Backfill Bayesian Ordered Lattice Design for Phase I Clinical Trials

This paper proposes the Backfill Bayesian Ordered Lattice Design (BF-BOLD), a straightforward Phase I clinical trial method that integrates activity evaluation into dose-finding to identify an optimal biological dose, thereby enhancing safety, reducing overdose rates, and improving the selection of the Recommended Phase II Dose without requiring complex statistical modeling.

The Gist

Imagine you are a chef trying to find the perfect recipe for a new, spicy dish. You have a range of spice levels from "mild" to "explosive." Your goal is to find the Maximum Tolerable Dose (MTD): the spiciest level a person can eat without getting sick (the "toxicity" limit).

In traditional cooking trials, chefs would taste a little bit, see if the diner gets a stomach ache, and then move up or down the spice ladder. But there's a catch: just because a dish isn't making people sick doesn't mean it's the best dish. Maybe the "mild" version is actually just as delicious (effective) as the "medium-hot" version, but with fewer side effects.

This is the problem the authors of this paper are solving. They are proposing a new way to run these "taste tests" (clinical trials) that doesn't just look for safety, but also finds the Optimal Biological Dose (OBD): the sweet spot where the drug works best without being too harsh.

Here is how their new method, called BF-BOLD, works, explained through simple analogies:

1. The Old Way vs. The New Way

• The Old Way (Standard Phase I): Imagine you are climbing a ladder to find the highest rung you can stand on without falling. You go up rung by rung. Once you find the highest safe rung, you stop. You assume that's the best rung to stand on. But what if the rung right below it is just as stable, but much more comfortable? The old method might miss that.
• The New Way (BF-BOLD): This method still climbs the ladder to find the highest safe rung (the MTD). But, while you are waiting for the results of the current rung, you send a few extra people down to try the lower rungs. This is called "Backfilling."

2. What is "Backfilling"?

Think of a busy restaurant kitchen. The head chef is testing the "Super Spicy" sauce (the current dose). While they wait to see if the first customer gets a stomach ache, the sous-chef takes a moment to serve the "Medium Spicy" sauce to a few other customers who are already waiting.

• Why do this? Because the "Medium Spicy" sauce is likely safe (since it's lower than what the chef is currently testing).
• The Benefit: By the time the chef finishes testing the "Super Spicy" sauce, they already have data on how the "Medium Spicy" sauce tastes. They can compare them immediately.

3. Finding the "Sweet Spot" (The OBD)

Once the trial is over, the team looks at two things:

1. Safety: Which was the highest dose that didn't make people sick? (The MTD).
2. Effectiveness: Did the lower doses work just as well?

If the "Medium Spicy" sauce tastes just as good as the "Super Spicy" one, but has fewer side effects, the team picks the "Medium Spicy" one as the winner. This is the Optimal Biological Dose (OBD). It's not just about surviving the dose; it's about finding the dose that gives the best benefit with the least risk.

4. Why is this better than the old way?

The paper compares their method to other methods (like BOIN or the old "3+3" design) and even to a two-step process (testing safety first, then testing effectiveness later in a separate trial).

• Safety First: Because they are testing lower doses while waiting for higher dose results, they catch safety issues faster. It's like having a safety net that catches you before you fall too far.
• No Wasted Time: Instead of finishing the safety trial, stopping, and then starting a whole new trial to test effectiveness, they do it all at once. It's like testing the engine and the brakes of a car simultaneously rather than one after the other.
• Avoiding the "Overdose" Trap: Sometimes, traditional methods pick a dose that is too high. This new method is smarter about backing off if the data suggests a lower dose is just as good.

5. The "Lattice" Concept

The authors use a fancy word, "Lattice," but imagine it as a staircase.

• In a normal staircase, you can only go up or down one step at a time.
• In their "Ordered Lattice," they have a map that helps them decide: "If Step 3 is too shaky, we know Steps 4, 5, and 6 are definitely too shaky too." This helps them make decisions faster and more logically.

The Bottom Line

This paper introduces a smarter, more efficient way to test new cancer drugs. Instead of just asking, "How much can we give before it hurts?", it asks, "What is the most effective amount we can give that doesn't hurt?"

By using "backfilling" (testing lower doses while waiting for higher dose results), they can find the perfect balance between safety and power, saving time, money, and most importantly, protecting patients from unnecessary side effects. It's like finding the perfect volume setting on a radio: loud enough to hear clearly, but not so loud that it hurts your ears.

Technical Summary

1. Problem Statement

Traditional Phase I clinical trials focus primarily on identifying the Maximum Tolerated Dose (MTD)—the highest dose with an acceptable rate of dose-limiting toxicity (DLT). However, this paradigm faces significant challenges in the era of molecularly targeted agents and immunotherapies, where toxicity often plateaus while efficacy may not increase linearly with dose.

• Limitations of Current Practice: Standard designs often select the MTD as the Recommended Phase II Dose (RP2D), which may be overly toxic or suboptimal for long-term treatment sustainability.
• Regulatory Context: The US FDA's "Project Optimus" calls for a reform in dose optimization, shifting focus from merely finding the MTD to identifying an Optimal Biological Dose (OBD). The OBD balances therapeutic activity with tolerability, often residing at a dose level lower than the MTD.
• The Gap: Conventional Phase I trials lack the sample size and design structure to simultaneously assess toxicity and activity (efficacy) across multiple dose levels. Traditional approaches often require separate, small-scale dose-expansion trials (Phase 1b) to evaluate activity, which delays development and may lock in an incorrect dose too early.

