Selecting The Optimal Biological Dose

Why Select a Biological Dose?

The maximum tolerated dose (MTD) answers a safety-oriented question: how high can dosing go while keeping dose-limiting toxicity within a prespecified limit? It does not, by itself, identify the dose that produces the most useful biological activity or clinical benefit. If a pharmacodynamic effect plateaus before toxicity approaches the limit, selecting the MTD can expose participants to additional toxicity without a meaningful gain in activity.

An optimal biological dose (OBD) is therefore defined relative to the treatment, population, outcome definitions, follow-up period, and decision criteria of a trial. The word “optimal” does not identify one universal mathematical rule. The rule should specify how pharmacodynamic (PD) evidence, clinical efficacy, and toxicity are combined, including any minimum efficacy and maximum acceptable toxicity. The earlier material on therapeutic windows and constrained dose selection provides useful context: an OBD rule makes the biological objective explicit while retaining clinical and safety requirements.

Define the Outcomes and the Decision Rule

Let \(d\) denote one of the tested dose levels. For each participant, define a PD outcome on a specified measurement scale and at a specified assessment time. For example, the outcome could be a change in a biomarker from baseline, with larger values representing greater target inhibition. Also define a clinical efficacy outcome, such as response by a fixed time, and an unacceptable toxicity outcome, such as a dose-limiting toxicity during a fixed observation window.

The estimands must refer to the same target population and be interpreted over their stated assessment periods. For illustration, let \(\mu_{\mathrm{PD}}(d)\) be the population mean PD activity at dose \(d\), with larger values favorable; let \(p_E(d)\) be the probability of clinical response; and let \(p_T(d)\) be the probability of unacceptable toxicity. If the PD endpoint is binary, a target-attainment probability can be used in place of a mean. The direction of benefit and the scale must be made explicit.

A transparent constrained rule can first define doses that meet clinical and safety requirements. For example, set a minimum response probability \(e_{\min}\) and a maximum acceptable toxicity probability \(t_{\max}\), then define

The set \(\mathcal{A}\) contains doses that are clinically promising enough and sufficiently safe under the chosen criteria. A PD target can also be required, such as \(\mu_{\mathrm{PD}}(d)\geq m_{\min}\). If no dose meets the requirements, the rule should permit a conclusion that no tested dose is suitable; it should not force selection of an unacceptable dose.

When \(\mathcal{A}\) is nonempty, one plateau-based rule selects a dose close to the greatest PD activity among acceptable doses. For a prespecified tolerance \(\delta\geq 0\), define

Here \(\delta\) is expressed in PD units. The rule chooses the lowest dose whose activity is no more than \(\delta\) below the best activity among doses meeting the clinical and safety criteria. A zero tolerance chooses a dose attaining the maximum estimated activity; a positive tolerance allows a small PD difference in return for choosing a lower dose. This is a decision criterion, not proof that the biological response is truly flat across those doses.

Worked Examples

Using Joint Evidence Correctly

PD response, clinical response, and toxicity may be related within a participant. For instance, the probability of clinical response among participants who achieve a PD target and avoid severe toxicity may differ from the marginal response probability. Earlier material on dependence between efficacy and toxicity and on joint outcome models establishes why this relationship can matter. If the decision rule concerns the probability of a joint outcome, estimate that joint probability from a joint model or joint data, or express it with conditional probabilities using the chain rule.

Multiplying marginal probabilities is justified by mutual independence of the events, but mutual independence is not necessary for the product to happen to equal the joint probability. Without independence, that equality may occur by coincidence; the marginals alone do not establish it. The following construction makes the distinction explicit.

Justification, Assumptions, and Diagnostics

The plateau rule has a direct decision-theoretic justification without assigning numerical utilities. Among doses meeting the prespecified efficacy and toxicity conditions, it avoids choosing a higher dose when a lower one is close enough to the best estimated PD activity. In symbols, if \(d_{\mathrm{OBD}}\in\mathcal{P}\), then by definition

This inequality follows by rearranging the condition that defines \(\mathcal{P}\). Since the selected dose is the minimum element of that set, no lower eligible dose is within the chosen tolerance. The rule therefore provides a precise meaning for “lowest dose near the acceptable PD maximum.” Its justification depends on the clinical team having chosen a meaningful PD scale and tolerance in advance.

Applying the rule assumes that the endpoint definitions and assessment times are clinically relevant; that dose-specific estimates target the population for which the recommendation is intended; and that the selected model or estimation method represents the observed dose-response patterns adequately. Monotonicity should not be assumed automatically for every PD or efficacy endpoint. If a monotone dose-response model is used, its fit and scientific rationale should be checked. Toxicity may be expected to increase with dose, but unexpected estimated reversals should be investigated rather than silently forced into a pattern.

Important diagnostics include checking whether:

• PD measurements are missing, delayed, or taken at different times across doses, which can make apparent dose differences incomparable;
• the selected dose meets the efficacy and toxicity criteria by a comfortable margin or only narrowly, given the uncertainty in the estimates;
• PD activity continues to rise substantially at the highest eligible dose, so a plateau-based rule may not identify a stable near-plateau dose;
• the estimated dose-response curves fit the observed data and whether conclusions change under plausible alternative models;
• the PD marker is a meaningful indicator of biological target engagement, rather than an unvalidated substitute for clinical benefit; and
• the joint probabilities used for a joint decision are estimated jointly or conditionally, rather than inferred from marginal products without an independence justification.

A dose can meet a PD target and still fail to improve clinical outcomes. PD activity is evidence about biological effect, not proof of clinical benefit or a causal surrogate for it. Conversely, a clinically active dose need not maximize a chosen biomarker. The prespecified efficacy requirement guards against selecting a dose solely for biological activity, but it cannot eliminate uncertainty or compensate for a poorly chosen endpoint.

How OBD Selection Differs From MTD Selection

Neither approach is automatically superior for every treatment. MTD selection can be appropriate when increasing dose is expected to improve treatment effect and toxicity is dose-limiting. OBD selection is especially relevant when biological activity can saturate below the MTD or when additional dose adds toxicity without a meaningful biological gain. A practical trial may use MTD information as a safety boundary while using the OBD rule to select among clinically acceptable doses. The two criteria should not be treated as interchangeable.

Common Mistakes and Limitations

• Calling the highest tolerated dose the OBD. The MTD is defined by a toxicity rule; an OBD also uses biological and clinical evidence.
• Choosing the dose with the largest PD estimate without constraints. A high PD estimate does not make a dose acceptable if its toxicity is excessive or its efficacy is inadequate.
• Declaring a plateau from visually similar point estimates. A tolerance \(\delta\) should be prespecified and interpreted in PD units; closeness of estimates is not proof of equivalence.
• Multiplying marginal probabilities as though dependence were required for equality or as though the product always gave the joint probability. Independence justifies the multiplication, but equality can also occur coincidentally. Without a valid independence assumption, use joint or conditional probabilities.
• Treating the selected dose as certain to be best. Dose selection is based on estimated quantities and may be sensitive to sample size, model assumptions, and thresholds.

The OBD rule is a recommendation within the tested doses and specified follow-up. It does not establish the best dose among untested levels, guarantee long-term safety, or resolve uncertainty about whether a PD marker predicts clinical benefit. If no tested dose satisfies the prespecified requirements, reporting that result is more informative than selecting a dose by default.