Four cases here, each for a different therapy area, showcase how the PoPS approach was applied at decision gates from preclinical candidate selection (Case 1) to FiH (Cases 2 and 3) and clinical proof-of-concept (PoC, Case 4). To protect proprietary information, the compounds involved have been anonymised. Nonetheless, we believe sufficient details are described to illustrate key aspects of situational application of the PoPS approach, from how the pharmacology endpoint is identified and how the success criteria are defined, to how uncertainties in the drug-disease system were considered and how PoPS was computed to capture these uncertainties.
Case 1: select a preclinical candidate for neurodegeneration [7]
A full report of this case is published [7]. Two lead compounds (A and B) for a neurodegenerative disease were ready for preclinical candidate selection. The target is expressed both in the brain and peripheral tissues; the elevation of the drug target activity in the brain is genetically linked to the disease. While target inhibition is a potential treatment, in-class data from animals had shown pharmacology-driven lung and kidney toxicity at high levels (> 90%) of peripheral target inhibition; therefore, the intervention need was defined as normalizing the brain target activity but preserving enough activity in peripheral tissues to mitigate the pharmacology-related safety risk. A decision was required to choose one of the two compounds for further preclinical development. However, before PoPS estimation, the two compounds could not be clearly differentiated, given both compounds demonstrated enough potency to inhibit the target activity to the normal level in the in vivo animal studies. Because the therapeutic success relied on balancing target inhibition in the brain versus in the periphery, the target activity, as measured by a blood biomarker, was the endpoint for both pharmacology and safety aspects of PoPS. The PoPS was specified as the probability for ≥ 80% of patients to achieve the desired balance while putting < 5% of patients at risk of over-inhibition.
One-compartment PK model in human, with first-order absorption and elimination was constructed for each compound, where key parameters including apparent oral clearance (CL/F) were predicted from multiple animal species. A PK/PD model for central and peripheral inhibition was further developed for each compound. The central and peripheral inhibitions were described using simple Emax models revealed by animal experiments, where the drug concentration leading to 50% of Emax, EC50, was substituted by estimates from in vitro human cell lines, the Emax was assumed to be 100% target inhibition, and the steady-state average free concentration considering brain-plasma unbound drug portion (kp,uu) was applied for the model. Translational uncertainty in kp,uu was addressed by applying a uniform distribution to the values estimated from multiple species: 0.45–0.75 for compound A and 0.35–0.50 for compound B. The key PK and PD parameters—CL/F, kp,uu, and EC50 were assumed to follow a log-normal BSV distribution, per established practice. Both compounds are Biopharmaceutical Classification System (BCS) class II compounds with limited solubility and high permeability; hence a moderate BSV (CV% = 30%) was assigned for CL/F, given 20–40% CV% usually estimated for this class [14,15,16]. The EC50 was assigned a 30% BSV, as estimated in vitro.
Clinical data were simulated for 1000 trials of 1000 subjects each using this model, including the BSVs and the relevant translational uncertainties described above. For PoPS computation, the uncertainty in the intervention needs—to normalise the target activity—was considered using the literature-reported range of the target elevation (1.5–3.0 folds) in the disease; whereas the range of kp,uu values determined from multiple species was the translation uncertainty. A unique combination of the two parameters was assigned to each simulated trial. In each of the 1000 simulations, the patients with desired benefit-risk profile were the ones with sufficient central inhibition (normalized target activity) and peripheral preservation (preserving ≥ 10% activity), and the patients with safety risk were the ones with insufficient peripheral preservation (Fig. 2 upper panels).
The results of this assessment showed the maximum proportions of patients achieving the required brain target activity inhibition and peripheral preservation were similar for the two compounds- 87% vs 79% (Fig. 2 upper panels), which still could not suggest a clear differentiation. However, when both efficacy (≥ 80% of patients with desired benefit-risk ratio) and safety criteria (< 5% of patients with safety risk) were considered together, the maximal PoPS was 61% for Compound A (at 56 mg daily dose, Fig. 2 lower left panel), and 21% for Compound B (at 18 daily dose, Fig. 2 lower right panel). Given the clearly higher PoPS of compound A, it was selected over compound B for further development.
This case study provided an example that PoPS led to an unambiguous decision at the stage of preclinical candidate selection, despite the fragmented nature of the drug-disease system knowledge and multiple uncertainties. The PoPS findings also provided insight for future compound discovery and guidance on clinical trial population selection; these were discussed in the original paper [7].
Case 2: commit to FiH of a molecule for a heart condition [17]
A compound for a heart condition was approaching FiH decision gate. Based on preclinical evidence, the molecule is expected to produce efficacy by inducing the production of a protein. The concentration–response of the protein’s mRNA was assessed in human blood ex vivo, while the exposure–response of the efficacy endpoint was evaluated in a mouse model. Toxicokinetic experiments were performed in three animal species, where PK and NOAEL were established. The PoPS was specified as the probability that ≥ 90% of patients achieve adequate pharmacology, and ≤ 5% exceed the safety limit.
