Skip to content

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

🕒 Published (v1): 2025-11-24 18:57 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Genie-CAT is a tool-augmented LLM agent that combines RAG over a metalloprotein literature corpus, PDB structural parsing, Poisson-Boltzmann electrostatics (APBS), and a symmetry-invariant ML redox potential predictor into a single agentic workflow for mechanistic enzyme hypothesis generation. The system uses a ReAct loop to dynamically select and orchestrate these tools in response to natural-language queries. In a proof-of-concept on ferredoxin 1CLF, it reproduces expert-derived residue-level mutation hypotheses in under 3 minutes.

Problem

Protein language models and diffusion-based design frameworks excel at sequence-structure mapping but cannot reliably capture electrostatic and redox mechanisms governing metalloenzyme function. Designing [Fe–S] cluster-containing proteins for specific redox potentials requires integrating literature knowledge, 3D structural context, physics-based electrostatics, and predictive modeling — a multi-modal evidence synthesis that no existing agentic or statistical system performs in a unified, mechanistically interpretable workflow.

Method

Genie-CAT implements a LangGraph ReAct agent with four typed tool modules:

  1. RAG: ~1,600 hydrogenase/metalloenzyme papers chunked into 500-character windows with 100-character overlap, encoded with all-MiniLM-L6-v2 (384-d embeddings), indexed in FAISS. Retrieved chunks are paired with document-level summaries (global + local context) before generation, following a multi-abstraction-level RAG design inspired by PaperQA2.
  2. Structural analysis: MDAnalysis parses PDB files; residues within a tunable spherical cutoff \(R_\text{cut}\) of Fe atoms are classified by physicochemical polarity, coordination, and solvent accessibility.
  3. Electrostatics: A three-stage pipeline — PDB2PQR charge assignment (Amber ff14SB + in-house Fe–S parameters) → APBS Poisson-Boltzmann solve → PyMOL surface visualization. Cross-protein electrostatic comparison tools are also registered as agent tools.
  4. Redox potential prediction: A symmetry-aware MLP (\(f_\theta: \mathbb{R}^{57} \to \mathbb{R}\)) trained on [4Fe–4S] cluster data. The 57-d feature vector encodes permutation-invariant geometric scalars (pairwise distances, bond-vector cosines, triple products across Fe and S subgroups) plus cluster-center electrostatic potential \(Q_i\) and electric field \(C_i\). Trained with AdamW (\(\text{lr}=10^{-3}\), \(\lambda=10^{-4}\), batch 1024), MSE loss, early stopping.

The ReAct loop iterates: thought → tool call → observation → synthesis, with tool selection determined dynamically per query.

Key Contributions

  • Unified agentic harness integrating heterogeneous computational modalities (literature, structure, physics, ML) for metalloenzyme design in a single containerized runtime
  • Multi-abstraction-level RAG design pairing chunk-level retrieval with document-level summaries for grounded hypothesis generation
  • Symmetry-invariant redox potential MLP enforcing \(D_{2d}\) permutation invariance over Fe and S subgroups of [4Fe–4S] clusters
  • Modular tool-registration architecture (LangGraph + Streamlit) allowing new physics engines (QM/MM, MD) to be added without modifying the agent core
  • Demonstrated end-to-end hypothesis generation for ferredoxin 1CLF reproducing expert-derived redox tuning predictions in <3 minutes vs. days–weeks manually

Results

  • RAG with Genie-CAT achieves mean LLM-as-a-judge correctness score 4.38 (95% CI: 4.35–4.42) vs. 4.01 (CI: 3.95–4.07) for GPT-4o-mini without retrieval, over 99 questions across 10 runs; win rate 0.30, tie rate 0.55
  • Ferredoxin 1CLF: predicted redox ordering consistent with known asymmetry — hydrophobic-cluster environment predicted at −425 mV vs. polar-environment cluster at −370 mV; within tens of mV of reported trends for related Fe–S proteins
  • Query latency: RAG/structural queries 1–5 s; ESP computation 120–180 s; redox inference ~20 s; full multi-step workflow <3 min per protein

Limitations

  • RAG corpus coverage limited to ~1,600 hydrogenase-focused publications; corpus gaps or curation biases propagate into hypotheses
  • Continuum electrostatics (APBS) neglects local polarization and quantum effects significant near metal centers
  • Redox predictor trained on a finite [Fe–S] dataset (scope not fully stated in provided text); generalization to out-of-distribution cofactors or mutations is uncertain
  • No wet-lab validation loop; hypotheses are not experimentally confirmed in this work
  • Evaluation is proof-of-concept on a single well-characterized ferredoxin; no systematic benchmark against alternative design methods

Relevance to Harnesses / Meta-Harnesses

Genie-CAT is a domain-specific tool-augmented agent harness: it implements the canonical harness pattern (ReAct loop + typed tool registry + modular backend) but applies it to a narrow vertical (metalloenzyme design) rather than general code or reasoning tasks. The architecture — independent tool modules that register input schemas and return structured results to a central LLM orchestrator — is a direct instantiation of the meta-harness design principle where the agent core is decoupled from domain backends. The explicit treatment of latency heterogeneity (1–5 s vs. 120–180 s tools) and the planned asynchronous HPC job submission (returning a job ID for later retrieval) anticipate patterns relevant to harnesses that coordinate long-running external processes. For researchers building meta-harnesses, Genie-CAT illustrates how tool-registration boundaries, multi-abstraction RAG, and symmetry-constrained ML predictors can be composed inside a standard ReAct scaffold.