From Pathogenesis Evidence to a Network
A disease mechanism is often described in words: a cell population produces a cytokine, the cytokine changes another cell’s behavior, and that behavior contributes to tissue pathology. A network diagram makes these proposed relationships explicit. Its value is not simply that it organizes many molecules on a page. It forces each arrow to answer two questions: what change is proposed to cause what other change, and how strong is the evidence for that claim?
For quantitative systems pharmacology, the diagram is a structured hypothesis, not a picture of established truth. A measured association can suggest an edge, but it does not by itself show that changing the proposed cause changes the proposed effect. The evidence label on each edge should make that distinction visible.
A node must be specific enough to interpret. “Myeloid cells” could mean cell abundance, activation, or production of a particular cytokine; these are not interchangeable quantities. Likewise, “cytokine” might mean production, extracellular availability, receptor signaling, or a downstream response. A useful diagram separates these when the evidence measures them separately. For example, it can represent “myeloid-cell IL-1β production” and “IL-1 signaling in fibroblasts” as different nodes, with a proposed link between them.
A sign is conditional on context. A positive edge does not mean the target rises under every disease state, dose, or time scale. It summarizes the direction of a causal effect in the experimental or clinical setting that supports the edge. Record that setting—such as cell type, tissue, perturbation, and observation window—alongside the diagram when it matters to interpretation.
What Makes an Edge Causal?
An intervention provides stronger evidence for a causal edge than a correlation does, provided the intervention changes the proposed cause and the outcome is measured appropriately. If selectively reducing \(A\) changes \(B\), that supports a causal effect of \(A\) on \(B\) in the tested conditions. Selectivity matters: if the intervention also changes other relevant processes, the result may not isolate the proposed link. A rescue experiment, in which the perturbed signal is restored and the response moves back toward its original level, can strengthen the interpretation.
Even a well-controlled intervention does not automatically establish every detail of the diagram. It may support an effect of \(A\) on \(B\) without showing that the effect is direct. The intervention could alter an unmeasured intermediary. Nor does a cytokine’s detection in a cell population establish that the cytokine is secreted in an active form or drives a target-cell response. These are separate claims and may need separate evidence.
Specify the measured or modeled quantity for each cell population, cytokine-related process, and disease response. Avoid using one vague node for several distinct quantities.
For each relationship, state which quantity is the proposed cause and which is the proposed effect. Use a sign only when the expected direction is meaningful in the stated context.
Mark whether an edge is supported by a relevant intervention, supported only in a limited experimental setting, or inferred from observational or mechanistic evidence.
Ask whether an unmeasured factor, a nonselective perturbation, or an untested intermediate could explain the observed relationship.
If arrows form a cycle, distinguish the supported links from the inferred ones. A drawn cycle is a hypothesis about connectivity, not proof of its strength or behavior over time.
Worked Examples: Building and Qualifying Edges
Worked Example: Separating Cytokine Source From Cytokine Action
In a hypothetical laboratory study, investigators examine activated myeloid cells and fibroblasts. A cell-specific secretion assay detects IL-1β release from the activated myeloid cells. In a separate coculture assay, fibroblasts produce 80 ng of a measured inflammatory product per \(10^5\) cells over 24 hours. Selective IL-1β immunodepletion from the coculture medium reduces the output to 50 ng per \(10^5\) cells over 24 hours. Adding IL-1β back to the depleted medium gives an output of 76 ng per \(10^5\) cells over 24 hours. What edges are supported?
The secretion assay supports the source statement: activated myeloid cells produce IL-1β under the tested conditions. The depletion comparison estimates a change of \(50-80=-30\) ng per \(10^5\) fibroblasts over 24 hours when IL-1β is removed from this assay system. The add-back changes output by \(76-50=26\) ng per \(10^5\) fibroblasts over 24 hours, toward the original value of 80.
A defensible diagram therefore includes an edge from activated myeloid-cell IL-1β production to IL-1β availability in the assay, supported by the secretion measurement, and a positive edge from IL-1β availability to fibroblast inflammatory-product output, supported by depletion and add-back under the assay conditions. The recovery is \(26/30 \approx 0.867\), or about 86.7% of the decrease after depletion. It strengthens the interpretation, but does not prove that IL-1β is the only active factor in the medium or that the effect is direct at the molecular level.
The diagram should not collapse these findings into an unqualified arrow from “myeloid cells” to “fibroblast activation.” The experiments address a specified cytokine and outcome in a defined assay. They do not establish that this route dominates in diseased tissue.
Worked Example: An Association Suggests, but Does Not Prove, an Edge
In a hypothetical tissue survey of 40 patients with autoimmune inflammation, investigators measure the abundance of a cytokine-producing cell signature and the abundance of a tissue-response signature. The two measurements are positively associated. No cells or cytokine signals are perturbed. How should the relationship appear in a network?
The diagram may include a provisional positive association between the measured signatures, clearly labeled as inferred rather than intervention-supported. It should not present the association as proof that the cell signature causes the tissue response. For example, a common disease-state variable could increase both measurements; the tissue response could also influence cell recruitment or activity.
The observation does not establish that the cells secrete the relevant cytokine, that the cytokine reaches or activates a target cell, or that changing this signal would change the tissue response. These are distinct links that require evidence. A useful next experiment would perturb the candidate cytokine pathway in an appropriate system and measure a prespecified target-cell response. Until then, the proposed causal arrow remains a hypothesis, even if its direction is biologically plausible.
Worked Example: Quantifying One Local Path Contribution
Consider a hypothetical model with three nodes: \(X\), pathogenic T-cell activity in arbitrary activity units; \(B\), macrophage cytokine-pathway signaling in signaling units; and \(Y\), stromal inflammatory output in output units. In a specified assay context, the estimated local slopes are \(dB/dX=0.20\) signaling units per activity unit and \(dY/dB=3.0\) output units per signaling unit. What is the local contribution along the path \(X \to B \to Y\)?