The protein mRNA, a biomarker reflecting signal transduction, was selected as the pharmacology endpoint, because it can be measured in healthy volunteers to inform dose selection in patients. A transduction PKPD model [18] was used to translate ex vivo data into an in vivo relationship by combining it with the predicted human PK (allometrically scaled from preclinical data).
To benchmark required pharmacology, an Emax relationship between PK and the efficacy endpoint was identified based on the mouse efficacy model and was scaled to human accounting for plasma partitioning and protein binding differences. The pharmacology and efficacy were bridged by PK: the required pharmacology was defined based on the exposure which was in turn corresponding to relevant efficacy.
Estimation uncertainty was included on key parameters, i.e., protein mRNA Emax and EC50. The percentage relative standard error (% RSE) of estimate derived ex vivo was used to define prior width; furthermore, estimation correlation between Emax and EC50 was incorporated into prior definition. Translational uncertainty was also included on PD BSV, by assuming three equally probable scenarios of low, medium, and high protein mRNA BSV.
Moderate PK BSV (CV% = 30%) was included on key parameters (central clearance, central volume of distribution, absorption rate constant). The BSVs in baseline (E0) and maximum increase in protein mRNA (Emax), both estimated ex vivo, were included.
The pharmacology need was defined as approximately 2–10 folds protein mRNA increase, based on medical and biological evidence that this level of pharmacology could produce meaningful efficacy. The safety limit was based on the molecule’s exposure being below the NOAEL from the three preclinical species.
Uncertainty was included on pharmacology need and safety limit. Data from animal models show that a 2–fourfold increase in protein mRNA was usually required to achieve meaningful efficacy, while higher values were only occasionally needed. Therefore, a beta distribution with mode at 2 fold (with decreasing probability towards values > 10 fold) was used to describe the required pharmacology. To reflect the knowledge at the time about the relevance of the three preclinical species (rat, dog, monkey), a weighted prior (60:30:10) was applied to the safety limit.
The PoPS was computed by simulating the mRNA level (for pharmacology) and plasma exposure (for safety) for 1000 virtual populations, each with 1000 patients per dose level, initially assuming priors outlined above. Subsequent toxicology assessments and expert consultations concluded that dose-limiting safety findings in rat and dog were unlikely to translate to human at the proposed clinical dose range. Therefore, the PoPS was re-estimated using the safety limit from monkey only. Scenario analysis was conducted, considering varying pharmacology requirement (2- to 8-fold biomarker increase) and protein mRNA variability level (low to high) (Fig. 3).
The first PoPS assessment provided a peak ~ 60%, at 300 mg QD dose (Fig. 3a). In the subsequent scenario analysis, the updated PoPS peaked at ≥ 90% (Fig. 3b).
In this case study, PoPS effectively integrated multi-source data, weighting their relevance through close interaction with experts, and outlined key assumptions and uncertainty sources. It provided an early understanding of the therapeutic potential of the molecule thus enhancing the confidence and clarity of the commit-to-FiH. Once clinical PK and protein mRNA data from the FiH are available, the PoPS will be updated to inform the decision to progress to subsequent patient trials.
Case 3: commit to FiH of a molecule for an auto-immune disease [19]
The molecule is a humanised IgG1 monoclonal antibody, for an autoimmune disease, approaching FiH decision gate. The binding of the drug to a soluble cytokine (TE) prevents the cytokine’s interaction with its receptors located on an autoimmune T-cell subset, reducing the expression of an intracellular protein (i.e., PP), and consequently decreasing the count of those autoimmune T-cells. The PoPS for the asset was specified as “the probability that > 50% of the subjects achieve required protein reduction in the target tissue, while > 50% have systemic drug exposure within the relevant safety limit”.
Modelling was performed using serum data of drug exposure, TE, and PP in monkeys. A target mediated drug disposition (TMDD) model was used to describe the TE. The PP was characterized by an indirect response model, where protein turnover in response to serum drug concentrations was characterized by an Imax model. The TE and PP models were translated to humans via allometric scaling of PK parameters and cross-species adjustment of baseline target level. The parameters describing protein turnover in monkeys were considered similar in humans. For PoPS computation, the PP endpoint was chosen over the TE endpoint because, being a downstream measure, it enabled the identification of the required pharmacology based on literature data of two compounds (one with a similar mechanism and for a similar indication; the other with a different mechanism though for the same indication). The PP model was extrapolated to target tissue by adjusting for drug’s distribution to the tissue.
