Reconciling Discordant Biomarkers in Phase I

Why Biomarkers Can Disagree

A Phase I biomarker panel may show a clear change in one measurement and little change—or a change in the opposite direction—in another. That pattern is not automatically evidence that the drug has no effect, nor is it proof that one assay is wrong. Biomarkers can represent different places in a biological pathway, respond on different timescales, and have different measurement quality. Several markers may also reflect much of the same underlying biology, so counting results as independent votes can exaggerate the strength of evidence.

The task is to identify what each measurement can support. A target-engagement biomarker, a downstream pharmacodynamic biomarker, and a surrogate biomarker answer different questions. A change in one does not by itself establish a change in the others. The earlier material on biomarker roles, assay performance, sampling windows, baseline and time-course controls, and exposure-response interpretation provides the relevant definitions; here, the emphasis is on reconciling results across a panel.

A Structured Interpretation

Before comparing directions, specify the claim being assessed. For example: “Does the drug engage its intended target?”, “Does target engagement alter a proximal pathway readout?”, or “Is there evidence of a downstream pharmacodynamic response?” A marker can be informative for one question and uninformative for another. A downstream measurement that does not change cannot, by itself, negate a well-measured proximal target-engagement result.

For each marker, establish its pathway position and expected response. A proximal marker is close to the drug target; a downstream marker reflects later biological steps that may involve feedback, compensation, or other influences. Record the expected direction, plausible lag, and whether the marker is expected to change at the dose and exposure observed. These expectations should be based on the intended mechanism and available prior evidence, not chosen after seeing which results look favorable.

Then assess data quality and timing. Confirm that the assay is fit for the observed concentration range, the samples were handled as intended, and baseline and control measurements are available. Compare the sampling times with the expected response course: a sample collected before a delayed response is expected cannot test that delayed response well. Finally, consider whether measurements are correlated and whether the same participants, time points, or sample-processing conditions could produce shared variation.

Worked Interpretation

Justifying the Correlation Adjustment

The variance calculation above follows from the rule \(\operatorname{Var}(aX+bY)=a^2\operatorname{Var}(X)+b^2\operatorname{Var}(Y)+2ab\operatorname{Cov}(X,Y)\). For an average of two estimates with equal variance 1 and correlation \(\rho\), their covariance is \(\rho\), so

When \(\rho=0\), this is \(0.50\), as for independent estimates. When \(\rho=0.80\), it is \(0.90\): the two measurements contain less independent information than two uncorrelated measurements. This calculation does not say that correlated biomarkers should never be analyzed together. It shows why an analysis that combines them must account for their covariance rather than treating each as a separate, independent confirmation.

Diagnostics, Alternatives, and Limits

Useful diagnostics follow the possible causes of discordance. Inspect assay precision and performance at the observed concentrations, sample handling records, baseline balance, and control-group time trends. Plot individual trajectories as well as group summaries: a mean can conceal participants moving in opposite directions, and repeated observations from one participant are not independent. Check the sampling schedule against the anticipated response lag and compare biomarker changes with exposure where that relationship is relevant. Earlier work on repeated biomarker measures, delayed responses, and model diagnostics supplies methods for these checks; the key here is to apply them to the particular disagreement rather than to treat the panel as a single outcome.

There are several reasonable approaches, each answering a different question. A pathway-based interpretation explains why markers may differ biologically, but depends on a credible mechanism and prior expectations. Marker-by-marker analysis preserves each outcome’s meaning, but can produce a fragmented picture. A composite score can summarize a pre-specified group of related measurements, but may hide opposite effects and must account for scale and correlation. A joint model can represent relationships among biomarkers, but requires enough data and a defensible model structure. None of these approaches makes an uncertain assay or mistimed sample informative by itself.

Common warning signs include a conclusion based on whichever marker is most favorable, treating “not statistically distinguishable from zero” as proof of no effect, and interpreting any opposite-direction result as biological failure. A small estimate may be imprecise; a null-looking value is not necessarily evidence of no meaningful response. Conversely, an apparently strong result can reflect assay artifacts, uncontrolled time trends, or selective interpretation. Results should be described with their units, uncertainty, and comparison group, not just as “positive” or “negative.”

Phase I biomarker panels are generally exploratory and may be small, sparse, or collected at only a few times. Thus, discordance may remain unresolved even after careful review. A defensible report separates the conclusion supported now—such as evidence of target engagement—from the open question, such as whether a downstream pathway response is reproducible. It also avoids inferring clinical benefit from a biomarker change unless that relationship is established for the intended use.