Generative AI for Sustainable Proteins: New Framework from King’s College London

By benchmarking fine-tuning against structured prompt engineering, the researchers demonstrate notable gains in accuracy (cosine similarity ≥0.94 for fine-tuned models), while highlighting trade-offs in generalisability and computational cost. The modular interface and open-source microservices provide a pathway towards integration with industrial R&D pipelines.

Kalian, A. D., Lee, J., Johannesson, S. P., Otte, L., Hogstrand, C., & Guo, M. (2025). From LLMs to sustainable proteins: Fine-tuning and prompt engineering for multi-agent AI in waste carbon-utilising microbial protein production. Carbon Capture Science & Technology, 100525. https://doi.org/10.1016/j.ccst.2025.100525

Here is a blog-style summary in UK academic English, based on the uploaded article (From LLMs to Sustainable Proteins: Fine-Tuning and Prompt Engineering for Multi-Agent AI in Waste Carbon-Utilising Microbial Protein Production).

• Year and geographical scope
  2025. Research conducted at King’s College London (UK), though the scope is global with applications in protein production and circular bioeconomy.

• Generative AI technology
  Multi-agent AI framework built on GPT-based LLMs. Two agents:
  (1) Literature search agent (GPT-4o) optimising microbial strain queries.
  (2) Information extraction agent (GPT-4.1) tested under fine-tuning vs. prompt engineering.
  Applications: automated retrieval of microbial protein data (protein yield, trophic mechanism, substrates, safety/toxicity screening).

• Research stage
  Proof-of-concept system tested in silico. Includes prototype web interface for interactive use. No industrial pilot yet.

Implementation insight

• Levels and barriers
  Technological: dependence on closed-source GPT models, trade-off between fine-tuning accuracy and generalisability.
  Organisational: need for integration into R&D pipelines, governance of AI-generated insights, domain expertise to curate prompts/training.
  Barriers: fragmented microbial protein literature, risk of hallucinations, computational costs of fine-tuning.

• Organisational conclusions
  Multi-agent AI can significantly accelerate discovery in CCU-integrated protein production by automating literature mining and synthesis. A dual-track optimisation strategy (prompt design plus targeted fine-tuning) is recommended. The modular design (APIs, dashboards, toxicity module) allows integration into existing research and industrial innovation pipelines.

• Economic/financial conclusions
  While no ROI data are presented, the framework promises:

  • Reduced costs of manual literature review and trial-and-error strain selection.

  • Potential acceleration of strain optimisation, lowering time-to-market for sustainable protein.

  • Future industrial use could reduce R&D overheads and improve process economics, though scaling costs and data governance remain challenges.