For the PoPS, it was crucial to include uncertainties at three levels i.e., estimation of PKPD parameters, translation of pharmacology from blood to tissue, and therapeutic need. Estimation uncertainties (% RSEs) were incorporated on PD parameters for drug’s efficacy (Imax) and potency (IC50) as well as the rate of protein degradation. Correlations among PK (clearances and volumes of distribution) and PD (Imax and IC50) parameters were included using the estimates from respective covariance matrices. The PP in the tissue was used as a metric for pharmacological effect. Because the drug effect occurs predominantly in the tissue but is measured in the blood, a blood-to-tissue translational uncertainty was included based on the estimated plasma-to-tissue distribution coefficient. The uncertainty in therapeutic need was incorporated via a uniform distribution (70–95% inhibition of the cytokine-driven protein), informed by the two benchmark drugs. The safety limit was derived from exposure at the NOAEL as observed in monkeys.
Clinical PP simulations were conducted for every-four-week (Q4W) dosing, including BSVs on PK and PD parameters as estimated from monkey data. In total, 500 trials (accounting for PKPD estimation and translational uncertainties) were simulated, each trial having 500 individuals (accounting for BSV). Using prior distributions for the pharmacology requirement, the overall PoPS was > 95% (Fig. 4a). Scenario analysis was also conducted (Fig. 4b). Under the most relaxed conditions, i.e., high potency in tissue and low requirement for pharmacology, PoPS was > 95%. Under the most stringent conditions, i.e., low potency in tissue and high requirement for pharmacology, a moderate PoPS of ~ 75% was estimated. Both for overall PoPS and scenario analysis, at doses > 5000 mg Q4W, a sharp decline in PoPS was expected, due to the exposure exceeding the safety limit.
Choosing the PP as the pharmacology endpoint for PoPS computation was a cornerstone for this case. It allowed benchmarking of required pharmacology using clinical data of drugs with a similar mechanism or for a similar indication (see above). The successful benchmarking in turn significantly expedited the program. The PoPS results established a high confidence in the drug’s ability to deliver required pharmacology within the desired safety limit, in this case leading to a firm decision of progression to FiH ahead of schedule.
Case 4: commit to clinical PoC of an oral anti-bacterial [20]
A candidate antibiotic was at the decision point for committing resource to clinical PoC, following FiH where human PK, safety and tolerability were assessed. The efficacy endpoint for the PoC would be the standard measure of bacteria decline in the relevant biological matrix. The method of PoPS computation was different from the other cases described above and followed less closely the process outlined in Fig. 1. It was inspired by the probability of target attainment (PTA) approach, established in antibiotic research [21,22,23]. The PTA approach maps out the cumulative distribution of an efficacy-driving PKPD index in a patient population. The usual index is AUC:MIC ratio, Cmax:MIC ratio or Time > MIC (the proportion of the dosing interval when plasma concentration of the drug is above MIC), where MIC is the minimum inhibitory concentration for pathogen growth. Achieving a critical value of the cumulative distribution observed for an effective treatment is then considered as the requirement for treatment success.
To choose the appropriate pharmacology endpoint for PoPS and establish the required response level, we identified a drug from the same class that had demonstrated varying degrees of efficacy in multiple studies. A population PK model was constructed for the benchmark drug and used to simulate PK profiles, sampling over all estimated BSVs in these positive studies (N = 3000 for each study). The MIC distribution found in literature was fitted by a logistic function, which was used to simulate a large population (N = 3000) of values to be paired randomly with the simulated PK. Meta-analysis including the multiple studies showed that the strongest efficacy driver was AUC:MIC [20], which was chosen as the endpoint for computing PoPS for the candidate drug. The distribution of the AUC:MIC for this benchmark drug at a target efficacy level served as the response level required for efficacy for the drug class.
To estimate the PoPS, a population PK model was constructed for the candidate drug using the FiH data. Variability in MICs was described by a log-normal distribution, fitted to proprietary data. The PK model was used to simulate a large population (N = 3000) of PK profiles for the highest proposed PoC dose, sampling over all estimated BSVs. MICs were subsequently sampled from the fitted log-normal distribution giving distribution of randomly paired AUC:MIC values (N = 3000). The distribution of AUC:MIC values for the drug was then compared to that for the reference drug at the (latter’s) target efficacy level.
Uncertainties were not explicitly included in this PoPS calculation for multiple reasons. Under time pressure, we were pragmatic to not include PKPD uncertainties: uncertainty (%RSE) for PK parameters was well below 20% for both candidate and benchmark drugs, as both models were built from frequently sampled rich data; and uncertainty in MICs was considered to be less than two-fold, inherently due to the serial-dilution experimental design for their estimation. Furthermore, uncertainty in translation was not relevant as we used clinical data of an in-class benchmark compound for the same infection. Finally, uncertainty in pharmacology requirement (AUC:MIC) was not needed when the PoPS at the highest proposed dose turned out to be negligible when benchmarked for the target efficacy (Fig. 5).
This was a new framework for predicting the probability of achieving target pharmacological effect levels, applicable to any anti-bacterial drug where in-class benchmark is available. Because the PoPS—the probability of the drug reaching benchmark AUC:MIC values corresponding to the target efficacy—was found to be negligible even at the highest proposed dose, the onward internal funding for the molecule was deprioritised.