Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Beyond Protein Language Models: An Agentic LLM Framework for Mechanistic Enzyme Design

Bruno Jacob, Khushbu Agarwal, Marcel Baer, Peter Rice, Simone Raugei

TL;DR

The paper introduces Genie-CAT, a domain-specific agentic LLM framework that combines retrieval-augmented reasoning, structural analysis, electrostatic calculations via PB/APBS, and a symmetry-aware redox predictor to generate mechanistically grounded hypotheses for metalloprotein design. Using ferredoxins, particularly a 1CLF ferredoxin with two [4Fe–4S] clusters, Genie-CAT demonstrates end-to-end multi-modal reasoning that ties sequence- and environment-level features to predicted redox shifts, achieving rapid hypothesis generation that aligns with expert intuition. The modular architecture supports extension to higher-fidelity simulations and broader cofactor spaces, with the potential to reduce design cycles from days to minutes while increasing interpretability and grounding. Overall, the work showcases how integrating literature grounding, structure-aware analysis, physics-based modeling, and targeted ML predictions in an agentic workflow can transform LLMs into productive partners for computational discovery in protein design.

Abstract

We present Genie-CAT, a tool-augmented large-language-model (LLM) system designed to accelerate scientific hypothesis generation in protein design. Using metalloproteins (e.g., ferredoxins) as a case study, Genie-CAT integrates four capabilities -- literature-grounded reasoning through retrieval-augmented generation (RAG), structural parsing of Protein Data Bank files, electrostatic potential calculations, and machine-learning prediction of redox properties -- into a unified agentic workflow. By coupling natural-language reasoning with data-driven and physics-based computation, the system generates mechanistically interpretable, testable hypotheses linking sequence, structure, and function. In proof-of-concept demonstrations, Genie-CAT autonomously identifies residue-level modifications near [Fe--S] clusters that affect redox tuning, reproducing expert-derived hypotheses in a fraction of the time. The framework highlights how AI agents combining language models with domain-specific tools can bridge symbolic reasoning and numerical simulation, transforming LLMs from conversational assistants into partners for computational discovery.

Beyond Protein Language Models: An Agentic LLM Framework for Mechanistic Enzyme Design

TL;DR

The paper introduces Genie-CAT, a domain-specific agentic LLM framework that combines retrieval-augmented reasoning, structural analysis, electrostatic calculations via PB/APBS, and a symmetry-aware redox predictor to generate mechanistically grounded hypotheses for metalloprotein design. Using ferredoxins, particularly a 1CLF ferredoxin with two [4Fe–4S] clusters, Genie-CAT demonstrates end-to-end multi-modal reasoning that ties sequence- and environment-level features to predicted redox shifts, achieving rapid hypothesis generation that aligns with expert intuition. The modular architecture supports extension to higher-fidelity simulations and broader cofactor spaces, with the potential to reduce design cycles from days to minutes while increasing interpretability and grounding. Overall, the work showcases how integrating literature grounding, structure-aware analysis, physics-based modeling, and targeted ML predictions in an agentic workflow can transform LLMs into productive partners for computational discovery in protein design.

Abstract

We present Genie-CAT, a tool-augmented large-language-model (LLM) system designed to accelerate scientific hypothesis generation in protein design. Using metalloproteins (e.g., ferredoxins) as a case study, Genie-CAT integrates four capabilities -- literature-grounded reasoning through retrieval-augmented generation (RAG), structural parsing of Protein Data Bank files, electrostatic potential calculations, and machine-learning prediction of redox properties -- into a unified agentic workflow. By coupling natural-language reasoning with data-driven and physics-based computation, the system generates mechanistically interpretable, testable hypotheses linking sequence, structure, and function. In proof-of-concept demonstrations, Genie-CAT autonomously identifies residue-level modifications near [Fe--S] clusters that affect redox tuning, reproducing expert-derived hypotheses in a fraction of the time. The framework highlights how AI agents combining language models with domain-specific tools can bridge symbolic reasoning and numerical simulation, transforming LLMs from conversational assistants into partners for computational discovery.

Paper Structure

This paper contains 15 sections, 2 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Protein language models and machine learning for protein design
  4. Physics-based modeling of enzyme electrostatics and redox chemistry
  5. Retrieval-augmented methods and Agentic LLMs
  6. Methods
  7. Literature Ingestion and Indexing
  8. Structural Preprocessing and Residue-Level Analysis
  9. Electrostatic Potential Computation and Visualization
  10. Redox Potential Model for Reasoning about Ligand Mutations
  11. Software Architecture
  12. Agentic Multi-Modal Reasoning on Metalloenzymes: A Ferredoxin as a Test Case
  13. Discussion
  14. Conclusion
  15. Acknowledgements

Figures (4)

  • Figure 1: Overview of Genie-CAT's LLM-guided agentic system workflow: The system identifies the different components in the workflow to be executed based on user query. Inputs (literature, PDB structures) are then processed through RAG, structural analysis, electrostatic calculations, and/or redox modeling. The evidence from different analysis is then summarized to the scientist enabling them to generate mechanistic hypotheses
  • Figure 2: RAG Example query and response with retrieved context.
  • Figure 3: Example PDB structure analysis: Agent analyses PDBs to determine residue compositions, maps each residue to associated polarities and summarizes results
  • Figure 4: Example interactive session with Genie-CAT on ferredoxin 1CLF. The user issues three sequential queries: (Q1) compute electrostatic potentials for 1CLF, (Q2) predict the redox potentials of its two [4Fe–4S] clusters, and (Q3) generate comparative figures and tables, including a structural plot of the protein. For each query, the agent selects and orchestrates the appropriate tools (APBS-based electrostatics, symmetry-aware redox predictor, and plotting utilities), then synthesizes the outputs into natural-language answers and visual artifacts, illustrating end-to-end, multi-modal reasoning within a single workflow.