Data and study design

The IPATential150 study was a Phase 3, randomized, double-blind, placebo-controlled, multicenter trial testing ipatasertib plus abiraterone plus prednisone/prednisolone, relative to placebo plus abiraterone plus prednisone/prednisolone in adult male patients with asymptomatic or mildly symptomatic, previously untreated mCRPC [15]. The study was approved by an ethics committee or institutional review board at each trial site and carried out in accordance with the International Conference on Harmonization Guideline for Good Clinical Practice. Written informed consent was obtained from all subjects before enrollment in the trial.

Patients who met the eligibility criteria were randomized in a 1:1 ratio to one of the two treatment arms; patients in the experimental arm received ipatasertib (400 mg QD) and patients in the control arm received matching placebo, each consisting of 28-day cycles of oral administration. The same formulation of ipatasertib (a film-coated tablet), which was also used in the randomized Phase 2 part of a Phase 1b/2 A.MARTIN study [13], was used throughout the study. In addition, all patients received abiraterone (1000 mg QD) plus prednisone/prednisolone (5 mg BID). Treatments were continued until disease progression (as assessed by Response Evaluation Criteria in Solid Tumors [RECIST] version 1.1 [17], or Prostate Cancer Working Group 3 [PCWG3] criteria [18], or both), intolerable toxicity, elective withdrawal or study conclusion. The co-primary endpoints were investigator-assessed rPFS in patients with PTEN-loss tumors and in the ITT population. Safety was evaluated in all patients who received any dose of ipatasertib, abiraterone, or placebo according to the Common Terminology Criteria for Adverse Events, version 4.0. Blood samples for PK assessments of ipatasertib and its major metabolite M1 were collected on days 1 (1–3 h post-dose) and 15 (pre-dose and 1–3 h post-dose) of cycle 1, day 1 of cycle 3 (pre-dose and 1–3 h post-dose), and day 1 of cycle 6 (pre-dose).

External validation of IPATential150 PK data and generation of exposure metrics

A previously developed population PK models of ipatasertib and M1 [16] were applied to the observed individual PK data from the IPATential150 study, which were not part of the data used for model development (i.e., served as external validation of the models), and empirical Bayes estimates (EBEs) of model parameters for each patient in the ipatasertib arm of the study were generated. Data from a total of 546 individuals who received at least one dose of ipatasertib and had at least one quantifiable PK observation were included in the analysis, comprising a total of 2561 ipatasertib observations and 2542 M1 observations. These EBEs were in turn used to simulate a concentration–time profile for the intervals required, from which the necessary exposure metrics were derived. A dosing interval and observation period of 24 h were assumed; concentration–time profile at 0–24 h after the initial dose and at 336–360 h after 2 weeks of QD dosing was used to derive exposure metrics after the first dose (i.e., single dose) and at steady state, respectively. The planned (per protocol) nominal dosing of 400 mg QD was always assumed and any dose adaptations were not taken into account to avoid potential post-randomization bias in the analysis (e.g., correlations between dose alterations and response). Generated exposure metrics included area under the concentration–time curve (AUC, calculated by the linear trapezoidal rule), maximum concentration (Cmax), assigned as the peak concentration within the specified interval, and trough concentration (Cmin), assigned as the concentration at 24 h (for the single dose) and at 360 h (for the steady state), after the single dose (AUCSD, Cmax,SD, Cmin,SD) and at steady state (AUCSS, Cmax,SS, Cmin,SS). Correlations between the simulated PK parameters (AUCSD, Cmax,SD, Cmin,SD, AUCSS, Cmax,SS, and Cmin,SS) as well as each patient’s AUCSS values for ipatasertib and M1 were evaluated; if these were highly correlated, only the representative exposure metrics (e.g., ipatasertib AUCSS) were to be used in subsequent analysis. Zero exposures were assigned to 554 patients in the control arm of the IPATential150 study.

Exposure–efficacy analysis on rPFS

The endpoint of interest for exposure–efficacy analysis was rPFS, assessed both in patients with PTEN-loss tumors and in the ITT population. Radiographic PFS was defined as the time from the date of randomization to the first occurrence of documented disease progression, as assessed by the investigators with the use of the PCWG3 criteria [18] (soft tissue by computerized tomography or magnetic resonance imaging scans according to the RECIST v1.1 [17], and bone metastasis by bone scan according to the PCWG3 criteria) or death from any cause, whichever occurs first. As an exploratory graphical analysis, rPFS data were explored using Kaplan–Meier plots with stratification by ipatasertib exposure groups; subjects in the placebo arm were compared with subjects from the active treatment arms who were grouped in exposure quartiles. Modeling of rPFS was performed using the Cox proportional hazards model, which allows the fitting of univariable and multivariable regression models with survival outcomes, as follows:

$$ h\left( } \right.} \right) = h_ \left( t \right) \times \exp \left( \times X_ + \cdots + \beta_ \times X_ } \right). $$

(1)

Here, h(t) is the hazard (the instantaneous rate at which events occur), h0(t) is the underlying baseline hazard, Xi is the set of explanatory covariates for individual i, β1…n are the coefficients describing the effects of explanatory covariates 1 − n, and Xi,1…i,n are explanatory covariates values 1 − n in individual i.

