MAPLE Records the Evidence and Modeling Decisions Behind QSP Calibration
MAPLE structures literature-derived data for quantitative systems pharmacology calibration. A pancreatic-cancer case shows why automatic checks and human judgment have separate roles: 18 monitored extractions required 50 retries, and scientific fields still needed substantial revision.
A calibration value needs its experimental context
Quantitative systems pharmacology (QSP) models often combine parameters and endpoints from many publications. A reported value may be extracted correctly yet remain unsuitable for the model because it comes from another species, disease or experimental condition. Researchers at Johns Hopkins University present MAPLE, Model-Aware Parameterization from Literature Evidence, as a framework for keeping the source evidence and the modeler's decisions together.
The framework uses large language models to find and draft structured records, with automated validation before scientific curation. Its application concerns one pancreatic ductal adenocarcinoma QSP model. The study follows the assembly and checking of calibration evidence for that model.
Two record types separate measured data from its modeled use
A SubmodelTarget record captures an isolated experiment that constrains individual parameters through a simplified forward model. Its input layer preserves the reported value, units, uncertainty, sample size, reference and quoted source passage. A separate calibration layer states how the model will use that measurement, including the forward model, prior distributions and error model.
A CalibrationTarget record captures a clinical or in vivo endpoint that constrains the complete model. It defines the observable calculated from model species, the empirical distribution and any treatment scenario. Shared parameter names allow measurements from several records to constrain the same quantity, rather than leaving their relationship to a later manual mapping step.
Both schemas include a source-relevance assessment. Species, indication and tumor-microenvironment differences require a justification and uncertainty allowance. The framework therefore distinguishes the fact reported by a source from the decision to transfer that fact into a particular model.
Automatic checks catch a defined set of failures
Validators check that internal references resolve, numeric values appear in their attached snippets, units have valid dimensions and generated observable code executes with mock inputs. DOI and title checks can detect a real identifier paired with an incorrect title. Additional checks examine denominators and whether calculated empirical summary statistics match the recorded data.
An external script can also check whether a snippet occurs in the cited paper. In this study that source-text check was applied selectively during curation, because access to full text varied. It was not a universal part of the automatic gate. A value matching its attached snippet therefore has a different verification scope from a snippet independently located in the original paper.
The case study preserves several denominators
The final corpus contained 37 SubmodelTargets and 45 CalibrationTargets. The latter comprised 26 baseline endpoints, 18 neoadjuvant immunotherapy endpoints involving GVAX with or without nivolumab, and 1 progression endpoint. The 37 parameter records mapped to 19 unique parameters; 11 parameters had more than one supporting target.
Detailed retry tracing was available for 18 batch SubmodelTarget extractions. None passed on its first attempt, and the validators triggered 50 retries before human review. These tracing results do not describe all 45 CalibrationTargets: equivalent per-target instrumentation was unavailable for that cohort.
Human curation changed the forward-model type in 65% of the 37 SubmodelTargets, prior parameters in 46% and source relevance in every record. For the 18 surviving batch-extracted CalibrationTargets with paired drafts, observable code changed in 39%. Another 27 CalibrationTargets were extracted interactively, with the modeler supplying decisions during extraction rather than revising a standalone initial draft.
The study reports fewer later changes in the interactive records, but the comparison was not controlled. The modes used different language models and occurred sequentially, after the modeler had learned from batch failures. The authors do not claim a measured time advantage over matched fully manual extraction.
Traceability does not settle evidence suitability
MAPLE v0.1.0 translated the 37 parameter records into a joint Bayesian inference script. All 19 parameters met the reported convergence criteria, although one inferred immune-division quantity remained below its expected range. Convergence therefore describes the inference computation, while biological suitability and identifiability still require examination.
The evidence base itself was mixed: 46% of parameter targets came from pancreatic-cancer-specific experiments, and many required transfer from another indication or species. The framework records those decisions and propagates translation uncertainty instead of making the source mismatch disappear.
This single-group, single-model evaluation characterizes a reproducible evidence workflow. Its practical contribution is an inspectable connection among a published measurement, the experiment that produced it and the assumptions that let it constrain a drug-development model. Broader performance, reduced modeler effort and improved downstream prediction remain separate questions.