Why Long-Term Persistence Needs More Than an Early Expansion Curve
After engineered cells expand and contract, their later counts depend on what happens to individual cells and on whether the environment supports replacement or proliferation. A declining count may be consistent with ongoing loss, but the shape of that decline matters: a constant loss rate predicts exponential decay, a changing rate predicts a different survival curve, and homeostatic support can produce a plateau or renewed increase. These alternatives can imply very different counts months after infusion.
The models below describe an idealized viable-cell count \(N(t)\), where \(t\) is time after a defined starting point and \(N(0)=N_0\). Counts are treated as continuous for modeling, even though individual cells are discrete. The models concern persistence in the modeled cell population; they do not, by themselves, explain every biological process that can alter a measured count.
Constant-Hazard Survival
If the hazard is constant, \(h(t)=k\), each surviving cell has the same instantaneous loss rate \(k\) throughout the interval. With no replacement or proliferation, the count obeys \(dN/dt=-kN\). The solution is exponential decay. A useful equivalent parameter is the half-life: the time required for the expected count to fall to one-half its initial value.
The half-life is constant in this model: every interval of length \(t_{1/2}\) multiplies the remaining count by one-half. This does not mean that the same number of cells is lost in each interval. The absolute number lost becomes smaller as the population becomes smaller.
Worked Example: Constant-Hazard Loss Over One Year
Suppose, hypothetically, that \(N_0=1.00\times10^6\) viable engineered cells are present at the start of follow-up and that their constant-hazard half-life is 180 days. Estimate the count after 365 days.
First convert the half-life to a hazard:
Using \(N(t)=N_0e^{-kt}\), the exponent at day 365 is \(kt=\ln(2)\times365/180\approx1.405548\), which rounds to 1.406 at three decimal places. Therefore,
The prediction is about 245,200 cells. The model assumes the same hazard throughout all 365 days and no cell production or influx. The arithmetic is consistent with the half-life: after 360 days, two half-lives have elapsed and the predicted count is \(2.50\times10^5\) cells; the additional five days bring it slightly lower.
Time-Varying Survival
A constant hazard is a specific assumption, not a requirement of survival modeling. If the hazard changes with time but cells are still neither produced nor replenished, the count follows \(dN/dt=-h(t)N(t)\). Separating variables and integrating from 0 to \(t\) gives the cumulative hazard \(H(t)\), and the surviving fraction is the exponential of its negative.
The integral is dimensionless: hazard has units of inverse time and the integration variable has units of time. For piecewise-constant hazards, the cumulative hazard is the sum of each period's hazard multiplied by its duration. The model still assumes no production or influx; time variation changes the loss pattern, not that structural assumption.
Worked Example: A Lower Hazard After the Initial Period
Consider a hypothetical population starting at \(1.00\times10^6\) cells. Suppose its hazard is \(\ln(2)/180\) per day for days 0–180, then falls to \(\ln(2)/360\) per day. Estimate the count on day 365 and compare it with the constant 180-day half-life model.
During the first 180 days, the cumulative hazard is \(\ln(2)\), so one-half of the starting cells remain: \(5.00\times10^5\) cells. During the next 185 days, the cumulative hazard is \(\ln(2)\times185/360\approx0.356201\). Thus,
The time-varying model predicts about 350,200 cells at day 365, compared with about 245,200 under a constant 180-day half-life. The difference follows from the lower hazard after day 180. This calculation does not establish that the hazard truly changes; it shows how that assumption affects the late-count prediction.
Homeostatic Support and a Persistent Equilibrium
Survival-only models cannot produce new cells. A different model is needed if proliferation is supported when the population is below a homeostatic level. One simple representation combines density-limited proliferation with cell loss:
Here \(r\) is the maximum per-capita proliferation rate, \(K\) is a density scale at which proliferation is limited, and \(\delta\) is a per-capita loss rate. All rates have units of inverse time and \(K\) has units of cells. The model assumes a well-mixed population with constant parameters and a proliferation rate that declines as \(N\) approaches \(K\). It treats homeostatic support indirectly through this density-dependent proliferation term; it does not model a particular cytokine, niche, or support-cell population.
When \(r>\delta\), define \(\lambda=r-\delta\) and \(N_*=K\lambda/r\). Then \(dN/dt=\lambda N(1-N/N_*)\). The positive equilibrium \(N_*\) is stable for positive counts: counts with \(0<N<N_*\) increase, while counts with \(N>N_*\) decrease. A count of exactly zero remains zero. To obtain the trajectory, let \(Y=1/N\). Differentiating gives \(dY/dt=-\lambda Y+\lambda/N_*\). Solving this linear equation with \(Y(0)=1/N_0\) gives the count formula below.
If \(r<\delta\), there is no positive equilibrium in this model and the count declines toward zero. If \(r=\delta>0\), the equation becomes \(dN/dt=-(r/K)N^2\), and a positive count tends toward zero more slowly than exponential decay. If \(r=\delta=0\), the count remains constant. Thus, a nonzero long-term plateau is a model consequence only under conditions that permit positive net growth at low density.