Covariate testing was performed in a stepwise fashion. Each putative covariate relationship was fitted in a univariable model first. All those with a p value for inclusion of < 0.15 were jointly included in a “full” model. Each was then subsequently removed one at a time and assessed using the Akaike information criterion (AIC) estimated as follows:

$$\mathrm=2k-2ln(\widehat).$$

(2)

Here, k is the number of estimated parameters and L̂ is the value of the likelihood function. When no further relationships met the retention criterion (i.e., removing a term in the model reduced the AIC relative to the model including the term), the model was considered final.

Time-varying covariate Cox modeling analysis for rPFS

In addition to the exposure–efficacy analysis using the standard Cox proportional hazards model, time-varying covariate Cox modeling analysis for rPFS was performed for ipatasertib-treated patients in the ITT population. The analysis was performed with Cox regression approach, but using the daily ipatasertib dose at each timepoint as a covariate instead of using one fixed value (i.e., nominal dose) per patient to calculate hazard. For the patients who discontinued ipatasertib due to adverse events (AEs), daily dose of 0 mg was assigned after the dose discontinuation. For the patients who discontinued ipatasertib due to other reasons (but before record of PFS event), PFS status was censored at the treatment discontinuation date.

Exposure–safety analysis on safety outcomes

Exposure–safety relationships were assessed for the following endpoints; serious AEs (SAEs), AEs leading to treatment discontinuation, AEs leading to dose reduction, and specific AEs of clinical interest (which were the following identified risks of ipatasertib) including diarrhea (Grade ≥ 2 and Grade ≥ 3), hyperglycemia (Grade ≥ 2), and rash (Grade ≥ 2). All safety events were binary outcomes (event or no event) and analysis was conducted using logistic regression as follows:

$$ Z_ = }\left( \right)}} \right)}}} \right) = \beta_ + \beta_ \times X_ + \cdots + \beta_ \times X_ . $$

(3)

Here, Zi is the log odds of the probability p that outcome y = 1 in individual i, β0 is the baseline log-odds that event y = 1, β1…n are the coefficients describing the effects of explanatory covariates 1 − n, and Xi,1…i,n are explanatory covariates values 1 − n in individual i.

The odds of outcome y = 1 may be recovered by exponentiating the log-odds. The probability Pi of the event in individual i can be calculated as follows:

Every putative covariate relationship was added in a single step. Each was subsequently removed one at a time and assessed in terms of the associated AIC. When no further relationships met the retention criterion (i.e., removal of the predictor reduced the AIC relative to the model including it), the model was considered final.

Covariate scope

Covariates included in the analyses for efficacy and safety, apart from ipatasertib treatment and exposure (e.g., AUC), included baseline age, baseline weight, race, geographic region, the Eastern Cooperative Oncology Group (ECOG) status at baseline, prior taxane-based therapy in hormone-sensitive prostate cancer setting (yes or no), factor(s) of progressive disease before initiation of the study treatment (PSA only or other), presence of visceral metastasis (yes or no), tumor PTEN-diagnostic status by immunohistochemistry assay (PTEN-loss: yes or no), baseline glucose (for hyperglycemia only), baseline HbA1c (for hyperglycemia only), and abiraterone trough concentration at steady state (pre-dose at day 15 of cycle 1; for efficacy analysis only).

Forest plots were used to illustrate the effects of covariates on parameters and generated for the model including all potential covariates of interest and for the final reduced model after stepwise reduction. For each covariate effect coefficient, the asymptotic standard errors were used to generate a 95% CI for the coefficient (defined as ± 1.96 × standard error [SE]). The point estimate and the upper and lower limits of the 95% CI for the covariate coefficient were used together with the covariate and the estimated value of the parameter to define a 95% confidence range for the parameter given the covariate relationship, expressed relative to the typical value of the population parameter in the population (such that “no effect” would be 1).

Software

Generation of EBEs from the population PK models was performed in the nonlinear mixed effect modeling software NONMEM version 7.4.3 (ICON Development Solutions, Ellicott City, MD, USA) [19], supplemented with Perl-speaks-NONMEM (PsN) version 4.9.0 (Uppsala University, Uppsala, Sweden) [20, 21]. R software version 4.0.0 (The R Foundation, Vienna, Austria) was used to derive exposure metrics based on EBEs generated by the population PK models (for ipatasertib and M1), and for general scripting, data management, Cox proportional hazards modeling, logistic regression modeling, goodness of fit analyses, and model evaluation.