Why Plausible Fits Can Still Be Wrong
Cell and gene therapy data often measure only part of a changing biological system. A dose record describes what was administered, a blood count describes cells in one sampled compartment, a vector assay describes genomes detected by that assay, and a protein assay describes product abundance. These quantities can be related, but they are not interchangeable. A model may fit the observations while assigning the wrong biological meaning to its states or rates.
The useful question is not simply whether a curve follows the data. It is whether the modeled state corresponds to the measurement, whether the assumed processes can produce the observed time course, and whether the available measurements distinguish those processes. The following checks target four recurrent failure modes.
Failure Mode 1: Administered Dose Is Not the Viable Starting State
The labeled or recorded dose may count cells before final handling, include nonviable cells, or include viable cells that do not enter the compartment represented by the model. The initial state for a viable-cell kinetic model should reflect the viable cells entering that modeled system, not automatically the nominal dose. The conversion from dose to starting conditions is a separate modeling step.
For a simple accounting model, let \(D\) be the nominal cell dose, \(v\) the viable fraction at administration, and \(r\) the fraction of viable cells recovered into the modeled compartment. If these fractions apply sequentially, the initial viable state is \(C_0=Dvr\). This calculation assumes the fractions are known or appropriately estimated, refer to the same product and administration, and are not already included in the reported dose.
Worked Example: Converting Dose to a Viable Starting State
Suppose a hypothetical cell product has a nominal dose of \(2.00\times10^8\) cells. The measured viable fraction is 0.72, and an independently justified recovery fraction into the modeled compartment is 0.65. What should the model use as its initial viable-cell state?
The viable administered count is \(2.00\times10^8\times0.72=1.44\times10^8\) cells. Applying the stated recovery fraction gives
Under these assumptions, \(9.36\times10^7\) viable cells enter the modeled compartment. Initializing the model at \(2.00\times10^8\) would overstate that starting state by a factor of \(2.00/0.936=2.137\). If early observations are lower than that nominal dose, a model initialized at the nominal dose may compensate by estimating implausibly rapid cell loss. This example does not imply that recovery is always a distinct known fraction; if it is unknown, it needs a justified estimate, a measurement, or a clear simplifying assumption.
A diagnostic is to compare the dose record with viability and recovery information and to check the time and compartment represented by the first count. An apparent early decline may reflect a mismatch between dose and initial state rather than biological death. Conversely, correcting the starting state does not prove that subsequent disappearance is survival loss: redistribution and measurement differences may also matter.
Failure Mode 2: Vector Genomes Are Not Transgene Expression
A vector-genome assay and a transgene-product assay answer different questions. Detected genomes need not all be transcriptionally active or productive, and product concentration also depends on formation and elimination. The distinction established in vector-kinetic and product-effect modeling can be represented with separate states.
Consider a deliberately minimal model with \(G(t)\), measured vector genomes in a relevant tissue, and \(P(t)\), transgene product concentration. Suppose genomes clear at first-order rate \(k_G\), a constant fraction \(q\) is productive, product forms at rate proportional to productive genome abundance, and product is eliminated at first-order rate \(k_{\mathrm{el}}\):
Here \(k_G\) and \(k_{\mathrm{el}}\) have units of inverse time; \(q\) is a dimensionless fraction; and \(k_{\mathrm{syn}}\) converts productive genome abundance into product concentration per unit time. These equations assume no genome input after the initial condition, constant rates and productive fraction, and no other product source. In particular, the genome state decays; a positive, nonzero steady-state \(G\) is not implied by this model.
Worked Example: Product Can Lag Behind Genome Abundance
For hypothetical initial conditions, take \(G(0)=100\) genome units and \(P(0)=0\) ng/mL, with \(k_G=0.10\ \text{day}^{-1}\), \(k_{\mathrm{el}}=0.20\ \text{day}^{-1}\), \(q=0.50\), and \(k_{\mathrm{syn}}=0.05\) (ng/mL)/day per genome unit. Calculate \(P(5)\).
