From Predicted Heterogeneity to an Enrollment Rule
A QSP model can predict that patients differ in pathway state and in their expected benefit from treatment. That heterogeneity can inform a practical development decision: define a biomarker rule for selecting trial participants. The central calculation is not just how many people pass a cutoff. It is also the treatment benefit predicted for those eligible, and how that benefit changes as the cutoff changes.
Enrichment is a choice about the population to enroll, not proof that a biomarker causes treatment response. The biomarker may be a useful predictor of modeled benefit because it reflects a relevant pathway state. The rule is credible only to the extent that the virtual population, biomarker mapping, and treatment predictions represent the intended clinical population.
Use a benefit scale with a clear clinical interpretation, such as points of improvement on a specified outcome scale over a defined period. Each \(\Delta_i\) should compare treatment and control predictions for the same virtual patient, under the stated simulation conditions. It is a model-based prediction, not an observed individual treatment effect.
Calculating Eligible Numbers and Expected Benefit
For a weighted virtual population, let \(w_i\) be patient \(i\)'s population weight, with weights summing to one. An equally weighted sample uses \(w_i=1/n\). For a cutoff \(c\), define eligibility by the indicator \(I(B_i \geq c)\), which equals one when the rule is met and zero otherwise. The eligible fraction and eligible mean benefit are:
The mean is defined only when \(q(c)>0\). If the target population contains \(N\) people, the expected number meeting the rule is \(Nq(c)\). The numerator of the mean formula is the population-weighted benefit total contributed by eligible people, per member of the full target population; dividing by \(q(c)\) converts that total to a mean per eligible person. This distinction matters: a stricter cutoff can raise mean predicted benefit while reducing the number of eligible candidates.
Worked Example: Comparing Broad and Strict Cutoffs
Hypothetical data. A QSP analysis represents a target population of 1,000 patients. The biomarker is pathway activation, and benefit is measured in predicted improvement points over 12 weeks. Assume each category lies wholly on one side of the cutoffs considered.
| Pathway activation \(B\) | Patients | Mean predicted benefit per patient |
|---|---|---|
| Below 2 | 400 | 1 point |
| 2 to below 5 | 350 | 3 points |
| 5 or above | 250 | 6 points |
For a cutoff of 2, 600 of 1,000 patients are eligible. The total predicted benefit across these eligible patients is \(350(3)+250(6)=1{,}050+1{,}500=2{,}550\) improvement points. The eligible mean is \(2{,}550/600=4.25\) points per patient. Thus \(q(2)=0.60\), corresponding to 600 eligible patients in this population of 1,000.
For a cutoff of 5, 250 patients are eligible, and their mean predicted benefit is 6 points per patient. The stricter rule raises the eligible mean from 4.25 to 6 points, but reduces the eligible count from 600 to 250. Neither number alone answers whether the cutoff is preferable; the decision also depends on the trial's development objective and whether enough eligible patients can be recruited.
Why Test Performance Alone Is Not Enough
A biomarker test may be summarized by sensitivity and specificity for a pathway-state category. Sensitivity is the probability of a positive test among people in the designated state; specificity is the probability of a negative test among people outside it. These quantities describe classification, not the distribution of treatment benefit within each state. To calculate mean benefit among test-positive patients, one must know how benefits are distributed among those who test positive.
A sufficient assumption for one common calculation is that, within each true-state group, test positivity is independent of predicted benefit. Equivalently for the mean calculation, the group-specific mean benefit among test-positive patients must equal the group-specific mean for the whole state group. This assumption may or may not be appropriate; it must be supported or replaced by direct joint predictions of state, test result, and benefit.
Worked Example: What Sensitivity and Specificity Can—and Cannot—Determine
Hypothetical data. Suppose 40% of a virtual population has high pathway state \(H\), and 60% has low state \(L\). The mean predicted benefit is 7 improvement points in \(H\) and 2 points in \(L\). A test has sensitivity 0.80 and specificity 0.90 for identifying \(H\). First suppose test positivity is independent of predicted benefit within each state.
The positive-test fraction is \(P(+)=P(H)P(+\mid H)+P(L)P(+\mid L)=0.40(0.80)+0.60(0.10)=0.38\). The false-positive probability in \(L\) is \(1-0.90=0.10\). Under the stated conditional-independence assumption, the positive patients' benefit total per member of the full population is \(0.40(0.80)(7)+0.60(0.10)(2)=2.24+0.12=2.36\) points. The positive-group mean is therefore \(2.36/0.38=6.21\) points per test-positive patient.
The calculation depends on that assumption. To see why, suppose instead that within \(H\), 20% of patients have predicted benefit 0 and 80% have benefit 8.75 points. Their state-group mean remains \(0.20(0)+0.80(8.75)=7\) points. Let the test be positive for all of the first subgroup and for 75% of the second. Sensitivity is still \(0.20(1)+0.80(0.75)=0.80\), but the mean benefit among positive patients within \(H\) is \([0.20(1)(0)+0.80(0.75)(8.75)]/0.80=5.25/0.80=6.5625\) points, not 7.
For simplicity, suppose all patients in \(L\) have benefit 2 points and 10% test positive. The overall positive fraction is still \(0.38\), but its benefit total per population member is now \(0.40(0.80)(6.5625)+0.60(0.10)(2)=2.10+0.12=2.22\) points. The positive-group mean is \(2.22/0.38=5.84\) points, rather than 6.21. Sensitivity, specificity, state prevalence, and state-group means have not changed; the within-state relationship between test positivity and benefit has. Thus those summary quantities alone do not determine the test-positive mean.
