Optimal designs for identifying effective doses in drug combination studies
summary
The gist
Optimal designs for identifying effective dose combinations that yield a prespecified effect are proposed to minimize the width of the confidence band of contour lines of interest.
In short
The paper proposes a new optimal design method to find drug dose combinations that produce a specific effect with high precision. It minimizes the variance of predictions at these effective doses using an Lq-norm criterion, outperforming traditional factorial and ray designs in simulations. This method offers more accurate estimation of effective dose combinations, potentially requiring fewer measurements.
Key concepts
- Nonlinear Response Surface Model
- This is a mathematical model used to describe how the effect of two drugs (C and D) changes based on their respective doses (c and d). The model includes terms for individual effects and an interaction term showing how the effect of one drug depends on the dose of the other, allowing for complex, non-straight-line relationships.
- Lq-norm Criterion
- This is a mathematical rule used to define what makes a design 'optimal.' The goal is to minimize this norm applied to the variance of predictions at the desired effective dose combinations. Minimizing this value directly leads to designs that provide the most precise estimates for those specific doses.
- MED-Optimal Design
- This is a specific type of optimal design derived from minimizing an Lq-norm criterion. It is designed to be highly efficient at pinpointing the effective dose combinations for a target effect size, often showing superior precision compared to standard designs like factorial or ray designs.
Terminology used across episodes
This episode discusses
The paper
Optimal designs for identifying effective doses in drug combination studies · Read on arXiv
Leonie Schürmeyer, Ludger Sandig, Jorrit Öhne, Leonie Theresa Hezler, Bernd-Wolfgang Igl, Kirsten Schorning
TU Dortmund University · Boehringer Ingelheim Pharma GmbH & Co. KG
We consider the optimal design problem for identifying effective dose combinations within drug combination studies where the effect of the combination of two drugs is investigated. Drug combination studies are becoming increasingly important as they investigate potential interaction effects rather than the individual impacts of the drugs. In this situation, identifying effective dose combinations that yield a prespecified effect is of special interest. If nonlinear surface models are used to describe the dose combination-response relationship, these effective dose combinations result in specific contour lines of the fitted response model. We propose a novel design criterion that targets the precise estimation of these effective dose combinations. In particular, an optimal design minimizes the width of the confidence band of the contour lines of interest. Optimal design theory is developed for this problem, including equivalence theorems and efficiency bounds. The performance of the optimal design is illustrated in different examples modeling dose combination data by various nonlinear surface models. It is demonstrated that the proposed optimal design for identifying effective dose combinations yields a more precise estimation of the effective dose combinations than ray or factorial designs, which are commonly used in practice. This particularly holds true for a case study motivated by data from an oncological dose combination study.
DOI: 10.1002/bimj.70186
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Optimal designs for identifying effective doses in drug combination studies".
Jane: Here's the gist of the paper. Optimal designs for identifying effective dose combinations that yield a prespecified effect are proposed to minimize the width of the confidence band of contour lines of interest.
Tom: This novel design criterion targets the precise estimation of these effective dose combinations by minimizing the Lq-norm of the variance of the prediction for the effect at those dose combinations.
Jane: Modeling Dose Combination Data The paper considers modeling the effect of two substances C and D using a nonlinear response surface model defined by Yij = η((ci, di), θ) + εij.
Tom: The regression function η is modeled additively including an interaction effect with the form η((c, d), θ) = θ0 + ηC (c, θC) + ηD(d, θD) + γηC (c, θC)ηD(d, θD).
Paper summary: Tom: So we're looking at this paper called "Optimal designs for identifying effective doses in drug combination studies". It tackles the problem of finding the best way to design experiments when you want to find dose combinations that get a specific effect from two different substances.
Jane: That sounds really practical, Tom. So, what's the main idea here? What's this paper actually claiming about designing those experiments?
Tom: The core thesis is about using an optimal design criterion to find those effective dose combinations more precisely. They want to minimize the width of the confidence band around those lines you're interested in, which basically means they're aiming for a really tight estimate of where that effect happens.
Lu: What they're doing is targeting the precise estimation of these effective dose combinations by minimizing the Lq-norm of the variance for that effect at those dose combinations. It’s about making sure you pinpoint those doses as accurately as possible, no matter what nonlinear model you use for your data.
Meng: So, if I'm running a trial and I know I need to find a certain synergy between drug C and drug D, this design method helps me narrow down where to test the doses most efficiently? How does it compare to the standard stuff?
Tom: The paper shows that this locally MED-optimal design outperforms established designs like factorial or ray designs when it comes to precision for estimating those effective dose combinations. The simulation study backs that up, showing the locally MED-optimal design has smaller root mean square error values and therefore higher precision for prediction at the set of effective dose combinations.
Jane: That's interesting because traditionally, we use those factorial or ray designs because they’re straightforward to set up, but this paper suggests they aren't always as accurate when you're dealing with nonlinear response surfaces. What kind of models are they looking at?
Lu: They model the effect of two substances C and D using a nonlinear response surface model defined by Yij equals eta parentheses c i d i, theta plus epsilon i j, where eta is modeled additively including an interaction effect with the form eta parentheses c d, theta). They even consider different functions for modeling how each substance responds individually, like linear or Emax models.
Meng: So they're not just assuming one simple way to model things; they're trying to be flexible with the underlying science, which is good for real-world drug interaction studies. But what is the actual mechanism behind this criterion? What makes it better than just picking a random design?
