Combination Dose-Finding Designs: A Practical Guide

Introduction

Early-phase trials evaluating two drugs in combination face a problem that does not arise in the same form for single-agent dose finding: there is no single ordered list of doses. Instead, investigators must learn about a two-dimensional dose-toxicity surface.

Suppose Drug A has three candidate dose levels and Drug B has three candidate dose levels. The trial therefore contains nine possible dose combinations. Increasing Drug A while holding Drug B fixed is one type of escalation; increasing Drug B while holding Drug A fixed is another.

The objective is usually not to identify the individually highest tolerated dose of each drug. Rather, the goal is to identify one or more dose combinations whose probability of dose-limiting toxicity (DLT) is close to a prespecified target while retaining clinically useful exposure to both agents.

Key idea: Combination dose finding is a two-dimensional decision problem. A dose combination can be acceptable even when neither drug is at its single-agent maximum, and the safest path through the dose matrix depends on the joint toxicity of the two agents.

Why Combination Dose Finding Is Different

For a single agent, dose levels can usually be ordered from lowest to highest:

d_1<d_2<\cdots<d_K

A conventional escalation design can therefore ask a relatively simple question: should the next cohort move from \(d_k\) to \(d_{k+1}\), remain at \(d_k\), or de-escalate?

With two agents, the candidate doses form a grid:

	Drug B: B1	Drug B: B2	Drug B: B3
Drug A: A1	(A1,B1)	(A1,B2)	(A1,B3)
Drug A: A2	(A2,B1)	(A2,B2)	(A2,B3)
Drug A: A3	(A3,B1)	(A3,B2)	(A3,B3)

There is no unique linear ordering of these nine combinations. For example, \((A2,B2)\) and \((A3,B1)\) are not naturally ordered without additional assumptions about their relative toxicity.

The Basic Dose-Finding Objective

Let \(p_{ij}\) denote the probability of a DLT when Drug A is given at dose level \(i\) and Drug B at dose level \(j\):

p_{ij}=P(\text{DLT}\mid A_i,B_j)

Let \(\phi\) denote the target DLT probability. A common early-phase objective is to identify combinations for which:

p_{ij}\approx\phi

For illustration, suppose the protocol specifies a target DLT probability of 25%:

\phi=0.25

The dose-finding problem is then to learn which combinations have toxicity probabilities near 0.25 while avoiding combinations that are substantially more toxic.

Important: The target toxicity probability is a design parameter, not a statistical estimate. It should be selected from clinical considerations, disease severity, available alternatives, treatment intent, and the consequences of toxicity.

The Dose-Combination Matrix

A useful first step is to explicitly define the dose matrix before the study begins. For example:

Combination	Drug A	Drug B	Conceptual position
C1	A1	B1	Lowest combination
C2	A1	B2	Increase B
C3	A1	B3	Highest B at A1
C4	A2	B1	Increase A
C5	A2	B2	Middle combination
C6	A2	B3	Higher combination
C7	A3	B1	Highest A at B1
C8	A3	B2	Higher combination
C9	A3	B3	Highest combination

Monotonicity: A Critical Assumption

Most combination dose-finding designs use some form of monotonicity assumption. The basic clinical expectation is that increasing either drug, while holding the other fixed, should not decrease toxicity.

In mathematical terms, one often expects:

p_{i+1,j}\ge p_{ij}\qquad\text{and}\qquad p_{i,j+1}\ge p_{ij}

This assumption does not mean that toxicity must increase by the same amount in every direction. It means that moving upward in either dose dimension should generally not make the treatment safer.

The assumption becomes especially important when data are sparse, because the trial may have observed only a small fraction of the possible dose combinations.

Practical implication: If the scientific setting makes monotonicity implausible—for example, because of unusual pharmacologic interactions—the design and model should explicitly account for that possibility rather than treating monotonicity as automatic.

What Makes a Combination an MTD?

The term maximum tolerated dose (MTD) can be more complicated for combinations than for single agents. In a two-dimensional matrix, there may be several combinations with toxicity near the target.

A useful conceptual definition is a set of dose combinations whose toxicity probabilities are sufficiently close to the target and that are not dominated by safer combinations offering greater or similar doses.

Thus, the result of a combination dose-finding trial may be a set of recommended combinations rather than one unique MTD.

How Escalation Works

At any point in the trial, the observed DLT data are used to determine which combinations remain plausible and which combinations should be avoided.

Start at a prespecified low combination: enroll patients at a combination judged sufficiently safe for initial treatment.

