From a Biological Explanation to a Model Structure
A mechanistic claim becomes useful in quantitative systems pharmacology when it specifies how biological quantities change and how those quantities connect to observations. “The drug suppresses inflammation through feedback,” for example, does not yet define a model. It leaves open which signal responds to the drug, what feeds back, whether feedback is immediate or delayed, and which measured quantity represents the relevant biology.
A model structure is the set of state variables, process equations, and observation equations that encode a proposed biological organization. Competing explanations should be written as explicit alternatives: a pathway is present or absent, a cell moves between compartments or is lost, or a measured signal represents free rather than total analyte. Parameter values can change the size or speed of a process within a structure; they do not, by themselves, create a missing state or a different pathway.
For a state vector \(x(t)\), a QSP model can be written generally as \(dx/dt=f(x(t),u(t),\theta)\), where \(u(t)\) denotes external inputs or interventions and \(\theta\) collects parameters. The observation for measurement type \(j\) can be represented as \(y_j(t)=g_j(x(t),\phi_j)+\epsilon_j(t)\). Here \(g_j\) is the measurement mapping, \(\phi_j\) its parameters, and \(\epsilon_j\) measurement error. A biologically distinct explanation can change \(f\), the components of \(x\), \(g_j\), or more than one of these.
Make the Structural Choices Explicit
Start by writing the biological claim in terms of entities and processes. For each candidate explanation, identify what is present in the system, what changes over time, and what causes each change. A useful process inventory asks whether each quantity is produced, consumed, activated, inhibited, transported, or measured. Then write a balance equation for each dynamic state: its rate of change must account for all included inputs and outputs.
The same biological noun does not guarantee the same model structure. “Mediator” might refer to free mediator, total mediator including bound complexes, or a latent activity that an assay detects indirectly. Likewise, “cell depletion” could mean death, movement out of sampled blood, or both. Those interpretations require different states or observation mappings, not merely different parameter values.
Write each explanation as a clear claim about which entities and processes exist, rather than as a vague label such as “stronger inhibition.”
Include a state when its amount or activity can change independently in a way relevant to the mechanism. Draw or list every process that adds to or removes from each state.
Convert the inventory into rate equations. Check that each term has units of state per unit time and that the signs match production, removal, or transfer.
Specify which states an assay detects, how they are combined or transformed, and the units of the reported measurement.
These steps keep three questions separate: what biology is proposed, how it evolves, and what the experiment records. They also make model assumptions inspectable. For example, a well-mixed compartment, constant clearance, or instantaneous binding approximation is a modeling assumption, not an observation.
Worked Example: Immediate Versus Mediated Feedback
Suppose a hypothetical stimulus \(u\) increases a circulating mediator \(M\). One explanation is that \(M\) directly suppresses its own production. A second explanation is that \(M\) induces a separate inhibitor \(F\), which then suppresses production. Treat \(u\) as a dimensionless constant input and concentrations as nM.
For direct feedback, let production be \(k_{\mathrm{in}}u/(1+M/K_f)\) and removal be \(k_{\mathrm{out}}M\). The model is
This structure assumes feedback is an immediate function of the current mediator concentration. Use hypothetical values \(k_{\mathrm{in}}=10\) nM/day, \(u=1\), \(K_f=5\) nM, and \(k_{\mathrm{out}}=1\)/day. At steady state, set \(dM/dt=0\):
The nonnegative solution is \(M=5\) nM; the other root, \(-10\) nM, is not a physically meaningful concentration. Without feedback, the same production and removal terms give \(0=10-M\), so the steady concentration is \(10\) nM. In this example, the direct feedback structure halves the steady concentration.
For mediated feedback, introduce \(F\) as a second state. One possible structure is \(dM/dt=k_{\mathrm{in}}u/(1+F/K_f)-k_{\mathrm{out}}M\) and \(dF/dt=k_FM-k_{\mathrm{deg}}F\). Now feedback has its own production and removal timescale. The added state is justified only if the proposed biological explanation requires a changing inhibitor with dynamics distinct from \(M\); it should not be added just to make a curve fit more closely.
