Agentic LLM Framework, Genie-CAT, Accelerates Mechanistic Enzyme Design Via Literature and Physics-Based Computation

The design of novel enzymes currently relies on painstaking cycles of experimentation and computational modelling, but a new framework promises to accelerate this process by combining the power of artificial intelligence with established scientific methods. Bruno Jacob, Khushbu Agarwal, and Marcel Baer, alongside Peter Rice and Simone Raugei, all from Pacific Northwest National Laboratory, present a system called Genie-CAT that functions as an intelligent agent for enzyme design. This innovative approach integrates literature analysis, structural modelling, and physics-based calculations into a unified workflow, allowing the system to generate and test hypotheses linking an enzyme’s sequence, structure, and function. Demonstrations using metalloproteins reveal that Genie-CAT can autonomously identify key modifications influencing enzyme behaviour, replicating expert-derived insights significantly faster, and establishing a new paradigm for AI-driven computational discovery.

AI System Grounds Knowledge in Scientific Literature

This research details the development of Genie, an AI system designed to accelerate scientific knowledge discovery by combining large language models with a robust retrieval system. Genie overcomes limitations of standard large language models, such as generating inaccurate information or relying on outdated knowledge, by grounding its responses in a vast and constantly updated corpus of scientific literature. The system employs a retrieval-augmented generation approach, accessing and analyzing scientific PDFs to provide informed answers to complex questions. A sophisticated retrieval system identifies relevant scientific passages based on a user’s query, focusing on directly answering the question rather than simply finding related topics, and utilizes multiple strategies to improve performance.

Genie then uses a powerful large language model to generate answers based on these retrieved passages, ensuring that all responses are supported by cited sources. Rigorous evaluation demonstrates that Genie significantly improves accuracy and reliability compared to standard large language models, while also enhancing the quality of citations. The system’s architecture is designed to be scalable and adaptable to different scientific domains, offering the potential to automate aspects of scientific knowledge discovery by identifying connections and insights from the vast scientific literature.

Tool-Augmented LLM for Metalloprotein Hypothesis Generation

The research team engineered Genie-CAT, a novel system designed to accelerate scientific hypothesis generation in protein design, specifically focusing on metalloproteins like ferredoxins. This work pioneers a tool-augmented large language model system that integrates literature review, structural analysis, and computational modeling into a unified workflow. The system begins by ingesting and indexing scientific literature, employing a retrieval-augmented generation approach that utilizes paper-level summaries to provide broader contextual understanding for the language model. Central to Genie-CAT’s functionality is the structural parsing of protein structures from the Protein Data Bank, allowing the system to analyze the three-dimensional environments surrounding key catalytic sites within proteins.

This structural analysis is coupled with physics-based electrostatic potential calculations, which determine charge distributions around metal clusters and predict how these distributions influence catalytic efficiency and selectivity. Furthermore, Genie-CAT incorporates machine learning prediction of redox properties, allowing the system to estimate how modifications to the protein sequence will affect its electrochemical behavior. By systematically combining these diverse data types, Genie-CAT delivers mechanistically interpretable hypotheses, positioning human experts to make informed design decisions based on comprehensive computational evidence.

AI Hypothesizes Protein Redox Tuning Changes

Genie-CAT represents a breakthrough in scientific hypothesis generation, specifically designed to accelerate protein design research. The system autonomously identifies residue-level modifications near iron-sulfur clusters that affect redox tuning, reproducing expert-derived hypotheses with significantly reduced time investment. This achievement stems from a unified workflow combining literature-grounded reasoning, structural parsing of protein data, electrostatic potential calculations, and machine-learning prediction of redox properties. Evaluations reveal that Genie-CAT consistently outperforms a standard large language model without retrieval augmentation, demonstrating a clear advantage from incorporating retrieved context.

The system utilizes a corpus of approximately 1600 publications on hydrogenases and related metaloenzymes, segmented and analyzed using advanced embedding techniques to enable context-aware retrieval. Genie-CAT’s structural analysis capabilities involve parsing protein structures from the Protein Data Bank, identifying key atoms, and computing distances to surrounding residues. The system then assigns physicochemical classes to each residue and generates summary statistics and visualizations to aid in analysis. Electrostatic calculations are performed using advanced computational methods to model charge distributions and calculate electrostatic potentials on molecular surfaces, providing detailed insights into the protein’s behavior.

Genie-CAT Accelerates Metalloprotein Hypothesis Generation

The research team developed Genie-CAT, an agent-based system that accelerates scientific hypothesis generation in protein design, specifically focusing on metalloproteins. This system uniquely integrates several capabilities, including literature review, structural analysis of protein data, calculations of electrostatic potential, and machine learning prediction of redox properties, into a unified workflow. By combining reasoning with data and physics-based computation, Genie-CAT generates testable hypotheses linking a protein’s sequence, structure, and function, successfully reproducing expert-derived insights at a significantly faster rate. The results demonstrate the need for specialized scientific agents in mechanistic protein design, as general-purpose language models lack access to the detailed structural, electronic, and thermodynamic evidence required for reliable biochemical hypotheses. Genie-CAT addresses this limitation by grounding its reasoning in quantifiable physical descriptors, such as electrostatic fields and redox features, thereby improving the mechanistic interpretability of its predictions and reducing inaccuracies. Future work will focus on expanding the system’s knowledge base, incorporating more advanced physics-based calculations, and integrating experimental feedback to further enhance its reliability and impact on protein design.