# LLMs for Protein Engineering at Cambrium ## Company Overview Cambrium is a biotech startup that combines AI and synthetic biology to create sustainable materials. Founded in 2020, they've achieved the remarkable feat of bringing a biotech product to market in less than two years - a record in the EU biotech space. Their first product is Novacoll, a vegan human collagen for cosmetics, created through genetic engineering and AI-driven protein design. ## Technical Infrastructure ### Protein Programming Language - Developed a novel protein programming language for designing proteins - Language transforms protein design into a mathematical optimization problem - Enables systematic exploration of protein sequence space (approximately 10^18 possibilities) - Allows specification of constraints like: ### LLM Integration - Built a Langchain-based research assistant for scientists - Features include: ### Data Infrastructure Challenges - High cost of experiments makes data extremely valuable - Critical focus on preventing data loss in pipelines - Two main data source types: - Key challenge: digitizing scientific processes ### Production System Architecture - Focus on reliability over latency - Main users are internal scientists rather than external customers - Less emphasis on high QPS or 24/7 availability - Strong emphasis on data integrity and experiment tracking - Integration with DNA synthesis APIs for automated ordering ## MLOps Practices ### Data Management - Strict protocols for experimental data collection - Integration of robotics to ensure data accuracy - Focus on complete data capture due to high cost per datapoint - Structured approach to managing both machine and human-generated data ### Model Development - Evolution from traditional protein engineering to LLM-based approaches - Current use of diffusion models for protein generation - Integration of protein folding predictions - Optimization for multiple constraints simultaneously ### Testing and Validation - High stakes for model predictions due to expensive validation - Focus on pre-experimental validation to reduce costs - Multiple checkpoints for verification - Comprehensive tracking of experimental outcomes ## Unique LLMOps Considerations ### Cost Structure - Emphasis on prediction accuracy over speed - High cost of experimental validation - Need to maximize value from each experiment - Balance between computational and wet-lab costs ### Regulatory Environment - Consideration of GMO regulations in different markets - Impact on product development strategy - Influence on technology stack choices ### Scientific Integration - Bridging gap between traditional lab work and AI - Training scientists in data-driven methodologies - Building systems that support scientific workflow ## Future Directions ### Technology Development - Expanding use of generative AI for protein design - Development of more sophisticated protein language models - Integration of new simulation capabilities ### Product Evolution - Starting with high-margin cosmetics market - Planning expansion into industrial materials - Long-term vision for sustainable alternatives to plastics - Exploration of specialized applications like space-grade adhesives ## Impact and Results - Successfully launched Novacoll in under two years - Demonstrated viability of AI-driven protein design - Created scalable platform for sustainable materials - Established new paradigm for biotech product development ## Challenges and Lessons Learned - Importance of data quality in scientific experiments - Need for balance between automation and human expertise - Value of domain-specific languages for complex problems - Critical role of proper data infrastructure - Importance of choosing appropriate market entry points