If the assay measures mediator concentration directly, a simple observation equation is \(y_M=M+\epsilon\), with \(y_M\) in nM and \(\epsilon\) representing assay error. An assay for downstream activity would need a different mapping. The equation for feedback alone does not establish what the assay measures.
Representing Redundant Pathways
Biological redundancy means that more than one route can contribute to a downstream process. A model that replaces several routes with one effective input may be a useful approximation, but it is a different structural claim: it does not separately represent the routes or their individual states. The choice depends on the biological question and on what quantities are measured.
Worked Example: One Driver Versus Two Redundant Ligands
Consider two ligands, \(L_1\) and \(L_2\), that can each activate a shared downstream response. For a single ligand, define fractional activation \(r_i=L_i/(K_i+L_i)\), where \(K_i\) and \(L_i\) have the same concentration units. A one-driver structure uses \(A=r_1\). A redundant-route structure assumes independent route contributions and defines combined activation as the probability that at least one route is active:
Suppose \(L_1=K_1=2\) nM and \(L_2=K_2=4\) nM. Then \(r_1=2/(2+2)=0.5\) and \(r_2=4/(4+4)=0.5\). Under the one-driver structure, \(A=0.5\). Under the redundant-route structure, \(A=1-(1-0.5)(1-0.5)=0.75\). These are dimensionless activation fractions.
If downstream mediator \(M\) is produced in proportion to activation, one possible process equation is \(dM/dt=k_AA-k_{\mathrm{out}}M\), where \(k_A\) has units of concentration per time. This equation is only appropriate if the biological claim supports proportional production over the modeled range. If the pathways have different downstream intermediates, those intermediates should be represented as states rather than hidden in \(A\).
The observation model also matters: a phospho-protein assay might report a signal proportional to \(A\), while an assay of the downstream mediator reports \(M\). They are not interchangeable observations. The assumed independence of route contributions is a biological simplification; if the ligands compete for one receptor or share a limiting component, the combination equation must reflect that instead.
Separate Cell Movement from Cell Loss
A decline in a sampled compartment is not, by itself, evidence that cells have died. Cells may leave the sampled compartment and accumulate elsewhere. Distinguishing these explanations requires states and pathways that represent the relevant locations, as well as an observation model that says which locations were measured.
Worked Example: Blood Decrease Through Trafficking
Let \(B\) and \(Q\) denote cell counts in blood and tissue, respectively, in cells. Consider constant production \(s\), movement from blood to tissue at rate \(k_{\mathrm{in}}B\), return at rate \(k_{\mathrm{out}}Q\), and death rates \(d_BB\) and \(d_QQ\). The balance equations are
Use hypothetical values \(s=100\) cells/day, \(k_{\mathrm{in}}=0.2\)/day, \(k_{\mathrm{out}}=0.1\)/day, and \(d_B=d_Q=0.1\)/day. At steady state the tissue equation gives \(Q=0.2B/(0.1+0.1)=B\). Substituting into the blood equation gives \(0=100+0.1B-0.3B\), so \(B=500\) cells and \(Q=500\) cells.
Now suppose a treatment changes only \(k_{\mathrm{in}}\) to \(0.4\)/day. The tissue steady-state equation gives \(Q=0.4B/0.2=2B\). The blood equation becomes \(0=100+0.1(2B)-0.5B\), so \(B=333.3\) cells and \(Q=666.7\) cells. Blood count falls by \(166.7\) cells, or \(33.3\%\) relative to baseline, while the combined count remains \(1000\) cells. In this model, that decrease is redistribution rather than an increase in cell death.
If only blood is sampled, a simple observation model is \(y_B=B+\epsilon_B\), with both \(y_B\) and \(B\) in cells. That measurement alone cannot show that tissue count increased. A tissue measurement or other information is needed to distinguish the trafficking interpretation from alternatives that remove cells. A blood-only model might summarize the decline as loss, but it cannot represent movement into an unmeasured tissue compartment.
