Integrating Evidence Across Assays and Biological Scales

Why Integrate Evidence Across Scales?

A molecular assay may show that a drug binds its target, a cellular assay may show a change in signaling, and a tissue assay may show altered permeability. A clinical endpoint may then change in a direction consistent with those findings. Taken together, these results can constrain one biological explanation more strongly than any single measurement. But they do so only if the connections between the measurements are made explicit.

The key is to distinguish the biological process being represented from the assay used to observe it. Measurements from different scales usually have different units, experimental conditions, and time points. Their numerical values should not be averaged or compared directly. Instead, a QSP model expresses how a shared set of biological states could generate each measurement.

A state is not automatically measured directly. For instance, an assay’s fluorescence may indicate target binding, but its value can also depend on calibration and assay conditions. Likewise, a tissue sample taken at one time does not by itself describe the full time course of a process. The model’s task is to connect the measured quantities to the states and to connect states across biological scales.

How a Cross-Scale Explanation Is Organized

A practical starting point is a chain of proposed links. For a hypothetical anti-inflammatory drug, the chain might be: drug concentration affects target occupancy; occupancy changes cellular pathway activity; pathway activity influences tissue leakage; and tissue leakage contributes to a clinical measure of fluid accumulation. Each link has to be stated clearly enough to make a prediction.

In general terms, let \(x(t)\) represent the model’s biological states at time \(t\), and let \(y_j\) represent the observation made by assay \(j\). The model connects them through an observation relationship:

Here, \(h_j\) is the relationship between relevant states and the measurement, \(t_j\) is the measurement time, \(q_j\) collects assay-specific conditions, and \(\epsilon_j\) represents measurement error or unexplained variation. Different assays can have different \(h_j\) and \(q_j\), even when they inform the same biological process. The model also specifies how the states change and influence one another over time.

This representation relies on several assumptions that should be checked for the intended use. The proposed states and links must be biologically relevant to the observations; the assay relationships must be appropriate for the experimental conditions; and any difference in timing must be represented or judged unimportant. A measurement from a cell system cannot simply be treated as a tissue measurement, and a tissue result cannot be substituted for a clinical endpoint. Each connection needs a stated biological and quantitative basis.

Use Cross-Scale Predictions, Not Raw-Value Comparisons

The most informative integration asks whether a proposed link predicts a measurement at another scale. The model can be calibrated or constrained by some observations, then checked against additional observations that were not used to set that particular link. This is a reason to keep the measurement units and assay relationships explicit: a cytokine concentration, a tissue permeability readout, and a clinical rate are not interchangeable quantities.

A useful short derivation follows from the chain structure. Suppose cellular activity \(S\) is linked to tissue leakage \(P\) by a specified function \(P=g(S)\), and tissue leakage is linked to a clinical quantity \(Y\) by \(Y=f(P)\). Then the model’s cross-scale prediction is the composition

This result follows by substituting \(P=g(S)\) into \(Y=f(P)\). It matters because it creates a constraint: a value measured at the cellular scale implies a prediction at the clinical scale, conditional on both links and the relevant context. If the clinical measurement disagrees, at least one link, assay relationship, context assumption, or measurement interpretation may be inadequate. The disagreement does not identify which one by itself.

Assessing Agreement and Recognizing Failure Modes

Cross-scale integration is useful when measurements constrain a shared explanation, rather than merely appearing in a plausible order. Before interpreting agreement, check whether the model predicts the direction and approximate magnitude of observations at scales not used to construct each link. Also check whether the measurements refer to compatible populations, exposure conditions, and times. When a mechanism operates with a delay, measurements taken at different times should not be treated as if they were simultaneous.

• Check the assay relationship. A readout may be an indirect proxy for a state, and the conversion from readout to state can depend on calibration and assay conditions.
• Check units and normalization. Ratios can help compare measurements within an assay, but a normalized ratio does not make different assays identical.
• Check timing. A short-lived molecular change may precede a slower tissue or clinical response. A mismatch in sampling times can create an apparent contradiction.
• Check uncertainty. Small discrepancies may be expected from measurement variation. A close match should not be described as exact, and uncertainty in the bridge relationships also matters.
• Check whether the prediction was genuinely cross-scale. A link fitted to the same observations it is later said to explain provides less independent support than a prediction checked against additional data.

A common mistake is to say, “The molecular assay changed, the tissue assay changed, and the clinical endpoint improved; therefore, the measured molecular change caused the clinical improvement.” This skips the bridge relationships and treats a pattern of observations as proof of causation. The correction is to state the proposed links, their assumptions, and the measurements they predict. Agreement across scales can support a coherent mechanistic account and constrain its parameters, but other explanations or omitted processes may also produce agreement.

Compared with fitting a separate empirical relationship to each assay, a cross-scale QSP model makes the connections between them explicit and can generate predictions across scales. Compared with relying on a narrative evidence chain alone, it requires the proposed links to be quantitatively consistent with measurements. Its limitations are equally important: model detail does not guarantee that an assay measures the intended state, and an incorrect bridge can propagate error from one scale to another. Integrating evidence is therefore a disciplined way to check coherence, not a substitute for sound experiments or evidence about each link.