The genome solution is \(G(t)=G(0)e^{-k_Gt}\). Substituting this into the product equation gives a linear differential equation with a decaying input. With \(P(0)=0\) and \(k_{\mathrm{el}}\ne k_G\), its solution is
At five days, the initial product formation rate is \(k_{\mathrm{syn}}qG(0)=0.05(0.50)(100)=2.50\) ng/mL/day. Therefore,
The product is positive even though it began at zero, and its value reflects both production and elimination over time. A genome measurement alone cannot establish this product concentration: the productive fraction, synthesis rate, and product elimination all matter. The calculation assumes the stated genome pool is not replenished; adding a genome input would require changing the genome equation and its solution.
A common diagnostic is a systematic mismatch between genome and product time courses—for example, genomes persist while product declines, or product rises after genome abundance has begun to fall. Such patterns do not by themselves identify the cause, but they warn against imposing a direct proportionality \(P(t)=aG(t)\) without justification. Compare candidate models using both types of measurement when available, and check whether the assay reports total, intact, or otherwise restricted genome material before assigning biological meaning to \(G\).
Failure Mode 3: Blood Counts Can Confound Trafficking and Loss
A blood sample observes cells in blood, not necessarily all cells in the body. If cells move between blood and tissue, a fall in blood abundance can occur without a fall in total abundance. A one-compartment loss model can therefore assign a clearance rate to what is actually redistribution. This is especially concerning when only one compartment is measured or when sampling begins during rapid distribution.
For illustration, let \(B(t)\) and \(T(t)\) be cell abundances in blood and tissue. In a short interval, suppose there is reversible first-order exchange but no proliferation or death:
The model assumes exchange rates \(k_{BT}\) and \(k_{TB}\) are constant and that no cells enter or leave the combined two-compartment system. Adding the equations gives \(d(B+T)/dt=0\), so total abundance is conserved even though the blood measurement can change substantially.
Worked Example: Apparent Blood Loss from Redistribution
Suppose a hypothetical infusion begins with \(B(0)=100\) cell units and \(T(0)=0\). Let \(k_{BT}=0.80\ \text{day}^{-1}\) and \(k_{TB}=0.20\ \text{day}^{-1}\), with no growth or death over the interval. Calculate blood abundance after one day and the apparent one-compartment loss rate inferred from the blood values alone.
The total \(N=B+T\) remains 100. At equilibrium, exchange balances, so \(k_{BT}B=k_{TB}T\); together with \(B+T=100\), this gives \(B_{\mathrm{eq}}=20\). The deviation of \(B\) from equilibrium decays at rate \(k_{BT}+k_{TB}=1.00\ \text{day}^{-1}\), hence
Thus \(B(1)=20+80e^{-1}=49.43\) cell units. If these two blood observations were forced into a no-input, one-compartment exponential-loss model \(B(t)=B(0)e^{-k_{\mathrm{app}}t}\), the apparent rate would be
That apparent loss rate is not a death rate in this example: total abundance is still 100 cell units. The calculation demonstrates how a good description of blood decline can misrepresent the underlying process. It relies on the stipulated two compartments and exchange-only dynamics; real systems may also have proliferation, death, additional tissues, or measurement error.
Useful diagnostics include comparing blood with tissue or whole-body measurements when feasible, checking whether blood decline is followed by return, and examining whether a fitted loss rate changes substantially with the sampling window. A single blood time course may not distinguish loss from movement. A model with more compartments is not automatically better: each additional state should correspond to a plausible process and, where possible, be informed by measurements that help distinguish it.
Failure Mode 4: Expansion and Loss Terms Must Have Compatible Meanings
A model can double-count loss or mix a net-growth description with a gross-process description. For example, logistic growth,
is often used as a phenomenological net expansion model: its observed rate slows as \(C\) approaches carrying capacity \(K\). If a separate death term \(-\delta C\) is added, the model becomes
This is coherent if \(rC(1-C/K)\) has been defined as a gross proliferation term and \(\delta C\) as a distinct loss process. But if the logistic term was fitted as net expansion already incorporating ordinary cell loss, adding \(-\delta C\) may count some loss twice. The equation alone does not resolve that biological interpretation; the parameter definitions and evidence must.
