Structured Schemas for LLM-Modeler Collaboration in Quantitative Systems Pharmacology Model Calibration

Quantitative systems pharmacology (QSP) models require calibration data from published literature, yet manual curation produces inconsistent documentation while large language model (LLM) extraction exhibits hallucination and fabrication errors unacceptable for quantitative modeling. We present MAPLE (Model-Aware Parameterization from Literature Evidence), a framework that uses structured validation schemas as a collaboration interface between LLMs and modelers. Two complementary schemas capture calibration data at different scales: one for isolated experiments that constrain individual parameters through simplified forward models, and one for clinical and in vivo endpoints that constrain the full model through species-level observables. Both schemas separate data extraction from modeling decisions, capturing literature values with full provenance in a machine-verifiable form. Targeted validators catch characteristic LLM errors: value-in-snippet matching detects hallucinated values, DOI resolution flags fabricated citations, and code execution catches malformed forward models. We evaluate MAPLE on 87 calibration targets for a pancreatic ductal adenocarcinoma (PDAC) QSP model, using two collaboration modes: batch LLM extraction followed by interactive curation, and interactive extraction where modeler and LLM collaborate in real time. Both modes required substantial modeler input: the modeler changed forward model types in 65% of SubmodelTargets, adjusted prior parameters in 46%, and revised source relevance assessments in all files. Interactively extracted targets embedded modeler effort in the extraction process, producing near-final output. The schemas ensure completeness and enable reproducible, provenance-rich calibration regardless of workflow.