.

Establishing a Stable Science Popularization Platform: Centering on the core theme of "AI-driven life sciences," we created and published a series of popular science content, attracting the attention of many science enthusiasts.

Obtaining Diverse Real-Time Feedback: Social media platforms became the most direct "public opinion litmus test" for our project. Questions, praise, and even criticism from netizens provided us with valuable first-hand feedback.

Achieving Mutual Inspiration and Joint Growth: Interaction with netizens is no longer one-way output but two-way inspiration. Technical suggestions and application ideas from netizens provided us with new perspectives beyond the team's thinking framework.

This ongoing public dialogue constantly reminds us of the ultimate goal of PROTEUS as an open-source project—not only to become a technical tool but also to serve as a bridge connecting science and society.