Background. Therapeutic drug monitoring (TDM) for linezolid, an essential MDR-TB drug, is strongly recommended but often impractical. Saliva sampling with limited sampling strategy presents a non-invasive and more feasible alternative to blood, but a validated saliva model for dosing guidance is lacking.
Aim. To develop a saliva population pharmacokinetic (PK) model for linezolid in Vietnamese MDR-TB patients.
Methods. Adults (n = 17) receiving linezolid for ≥7 days had blood and saliva samples collected at 0, 2, and 5 hours post-dose at steady state (ACTRN12620000681954). Linezolid concentrations (102 paired saliva-plasma samples) were quantified using HPLC-MS. One and two-compartment models with additive, exponential, and combined error models were tested through baseline model development, and potential covariates were tested on the parameters using stepwise covariate modelling (scm) and diagnostics checks (NONMEM 7.4.4 and PsN 5.3.1). After validating plasma model, saliva component was integrated either as a scaling factor or as a distinct compartment. Sampling importance resampling technique (SIR) was used to assess the robustness of the final parameter estimates. The model is being applied for Monte Carlo simulations (n = 1000) to explore effective limited sampling strategies, both with and without using Bayesian software for a widespread applicability in various settings. AUC24h was calculated using standard 12-point sampling per patient with post-hoc analysis and NCAPPC package in R, and limited sampling strategies (1, 2, or 3 points) were compared to this standard using paired t-tests (α = 0.05).
Results. One-compartment population PK model with 1st order absorption and elimination, in which a saliva compartment was integrated, best described the PK in both plasma and saliva. Distinct compartment method outperformed scaling method in terms of objective function (dOFV = ̶ 43.084, p < 0.05 at α = 0.05 with 2 degrees of freedom, χ2 distribution), covariance/correlation steps, model stability and precision. The final model was optimised using the FOCE-I with PRIOR approach, yielding population parameters: Ka = 1.5 h-1, F = 1, CL/F = 4.3 L/h, V/F = 46.8 L, K23 (transfer rate from plasma to saliva) = 4.9 h-1, K32 (transfer rate from saliva to plasma) = 1.8 h-1, K30 (elimination rate from saliva) = 2.1 h-1, IIVCL = 44.6%. Body weight with estimated exponent (θCOV = 1.1) was identified as significant covariate for V, resulting in respective proportional residual errors of 25.8% (plasma) and 35.9% (saliva). Visual diagnostics confirmed a good predictive performance (Fig. 1). For the model’s application, preliminary results suggest that six saliva samples per patient without requiring Bayesian software may be sufficient—compared to the typical 12 samples—for estimating AUC24h using non-compartmental analysis. Using Bayesian software, 2-point sampling (0h and 2h) produced AUC24h estimates that were statistically comparable to 3-point observed sampling (p < 0.05), whereas 1-point (0h) sampling yielded significantly different results.
Conclusion. By leveraging Bayesian-guided estimation with limited samples, the validated saliva model offers a promising framework for enabling saliva-based TDM of linezolid such as guiding TDM implementation strategies or future clinical trials.
Figure 1. a) Diagnostic goodness-of-fit plot of the final linezolid saliva model: observed versus individual predicted concentrations; and b) Visual predictive check of the model for plasma (CMT:2) and saliva (CMT:3).