A multiple-dose, randomized, double-blind, placebo-controlled phase IIa trial (ClinicalTrials.gov #NCT03237728) was conducted in Chinese sepsis patients. In the pre-trial, a total of eight patients were randomized at a 3:1 ratio to the 0.06 mg/kg drug group and the placebo group. In the formal trial, a total of 36 patients were randomized at a 1:1:1 ratio to the 0.12 mg/kg, 0.24 mg/kg, and placebo group. In the placebo group, enrolled sepsis patients received the standard care therapy required by the sepsis bundle such as antibiotics and vasopressor, etc. (Levy et al., 2018). In the drug groups, all enrolled subjects were administrated for 1 h with an intravenous infusion of kukoamine B every 8 h for seven days besides the standard care therapy by the sepsis bundle. The dose of kukoamine B was calculated according to the patient’s bodyweight. SOFA score was evaluated prior to the dose and each day for eight days following dosing. In the pre-trial, blood samples were collected pre-dose and 0.5, 1, 2, 4, 6, and 8 h after the first starting dose, before the last three doses, and 0.5, 1, 2, 4, 6, 8, 12, 16, and 24 h after the last dose. In the formal trial, blood samples were collected pre-dose and 0.5, 1, 2, 6, and 8 h after the first starting dose, before the last three doses, and 0.5, 1, 2, 6, and 12 h after the last dose. The study was conducted in accordance with the ethical principles stated by the Declaration of Helsinki, International Conference on Harmonization Technical Requirements for the Registration of Pharmaceuticals for Human Use of Good Clinical Practice. All participants or their guardians were required to sign a subject information agreement and the protocols of the study were approved by the Peking Union Medical College Hospital (PUMCH) Investigational Review Board (IRB).

Kukoamine B concentrations in the plasma of sepsis patients were analyzed using a validated ultra-performance liquid chromatography tandem mass spectrometry (UPLC-MS/MS) method with a limit of quantification (LLOQ) of 0.1 ng/ml (Wang et al., 2017 ). Concentrations vs. time profiles per group were plotted by Phoenix WinNonlin (version 7.0, Certara, NJ, United States). Area under curve (AUC_0–8 h) of the first dose and the last dose of 19 patients were calculated by non-compartment analysis (NCA) using Phoenix WinNonlin (version 7.0, Certara, NJ, United States), respectively. AUC_0–8 h of the first dose of five patients were calculated by NCA because of the missing day 7 PK samples. The exposure-response model was built by a nonlinear mixed-effect modeling approach using NONMEM (version 7.3; Icon Inc., North Wales, PA, United States) in conjunction with Pirana (version 2.8.0, Pirana Software and Consulting BV, Amesterdam, The Netherlands) and the PSN (version 1.1.453) toolkit. Post processing of the NONMEM output was undertaken with Xpose (Version 4.0 preview, release 6), programmed in the statistics package R (version 3.5.1). The FOCEI algorithm was employed in the model-building process. The baseline SOFA as a variable was directly included in the model. The structure of the model was evaluated based on objective function value (OFV), goodness-of-fit plots, the precision of parameter estimates, prediction-corrected visual predictive check (pcVPC, Bergstrand et al., 2011), physiological/pathological plausibility, and shrinkage (Savic et al., 2009). Conditional weighted residuals were calculated.

The data of the placebo and drug groups were pooled for modeling. Pharmacological effect in the drug groups was affected by two factors: i) pharmacological effect of the standard care therapy (Zhang et al., 2019) and ii) pharmacological effect of kukoamine B. The exposure-response model was modeled in Eq. 3 (Hu, 2014; Hu et al., 2014; Hu and Zhou, 2016): SOFA ( t ) = Base − F placebo × Base × ( 1 − e − k t ) − ( 1 − F placebo ) × Base × ( 1 − R ( t ) ) + ε , (3) Where S O F A (t) is time-dependent SOFA, F placebo (0≤ F placebo ≤ 1) is the maximum fraction of the standard care therapy. ( 1 − F placebo ) is the maximum fraction of drug effect. Base indicated the baseline SOFA, Base × ( 1 − e − k t ) is the change of SOFA which is productive by the standard care therapy, IIV for k is assumed to be normally distributed and is described by an addition model,     R ( t ) is a latent variable which is productive by kukoamine B. ε   ∼ N (0, σ²) represents the residual unexplained variability, which is described by an additional function.

The pharmacological effect of kukoamine B is assumed to be driven by a latent variable as follows:

R (t) governed by: d R ( t ) d t = k in × ( 1 − A U C A U C + E A U C 50 ) − k out × R ( t ) , (4) Where AUC is steady state AUC for kukoamine B: k in , E A U C 50 , and k out are the parameters of the indirect response (IDR) model.   E A U C 50   is the steady state AUC of kukoamine B that produces half of the maximum attainable inhibition. k in is the zero-order production rate of the residual percentage of the SOFA score, k out is the first-order elimination rate of the residual percentage of the SOFA scores. It is assumed that at the baseline, R (0) = 1, yielding k in = k out . IIV for k in and E A U C 50 is fixed to achieve successful OFV minimization. For the placebo group, R (t) = 1 (AUC = 0).

