Accelerated Clinical Development Planning

Accelerated CDPs: context and challenge

When are accelerated plans considered?

  • Strong clinical rationale (well understood MoA, target)
  • Strong unmet need (e.g. rare diseases)
  • Strategic considerations (competitive pressure)

How are CDPs accelerated?

By bypassing traditional steps:

  • Skipping Ph2a PoC →\rightarrow straight to Ph2b
  • Skipping Ph2b →\rightarrow straight from PoC to Ph3
  • Skipping Ph2 entirely

Quantifying risk

Acceleration entails risk — the key challenge for statisticians is to quantify and communicate the level of risk

Concepts

Frequentist metrics

In an idealized situation where we evaluate the treatment effect for the same population, endpoints, control arms at each stage …

Given we are starting stage xx, what is P(pass all remaining stages)P(\text{pass all remaining stages})?

We can answer this question only with “if-then” analyses, where the “if” is an assumption about the true treatment effect.

Shortcoming: how plausible are these scenarios?

Bayesian extension

Allows us to incorporate beliefs and uncertainty about treatment effect through a prior, which (in the idealized situation) can be easily updated as data accumulates

If we begin (before phase-2) with a uniform prior for the treatment effect (from zero to 150% of the TPP):

Example

Straight to Phase-3: how to evaluate risk?

  • Setting: Lifecycle management program going straight to Phase 3 — no separate proof-of-concept study
  • Rationale: Positive results in adjacent indications; accelerated timelines due to competitive pressure
  • Key risk: High uncertainty about the treatment effect at the pivotal decision point
  • Risk mitigation: Two parallel Phase-3 studies will include a “PoC-like” interim futility analysis, with a stringent boundary to discharge risk early

Two key risk metrics for decision makers:

  1. Today: what is the probability of success (PoS), given what we know?
  2. Tomorrow: if the program passes the futility interim, how much does this increase the PoS estimate?

PoS today and after futility

Assurance=∫P(θ̂int>c,θ̂fin∈𝒮|θ)π(θ)dθAssurancepost=AssuranceP(θ̂int>c) \text{Assurance} \;=\; \int P\!\Big(\hat\theta_{\text{int}} > c,\;\; \hat\theta_{\text{fin}} \in \mathcal{S} \;\Big|\; \theta\Big) \; \pi(\theta)\; d\theta \qquad\qquad \text{Assurance}_{\text{post}} \;=\; \frac{\text{Assurance}}{P\big(\hat\theta_{\text{int}} > c\big)}

with (θ̂int,θ̂fin)∣θ∼BVN((θ,θ),σint,σfin,ρ)\big(\hat\theta_{\text{int}},\, \hat\theta_{\text{fin}}\big) \mid \theta \;\sim\; \text{BVN}\!\big((\theta, \theta),\; \sigma_{\text{int}} , \sigma_{\text{fin}} , \rho\!\big), ρ=τ\;\rho = \sqrt\tau, futility boundary cc, and success region 𝒮\mathcal S.

PoS=Joint assurance for all Phase-3 trials×P(approval | all Phase-3 success)⏟separate assessment \text{PoS} \;=\; \text{Joint assurance for all Phase-3 trials} \; \times \underbrace{\text{P(approval | all Phase-3 success)}}_{\text{separate assessment}}

Ingredients:

  1. A suitable prior π(θ)\pi(\theta) for the treatment effect — difficult yet crucial when Phase 2 data is absent (next slide)

  2. Interim and final readouts share patients ⇒\Rightarrow correlated at square root of information fraction

  3. Monte Carlo evaluation: draw θ(m)∼π(θ)\theta^{(m)} \sim \pi(\theta), then jointly simulate (θ̂int(m),θ̂fin(m))\big(\hat\theta_{\text{int}}^{(m)},\, \hat\theta_{\text{fin}}^{(m)}\big) from the BVN

From MC draws to PoS:

  • PoS today = fraction of all MM draws where interim passes and final succeeds, and regulatory approval is obtained

  • PoS post futility = based on restriction to draws with passed futility — equivalent to PoS/P(pass)\text{PoS}/P(\text{pass}) (Dragalin 2026)

  • Passage of futility is a pre-posterior Bayesian update with a censored observation.

