Epithelial–Immune Cytokine Loops in Inflammatory Bowel Disease

Why Model the Epithelial–Immune Loop?

In inflammatory bowel disease, epithelial injury and immune activation can reinforce one another. Barrier damage can increase epithelial signals that stimulate cytokine production; cytokines can, in turn, contribute to further barrier damage. The resulting feedback is important to represent explicitly: a cytokine concentration alone does not show whether its source is epithelial damage, immune-cell activity, or an external injury process.

A compact model can clarify how the same network responds differently to a brief insult and to an ongoing one. The model below is a mechanistic hypothesis, not a claim that one cytokine or cell type explains all IBD. It focuses on three linked states and makes its assumptions visible so that they can be challenged against data.

A Minimal Model and Its Assumptions

First consider a linear model near a reference state:

Here \(u(t)\) is an external injury input, such as a transient epithelial insult or a sustained injurious drive. The positive parameters \(a\), \(b\), and \(c\) represent, respectively, damage-driven epithelial signaling, signal-driven cytokine production, and cytokine-associated damage. The positive parameters \(k_E\), \(k_C\), and \(k_D\) represent loss of signal, cytokine clearance, and barrier repair. With normalized states, the parameters have units of inverse time; \(u(t)\) has units of normalized damage per time. In the examples, time is in days.

The model assumes that the relevant effects are positive and approximately proportional over the range considered; that the states are well mixed within one tissue compartment; and that loss or repair is first order. It treats cytokine-associated damage as a net effect, not as proof that a measured cytokine directly injures epithelium. It also omits explicit cell populations, multiple cytokines, and transport between tissue and blood. These choices make the feedback interpretable, but restrict the range where the linear model is credible.

The feedback path is \(D\rightarrow E\rightarrow C\rightarrow D\). Each link has a positive gain, while the \(k\) terms oppose accumulation. Epithelial signals and cytokines need not rise and fall together: each is a state with its own production and removal. This is consistent with the source–sink and innate–adaptive coupling ideas used in earlier material, without assigning the loop to a single mediator.

Feedback Gain and Resolution After Injury

The linear model’s central quantity is the dimensionless loop gain \(L=\frac{abc}{k_Ek_Ck_D}\). It compares the product of gains around the feedback loop with the product of the three loss rates. To see why it matters, set the input to zero and seek a steady state. The first two equations imply \(E=(a/k_E)D\) and \(C=(b/k_C)E=(ab/(k_Ek_C))D\). The damage equation at steady state then gives \(0=cC-k_DD\), or \((abc/(k_Ek_C)-k_D)D=0\). Unless \(L=1\), the only steady state is \(D=E=C=0\); the question is whether small deviations decay toward it.

For a stability check, let a small perturbation vary as \(e^{\lambda t}\). Substitution into the zero-input equations gives the characteristic equation \((\lambda+k_E)(\lambda+k_C)(\lambda+k_D)-abc=0\). Its constant term is \(k_Ek_Ck_D-abc\). If \(L>1\), the left-hand side at \(\lambda=0\) is negative and it tends to a positive value as real \(\lambda\) grows, so there is a positive real root and the zero state is unstable. If \(L<1\), all coefficients of the cubic are positive. The remaining Routh–Hurwitz condition also holds: the product of the coefficient of \(\lambda^2\) and that of \(\lambda\) exceeds the constant term, since \((k_E+k_C+k_D)(k_Ek_C+k_Ek_D+k_Ck_D)>k_Ek_Ck_D>k_Ek_Ck_D-abc\). Thus the zero state is stable. At \(L=1\), the characteristic equation has a zero root, marking the threshold in this linear analysis.

This stability result is a local property of the stated linear model. In particular, \(L<1\) means a disturbance decays once \(u(t)\) is removed; it does not mean that no injury can occur. A positive continuing input can maintain positive states even when the feedback loop is stable.

Ongoing Injury and the Meaning of Persistence

For constant injury \(u(t)=u_0>0\), the linear model has a steady state when \(L<1\). Setting each derivative to zero gives \(E=(a/k_E)D\), \(C=(ab/(k_Ek_C))D\), and \(D=u_0/[k_D(1-L)]\). The denominator shows that feedback amplifies the steady-state effect of the continuing input. As \(L\) approaches 1 from below, the predicted steady state grows sharply, which is also a warning that the linear approximation may leave its valid range.

Diagnostics, Alternatives, and Limitations

A model should be checked against both its mathematics and the biology it is intended to represent. Useful diagnostics include:

• Sign and boundary checks. Verify that each production and loss term has the intended sign. If numerical integration produces negative cytokine levels or damage outside its allowed range, check the solver, step size, and model formulation.
• Massive or implausible predictions. A very large steady state when \(L\) is near 1 can indicate that a linear approximation is being used beyond its range, not that tissue damage can grow without bound.
• Timing and measurement checks. Compare predicted state changes with measurements on a time scale appropriate to signal production, cytokine clearance, and tissue repair. A mismatch may reflect unmeasured sources, compartment mixing, or incorrect rate assumptions.
• Parameter identifiability. Similar trajectories can arise from different combinations of production and clearance rates. A fit to cytokine observations alone may not identify all three links in the loop; measurements of epithelial injury and barrier damage help constrain the mechanism.

The linear model is useful for local reasoning and transparent calculations, but it has no upper bound on \(E\), \(C\), or \(D\). A bounded alternative can make damage a fraction and represent saturation:

Here \(K_E\) and \(K_C\) are positive half-saturation constants on the scales of \(E\) and \(C\). The factor \(1-D\) limits additional damage as the damaged fraction approaches 1, and the saturating terms prevent production from increasing indefinitely. This alternative is not automatically better: it introduces parameters that may be hard to estimate, and saturation alone does not guarantee multiple stable states or persistent inflammation after injury stops. Those behaviors require appropriate nonlinear feedback and must be demonstrated by analysis or simulation, not assumed from the equations’ appearance.

Compared with a one-way cytokine-to-damage model, the coupled model can represent how damage feeds back into cytokine production. Compared with a detailed cell-population model, it is easier to analyze but cannot distinguish which epithelial or immune populations generate the signals. The choice depends on the question and available data. Neither model establishes causation from cytokine covariance alone; mechanistic links require supporting measurements or perturbations.