Delayed Tumor Response After Treatment Initiation

Why Treatment Effects May Appear Late

Treatment can begin before its effect on viable tumor burden is measurable. Drug exposure may need to trigger downstream processes before cells are killed, and tumor volume may respond only after enough cells have been lost. A model that applies the full killing effect immediately at treatment initiation can therefore predict early shrinkage when the observed tumor is still growing. A transit chain represents this delay by passing a treatment-effect signal through sequential stages before it acts on the tumor.

The chain is a phenomenological representation: it captures the timing and shape of a delayed response without asserting that each transit stage is a separately measured biological process. The model below isolates the delay in the killing signal. Tumor growth is exponential, and the delayed signal scales a constant maximum killing rate. This makes the effect of the lag explicit; more complex growth laws or exposure-effect relationships can be substituted when justified by the question and data.

Definition: A transit chain is a sequence of $n$ first-order stages through which a treatment-effect input passes before reaching its output. Let $u(t)$ be the normalized input, $z_i(t)$ the signal in stage $i$, $k_{\mathrm{tr}}$ the transit rate, and $q(t)=z_n(t)$ the output. For $n\geq 1$, $$ \frac{dz_1}{dt}=k_{\mathrm{tr}}\bigl(u(t)-z_1(t)\bigr), \qquad \frac{dz_i}{dt}=k_{\mathrm{tr}}\bigl(z_{i-1}(t)-z_i(t)\bigr),\quad i=2,\ldots,n. $$ Here $k_{\mathrm{tr}}$ has units of inverse time. If $u(t)$ is a unitless normalized effect, each $z_i(t)$ and $q(t)$ is unitless as well.

Step Input and the Shape of the Lag

Suppose treatment starts at time zero, the normalized input changes from zero to one, and every transit stage initially has value zero. The response is not a simple shift of the input: the output rises gradually. For one stage, $q(t)=1-e^{-k_{\mathrm{tr}}t}$. Additional stages make the early rise more gradual, even when the input and transit rate are unchanged.

The result can be derived by viewing each stage as a first-order delay. The time spent passing one stage has an exponential distribution with rate $k_{\mathrm{tr}}$; the total transit time through $n$ serial stages is the sum of $n$ such times. Its density at time $s\geq 0$ is $k_{\mathrm{tr}}^n s^{n-1}e^{-k_{\mathrm{tr}}s}/(n-1)!$. The output after a unit step is the fraction that has completed transit by time $t$, so it is the integral of this density from zero to $t$. Repeated integration by parts gives the step response:

q(t)=1-e^{-k_{\mathrm{tr}}t} \sum_{j=0}^{n-1}\frac{(k_{\mathrm{tr}}t)^j}{j!},\qquad t\geq 0.

This expression starts at zero and approaches one. The mean time to pass through the chain is $n/k_{\mathrm{tr}}$: each of the $n$ stages has mean transit time $1/k_{\mathrm{tr}}$, and the total time is their sum. This mean is useful, but it does not fully specify the response. Chains with the same mean can have different onset and rise shapes because their numbers of stages differ.

Worked Example: Checking the Early Output of a Two-Stage Chain

Suppose a unit treatment-effect input begins at time zero. Use a two-stage chain with $k_{\mathrm{tr}}=0.5$ per day. What fraction of the eventual effect has reached the tumor-killing output by day 1?

For $n=2$, the step response is $q(t)=1-e^{-k_{\mathrm{tr}}t}(1+k_{\mathrm{tr}}t)$. At one day, $k_{\mathrm{tr}}t=(0.5\ \text{per day})(1\ \text{day})=0.5$, so $$ q(1)=1-e^{-0.5}(1+0.5) =1-(0.6065)(1.5) \approx 0.0902. $$

Thus about 9.02% of the eventual normalized effect has reached the output by day 1. The two-stage chain has mean transit time $2/(0.5\ \text{per day})=4$ days. The day-1 output is not 25% merely because one day is one-quarter of that mean; the response follows the chain’s curved time course, not a linear ramp.