Calibrated benchmark priors

Problem: Before we’ve seen Phase-2 data — what prior should we use for the treatment effect?

Idea Hampson et al. (2022): Build a two-component mixture prior :

  1. Null component centered at zero (no effect)
  2. TPP component centered at the TPP

Calibration: choose the mixture weight ww so that the prior-predictive probability of a program succeeding across a standardized series of remaining trials matches the industry-wide historical success rate for similar programs.

Result for the example project

How much does the futility analysis de-risk? Try adjusting the design:

The fundamental trade-off

More stringent futility boundaries lower PoS today (“PoS loss”), but they raise conditional PoS (greater de-risking).
  • Stricter boundary →\rightarrow more risk discharged if the program continues, but more PoS lost overall
  • Quantifying this trade-off is the core contribution, to create transparency for decision makers

Value optimization

The natural extension: eNPV

%%{init: {'theme': 'base', 'themeVariables': {'primaryColor': '#f5f5f5', 'primaryTextColor': '#161616', 'primaryBorderColor': '#dadada', 'lineColor': '#161616', 'secondaryColor': '#dadada', 'tertiaryColor': '#fcfcfc', 'fontFamily': 'system-ui, sans-serif'}}}%%
flowchart LR
  A["Start Ph3<br/>Cost C₁"] --> B{"Pass futility?"}
  B -- "Yes · P(pass)" --> C["Continue<br/>Cost C₂"]
  B -- "No · 1−P(pass)" --> D["Stop early<br/>Save C₂ + C₃"]
  C --> E{"Final success?"}
  E -- "Yes · P(pivotal success|pass)" --> F["Submit<br/>Cost C₃"]
  E -- "No" --> G["Write-off<br/>C₁ + C₂"]
  F --> H{"Approved?"}
  H -- "Yes · P(approval|pivotal success)" --> I["Revenue V"]
  H -- "No" --> J["No approval<br/>Write-off C₁ + C₃"]

  • eNPV is an expectation of costs and revenue over this probability tree
  • Futility affects eNPV through two channels: cost savings from early stopping and reduced probability of reaching revenue
  • The same simulation framework that produces conditional PoS provides all stagewise inputs to eNPV

Takeaways

Key messages

Methodological

  1. Conditional assurance is the natural prospective metric for futility-gated programs — it handles uncertainty about θ\theta without unblinding

  2. The benchmark-calibrated prior is a principled fallback when Phase 2 data is absent; data sharpens it when available

  3. For straight-to-phase-3 accelerations with stringent de-risking via futility rules: joint simulation approach (BVN with ρ=τ\rho = \sqrt\tau) provides a complete picture: P(pass), conditional PoS, and overall PoS in a single framework

Practical

  1. No single metric suffices for communicating about risk — but PoS is central because it feeds both expectations and portfolio valuation

  2. Transparent quantification of trade-offs (PoS loss vs. risk discharged) enables informed design decisions across diverse stakeholders

Acknowledgements

  • Zhenwei Yang
  • Francesca Gasperoni
  • Jennifer Ng
  • Markus Lange
  • Björn Holzhauer
  • Joseph Kahn
  • Alex Przybylski
  • David Ohlssen
  • Nirav Ratia
  • Andrew Wright
  • Robert Grant
  • Steffen Ballerstedt
  • Insa Sommer
  • Juergen Ruth

References

Dragalin, Vladimir. 2026. “Informative Futility Rules Based on Conditional Assurance.” Statistics in Medicine 45 (1-2): e70330. https://doi.org/https://doi.org/10.1002/sim.70330.
Hampson, Lisa V., Björn Bornkamp, Björn Holzhauer, Joseph Kahn, Markus R. Lange, Wen-Lin Luo, Giovanni Della Cioppa, Kelvin Stott, and Steffen Ballerstedt. 2022. “Improving the Assessment of the Probability of Success in Late Stage Drug Development.” Pharmaceutical Statistics 21 (2): 439–59. https://doi.org/https://doi.org/10.1002/pst.2179.