Worked Example: Homeostatic Support Changes the Late Count
Suppose, hypothetically, that \(N_0=1.00\times10^5\) cells, \(K=1.00\times10^6\) cells, \(r=0.006\ \text{day}^{-1}\), and \(\delta=0.004\ \text{day}^{-1}\). Find the equilibrium and predict the count after 365 days.
Since \(r>\delta\), the net low-density rate is \(\lambda=0.006-0.004=0.002\ \text{day}^{-1}\), and
Substitute into the trajectory. The exponent is \(-\lambda t=-0.002\times365=-0.730\), and \(N_*/N_0-1=3.333-1=2.333\):
The predicted count rises from 100,000 toward the equilibrium of about 333,300 cells. Under survival-only loss at the same \(\delta=0.004\ \text{day}^{-1}\), the prediction would instead be \(1.00\times10^5e^{-0.004\times365}\approx2.322\times10^4\) cells. The contrast is not a small adjustment: it comes from adding density-dependent proliferation to the model.
Comparing Model Assumptions and Late-Count Implications
| Model | Core assumption | Typical late-count behavior |
|---|---|---|
| Constant-hazard survival | Fixed per-cell loss rate; no production or influx | Exponential decline toward zero |
| Time-varying survival | Loss rate changes with time; no production or influx | Decline shaped by cumulative hazard; may slow or accelerate |
| Homeostatic support | Density-dependent proliferation offsets loss | May approach a positive equilibrium if low-density proliferation exceeds loss |
The models answer different biological questions. Use constant-hazard survival when a stable loss rate is a defensible approximation over the prediction interval. Use time-varying survival when there is a reason to represent changing loss, such as a specified change in the modeled cell-loss process. Use a homeostatic-support model when population-dependent proliferation is part of the mechanism being considered. A more flexible model is not automatically better: its additional parameters require biological justification and information in the observations.
Diagnostics, Limitations, and a Common Mistake
Before interpreting a long-term prediction, check that units and limiting behavior make sense. A hazard must be nonnegative and have units of inverse time; its integral must be dimensionless. A survival-only prediction must not increase above \(N_0\). In the homeostatic model with \(r>\delta\), a prediction should move toward \(N_*\), not away from it. Calculated counts must remain nonnegative, and the implied loss or proliferation rates should be biologically plausible over the whole prediction period.
A model can fit observed counts while its mechanism remains uncertain. For example, a slowly declining count could reflect a declining loss hazard, production that partly offsets loss, or processes outside the model. Counts alone do not necessarily distinguish these explanations. The equations here also omit cell-state heterogeneity and changes in support, and they assume parameters remain fixed except where a time-varying hazard is explicitly specified. Long extrapolations are especially sensitive to those assumptions.
- Common mistake: “The late count is higher than the constant-hazard prediction, so the hazard must have fallen.” Correction: a lower hazard is one possible explanation, but proliferation or another process could also raise the count relative to a survival-only prediction. State which mechanism the model includes, and do not infer it from the count trajectory alone.
- Common mistake: using the time-varying survival formula while allowing homeostatic proliferation to replenish cells. Correction: the integrated-hazard formula assumes no production or influx. Include a production or proliferation term if that mechanism is part of the model.
- Common mistake: interpreting \(K\) itself as the equilibrium count. Correction: in the model with cell loss, the positive equilibrium is \(N_*=K(r-\delta)/r\), which is below \(K\) when \(\delta>0\).
These models are alternatives for representing persistence, not interchangeable curve shapes. A robust prediction should make the assumed loss and support mechanisms explicit and should show how late counts change when those assumptions change.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A survival-only population has a constant half-life of 100 days. What fraction remains after 300 days?
- A time-varying survival model has hazard \(0.002\ \text{day}^{-1}\) for 50 days and \(0.001\ \text{day}^{-1}\) for the next 50 days. What fraction of the initial cells remains at day 100?
- In the homeostatic model, \(r=0.005\ \text{day}^{-1}\), \(\delta=0.006\ \text{day}^{-1}\), and \(K=1.00\times10^6\) cells. Is there a positive equilibrium?
- Why can an observed count above a constant-hazard prediction not, by itself, establish that the loss hazard decreased?
Solutions. (1) Three half-lives have elapsed, so the remaining fraction is \((1/2)^3=0.125\), or 12.5%. (2) The cumulative hazard is \(0.002(50)+0.001(50)=0.150\); the fraction remaining is \(e^{-0.150}\approx0.8607\), or 86.07%. (3) No. Since \(r<\delta\), low-density net growth is negative, so this model has no positive equilibrium. (4) A count above the survival-only prediction could result from proliferation or another process as well as from a decreased loss hazard. The count alone does not identify which mechanism produced the difference.