DDI Manager: a Dynamic Model-Informed Precision Dosing Tool for Managing Tacrolimus-Voriconazole and Azole Drug-Drug Interactions
Berrah R, Woillard JB, Labriffe M, Baklouti S, Gandia P, Couzi L, Bouchet S, Marquet P, Monchaud C.
Abstract
Tacrolimus has a narrow therapeutic index and interacts strongly with azole antifungals, yet conventional dose‑adjustment rules provide only steady‑state estimates and cannot capture the time course of the interaction, including its onset and the delayed recovery of enzyme activity after the azole is withdrawn. We developed DDI Manager, a model‑informed precision dosing web application combining a tacrolimus population pharmacokinetic model, mechanistic interaction equations, and maximum a posteriori Bayesian estimation to forecast tacrolimus exposure over time.
We report the development and external clinical validation of the tacrolimus–voriconazole module. Four candidate interaction structures were evaluated against retrospective and prospective multicenter cohorts; the retained module combines time‑dependent CYP3A4 inactivation with competitive CYP3A5 inhibition, with apparent in‑vivo constants calibrated so that the assembled model reproduces the exposure ratio of the prescribing information (predicted AUC ratio 3.20 versus 3.2). Performance was resolved jointly by interaction phase (baseline, wash‑in, plateau, withdrawal) and by the information available at the time of prediction, from no concentration at all to every earlier concentration. Phase boundaries were read off the model rather than chosen. Each cell was compared, on the same observations and with the same information, against three sourced static rules, a no‑interaction negative control, and the last measured trough rescaled for dose.
The evaluation set comprised 394 trough concentrations from 30 transplant recipients in three cohorts. The benefit of dynamic forecasting was real, statistically supported, and confined to the wash‑in period: 80.6% of predictions fell within 2‑fold and the model beat all five comparators with confidence intervals excluding zero (+25.0 points versus the no‑interaction control, +30.6 versus the ×3.03 rule of the prescribing information, +50.0 versus the ×4.40 observed ratio, +58.3 versus the ×9.24 DDI‑Predictor prediction, and +38.9 versus the last measured trough). At plateau it held no advantage (0.0 points versus the ×3.03 rule, 95% CI -20.6 to +16.5) and over‑predicted exposure about 1.7‑fold; after withdrawal, within the 14‑day re‑escalation window, the adaptive mode reached 66.7% versus 33.3% within 2‑fold. Neither baseline nor sequential Bayesian updating corrected the plateau bias, identifying it as structural rather than a failure of individualization. Pooling phases averages a period where the model genuinely wins with one where it over‑predicts, and returns a near‑unity ratio by cancellation.
Given the narrow therapeutic index of tacrolimus, the tool is intended to support, alongside close therapeutic drug monitoring, anticipatory individualized dose adjustment during complex interaction transitions.
Study context
The clinical evaluation described here concerns the tacrolimus–voriconazole module, not every interaction available in the application. The manuscript reports 394 trough measurements from 30 transplant recipients and compares forecasts using the same observations and available information.
The clearest supported advantage is during wash-in. At plateau, the model has no advantage over a static rule and overpredicts exposure. The smaller withdrawal analysis must be interpreted separately. These research findings do not replace close therapeutic drug monitoring.
Source: author-provided manuscript; the working presentation was used for context only. This is not the official conference abstract. Text and results may change before publication.
Open TacDDIProject overview
Link to this abstract