Observe DLTs: collect the prespecified safety endpoint during the dose-limiting window.

Update the dose-toxicity information: use the accumulated data to classify or estimate candidate combinations.

Select the next combination: escalate, stay, or de-escalate according to the prespecified design rules.

Continue until the stopping criterion is met: the trial may stop when adequate information has been collected or when the relevant dose combinations have been identified.

Worked Example: A 3 × 3 Dose Matrix

Consider a hypothetical Phase I trial evaluating Drug A and Drug B in combination. Each drug has three dose levels, producing nine candidate combinations.

Planning parameter	Value
Drug A dose levels	A1, A2, A3
Drug B dose levels	B1, B2, B3
Candidate combinations	9
Target DLT probability \(\phi\)	25%
Starting combination	(A1, B1)
Cohort size	3 patients

Step 1: Start at the Lowest Combination

The first cohort receives \((A1,B1)\). Suppose none of the three patients experiences a DLT.

x=0,\qquad n=3,\qquad \hat p=\frac{0}{3}=0

The observed toxicity is below the 25% target, so the lowest combination does not provide evidence that the combination is excessively toxic.

Step 2: Escalate in One Dimension

The next combination might increase Drug B while keeping Drug A fixed:

(A1,B1)\longrightarrow(A1,B2)

Suppose the next three patients produce one DLT:

x=1,\qquad n=3,\qquad \hat p=\frac{1}{3}=0.333

The observed rate is above the 25% target, but three patients provide very little information. A modern dose-finding design would not generally treat \(1/3\) as proof that the combination is unacceptable. Instead, the prespecified decision rule would account for the uncertainty created by the small sample.

Step 3: Use the Accumulated Information

Suppose additional patients are treated at \((A1,B2)\), producing a total of two DLTs among six patients:

x=2,\qquad n=6,\qquad \hat p=\frac{2}{6}=0.333

The observed toxicity remains above the target. If the design concludes that \((A1,B2)\) is too toxic, that combination—and combinations expected to be at least as toxic—may be excluded from further treatment.

Notice the importance of the two-dimensional structure. If \((A1,B2)\) is considered too toxic, the design does not necessarily conclude that \((A2,B1)\) is too toxic. These are different combinations and may have different toxicity probabilities.

Step 4: Move Along a Different Path

The trial may instead explore \((A2,B1)\). Suppose the observed data at this combination are one DLT among six patients:

x=1,\qquad n=6,\qquad \hat p=\frac{1}{6}=0.167

This combination appears less toxic than \((A1,B2)\) in this hypothetical dataset. The design may therefore continue upward from \((A2,B1)\), subject to its prespecified escalation rules.

The key lesson: In a combination trial, there may be multiple routes through the dose matrix. The observed safety of one combination does not automatically determine the safety of another combination unless the statistical model explicitly connects them.

Model-Assisted Combination Designs

Model-assisted methods attempt to obtain the advantages of a statistical dose-toxicity model while retaining relatively transparent decision rules. One important example is BOIN-combo, which extends the Bayesian optimal interval design framework to dose combinations.

The basic BOIN philosophy is to define an interval around the target toxicity probability. A combination is treated as too safe, approximately at the target, or too toxic according to prespecified boundaries.

For a target \(\phi\), one can conceptually define:

\lambda_1<\phi<\lambda_2

where \(\lambda_1\) and \(\lambda_2\) are the lower and upper toxicity boundaries used by the decision rule.

Estimated toxicity relative to interval	Conceptual action
Below \(\lambda_1\)	Combination appears too safe → consider escalation
Between \(\lambda_1\) and \(\lambda_2\)	Combination is near target → remain or consider selection
Above \(\lambda_2\)	Combination appears too toxic → de-escalate or eliminate

For a true combination design, these decisions are applied within the two-dimensional dose matrix rather than to a simple one-dimensional sequence.

Model-Based Combination Designs

Model-based methods specify a mathematical model for the probability of DLT across the dose matrix and update the model as data accumulate.

A generic logistic formulation might be written as:

\operatorname{logit}(p_{ij}) = \beta_0+\beta_A z_i+\beta_B w_j+\beta_{AB}z_iw_j

where \(z_i\) and \(w_j\) represent scaled dose levels and \(\beta_{AB}\) represents an interaction term.

The interaction term is important because the effect of one drug may depend on the dose of the other. A model without an interaction term assumes a particular form of additivity on the chosen model scale.

Model caution: A more complicated model is not automatically a better model. With only a small number of patients and few observed combinations, interaction parameters can be weakly identified. The model must therefore be compatible with the amount and quality of information expected during the trial.