Worked Example: Checking a Logistic Fit Before Adding Loss
A hypothetical cell count rises from 100 to 400 units in four days. Suppose a no-loss logistic model uses \(K=1000\) units. Calculate the net expansion rate \(r\) needed for that endpoint, then check what happens if a loss term \(\delta=0.20\ \text{day}^{-1}\) is added while \(r\) and \(K\) are left unchanged.
For the no-loss model, the solution is \(C(t)=K/[1+(K/C(0)-1)e^{-rt}]\). Rearranging at time \(t\) gives
With the added loss term, the equation can be rewritten as \(dC/dt=(r-\delta)C[1-C/K_{\mathrm{eff}}]\), where \(K_{\mathrm{eff}}=K(1-\delta/r)\). The unchanged parameters imply \(r-\delta=0.2479\ \text{day}^{-1}\) and \(K_{\mathrm{eff}}=1000(1-0.20/0.4479)=553.5\) units. At day 4, this model predicts
It no longer reproduces the 400-unit endpoint. The result does not show that adding loss is inherently wrong; it shows that a rate fitted as net expansion cannot simply be retained unchanged after altering the model. If loss is biologically distinct and supported by evidence, parameters must be interpreted and estimated under the expanded model. If it is already folded into the net rate, an extra loss term may be redundant or double-counting.
Practical Model Checks and Alternatives
When a model produces an implausible rate or a systematic residual pattern, first check state definitions and model structure before adding parameters. A useful review sequence is:
State what the assay measures, its compartment, units, and any conversion from administered dose.
For each source and loss term, specify the biological process, units, and whether its rate describes a gross or net change.
Calculate initial slopes, long-term behavior, conservation relations, and predictions at observed times; reject interpretations inconsistent with those implications.
Consider a viable-dose conversion, separate genome and product states, or explicit compartments when the biology and measurements support them.
These alternatives answer different questions. A dose-to-viable-state conversion corrects the initial condition but does not explain later redistribution. Separate genome and product states represent production and elimination but do not establish which genomes are productive unless \(q\) or related quantities are supported. A trafficking model distinguishes compartment movement from total loss only to the extent that data inform the compartments. A gross proliferation-plus-death model can be more mechanistic than a net expansion curve, but it needs defensible definitions and enough information to estimate its processes.
Diagnostics include residuals that depend on time or compartment, predicted states that become negative, impossible long-term behavior, and parameters that change markedly when an early time point or assay is omitted. A fit that is numerically successful is not a validation of its biological interpretation. When data do not distinguish two explanations, report that limitation and avoid presenting one fitted rate as a uniquely identified mechanism.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A nominal dose is \(1.50\times10^8\) cells, viability is 0.80, and the justified recovery fraction into the modeled compartment is 0.50. What initial viable state follows from these assumptions?
- In a model with \(dG/dt=-k_GG\) and \(k_G>0\), can a positive constant genome abundance be a steady state without an input? Explain.
- In an exchange-only blood–tissue model, what does \(d(B+T)/dt=0\) imply, and what does it not imply about blood counts?
- When is it reasonable to add a separate loss term to a logistic expansion model, and what must be checked first?
Solutions. (1) \(C_0=(1.50\times10^8)(0.80)(0.50)=6.00\times10^7\) viable cells. (2) No: the only steady state satisfies \(0=-k_GG\), so \(G=0\). A positive constant state requires a balancing input or a different assumption. (3) Total abundance \(B+T\) is conserved under the stated exchange-only model, but \(B\) can rise or fall as cells move between compartments. (4) It is reasonable when the expansion term represents a gross proliferation process and the added loss is a distinct process. Check that the original fitted expansion term did not already represent net growth with that loss incorporated.