Putting the Delayed Signal Into Tumor Dynamics

Let $V(t)$ be viable tumor-cell burden, $r$ the untreated exponential growth rate, and $k_{\mathrm{kill}}$ the maximum killing rate under a fully developed normalized effect. Both rates have units of inverse time. If the output $q(t)$ scales killing, the model is

\frac{dV}{dt}=\bigl(r-k_{\mathrm{kill}}q(t)\bigr)V(t).

With $V(0)=V_0$, integration gives the log-burden trajectory. This form also shows why the lag matters: tumor size depends on the accumulated killing signal, not just its value at the final observation time.

\log\!\left(\frac{V(t)}{V_0}\right) =rt-k_{\mathrm{kill}}\int_0^t q(s)\,ds.

For a two-stage chain with a unit step input, $q(t)=1-e^{-k_{\mathrm{tr}}t}(1+k_{\mathrm{tr}}t)$. Integrating this response gives

\int_0^t q(s)\,ds =t-\frac{2-e^{-k_{\mathrm{tr}}t}(k_{\mathrm{tr}}t+2)} {k_{\mathrm{tr}}}.

The integral has units of time, so multiplying it by $k_{\mathrm{kill}}$ produces a dimensionless contribution to log tumor burden. At any instant, the sign of $r-k_{\mathrm{kill}}q(t)$ determines whether burden is increasing or decreasing then. A positive early slope does not imply that treatment has no effect: it can occur while the delayed signal is still too small to offset growth.

Worked Example: Early Growth Despite an Active Delayed Effect

Consider hypothetical tumor dynamics with $V_0=100$ arbitrary burden units, $r=0.03$ per day, $k_{\mathrm{kill}}=0.08$ per day, and a two-stage chain with $k_{\mathrm{tr}}=0.5$ per day. The treatment input is a unit step at time zero. Calculate the model-predicted burden at day 1 and compare the initial response with an immediate killing effect.

From the previous calculation, $q(1)\approx0.0902$. The instantaneous net rate at day 1 is therefore $$ r-k_{\mathrm{kill}}q(1) =0.03-(0.08)(0.0902) \approx0.0228\ \text{per day}. $$

The burden is still increasing at day 1, because the delayed killing signal has not yet overcome the growth rate. To calculate the burden, use the integrated two-stage response. At day 1, $k_{\mathrm{tr}}t=0.5$, so $$ \int_0^1 q(s)\,ds =1-\frac{2-e^{-0.5}(2.5)}{0.5} \approx0.03265\ \text{days}. $$

Consequently, $$ \log\!\left(\frac{V(1)}{100}\right) =(0.03)(1)-(0.08)(0.03265) \approx0.02739, \qquad V(1)=100e^{0.02739}\approx102.8 $$ arbitrary burden units. In contrast, applying the full killing rate immediately would give $V(1)=100e^{(0.03-0.08)(1)}\approx95.12$ units. The contrasting predictions arise from the assumed timing of killing, not from a change in the baseline growth or maximum killing rates.

Same Mean Lag, Different Early Trajectories

The number of stages and the transit rate jointly determine the response shape. Increasing $n$ while increasing $k_{\mathrm{tr}}$ to keep $n/k_{\mathrm{tr}}$ fixed does not generally preserve early output. This distinction matters when interpreting early tumor measurements: a mean lag alone does not tell the model how abruptly or gradually the effect begins.

Worked Example: Comparing Chains With a Four-Day Mean Transit Time

Compare two hypothetical chains, both with mean transit time four days: a one-stage chain with $n=1$, $k_{\mathrm{tr}}=0.25$ per day, and a four-stage chain with $n=4$, $k_{\mathrm{tr}}=1$ per day. Find their outputs at day 2. Then use $r=0.03$ per day and $k_{\mathrm{kill}}=0.08$ per day to determine the instantaneous net tumor rate at that time.