Copula and Other Surface Models

Another class of approaches models the joint dose-toxicity surface using structures designed specifically for combination trials. Copula-based approaches, for example, can model the marginal toxicity behavior of each agent together with their dependence or interaction structure.

The conceptual goal is the same: estimate or rank the toxicity of candidate combinations while borrowing information across the dose matrix.

The advantage of a surface model is that information from one combination can inform expectations about neighboring combinations. The trade-off is that the operating characteristics depend on the assumptions built into the model.

Why Simple 3+3 Is Challenging for Combinations

The traditional 3+3 design is easy to explain for a single ordered dose sequence, but its logic does not naturally solve the two-dimensional problem.

With combinations, investigators must decide:

Which dimension should be escalated?
Can both agents be increased simultaneously?
How should diagonal moves be treated?
What happens when one combination is too toxic but a neighboring combination may still be appropriate?
How should information be borrowed between combinations?
How are multiple acceptable combinations selected at the end?

Simply applying 3+3 independently to each drug can discard important information about the joint toxicity of the combination and can lead to inefficient exploration of the dose matrix.

De-Escalation and Elimination

A critical safety feature of modern combination designs is the ability to eliminate combinations that are sufficiently likely to exceed the target toxicity level.

Conceptually, if a combination has:

P(p_{ij}>\phi\mid\text{data})\text{ sufficiently large}

then the design may classify the combination as overly toxic.

Under a monotonicity assumption, combinations that are higher in both dose dimensions than a known overly toxic combination can also be excluded. This creates a useful toxicity contour across the dose matrix.

The MTD Contour

Imagine the true toxicity surface contains the following conceptual pattern:

	B1	B2	B3
A1	Low	Near target	High
A2	Near target	Near target	High
A3	Near target	High	Very high

The combinations labeled "near target" form an approximate MTD contour. There may be several clinically reasonable choices rather than one uniquely correct dose.

This is one of the most important conceptual differences between single-agent and combination dose finding: the answer may be a curve or set of combinations, not a single dose.

Worked Example: Interpreting a Candidate Final Matrix

Suppose the final model-assisted analysis produces the following estimated toxicity probabilities:

	B1	B2	B3
A1	0.08	0.19	0.34
A2	0.16	0.24	0.41
A3	0.25	0.39	0.57

With a target of 25%, the combinations closest to the target include \((A2,B2)\) and \((A3,B1)\). Both have estimated toxicity of approximately 25%.

However, they are not necessarily interchangeable. The final recommendation may depend on efficacy, pharmacokinetics, pharmacodynamics, convenience, cumulative exposure, drug-drug interactions, and other clinical considerations.

Important: Dose finding establishes a safety-based dose recommendation. It does not by itself establish that the selected combination has the best efficacy, tolerability, or benefit-risk profile.

Safety vs. Efficacy

Combination dose-finding studies are usually driven primarily by toxicity, but the ultimate goal is a clinically useful dose. A dose that is perfectly aligned with the toxicity target may not be the preferred dose if a lower combination provides similar efficacy with substantially better tolerability.

Therefore, later-stage dose optimization may consider:

DLTs and other adverse events
Exposure and pharmacokinetic measurements
Pharmacodynamic biomarkers
Preliminary antitumor or disease-response activity
Drug-drug interactions
Dose intensity and treatment interruptions
Patient convenience and administration burden

Operating Characteristics

Simulation is essential when evaluating a combination dose-finding design. Because the true toxicity surface is unknown, the design should be tested under many plausible surfaces rather than only one assumed scenario.

Important operating characteristics include:

Operating characteristic	Question answered
Probability of selecting the target combination	How often does the design identify an appropriate combination?
Probability of selecting an overly toxic combination	How often does the design make an unsafe final recommendation?
Patient allocation to overly toxic doses	How much exposure occurs above the target?
Patient allocation to subtherapeutic doses	How much trial capacity is spent at doses that are too safe?
Average number of DLTs	How much toxicity is expected during dose finding?
Number of combinations explored	How efficiently does the design learn the dose surface?

Simulation of a Combination Design

Suppose the investigators want to compare two candidate designs. They can define several plausible true toxicity matrices and repeatedly simulate complete trials under each matrix.

For each simulated trial, record:

The sequence of combinations assigned.
The number of patients treated at each combination.
The number of DLTs observed.
The final recommended combination or combination set.
Whether the final recommendation is near the true target contour.

