Why Insulin-Action Models Can Count One Effect Twice
Insulin can lower circulating glucose through more than one physiological route. It can suppress hepatic glucose production and stimulate glucose uptake in peripheral tissues, among other effects. Representing both hepatic and peripheral actions is not, by itself, double counting: these are distinct fluxes. The problem arises when the same biological action is encoded more than once within a route—for example, as both a direct insulin multiplier on hepatic production and a downstream signaling effect that suppresses that same production.
This risk is common when a model is assembled from modules. One module may describe an insulin-to-flux relationship, while another describes a signaling pathway that also affects the flux. If both are applied without a clear causal link, the model can predict more glucose suppression or uptake than the mechanisms and data support. The remedy is not to remove insulin action indiscriminately. It is to make each modeled effect traceable from its input signal to the flux it changes.
Start With Fluxes and Their Causal Links
A useful audit begins with the glucose balance. Let \(G\) be glucose amount in a modeled pool, \(R_a\) the rate at which glucose appears from outside that pool, \(H\) hepatic glucose production, \(U_p\) peripheral glucose uptake, and \(E\) other glucose removal, such as renal loss. Each rate must use the same amount-per-time units. A simple balance is:
This equation makes the bookkeeping visible. Hepatic production contributes glucose to the pool; peripheral uptake and other removal subtract it. If a model reduces \(H\), it has already represented a glucose-lowering contribution. If it also adds that same hepatic suppression as a separate removal term, or applies it again through a signaling multiplier on \(H\), the balance includes the action twice.
The two common structural choices below can both be valid. The key is not to combine them for the same effect.
Mediated structure: Specify an insulin-driven signaling state \(S_H\), then let that state affect the flux, such as \(H=H_0 f_H(S_H)\). If \(S_H\) is intended to mediate the insulin effect on \(H\), do not also apply a second direct insulin response to \(H\) for that same pathway.
The same distinction applies to peripheral uptake. A reduced-form insulin response can act directly on \(U_p\), or a modeled downstream signal can control it. A model may include both direct and mediated contributions only when they represent distinct, justified mechanisms rather than two encodings of one response.
A Short Derivation: How Repeated Effects Distort the Balance
Suppose the true insulin-related hepatic suppression in a particular condition is a fraction \(r\) of baseline hepatic production \(H_0\), with \(0\leq r\leq 1\). A single application gives \(H=H_0(1-r)\). If the same suppression is applied a second time as another multiplier, the model instead gives:
The extra suppression is the difference between the single-application and repeated-application predictions:
For \(0<r<1\), this difference is positive. Thus the repeated multiplier predicts lower hepatic production than the single effect warrants. In the glucose balance, that error lowers \(dG/dt\) by the same amount, all else equal. The calculation is not a claim that every pair of insulin effects should be combined additively. It shows what happens when an identical fractional effect is mistakenly applied twice.
Worked Example: Repeated Suppression of Hepatic Production
Hypothetical data. In a model, baseline hepatic glucose production is \(H_0=2.00\) mg/kg/min. A specified insulin condition produces a 40% suppression, so \(r=0.40\). Suppose a direct insulin term and a downstream signal are both intended to encode this same suppression.
Applying the response once gives:
Applying the same 40% reduction a second time gives:
The duplicate representation lowers predicted production by \(1.20-0.720=0.480\) mg/kg/min beyond the single specified effect. The correction is to retain one response for this hepatic action, or to define distinct, evidence-supported mechanisms and parameterize their separate contributions explicitly.
Trace Signaling to Peripheral Uptake
A signaling variable is not automatically a separate physiological effect. It may be an intermediate representation of the insulin response already encoded in a flux equation. For example, let \(u\) be a normalized insulin stimulus and \(q\) a downstream signal. If \(q=u\) in the current model and both are used to stimulate the same uptake flux, the two terms may be duplicates. By contrast, a hepatic signal and a peripheral signal can legitimately differ in their dynamics or sensitivity, provided the model treats them as tissue-specific pathways and does not also include the same response through another term.
Worked Example: A Direct Term and a Signaling Term on Uptake
Hypothetical data. Basal peripheral uptake is \(U_0=1.00\) mg/kg/min. The insulin stimulus is \(u=0.50\), the downstream signal is \(q=0.50\), and a response coefficient of \(0.80\) is used. Suppose the model applies both terms multiplicatively:
Substitution gives:
If \(q\) represents the same insulin stimulus rather than a separate mechanism, one response term is appropriate:
The duplicated structure adds \(1.96-1.40=0.560\) mg/kg/min of uptake beyond the single response. If the model instead has evidence for two distinct contributions, their relationship should be stated and supported; the numerical example alone cannot establish that they are biologically identical.
Check the Whole Glucose Balance
An error in either hepatic production or peripheral uptake changes the predicted glucose balance, and errors in both can compound. The following example keeps the flux units consistent and shows how a pathway audit can reveal an implausible net prediction.
