AXIS: Agentic Extraction and Integration of Simulation Scenarios for Quantitative Systems Pharmacology

Abstract

Quantitative Systems Pharmacology (QSP) models are executable and mechanistic representations of biological systems that are increasingly used to design, evaluate, and help enrich clinical trials. Reusing a published model, however, requires knowing what each entity represents biologically (often inferable only from its relations to other entities) and what value it must take for a given scenario, information carried in the accompanying publication, not the model file. Agentic systems automate simulation setup from user-specified tasks over standardized parameter sets. In parallel, literature-mining tools extract reported quantities from publications. However, closing the loop, going from paper and model to an executable configuration that reproduces the published result, remains a manual curation task. Here, we present AXIS, a modular multi-agent framework that converts a published QSP model together with its companion publication into simulation-ready configurations. The framework integrates dependency-aware entity grounding, evidence-linked scenario extraction, semantic and structural entity search, sandboxed deterministic computation, and expert input on ambiguity. AXIS comprises three modules that ground the model's entities, recover the scenarios described in the paper, and resolve each into concrete parameter assignments. We demonstrate the framework end-to-end on a published disease model, reconstructing the figure reported in its companion publication. AXIS recovers the reported scenarios, produces parameter assignments matching an expert-curated ground truth, and defers to an expert where a required quantity appears in neither source.