Objective: To characterize protein binding in patients with epilepsy who achieve transient high (>150 mg/L) total plasma concentrations following rapid valproate infusion at very high doses. Methods: Patients with epilepsy (n = 40) were administered 20 or 30 mg/kg loading doses (6 or 10 mg/kg/min) of undiluted valproate sodium injection. Total and unbound valproic acid (VPA) concentrations were used to assess VPA binding to plasma albumin. One- and two-binding site models were explored in a nonlinear mixed effects population analysis framework. The relative importance of weight, age, sex, race and enzyme-inducing comedications on the binding site association constant (K) was examined using the likelihood ratio test. Intersubject and intrasubject variabilities were characterized using exponential or proportional error models. Results: Optimal characterization of the data was achieved using the one-binding site model. Population binding parameter estimates (standard error) for number of binding sites (N) and K were 1.98 (0.0865) and 15.5 [2.28 (1/mM)], respectively. No significant covariates were identified for VPA protein binding. The intersubject and intrasubject coefficients of variation were 32% and 14%, respectively. Conclusions: A one-binding site model without any significant covariates for binding constants optimally described VPA protein binding. As the estimated dissociation constant (1/K, 64.5 μm or 9.3 mg/L) was within the therapeutic range (5-15 mg/L) for unbound VPA concentrations, protein binding was nonlinear. Although the range of unbound fraction and VPA concentrations were much higher than previous studies, the dissociation constant was consistent with historical data in normal healthy adults and epilepsy patients receiving lower doses. © 2007 Blackwell Publishing Ltd.