Worked Example: Comparing Corrected and Duplicated Net Balance
Hypothetical data. During a specified interval, glucose appearance is \(R_a=0.300\) mg/kg/min, other removal is \(E=0.100\) mg/kg/min, and the correct insulin-responsive fluxes are \(H=1.20\) mg/kg/min and \(U_p=1.40\) mg/kg/min. Use the balance \(dG/dt=R_a+H-U_p-E\).
With each insulin action represented once:
Now suppose the duplicate hepatic and uptake effects from the previous examples are both included: \(H=0.720\) mg/kg/min and \(U_p=1.96\) mg/kg/min. Then:
The predicted net balance shifts downward by \(1.04\) mg/kg/min. This equals the sum of the two extra changes in the balance: \(0.480\) mg/kg/min less hepatic production and \(0.560\) mg/kg/min more peripheral uptake. The example illustrates bookkeeping, not a clinical prediction: the hypothetical values are chosen to demonstrate the arithmetic.
Practical Audit, Diagnostics, and Alternatives
When a model produces unexpectedly strong glucose lowering after an insulin perturbation, inspect the causal path to each flux before adjusting fitted parameters. A structured audit can make hidden duplication easier to find.
List every term that changes hepatic production, peripheral uptake, or another glucose input or removal. Keep production and removal signs explicit in the balance.
For each insulin-responsive term, record whether it is driven directly by insulin, by a modeled signal, or by both. State whether the two inputs represent distinct mechanisms.
Hold other inputs fixed and turn on one proposed insulin effect at a time. Compare the flux change with the intended response and with any available flux-specific observations.
Calculate the signed contribution of each changed flux to \(dG/dt\). Confirm that a pathway is not represented both inside a flux and again as a separate balance term.
Useful warning signs include a flux changing twice when a single signal is activated, an insulin perturbation producing a larger effect than the specified component responses, or fitted pathway coefficients becoming unusually small or compensatory. These are diagnostics, not proof of double counting: strong parameter trade-offs, incorrect units, missing feedback, or an unsuitable response function can produce similar symptoms. Flux-specific measurements help distinguish those explanations. A glucose concentration trajectory alone may constrain the net balance without uniquely revealing how much of the change came from hepatic production versus peripheral uptake.
Two modeling strategies are common. A reduced-form model maps insulin exposure directly to each flux and is often economical when the available data support net responses but not intermediate signaling. A mechanistic mediated model represents a signal between insulin and the flux and can express delays or pathway-specific dynamics, but requires a clear causal structure and enough information to support its additional parameters. Neither approach is inherently immune to double counting. The crucial rule is to avoid applying a direct response and its mediated representation as though they were independent effects.
Tissue-specific insulin resistance also does not justify collapsing every response into one common signal. Hepatic and peripheral sensitivities or time courses may differ. They can be modeled separately when that distinction is supported, while preserving the rule that each specified action enters its target flux once. A mechanistic pathway is a hypothesis about causal organization; a good fit alone does not prove that the chosen organization is correct.
Common Mistake and Key Takeaway
Incorrect reasoning: “The model has a direct insulin coefficient and a downstream signaling state, so both should multiply the flux to make the pathway more mechanistic.” This assumes that two model terms describe two biological actions. They may instead describe the same action at different levels of the causal chain.
Correction: First identify whether the signal mediates the direct insulin effect or represents a distinct mechanism. If it is a mediator of that same effect, use the mediated path to determine the flux response, rather than applying both paths. If the mechanisms are distinct, define their separate contributions and check that their combination is consistent with flux data and the glucose balance.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Baseline hepatic production is \(1.50\) mg/kg/min and one insulin action suppresses it by 20%. What is the single-response prediction? What would applying the same fractional suppression twice predict?
- A model includes a direct insulin effect on peripheral uptake and a downstream signal that is defined to be the same normalized insulin response. What information is needed before treating these as two separate contributions?
- In a glucose balance, hepatic production falls by \(0.30\) mg/kg/min and peripheral uptake rises by \(0.20\) mg/kg/min. What is the resulting change in \(dG/dt\), with other fluxes fixed?
- Why might a close fit to glucose concentrations fail to establish that hepatic and peripheral insulin effects have been correctly separated?
Solutions. For Question 1, a single suppression gives \(1.50(1-0.20)=1.20\) mg/kg/min. Applying it twice gives \(1.50(0.80)(0.80)=0.960\) mg/kg/min, an additional decrease of \(0.240\) mg/kg/min. For Question 2, the model needs a biological rationale and supporting evidence that the direct term and the signal represent distinct actions; if the signal mediates the direct response, applying both to the same flux duplicates it. For Question 3, the hepatic change contributes \(-0.30\) mg/kg/min to the balance and the uptake change contributes another \(-0.20\) mg/kg/min, so \(dG/dt\) changes by \(-0.50\) mg/kg/min. For Question 4, glucose concentration reflects the net balance of inputs and removals, so different combinations of hepatic production and peripheral uptake can produce similar concentration trajectories. Flux-specific data or other independent constraints are needed to separate their contributions.