Structured Schemas for Provenance-Rich, LLM-Assisted QSP Model Calibration

Area of Science:

• Computational biology
• Pharmacology
• Systems biology

Background:

• Quantitative systems pharmacology (QSP) models rely on literature data for calibration.
• Manual data curation is inconsistent, and large language model (LLM) extraction can lead to errors like hallucinated values and fabricated citations.

Purpose of the Study:

• To introduce MAPLE (Model-Aware Parameterization from Literature Evidence), a novel framework designed to enhance the accuracy and reliability of parameter extraction for QSP models.
• To provide a collaborative interface between LLMs and modelers using structured validation schemas for improved data curation.

Main Methods:

• MAPLE employs two distinct schemas: SubmodelTarget for individual parameter constraints and CalibrationTarget for full model calibration with clinical/in vivo endpoints.
• Structured validation schemas separate data extraction from modeling decisions, ensuring full provenance for every extracted value.
• Automated validators identify and correct characteristic LLM errors by cross-referencing values with source snippets, resolving DOIs, and executing code.

Main Results:

• MAPLE successfully extracted and curated 37 SubmodelTargets and 45 CalibrationTargets for a pancreatic ductal adenocarcinoma QSP model.
• Automated validators triggered 50 retries before human review, ensuring verified citations and direct quotes for all values.
• 11 out of 19 model parameters were supported by multiple independent literature sources, increasing model robustness.

Conclusions:

• MAPLE facilitates the creation of more reliable QSP models by integrating LLM capabilities with essential human oversight and scientific judgment.
• The framework records modeler reasoning for reproducibility and independent verification, crucial for long-term project continuity and knowledge transfer.
• While evaluated on one model, MAPLE offers a promising approach to address challenges in literature-based data extraction for complex biological models.