Worked Analysis of a Phase I Biomarker Dataset

From Serial Observations to a Pharmacology Conclusion

A Phase I dataset can contain many concentration and biomarker measurements but very few participants. A useful analysis must preserve the timing in those records, define an exposure measure and response endpoint before fitting, and avoid making the number of observations look larger than the number of independent participants. The aim is not to make a small study answer every development question. It is to estimate the pattern supported by the observed exposure range, examine whether the model is credible, and state the remaining uncertainty.

The introductory material on biomarker exposure-response establishes why exposure and response should be linked at the participant level and why timing matters. Here, a compact hypothetical dataset is used to carry that analysis through to a bounded conclusion. A linear model is chosen as a simple empirical description over the observed range; it is not assumed to be the drug’s true response law at all concentrations.

Define and Summarize the Data

Suppose a hypothetical early-phase study collected plasma concentrations and biomarker measurements before dosing and at 2, 4, and 8 hours. The assay and sampling windows were considered suitable for this planned summary, and the 8-hour biomarker endpoint was chosen before examining exposure-response results. For each participant, serial plasma concentrations are summarized by the linear trapezoidal rule. For adjacent sampling times \(t_j\) and \(t_{j+1}\), this rule approximates the area as the interval duration multiplied by the average of its endpoint concentrations:

Here, concentration is in mg/L, time is in hours, AUC is in mg·h/L, and \(X\) is in mg/L. The calculation assumes the measured concentrations adequately represent the concentration-time profile between samples; sparse or poorly timed samples can weaken that assumption. The 8-hour biomarker reduction is calculated relative to that participant’s baseline and reported in percentage points, not as a relative percentage change unless that is how the endpoint was specifically defined.

The following hypothetical dataset contains seven participants, including one with no measurable drug exposure. The exposure column summarizes each participant’s serial PK samples as above. Biomarker measurements were also collected serially; the prespecified 8-hour baseline-adjusted reduction is shown here because it is the endpoint used for this particular fit.

Fit a Participant-Level Exposure-Response Model

For this illustration, fit the ordinary least-squares linear model \(Y_i=\beta_0+\beta_1X_i+\epsilon_i\). Its assumptions for inference are that the conditional mean response is linear over the analyzed exposure range, participant-level errors are independent with mean zero, their variance is constant, and the errors are approximately normally distributed when using the small-sample \(t\)-based interval and test. The exposure values are treated as measured adequately for ordinary least squares. These assumptions require scrutiny, especially in a small, potentially dose-escalated study.

The estimated slope is the sample covariance of exposure and response divided by the sample variance of exposure. The intercept is the fitted response at zero exposure. For this dataset, \(\bar{x}=1.50\) mg/L and \(\bar{y}=15.0\) percentage points. The sums needed for the fit are \(S_{xx}=\sum_i(X_i-\bar{x})^2=7.00\) (mg/L)\(^2\) and \(S_{xy}=\sum_i(X_i-\bar{x})(Y_i-\bar{y})=56.0\) (mg/L)·percentage points.

A short derivation explains why this line is the least-squares fit. For any proposed intercept \(a\) and slope \(b\), least squares minimizes \(\sum_i(Y_i-a-bX_i)^2\). Setting the derivative with respect to \(a\) to zero gives \(a=\bar{y}-b\bar{x}\). Substituting this into the sum of squares and setting its derivative with respect to \(b\) to zero gives \(\sum_i(X_i-\bar{x})(Y_i-\bar{y})=b\sum_i(X_i-\bar{x})^2\). Therefore \(b=S_{xy}/S_{xx}\), provided \(S_{xx}>0\), as it is here.

The fitted values at the seven exposures are 3, 7, 11, 15, 19, 23, and 27 percentage points. The corresponding residuals, observed minus fitted, are −2, 1, 1, 1, 0, 0, and −1 percentage points. Their squared values sum to \(\mathrm{SSE}=8.00\) (percentage points)\(^2\). With two fitted regression parameters, the residual degrees of freedom are \(df=n-2=5\). The residual standard deviation is \(s_e=\sqrt{\mathrm{SSE}/df}=\sqrt{8.00/5}=\sqrt{1.60}=1.26\) percentage points.

Check the Fit Before Interpreting It

The numerical result is only useful if the data and model support it. In this example, residuals are small and show no obvious curvature by inspection, but seven participants provide little power to detect a pattern. A practical diagnostic review should include the following:

• Plot response against exposure and inspect residuals against fitted values and exposure. Curvature can suggest that a straight line is inadequate. A fan shape can signal nonconstant residual variance. Clusters by cohort or time can reveal structure that the single-predictor fit misses.
• Check influential participants. A participant at the edge of the exposure range can strongly affect the slope, particularly with a small sample. Compare the fit with and without influential observations, while retaining scientifically justified data rather than removing observations just to improve the result.
• Check the sampling and endpoint definition. Confirm baseline adjustment, assay performance, time alignment, and the exposure calculation. Missing or anomalous concentrations, an endpoint measured at inconsistent times, or substantial exposure measurement error can distort the relationship.
• Respect the data hierarchy. Serial samples from one participant are not independent participants. Regressing every concentration or biomarker sample as though it came from a new person would understate uncertainty. If the question requires the full time course rather than a prespecified endpoint, a model that represents repeated measures and their timing is needed.

The independent-error assumption also deserves attention in an escalation study. If exposure mostly reflects assigned dose, dose cohort may be associated with other changes across cohorts, such as study procedures or participant characteristics. A slope can then reflect more than a direct pharmacological effect. The analysis should describe the design and avoid treating an unadjusted association as a causal estimate.

Choose the Conclusion to Match the Evidence

A linear model is a transparent starting point when the question is whether the observed data show a directional association across a limited exposure range. It is not interchangeable with every exposure-response approach. A direct-effect Emax model is more appropriate when a saturating relationship is scientifically motivated and the data cover enough of the curve to estimate its parameters. Turnover or effect-compartment models may be needed when response timing is delayed relative to plasma concentrations or when the time course shows hysteresis. A repeated-measures approach is relevant when the question concerns serial responses rather than one participant-level endpoint. Each alternative requires enough information in the data to support its additional structure; a more complex model does not automatically resolve sparse or poorly timed observations.

In this example, a defensible pharmacology statement is: Among these seven hypothetical participants, higher average plasma exposure from 0 to 8 hours was associated with a larger 8-hour baseline-adjusted biomarker reduction over the observed exposure range of 0–3 mg/L. The fitted linear slope was 8.00 percentage points per mg/L, with a model-based 95% confidence interval of 6.77 to 9.23. The small dataset supports a positive association under the stated assumptions, but does not establish a plateau, characterize effects beyond the observed range, or demonstrate target engagement or clinical benefit. The conclusion is deliberately limited to the endpoint, interval, range, and model actually analyzed.

A frequent mistake is to read a strong slope estimate and a small p-value as proof that the biomarker response is causal, clinically meaningful, and saturating at higher exposures. The correction is to separate those claims: the regression estimates an association and its uncertainty; the design and diagnostics qualify its credibility; biological relevance and clinical meaning require additional evidence. Extrapolating the fitted line beyond 3 mg/L is especially risky because a linear fit imposes no maximum response.