A Justification for the Weighted Calculation
The eligible mean is a conditional population average. The numerator adds each patient's benefit weighted by their representation in the target population, retaining only those who meet the rule. The denominator is the total weight of that eligible subset. Dividing therefore renormalizes the eligible weights to sum to one:
This also explains why the calculation should use joint patient-level or stratum-level predictions. If only separate summaries of biomarker frequency and state-specific mean benefit are available, their combination may silently assume that biomarker eligibility does not select patients with different benefits within a state. That assumption is not a mathematical consequence of the summaries.
State who the virtual population represents and define the direction, units, and time horizon of predicted benefit.
For each virtual patient or weighted stratum, retain the biomarker value, population weight, and predicted treatment benefit together.
Calculate the eligible fraction, eligible count for a specified population size, and weighted mean benefit for every cutoff.
Assess whether the eligible pool is recruitable and whether conclusions persist under plausible changes in model assumptions, population composition, and biomarker measurement.
Worked Example: Comparing Several Cutoffs in a Weighted Population
Hypothetical data. A virtual population represents 2,000 potential patients. Predicted benefit is measured in improvement points over 12 weeks. Each biomarker category is fully included or excluded by the candidate cutoffs.
| Biomarker category | Patients | Mean predicted benefit per patient |
|---|---|---|
| Below 1 | 800 | 1 point |
| 1 to below 3 | 600 | 2.5 points |
| 3 to below 6 | 400 | 5 points |
| 6 or above | 200 | 8 points |
With no enrichment, the population mean is \([800(1)+600(2.5)+400(5)+200(8)]/2{,}000=(800+1{,}500+2{,}000+1{,}600)/2{,}000=2.95\) points per patient.
At cutoff 1, 1,200 patients are eligible. Their predicted benefit total is \(600(2.5)+400(5)+200(8)=5{,}100\) points, so their mean is \(5{,}100/1{,}200=4.25\) points. At cutoff 3, 600 are eligible, with total \(400(5)+200(8)=3{,}600\) points and mean \(3{,}600/600=6\) points. At cutoff 6, 200 are eligible, with mean 8 points. The corresponding eligible fractions are 0.60, 0.30, and 0.10.
The cutoffs therefore trade pool size against predicted benefit per eligible patient. For example, cutoff 6 has the highest mean benefit but yields only 200 potential candidates in the represented population. Whether that is adequate depends on recruitment needs and on how well the virtual population represents the actual recruitment setting. The calculations describe predicted averages; they do not guarantee that the trial's realized benefit will match them.
Choosing and Checking an Enrichment Rule
A useful comparison includes the no-enrichment population and several clinically plausible cutoffs, not only the cutoff with the largest predicted mean. Report the eligible fraction, the expected number eligible in a stated target population, and the mean benefit on the same scale for every option. If the QSP model predicts a non-monotone relationship between biomarker and benefit, a simple one-direction threshold may discard people with substantial predicted benefit; inspect the patient-level predictions rather than presuming that more biomarker always means more benefit.
The main alternative to excluding biomarker-negative patients is to enroll a broader population and use the biomarker to characterize or stratify the enrolled group. That approach preserves access to a wider candidate pool but may yield a lower average predicted benefit if many enrollees are expected to benefit less. Enrichment concentrates the modeled population but narrows generalizability and may slow recruitment. These are distinct design choices; a favorable predicted mean alone does not resolve them.
Before using a cutoff, check whether the simulated biomarker distribution matches available information about the target population, whether the biomarker is measurable with acceptable reliability, and whether the model preserves the joint relationship between biomarker and treatment benefit. Examine sensitivity to threshold choice and to plausible changes in pathway-state prevalence. Small changes in cutoff that cause large changes in eligible fraction or mean benefit can signal an unstable rule, especially if few virtual patients lie near the boundary. Measurement error can also move real patients across the cutoff, so predictions based on error-free values may overstate enrichment.
Enrichment is most useful when predicted heterogeneity is large enough to matter, the biomarker rule can be implemented reliably, and the resulting eligible population remains feasible. Its limitations should remain visible: model predictions inherit uncertainty in structure, parameters, population composition, and translation from pathway state to clinical benefit. An enrollment rule can prioritize a population for evaluation, but it cannot by itself establish clinical treatment effects or eliminate the need to assess model credibility.
Check Your Understanding
Use the definitions and calculations in this tutorial. Benefits are positive improvements in the stated units.
- A cutoff admits 300 of 1,000 patients. Their weighted benefit total is 1,350 improvement points. What are the eligible fraction and eligible mean benefit?
- In a target population, 25% are in state \(H\) and 75% in state \(L\). A test has sensitivity 0.80 and specificity 0.92. Under conditional independence of test positivity and benefit within state, what fraction tests positive?
- For the population in question 2, mean benefits are 6 points in \(H\) and 1 point in \(L\). Under the same conditional-independence assumption, calculate mean benefit among test-positive patients.
- Explain why a cutoff can raise mean predicted benefit while reducing the number of eligible patients, and name one additional feasibility check needed before adopting it.
Solutions. (1) The eligible fraction is \(300/1{,}000=0.30\), or 30%; the eligible mean is \(1{,}350/300=4.50\) improvement points per eligible patient. (2) The false-positive probability is \(1-0.92=0.08\), so the positive fraction is \(0.25(0.80)+0.75(0.08)=0.20+0.06=0.26\), or 26%. (3) The positive-group benefit total per population member is \(0.25(0.80)(6)+0.75(0.08)(1)=1.20+0.06=1.26\) points. Dividing by the positive fraction gives \(1.26/0.26=4.85\) points per test-positive patient. The conditional-independence assumption is needed to use the state means in this calculation. (4) A stricter rule can remove lower-benefit patients from the average, increasing the eligible mean, while also reducing the eligible fraction. Check, for example, whether the remaining eligible pool is large enough and reachable enough to recruit the intended trial population.