GRASP: Graph Reasoning Agents for Systems Pharmacology with Human-in-the-Loop

TL;DR

GRASP tackles the bottleneck of Quantitative Systems Pharmacology model construction by integrating a graph-structured knowledge representation with four specialized agents and a human-in-the-loop interface. It operates in two phases—Understanding and Action—preserving mass balance, units, and physiological constraints through constraint-aware graph manipulation and BFS-based parameter alignment. Across LLM-based evaluations, GRASP outperforms SME-guided baselines in biological plausibility, mathematical correctness, structural fidelity, and code quality, while enabling efficient, iterative, natural-language-driven model modification. A progressive case study demonstrates GRASP handling increasingly complex pharmacological mechanisms (TMDD, multi-receptor interactions, cooperative binding) without compromising constraint satisfaction, highlighting its potential to democratize high-fidelity QSP modeling.

Abstract

Quantitative Systems Pharmacology (QSP) modeling is essential for drug development but it requires significant time investment that limits the throughput of domain experts. We present \textbf{GRASP} -- a multi-agent, graph-reasoning framework with a human-in-the-loop conversational interface -- that encodes QSP models as typed biological knowledge graphs and compiles them to executable MATLAB/SimBiology code while preserving units, mass balance, and physiological constraints. A two-phase workflow -- \textsc{Understanding} (graph reconstruction of legacy code) and \textsc{Action} (constraint-checked, language-driven modification) -- is orchestrated by a state machine with iterative validation. GRASP performs breadth-first parameter-alignment around new entities to surface dependent quantities and propose biologically plausible defaults, and it runs automatic execution/diagnostics until convergence. In head-to-head evaluations using LLM-as-judge, GRASP outperforms SME-guided CoT and ToT baselines across biological plausibility, mathematical correctness, structural fidelity, and code quality (\(\approx\)9--10/10 vs.\ 5--7/10). BFS alignment achieves F1 = 0.95 for dependency discovery, units, and range. These results demonstrate that graph-structured, agentic workflows can make QSP model development both accessible and rigorous, enabling domain experts to specify mechanisms in natural language without sacrificing biomedical fidelity.

Figures (7)

• Figure 1: GRASP Multi-Agent System Architecture and Workflow. (A) Understanding: Multi-agent QSP model understanding demonstrating iterative model understanding and reproduction, where agents collaboratively extract knowledge graphs from original QSP code, generate equivalent MATLAB code, and validate with feedback loops until convergence. (B) Action: illustrating interactive modification workflow, where natural language user prompts are processed to update knowledge graphs, regenerate MATLAB code, and perform automated debugging cycles until successful execution, with versioned output management for traceability.
• Figure 2: GRASP vs CoT and ToT with SME guided prompts with LLM as a judge.
• Figure S1: Initial two-compartment pharmacokinetic model serving as the baseline for progressive modifications. This represents the starting point before any conversational modifications, showing basic drug distribution between central (V_c) and peripheral (V_p) compartments with linear elimination kinetics.
• Figure S2: Model evolution after Prompt 1 (C.2): Addition of full TMDD system with R1 receptor. The figure demonstrates GRASP's capability to integrate target-mediated drug disposition mechanisms including receptor binding, internalization, degradation, and soluble receptor shedding processes while preserving the original pharmacokinetic structure.
• Figure S3: Model expansion after Prompt 2 (C.3): Integration of dual-target TMDD system with R2 receptor. This figure illustrates the system's ability to handle complex multi-target pharmacology with independent receptor dynamics, competitive drug binding, and parallel TMDD pathways while maintaining mathematical consistency across all biological processes.