For the one-stage chain, $$ q_1(2)=1-e^{-(0.25)(2)} =1-e^{-0.5} \approx0.3935. $$ For the four-stage chain, $k_{\mathrm{tr}}t=2$, and $$ q_4(2) =1-e^{-2}\left(1+2+\frac{2^2}{2!}+\frac{2^3}{3!}\right) =1-e^{-2}(6.3333) \approx0.1429. $$

The instantaneous net rates are $0.03-(0.08)(0.3935)\approx-0.00148$ per day for the one-stage chain and $0.03-(0.08)(0.1429)\approx0.01857$ per day for the four-stage chain. Thus the one-stage model predicts slight shrinkage at day 2, while the four-stage model predicts continued growth. Both chains have a four-day mean transit time, yet their early outputs differ substantially.

Assumptions, Diagnostics, and Alternatives

The equations are useful only when their assumptions suit the modeling question. In this formulation, the treatment input is a normalized step, the transit chain is linear with the same rate in each stage, and the output scales a constant maximum killing rate. Tumor burden follows exponential growth in the absence of killing, and the treated population is represented as a single viable burden. These assumptions make the lag’s effect easy to isolate; they do not establish that real tumors obey these mechanisms exactly.

When applying the model to data, check whether the implied timing is credible and whether the observations are informative about it. Early measurements are particularly valuable for distinguishing immediate from delayed effects. If observations begin only after the predicted response has developed, different combinations of growth rate, killing rate, and transit parameters may produce similar trajectories. Likewise, sparse measurements around treatment initiation can make the onset time poorly determined. Inspect observed and predicted trajectories together, including the early period, and ask whether systematic early underprediction or overprediction suggests that the chosen delay shape is inadequate.

A fitted transit chain should not be interpreted as proof of a specific sequence of biological events. Its parameters can also be confounded with an assumed growth law, the exposure-effect relationship, and how tumor-cell burden maps to measured tumor volume. The model describes delayed killing in the chosen state variable; measurement changes need not equal changes in viable-cell burden at every time point.

The main alternatives encode different timing assumptions. An immediate-effect model applies killing as soon as its input changes, which is appropriate only when that approximation is adequate for the timescale of interest. A one-stage effect compartment gives a gradual response with a broad first-order delay; it is exactly the $n=1$ transit chain. A fixed-delay model shifts the input by a specified time and then applies it, producing a sharp onset rather than the gradual rise of a multi-stage chain. The transit chain offers adjustable response shape between these choices, but additional stages also add assumptions and parameters. Select among them based on the biological question, sampling times, and whether the observed data can distinguish their predicted early trajectories.

Common pitfall: “The tumor grew after treatment began, so the treatment had no effect.” That conclusion does not follow when killing is delayed. In the model, early growth is possible whenever $r>k_{\mathrm{kill}}q(t)$, even though $q(t)$ is already positive. The relevant check is the modeled balance of growth and delayed killing over time, together with whether the observations support the proposed timing—not the direction of a single early size change alone.

Check Your Understanding

Try each question before reading its solution below. All numerical values are hypothetical.

A three-stage chain has $k_{\mathrm{tr}}=0.6$ per day. What is its mean transit time?
For a one-stage chain with $k_{\mathrm{tr}}=0.4$ per day and unit-step input, calculate $q(2)$.
At a particular time, $r=0.025$ per day, $k_{\mathrm{kill}}=0.10$ per day, and $q(t)=0.20$. Is burden increasing or decreasing at that instant?
Why can two chains with the same mean transit time predict different tumor sizes early after treatment?

Solutions. (1) The mean is $n/k_{\mathrm{tr}}=3/(0.6\ \text{per day})=5$ days. (2) $q(2)=1-e^{-(0.4)(2)}=1-e^{-0.8}\approx0.5507$. (3) The net rate is $0.025-(0.10)(0.20)=0.005$ per day, so burden is increasing at that instant. (4) The mean specifies the average total transit time, but not the response shape; different stage counts and transit rates can yield different fractions of the effect at early times, and therefore different accumulated killing and tumor sizes.