For a small change in \(X\), the change transmitted through \(B\) is approximately \(\Delta B \approx (dB/dX)\Delta X\). The corresponding change in \(Y\) along this route is \(\Delta Y_{\text{route}} \approx (dY/dB)\Delta B\). Substituting the first relation gives the chain-rule result:
Thus, under the stated local-slope and pathway assumptions, a one-activity-unit change in \(X\) corresponds to an approximate 0.60-output-unit change in \(Y\) through this path near the operating point. The units cancel correctly. This is not necessarily the total effect of changing \(X\): other routes from \(X\) to \(Y\) may contribute. Nor does the product establish a causal path if either edge is only observational. If the slopes had not been specified or estimated, the route contribution would have to remain \((dY/dB)(dB/dX)\); it could not be assigned a numerical value from unrelated change measurements.
Justifying a Path Contribution
The chain-rule calculation above is a local justification, not a general shortcut for turning any two associations into a causal effect. Suppose \(B\) depends on \(X\), and the route-specific component of \(Y\) depends on \(B\). For a sufficiently small change \(\Delta X\), the first-order approximations are \(\Delta B \approx (dB/dX)\Delta X\) and \(\Delta Y_{\text{route}} \approx (dY/dB)\Delta B\). Substitution yields \(\Delta Y_{\text{route}} \approx (dY/dB)(dB/dX)\Delta X\). Dividing by \(\Delta X\) gives the product of local slopes.
This argument assumes that the slopes are appropriate for the same context and operating point, that the proposed direction of influence is justified, and that the route can be interpreted separately from other pathways. If there are interactions, strong nonlinearities, or important unmeasured causes, the simple local product may not represent a causal route contribution. A network diagram should preserve the assumptions and evidence behind an edge rather than implying more certainty than the analysis supports.
Representing Feedback Without Overclaiming
A feedback hypothesis appears when directed arrows form a cycle. For example, a diagram might propose that a tissue response increases immune-cell activity, which increases cytokine signaling, which then contributes to the tissue response. The cycle is a statement about proposed connectivity. It does not, by itself, show that the cycle is active in a patient, that every edge is causal, or that the cycle amplifies disease.
Use evidence labels on each arrow in the cycle, not just one label for the loop. If one link is supported by a selective intervention and two are inferred from expression patterns, the diagram should show that uneven evidence. It is also useful to annotate the relevant cell types, tissue, and time window, because apparent cycles may combine findings from different experimental settings that do not jointly establish one operating mechanism.
Diagnostics, Alternatives, and Limitations
Before treating a diagram as a useful mechanistic summary, check whether each edge has a clear operational meaning and whether its sign is justified. A mismatch between the measured node and the claimed node is a warning: transcript abundance is not automatically secretion, and secretion is not automatically target-cell signaling. Likewise, a perturbation that changes several processes cannot uniquely identify one arrow without additional controls or experiments.
| Evidence type | What it can support | What it does not establish alone |
|---|---|---|
| Co-occurrence or correlation | A relationship worth representing as an association or hypothesis | Direction of causation or a causal edge |
| Cell-specific expression or secretion measurement | A candidate source in the measured setting | That the signal causes a target-cell response |
| Selective intervention with an outcome measurement | A causal effect in the tested conditions, if important alternative effects are controlled | That the effect is direct, universal, or dominant in disease |
| Intervention plus rescue | Stronger support that the perturbed signal contributes to the measured response | That all other mediators or pathways are irrelevant |
A common mistake is to draw a solid causal arrow because two measurements rise together. The correction is to label the observation as an association or inference, then reserve a causal label for evidence that changes in the proposed cause alter the proposed effect under an appropriate intervention. Another frequent error is to infer a direct cell-to-cell effect from a perturbation that changes an outcome. The measured result may support an overall effect while leaving intermediates unresolved; represent the known effect and mark the unknown mechanism rather than inventing a direct link.
Network diagrams also simplify. They may omit unmeasured mediators, competing pathways, context dependence, and uncertainty in signs or magnitudes. A diagram is most useful when it exposes those gaps and guides discriminating experiments. It is not a substitute for a specified model, and connectivity alone does not determine the time course or magnitude of disease behavior.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A tissue survey finds that a cytokine transcript and a pathology score are positively associated. What kind of edge, if any, is justified?
- In a controlled assay, blocking a cytokine receptor lowers a target-cell response. Name one causal conclusion supported by this result and one conclusion it does not establish by itself.
- For a hypothetical path \(A \to B \to C\), the local slopes are \(dB/dA=0.5\) units of \(B\) per unit of \(A\) and \(dC/dB=-2\) units of \(C\) per unit of \(B\). Calculate the local route contribution and its sign.
- A proposed three-edge cycle contains two intervention-supported edges and one edge inferred from co-occurrence. How should the cycle be labeled?
- Why should a diagram distinguish cytokine production from cytokine signaling in a target cell?
Solutions. (1) The survey supports a positive association and may motivate a provisional inferred edge, but it does not establish direction or causation. (2) It supports a causal effect of receptor blockade on the measured response in that assay, assuming the intervention is sufficiently selective; it does not by itself show that the cytokine acts directly or that the same effect dominates in diseased tissue. (3) The product is \((-2)(0.5)=-1\) unit of \(C\) per unit of \(A\), so the local route contribution is negative under the stated assumptions. (4) Label the two supported edges and the inferred edge separately; the whole cycle remains incompletely established. (5) Evidence that a cell produces a cytokine does not alone show that the cytokine reaches, activates, or changes the response of a particular target cell.