After thousands of simulated trials, the investigators can estimate probabilities such as:

P(\text{select a target combination})

and:

P(\text{select an overly toxic combination})

Why simulation matters: A design can look reasonable on paper but perform poorly when the true toxicity surface differs from the assumptions used to construct it. Simulation reveals those weaknesses before patients are enrolled.

Common Mistakes

Treating the dose matrix as one linear sequence. Two-dimensional combinations generally cannot be ranked without assumptions about their relative toxicity.
Escalating both drugs automatically. A diagonal move can produce a much larger increase in combination exposure than a one-dimensional escalation.
Ignoring interaction. The joint effect of two agents may differ from what would be expected from their individual effects.
Using single-agent MTDs as the combination starting point. The combination may have substantially different toxicity because of pharmacologic interaction.
Selecting one MTD when several combinations are reasonable. A final recommendation may appropriately consist of an MTD contour or several clinically prioritized combinations.
Evaluating only the final selection probability. Patient exposure to overly toxic combinations is also an important safety operating characteristic.
Failing to simulate unusual toxicity surfaces. The design should be challenged under scenarios with strong interaction, weak interaction, steep toxicity gradients, and broad or narrow MTD contours.
Changing the design rules after observing the data. Adaptive decisions should be specified in advance or handled through an appropriate prespecified adaptation framework.
Confusing dose finding with dose optimization. The safest combination near the target toxicity probability is not automatically the combination with the best clinical benefit-risk profile.

Combination Dose Finding vs. Single-Agent Dose Finding

Feature	Single agent	Combination
Dose structure	Usually one-dimensional	Two- or higher-dimensional
Candidate treatments	Ordered dose levels	Dose-combination matrix
Escalation	Usually up/down one sequence	Multiple possible directions
Interaction	Not applicable between agents	Potentially critical
Final target	Often one MTD	May be an MTD contour or set
Simulation burden	Moderate	Often substantially greater

Design Selection: Which Approach?

No single combination dose-finding design is universally best. The appropriate approach depends on the amount of prior information, the number of dose levels, the plausibility of monotonicity, the expected strength of interaction, operational constraints, and the importance of model transparency.

Approach	Strength	Trade-off
Rule-based	Simple and transparent	Can be inefficient in a two-dimensional matrix
Model-assisted	Transparent decisions with statistical calibration	Requires careful construction for the dose matrix
Model-based	Can borrow information across combinations efficiently	More dependent on model assumptions
Surface/interaction models	Can explicitly represent combination effects	May require more information and stronger assumptions

The Most Important Concept

The most important conceptual point is that a combination dose-finding trial is not simply a single-agent dose-escalation study with two drugs listed in the protocol.

The treatment space is multidimensional, and the trial must learn a dose-toxicity surface. The design must therefore specify how information is borrowed across combinations, how unsafe regions are eliminated, how escalation moves through the matrix, and how one or more final combinations are selected.

A well-designed study balances:

Protection from excessive toxicity
Efficient exploration of the dose matrix
Accurate identification of combinations near the target toxicity probability
Reasonable exposure to subtherapeutic combinations
Robustness to uncertainty about the true interaction between agents

Bottom line: Combination dose-finding designs extend the dose-finding problem from one dimension to a dose-combination surface. The target is typically a set of combinations with toxicity near a prespecified level, and modern approaches—including model-assisted and model-based methods—use the accumulated data to navigate the matrix while controlling the risk of treating patients at excessively toxic combinations. Simulation under plausible toxicity surfaces is a central part of evaluating whether the design is fit for purpose.

References

Thall, P.F., Millikan, R.E., Mueller, P. & Lee, S.J. (2003). Dose-finding with two agents in Phase I oncology trials. Biometrics, 59, 487–496.

Braun, T.M. & Wang, S. (2010). A Bayesian model for identifying the maximum tolerated combination dose for drug combinations. Statistics in Medicine, 29, 1513–1525.

Liu, S. & Yuan, Y. (2015). Bayesian optimal interval designs for phase I clinical trials. Journal of the Royal Statistical Society: Series C, 64, 507–523.

Mu, R., Yuan, Y., Chen, S., Li, X. & Li, Y. (2019). BOIN-combo: Bayesian optimal interval design for dose finding in drug-combination trials. Statistics in Medicine.

Huang, X., Biswas, S., Oki, Y., Issa, J.-P. & Berry, D.A. (2007). Determining maximum tolerated dose of combined drugs from phase I clinical trials. Clinical Cancer Research, 13, 4458–4464.