Observation Models Are Part of the Mechanism Account
A model’s latent states are not automatically the quantities reported by assays. An observation equation should specify whether the assay detects one state, a sum of states, or a transformed signal. This distinction can change how biological evidence is interpreted even when the process equations are unchanged.
Worked Example: Free and Total Analyte Measurements
Suppose a ligand \(L\) binds a protein, creating complex \(C\). Under a rapid equilibrium approximation, let total binding protein be \(R_{\mathrm{tot}}\) and the dissociation constant be \(K_d\), both in nM. The complex concentration is \(C=R_{\mathrm{tot}}L/(K_d+L)\). A free-ligand assay reports \(L\); a total-ligand assay reports \(L+C\).
For hypothetical values \(L=2\) nM, \(R_{\mathrm{tot}}=8\) nM, and \(K_d=2\) nM, the complex is \(C=8(2)/(2+2)=4\) nM. The free assay reads \(2\) nM, whereas the total assay reads \(2+4=6\) nM, before measurement error. Treating the total result as if it were a free concentration would therefore misstate the observed quantity by \(4\) nM in this example.
The equilibrium formula assumes binding is fast relative to the timescale of interest and that the stated total protein is available for binding. If those assumptions do not hold, a dynamic complex state may be needed. The observation mapping should match the assay definition, not be selected only because it improves numerical agreement.
Checks, Alternatives, and Limitations
A structural formulation should survive basic checks before calibration or interpretation. Check that units agree in every equation, that each state has a clear biological meaning, and that mass or cell balance is consistent with the proposed sources, sinks, and transfers. For nonnegative biological quantities, inspect whether the equations can generate negative concentrations or counts under plausible inputs. A negative state may indicate an unsuitable functional form, missing constraints, or a numerical problem.
Also inspect the observation model against assay definitions and sampling compartments. A blood count is not a whole-body cell count; total analyte is not necessarily free analyte; and a downstream activity assay need not equal a mediator concentration. Residual patterns from calibration, source-specific measurement differences, and lineage records can help locate inconsistencies, but a good fit alone does not prove that a proposed pathway exists.
Alternative formulations are often useful for different purposes. A detailed compartment model represents transport explicitly; a reduced model may combine compartments when only aggregate behavior matters. A pathway-resolved model can retain distinct mediators; an effective-input model can summarize them when the data and question do not require separate states. These alternatives trade biological detail against parsimony and interpretability. A reduced structure should be described as an approximation, not as evidence that omitted biology is absent.
A common mistake is to treat every mismatch as a reason to add a new state. The reasoning “the fitted curve is poor, therefore there must be an unmodeled inhibitor” is not sufficient: the issue could instead be an incorrect observation mapping, a misspecified existing process, poor data quality, or an inappropriate assumption. The correction is to identify the specific biological claim represented by the added state, write its inputs and outputs, and ask whether the available measurements can inform it. Conversely, a model that fits blood counts by increasing death should not be called a trafficking model unless it includes a compartment-transfer process.
Check Your Understanding
Answer each question by identifying the structural choice involved. Brief solutions are provided after the questions.
- A measured blood-cell count falls after dosing. What extra state and process would represent redistribution to tissue rather than death?
- Why is an equation for a total-analyte assay not necessarily the same as one for a free-analyte assay?
- If two ligands activate a response independently, and each has fractional activation \(0.4\), what combined activation follows from \(A=1-(1-r_1)(1-r_2)\)?
- What is the difference between changing a clearance parameter and adding a delayed feedback state?
- Name two checks that can identify a poorly specified QSP structure before interpreting its biological meaning.
Solutions. (1) Add a tissue cell-count state \(Q\) and transfer terms from blood to tissue, and optionally back; death remains a separate removal term. (2) Total assay signal includes free and bound analyte, while a free assay reports only free analyte; their observation mappings differ. (3) \(A=1-(0.6)(0.6)=0.64\), dimensionless. (4) A clearance change alters the rate of an existing process; a delayed feedback state adds a distinct time-varying quantity with its own production and removal dynamics. (5) Examples include checking units and signs in each balance equation, checking consistency with measured compartments and assay definitions, and checking whether nonnegative states remain biologically plausible.