2. Methodology: BF-BOLD Design

The authors propose the Backfill Bayesian Ordered Lattice Design (BF-BOLD), an extension of the existing Bayesian Ordered Lattice Design (BOLD). This method integrates a "backfilling" strategy to evaluate drug activity at lower dose levels while the trial is simultaneously escalating to find the MTD.

Core Components:

• Dual Objectives: The design simultaneously estimates the DLT rate ($\pi_j$) and the activity rate ($\pi^a_j$) for each dose level $j$. Both are treated as binary outcomes.
• The Backfill Mechanism:
  • While the trial escalates doses to find the MTD, patients are "backfilled" into lower dose levels (specifically the "anti-cover" dose, i.e., the dose immediately below the current dose) to gather activity data.
  • Safety Constraints: Backfilling only occurs if the lower dose has demonstrated sufficient safety (via Conjugate Posterior Probability of Activity being Above Target, $CPAT^a$) and if the current dose is not deemed excessively toxic.
  • Eligibility: A dose is eligible for backfill if the probability of its activity rate exceeding a minimum threshold ($\phi_a$) is above a specific cutoff (default $\gamma_a = 0.2$).
• Statistical Framework:
  • Priors: Beta distributions are used for both toxicity and activity rates. Non-informative priors are set with a Prior Effective Sample Size (PESS) matching standard cohort sizes.
  • Order Constraints: The Pool-Adjacent-Violators Algorithm (PAVA) is applied to enforce monotonicity (toxicity increases with dose; activity is assumed to plateau or increase). This ensures estimates are isotonic.
  • Dose Selection: Dose escalation/de-escalation decisions are driven by the PPAT (PAVA-adjusted Posterior Probability of DLT rate being Above Target), aiming to keep the probability close to a target threshold $\tau$ (default 0.5).
• Stopping Rules & Identification:
  • MTD Identification: Selected based on the dose with the PAVA-adjusted posterior mean DLT rate closest to the target toxicity rate ($\phi$).
  • OBD Identification: After the MTD is identified, the OBD is selected from doses $\le$ MTD. The criteria are:
    1. Activity rate $\ge$ minimum acceptable threshold ($\phi_a$).
    2. Activity rate is within a pre-specified "trade-off" margin (e.g., 10%) of the MTD's activity rate.
    3. The lowest dose satisfying these conditions is chosen.

3. Key Contributions

• Integration of Activity into Phase I: BF-BOLD allows for the simultaneous assessment of toxicity and activity within a single Phase I trial, eliminating the need for a separate, sequential dose-expansion phase in many scenarios.
• Optimal Biological Dose (OBD) Framework: It provides a rigorous statistical procedure to identify a dose that maximizes the therapeutic window (activity vs. toxicity) rather than just the safety limit.
• Overdose Control: The design inherently reduces overdose rates. By backfilling lower doses and utilizing the BOLD overdose control parameter ($\tau$), the probability of treating patients with doses significantly higher than the true MTD is minimized.
• High Specificity for Futility: The design demonstrates high specificity in detecting when no OBD exists (i.e., when activity rates are below the acceptable threshold for all safe doses), potentially saving resources by halting futile development early.
• Computational Efficiency: Unlike complex simulation-based adaptive designs, BF-BOLD relies on conjugate Bayesian updates and PAVA, making it computationally fast and straightforward to implement without extensive pre-trial simulations.

4. Simulation Results

The authors conducted extensive simulations (using R) comparing BF-BOLD against standard BOLD, BF-BOIN (Backfill BOIN), and a hybrid Phase 1a/1b design (BOLD-1b).

• Accuracy: BF-BOLD achieved high accuracy in identifying both the MTD and OBD.
  • MTD accuracy remained stable regardless of the "gap" (distance) between the true MTD and true OBD.
  • OBD accuracy improved as the number of backfilled patients increased (1–3 patients per stage).
• Safety:
  • Overdose Rates: BF-BOLD significantly reduced overdose rates compared to standard BOLD and the BOLD-1b design.
  • Specificity: When no OBD existed (activity rates were too low), BF-BOLD correctly identified the absence of an OBD with high specificity (0.93–0.95 for $N_{max}=21/30$).
• Comparative Performance:
  • vs. BF-BOIN: BF-BOLD generally outperformed BF-BOIN in MTD accuracy, particularly when the discrimination between dose levels was high. BF-BOLD required slightly fewer patients on average.
  • vs. BOLD-1b (Phase 1a + Expansion): BF-BOLD outperformed the sequential 1a/1b design in both MTD and OBD accuracy across various scenarios. It achieved similar total patient counts but with better dose selection flexibility.
  • Exception: BOLD-1b performed slightly better only in the "bottom state" scenario (where all doses are too toxic), as the concentrated expansion phase helped correct initial misclassifications.

5. Significance and Conclusion

The BF-BOLD design represents a paradigm shift aligned with the FDA's Project Optimus. By integrating backfilling into a Bayesian lattice framework, it addresses the critical need to find the Optimal Biological Dose rather than just the Maximum Tolerated Dose.

• Clinical Impact: It enhances patient safety by reducing the likelihood of overdosing and improves treatment sustainability by identifying lower, safer doses that maintain efficacy.
• Efficiency: It streamlines drug development by collapsing the toxicity and activity assessment phases, potentially reducing the time and cost associated with traditional Phase 1b expansion studies.
• Practicality: The method is computationally simple, adaptable to various trial settings (e.g., neoadjuvant trials where early efficacy markers are available), and provides a robust statistical basis for decision-making in early-phase oncology trials.

In summary, BF-BOLD offers a flexible, safe, and statistically rigorous approach to modernizing Phase I trials, ensuring that the transition to Phase II is based on a dose that is not only safe but also optimally effective.