When Spatial Heterogeneity Changes Model Predictions

Why Spatial Heterogeneity Can Change a Prediction

A well-mixed model represents a tissue or cell population with one concentration and one state for each modeled quantity. This can be efficient and sufficient when transport makes concentration differences small on the timescale relevant to the question. But a single average concentration does not necessarily predict the average biological response. If occupancy or downstream response is nonlinear in concentration, applying the response function to the average can differ from calculating the local response first and then averaging it.

The decision is therefore not simply whether a tissue is spatially heterogeneous. It is whether spatial differences persist over relevant distances and times, and whether those differences change the model output needed for the biological question. Diffusion timescale estimates, the expected degree of nonlinear response, and spatially resolved observations provide complementary evidence.

Transport Time as a First Screening Test

For diffusion with coefficient \(D\), a useful order-of-magnitude estimate of the time to diffuse across a characteristic distance \(L\) is

Here, \(L\) is a distance in the tissue and \(D\) has units of distance squared per time, so \(\tau_D\) has units of time. This estimate is a screening calculation, not an exact prediction of equilibration. Geometry, boundary conditions, spatially varying properties, and local removal or binding can alter how quickly concentrations become similar.

Compare \(\tau_D\) with the timescale of the biological process and the observation. If diffusion across the relevant distance is much faster than binding, signaling, or the time resolution of interest, substantial concentration differences may dissipate before they affect the measured output. If transport is comparable to or slower than those processes, local responses can develop before the tissue equilibrates. This comparison indicates when spatial effects are plausible; it does not by itself establish that a spatial model is necessary.

For reversible binding in a region with approximately constant free concentration \(C\), a simple binding timescale is \(1/(k_{\mathrm{on}}C+k_{\mathrm{off}})\), where \(k_{\mathrm{on}}\) and \(k_{\mathrm{off}}\) are the association and dissociation rate constants. Comparing that local timescale with \(\tau_D\) helps determine whether appreciable binding can occur before concentrations change through transport. The estimate assumes the simple binding process and a locally stable \(C\); it should not be treated as a full model of a tissue with depletion or complex kinetics.

Nonlinear Responses Do Not Generally Commute With Averaging

A common way to expose the effect of heterogeneity is to compare the response at the mean concentration with the mean of local responses. For a simple equilibrium occupancy model, the local fractional occupancy at free concentration \(C\) is

where \(K_d\) and \(C\) have the same concentration units. If regions have concentrations \(C_i\) and occupy fractions \(w_i\) of the relevant population, with \(w_i\geq 0\) and \(\sum_i w_i=1\), the mean local occupancy is \(\sum_i w_i g(C_i)\). A single-concentration model instead predicts \(g(\sum_i w_i C_i)\). These are not generally equal.

For \(C\geq0\), the second derivative of the occupancy function is

Thus \(g\) is concave over nonnegative concentrations. Jensen's inequality gives \(\sum_i w_i g(C_i)\leq g(\sum_i w_i C_i)\) under the assumptions of this occupancy function and common \(K_d\). In this case, using mean concentration tends to overpredict the average local occupancy when concentrations vary. Other response functions can have different curvature, so the direction and magnitude of the discrepancy must be checked for the response of interest.

Use Spatial Observations and the Relevant Aggregate

The appropriate average depends on the model output. If the question concerns the fraction of receptors occupied across tissue, regions with more receptors contribute more bound receptors. An unweighted mean of regional occupancies is appropriate only when the regions contribute equally to the receptor population. Likewise, an assay signal may weight regions by tissue volume, cell number, receptor abundance, or signal-generation properties. A bulk measurement can conceal gradients and does not automatically equal the average of the biological quantity being modeled.

Spatially resolved measurements can test whether the gradients relevant to a prediction are present. Examples include measuring drug concentration or target engagement at multiple tissue locations, or comparing regions expected to differ in delivery. Measurements should be interpreted at their actual spatial and temporal resolution: a coarse assay can average away a gradient, and an indirect biomarker may not measure local free concentration or occupancy directly. The earlier material on direct occupancy assays and indirect target-engagement biomarkers is useful for distinguishing what the observed signal represents.

Choosing a Model and Diagnosing Failure

Start from the biological question and the output to be predicted. A well-mixed model is often a reasonable choice when evidence supports rapid transport relative to the relevant response times, spatial variation is small at the scales that matter, and predictions are insensitive to plausible heterogeneity. If measurements show persistent gradients, or a nonlinear local response makes those gradients consequential, a spatial representation may be required. An intermediate approach is to represent a small number of distinct regions or compartments; it can capture some heterogeneity without resolving position continuously, but its adequacy must also be checked against the question and data.

Useful diagnostics include comparing predicted outputs from a well-mixed model with regional predictions under plausible concentration profiles; checking whether the measured spatial pattern changes over the response time; and verifying that the average used for comparison matches the measured or clinically relevant quantity. Warning signs include systematic errors by tissue location, a bulk prediction that matches only after local overprediction and underprediction cancel, or a response that changes materially when the same mean concentration is distributed differently across regions.

A timescale calculation can fail as a decision rule if the chosen distance is irrelevant, the effective diffusion coefficient is poorly known, or uptake and binding act as local sinks. A spatial model can also be weakly identified if only bulk data are available: many different spatial profiles may produce the same bulk average. In that case, further spatial measurements or sensitivity analysis may be more informative than adding model detail. A well-mixed model is not inherently incorrect, and a spatial model is not automatically more credible; the choice is conditional on the question, evidence, and prediction robustness.

Common Mistakes

• Comparing only mean concentrations. Equal means do not imply equal average responses when the response is nonlinear. Compare local responses and the appropriately weighted aggregate.
• Treating \(\tau_D\) as an exact equilibration time. It is an order-of-magnitude estimate. Geometry, boundaries, binding, and removal can matter.
• Using the wrong weights. Volume-weighted concentration and receptor-weighted occupancy answer different questions when receptor abundance varies.
• Assuming that a gradient automatically requires a spatial model. If the requested prediction is insensitive to that gradient, a well-mixed model may still answer the question adequately.
• Assuming a bulk assay rules out spatial variation. A pooled observation can conceal local differences, especially when the assay does not measure the local quantity directly.