Goodness-of-fit (GOF) plots were used to evaluated the predictive performance of the exposure-response model, including dependent variable vs. individual prediction (IPRED) or population prediction (PRED), individual weighted residuals (IWRES) vs. PRED, and CWRES vs. time. pcVPC was used for model evaluation (n = 1,000) and stratified by the dose to ensure the models performed adequately across subgroups (U.S. Food and Drug Administration, 2019).

To characterize the profiles of day 7 SOFA vs. AUC, the simulation was performed based on the built exposure-response model. The method was as follows: time in the exposure response model was fixed at 7 (t = 7), day 7 SOFA scores were simulated at the baselines 15, 10, and 5 1,000 times when AUC ranged from 0 ng*h/mL to 5,000 ng*h/mL. The 30th, 50th, and 70th percentiles (prediction intervals) of day 7 SOFA scores were plotted against AUC. The profiles of SOFA vs. time were also simulated, and the plot was shown with different colored shaped areas representing the placebo group and the recommended 0.24 mg/kg group in a panel.

The data of 10 sepsis patients in the placebo group and 24 sepsis patients in the drug groups contributing 245 measurable SOFA scores were finally included into our model dataset. Ten patients were excluded from the SOFA dataset according to the exclusion criteria or due to their withdrawal from the clinical trials. There were 43 missing PK samples due to subjects withdrawing from clinical trials or because the drug was not successfully administrated. The results are shown in Supplementary Table S1. The PK concentrations of all samples except 0 h were more than LOQ. For 19 patients, we calculated the AUC_0–8 h of the first dose and the last dose, and the average of them was used as the exposure measure. Five patients dropped out of the clinical trial, so the PK sample of the last dose was missing, and the AUC_0–8 h of the first dose was used. AUC (mean ± SD) for the 0.06 mg/kg group was 295.21 ± 47.16 (mean ± SD) h*ng/ml; AUC (mean ± SD) for the 0.12 mg/kg group was 827.85 ± 334.10 (mean ± SD) h*ng/ml; AUC (mean ± SD) for the 0.24 mg/kg group was 1,482.51 ± 378.34 (mean ± SD) h*ng/ml. Supplementary Figure S1 shows the concentration vs. time profiles per group, Supplementary Figure S2 shows the SOFA score vs. time profiles per group. Baseline SOFA (mean ± SD) for the placebo group was 12.10 ± 4.36 (mean ± SD); baseline SOFA (mean ± SD) for the 0.06 mg/kg group was 9.25 ± 2.22 (mean ± SD); baseline SOFA (mean ± SD) for the 0.12 mg/kg group was 9.56 ± 3.09 (mean ± SD); and baseline SOFA (mean ± SD) for the 0.24 mg/kg group was 10.09 ± 3.86 (mean ± SD).

A latent-variable approach in conjunction with an inhibitory indirect response model was used to link the standard care therapy effect and drug exposure to SOFA score. The results of the parameter estimation are summarized in Table 1. The OFV value was 526.34.

The validity of the exposure-response model was assessed by graphical analysis (goodness-of-fit plots) and pcVPC. Cross-validation or external validation was not used due to the limitation of sample sizes (Colby and Bair, 2013). The goodness-of-fit diagnostic plots are shown in Figure 1. The SOFA score of some subjects did not decrease, indicating that these subjects had no response to the standard care therapy and kukoamine B. Most dots of CWRES were shown to lie within an acceptable range (−2 to 2), which suggested little to no bias over time. Most dots of absolute value of IWRES lay within an acceptable range (0–2), indicating no bias with the predicted value. pcVPC stratified on dose was carried out with 1,000 simulations, the plot with 80% confidence interval (low sample size) is shown in Figure 2. Most of the observed SOFA scores were within the 80% CI of the model-predicted SOFA score, demonstrating adequate model prediction. In the 0.06 mg/kg group and 0.12 mg/kg group, the 10th and 90th percentiles of the predicted values were wider than the 10th and 90th percentiles of the observed values. It was elucidated that the random effect predicted was higher than the observed values because of a low sample size. The 50th percentiles of the observed values and predicted values were generally consistent. In the placebo group, the standard care effect was somewhat overestimated, which was assumed to be related to the high heterogeneity of sepsis patients.

Day 7 SOFA vs. AUC profiles were simulated, which are shown in Figure 3. It was observed that day 7 SOFA decreased to a plateau when AUC was 1,500 h* ng/mL and the patients were administrated for 1 h an intravenous infusion of kukoamine B every 8 h for seven continuous days. Furthermore, SOFA score decreased more significantly when the baseline SOFA was high. It provided some important information for the inclusion criteria of subjects in the following Phase IIb clinical trial. Besides, SOFA score vs. time profiles were also simulated, and the plots overlaying the effect of the placebo and the recommend 0.24 mg/kg dosing regimen are shown in Figure 4. A difference in the efficacy of the placebo group and the 0.24 mg/kg group was observed. In the 0.24 mg/kg dosing regimen, no serious adverse effect was investigated in the Phase IIa trial. For 80% of patients, AUC can achieve 1,500 h*ng/mL. Based on the simulated results and good safety, we recommended the 0.24 mg/kg dosing regimen for the Phase IIb clinical trial.