Paper summary: Tom: The core of the proposed criterion involves minimizing an Lq-norm of a function phi, where q is some number between one and infinity. This minimization happens over all approximate designs on the design space, subject to a constraint that the gradient of eta parentheses c zero d zero, theta) is within the range of the information matrix m(xi, theta).
Jane: That sounds mathematically intensive. So what does that constraint actually mean for someone who isn't deep in optimization? What's it telling us about choosing a design?
Lu: They use an equivalence theorem as a central tool to check if the designs they find numerically are truly optimal. It’s a way to verify the numerical solutions you get from the minimization process.
Meng: If this is about minimizing variance, what does that translate to for deployment? Does it mean fewer total experiments or just better results from the same number of experiments? I need to know if this saves time or just gives me cleaner data.
Tom: The paper suggests using this MED-optimal design allows for more precise estimation of effective concentrations compared to other designs, and it can do that even when alternative designs would require a larger number of observations to achieve the same level of precision.
Jane: So, we're talking about getting high accuracy with potentially smaller sample sizes. That ties into the 3R framework, which is important for research and animal experiments. What about uncertainty? If you're unsure about the interaction effect itself, how does this approach handle that?
Lu: They propose a robust version of the criterion called the Bayesian MED-optimal design. This accounts for uncertainty about the interaction effect by incorporating a prior distribution of that unknown parameter theta. That makes it more robust when prior knowledge about the interaction is not very informative.
Meng: A Bayesian approach sounds safer, but does it make things computationally expensive for production? How much extra overhead are we talking about when we add that prior distribution to the optimization?
Tom: The simulation study on robustness shows that both locally MED-optimal designs and the Bayesian MED design show a similar ability to estimate those contour lines for p one ten and p two thirty in a setting with an additive model. The Bayesian version performs nearly as well as the locally optimal design in that context.
Paper summary: Jane: So, it seems like they've got a solid method for handling both the precision and the uncertainty, even when things get complex with different interaction types. What's the final word on what this means practically for people doing these kinds of studies?
Tom: The paper concludes that using this MED-optimal design allows for more precise estimation of effective concentrations compared to other designs because alternative designs would need a larger number of observations to get the same precision. It also highlights that even in one dimensional cases, the confidence interval for those effective doses can be quite large in practical settings because there are various uncontrollable factors involved.
Lu: They emphasize that this approach is very flexible, allowing adaptation to any nonlinear surface model. The criterion itself is based on an Lq-norm where q is fixed beforehand. For example, they found that q equals two outperformed q equals one in their analysis of the optimal designs for various selected values of q.
Meng: So it's adaptable to different scientific models, which is good for flexibility, but we still have to pick that fixed value of q upfront. That’s a design choice we have to make before we even start the experiment.
Jane: It sounds like this method offers a way forward for making dose combination studies more efficient and reliable by focusing on minimizing the asymptotic variance of those effective dose combinations, which means less variability in how precisely you estimate the doses you need.
Tom: So, to wrap up on "Optimal designs for identifying effective doses in drug combination studies," this paper proposes a design approach that uses an Lq-norm criterion to find the best way to identify effective dose combinations within drug combination studies where two drugs interact.
Jane: And it claims this method provides higher precision than traditional methods like factorial or ray designs when estimating those effective dose combinations, and it does so with potentially fewer measurements.
Lu: The paper also suggests that this framework can handle uncertainty about the interaction effect using a Bayesian approach, making it more robust when prior information is limited.
Meng: From an engineering standpoint, the main point is that this design method gives you higher precision for your predictions and helps you understand how much data you actually need to collect.
Tom: It’s about moving beyond just picking a fixed design type and instead finding a criterion that adapts to any nonlinear surface model while aiming for the tightest possible estimate of those effective dose combinations.
Conclusion: Tom: So, we've been looking at this paper called "Optimal designs for identifying effective doses in drug combination studies" by Schürmeyer and her team, and it boils down to finding the best experimental design when you're testing two drugs together.
Jane: It’s about using a specific math rule, an Lq-norm criterion, to figure out exactly where the optimal dose combinations are so you don't waste time testing things that won't show any effect.
Tom: They show this approach beats traditional designs like factorial or ray designs when it comes to getting a tighter estimate of those effective dose combinations.
Jane: So, what does that mean for the actual science? If you use this method, are you getting a better picture of how those drugs work together than if you just run a standard experiment?
Tom: It means the precision for your predictions is higher because the design is specifically optimized to minimize the error in estimating those dose combinations.
Jane: That sounds useful for drug discovery because it helps you narrow down which combinations actually matter, even when your response surface model is complex and nonlinear.
Tom: Exactly. The paper also introduces a way to make this design more robust by using a Bayesian version that accounts for uncertainty about the interaction effect itself, which is important when you don't have perfect prior knowledge.
Jane: So, while the main point is better precision, they’ve also thought about making it safer when the underlying scientific assumptions aren't crystal clear.
Tom: Right. It’s not just a new way to run an experiment; it’s a smarter way to choose the experiment itself based on what you already know about how these substances interact.
Jane: It definitely points toward more efficient research, potentially letting us get better results with smaller sample sizes for animal studies or human trials.
Tom: It suggests that instead of sticking to old methods blindly, you can use this criterion to adapt your experimental plan to any nonlinear model you throw at it.
Jane: It’s a flexible tool for figuring out those precise dose combinations that lead to the desired effect size. Next up, we're going to look at how they actually tested this on different types of drug data.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates