Arno Strouwen 1,2
1 PumasAI (Dover, USA), 2 KULeuven BIOSYST (Leuven, Belgium)
Introduction: Establishing a robust in vitro-in vivo correlation (IVIVC) remains central to formulation development, clinically relevant dissolution specification setting, and post-approval change management. Level A IVIVC, the most informative correlation level, establishes a point-to-point relationship between in vitro dissolution and in vivo absorption, typically through deconvolution of plasma concentration-time profiles followed by regression of cumulative absorbed fraction against cumulative dissolved fraction. In practice, implementing a complete Level A workflow requires careful handling of asynchronous in vitro and in vivo sampling schedules, selection of unit impulse response (UIR) models, and external validation on held-out formulations. Despite the regulatory importance of IVIVC [1,2], few software tools offer a scriptable, end-to-end pipeline that is both flexible and easy to integrate into existing data analysis workflows.
Objectives: To present InVitroInVivoCorrelation.jl, a commercial Julia package that implements a modular, scriptable Level A IVIVC workflow, and to evaluate its predictive performance on qualification datasets.
Methods: The package implements a five-step workflow: (1) build a dissolution source from observed in vitro data using interpolation or user-provided dissolution models; (2) fit UIR functions from a fast-release reference formulation (intravenous or immediate-release data); (3) deconvolve oral concentration-time profiles via numerical deconvolution to estimate in vivo absorbed fractions; (4) fit the Level A correlation model relating cumulative dissolution to cumulative absorption with configurable timepoint strategies and optional held-out validation formulations; and (5) predict absorption and plasma concentrations via convolution for new formulations and compute validation diagnostics including percent prediction error (%PE) for AUC and Cmax. The design is modular: the default Level A model can be replaced with user-defined correlation functions without altering the surrounding workflow, and dissolution input accepts either interpolated measurements or custom models. The package is DataFrames-native throughout, supporting direct ingestion of long-format tabular inputs and returning standardized tabular outputs for reporting and visualization. Timepoint handling supports observed, uniform-grid, or user-defined strategies, enabling pooled fitting when in vitro and in vivo sampling schedules differ across formulations. Qualification was performed by fitting correlation models to subsets of formulations and computing external prediction errors on held-out formulations under both uniform and ragged (asynchronous) sampling conditions.
Results: Under uniform sampling, the software recovered expected dissolution-to-absorption timing relationships and produced stable UIR estimates from the fast-release reference formulation step, supporting reliable downstream deconvolution and prediction. External prediction accuracy for held-out formulations was strong: absolute %PE values for formulation B were below 2.0% for both AUC and Cmax, and absolute %PE values for formulation D were below 2.5% for both metrics. These results satisfy the acceptance criteria outlined in FDA guidance [1] (absolute %PE below 10% for individual formulations). Under ragged sampling with asynchronous in vitro and in vivo timepoints, practical predictive utility was retained with graceful degradation: absolute errors remained below 9.0% for formulation B and below 10.0% for formulation D for both AUC and Cmax, indicating robustness to realistic nonuniform sampling rather than catastrophic loss of predictive performance. To illustrate the simulation capabilities of the workflow, a sensitivity analysis was performed by systematically varying the dissolution profile shape and propagating each variant through the fitted IVIVC model via convolution. The resulting predicted AUC and Cmax values responded smoothly to changes in dissolution rate, confirming that the pipeline can be used to explore formulation design space in silico.
Conclusions: InVitroInVivoCorrelation.jl provides a reproducible, scriptable, and extensible Level A IVIVC workflow suitable for research and pre-regulatory model development in industrial settings. The package combines deconvolution-convolution workflows, flexible Level A fitting, and validation-first prediction utilities in a single Julia-native framework, while allowing customization of both dissolution representation and correlation model structure. The scriptable design also supports simulation-based applications such as dissolution sensitivity analysis, enabling formulation teams to explore design space in silico before committing to new batches. Its DataFrames-based design integrates directly into existing tabular analysis pipelines, lowering the barrier to adopting scriptable IVIVC in routine practice.
References:
— References window —
References:
[1] Food and Drug Administration. Guidance for Industry: Extended Release Oral Dosage Forms: Development, Evaluation, and Application of In Vitro/In Vivo Correlations. 1997.
[2] Cardot JM, Beyssac E, Alric M. In vitro-in vivo correlation: importance of dissolution in IVIVC. Dissolution Technologies. 2007;14(1):15-19.
Reference: PAGE 34 (2026) Abstr 11907 [www.page-meeting.org/?abstract=11907]
Poster: Methodology - Other topics