TL; DR

The EM serves to run in silico clinical trials cheaper and faster than in vivo clinical trials. EM can be applied from discovery (target selection) to market access (translating trial efficacy data into real-world outcomes).

The EM serves to run in silico clinical trials cheaper and faster than in vivo clinical trials [1] [2]. While its primary purpose is to inform in vivo trial design (e.g. profiling optimal responders or selecting the optimal dose-effect relationship) in order to de-risk late-stage development, the scope of applications extends from discovery (target selection) to market access (translating trial efficacy data into real-world outcomes).

In this framework, taking a hypothetical case, comparing a PD-1 inhibitor, a CTLA-4 inhibitor and a TIM3 inhibitor for melanoma becomes a matter of benchmarking these combinations’ relative Number of Prevented Events (NPE) over the same Virtual Population (VP), e.g. representative of the US melanoma population (see Boxes 7 and 8).

Box 7: Applying the EM to combination exploration

Treatment effect is represented as an alteration of the target site on the selected pathway: inhibition ofPD-1, CTLA-4 and TIM3, respectively.

The Virtual Population combines parameters from the disease model (e.g. expression status of PD-1 gene, concentration of T-cell in the TME) with “real patient” data drawn from selected databases to represent the US population. As shown in Box 4, each dot in the Rc,Rt plane represents a (virtual) patient. Only a subset of the VP is charted on the Rc,Rt planes in the right-hand panels for clarity purposes (i.e. not all the US melanoma population is charted).

The simulation consists of:

Run I Applying the melanoma model to the VP to generate the distribution of the clinical outcome rates without treatment, Rc, over the VP.

 

Run II Inhibiting CTLA-4 with treatment X in the melanoma model and applying the modified disease model to each of the same virtual patients to generate the distribution of the clinical event risk modified by treatment X, Rt, and sum all resulting ABs to get the X-related NPE (plot 1 in the right-hand panel).

 

Run III Inhibiting PD-1 with treatment Y in the melanoma model and applying the modified disease model to each of the same virtual patients to generate the distribution of the clinical event risk modified by treatment Y, Rt, and sum all resulting ABs to get the Y-related NPE

 

Run IV Blocking CTLA-4 with treatment X and PD-1 with treatment Y in the melanoma model and applying the modified disease model to each of the same virtual patients to generate the distribution of the clinical event risk modified by the combination of treatments X and Y, Rt, and sum all resulting ABs to get the X+Y-related NPE (plot 3 in the right-hand panel)

Compare the three NPEs (in this hypothetical case, respectively, 35, 53 and 84). The best scenario is the one corresponding to the highest NPE, i.e. 84. Thus, the combination of the PD-1 and CTLA-4 inhibitors exhibits synergistic efficacy in reducing the disease burden compared to each as monotherapy.

 

Run V The same principles apply to exploring the combination of a PD-1 inhibitor Y and TIM3 inhibitor Z (plot 4). The simulation indicates that the PD-1/CTLA-4 combination (NPE=84) yields a higher predicted clinical efficacy than the PD-1/TIM3 combination (NPE=60).

Conclusions i) Treatment Y is more efficacious than treatment X as monotherapy;

ii) The total effect of combining X and Y is less than what would have been observed if the effects of X and Y had been additive. The interaction between the two treatments is negatively synergistic;

iii) The PD-1/CTLA-4 combination compares favourably to the PD-1/TIM3 combination and should therefore be prioritized.

Using the same disease model and VP, optimal responders are defined as those virtual patients whose AB exceeds a pre-specified threshold, which could be set in terms of relative efficacy (compared to competing drugs) or cost to the payer (thus defining a unit cost per prevented event). These optimal responders are characterized by the set of biological/clinical parameters they share, e.g. cytokine expression profile, degree of immune system reactivity, mutation profile. Theranostic biomarkers are patient descriptors (parameters from the disease model) most correlated with the degree of clinical benefit (AB).

Box 8: Responder identification

Taking plots 1 and 2 from Box 6 (PD-1 and CTLA-4 inhibitors) and applying an arbitrary threshold (the dotted grey line) of 0,25 for AB to define responders (to treatments X and Y, respectively).

Were these responders all recruited in a trial comparing both compounds, the application of the EM would enable the characterization of these responders (i.e. virtual patients with an AB > 0,25) and the subsequent recruitment of the optimal profiles for a phase 3 trial. A significant difference in favour of inhibitor Y would result from such a trial designed with the support of the EM.

 

 

Conversely, a conventional trial where patients are eligible mainly on the basis of the stage of their disease (low risk or high risk areas in the plots) would likely be inconclusive in establishing these two drugs’ relative efficacy.

The EM has been applied in immuno-oncology on previous occasions, one where clinical trial efficacy results on US patients were to be translated into a prediction of real-world outcomes in another population treated with a CTLA-4 checkpoint inhibitor. In cardiovascular diseases, the EM has been applied to evaluate efficacy of heart rate reduction in the prevention of stable effort angina pectoris attacks[2:1] [3]. In the chronic obstructive pulmonary diseases (COPD) space, the EM is currently applied to identify both biomarkers and novel targets to prevent Chronic Lung Allograft Dysfunction (CLAD) [4].