📄 cancer biology

Explaining and predicting pseudoprogression in immunotherapythrough joint modeling of immune infiltration and ctDNA dynamics

This study develops a mechanistic mathematical model integrating tumor-immune dynamics and ctDNA fluctuations to explain pseudoprogression and trains a random forest classifier that demonstrates superior early discrimination between true progression and pseudoprogression compared to standard imaging criteria.

Original authors: Li, A. K., Lou, E., Leder, K., Foo, J.

Published 2026-09-22

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Li, A. K., Lou, E., Leder, K., Foo, J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

When a patient begins treatment with modern immunotherapy, the goal is to wake up the body's own immune system to hunt down cancer. These drugs, known as immune checkpoint inhibitors, remove the brakes that tumors often use to hide from immune cells. However, this awakening can create a confusing signal for doctors. As immune cells rush into a tumor to attack it, they cause swelling and inflammation, much like a crowd gathering at a scene. On a standard medical scan, this influx of cells can make the tumor look larger than it was before treatment started. This phenomenon is called pseudoprogression. It is a critical puzzle in cancer care because a tumor that appears to be growing might actually be dying, while a tumor that is truly growing might be mistaken for a temporary flare-up. If doctors mistake a shrinking tumor for a growing one, they might stop a life-saving treatment too soon. If they mistake a growing tumor for a shrinking one, they might delay switching to a different therapy, allowing the cancer to spread.

For decades, doctors have relied on measuring the size of tumors on scans to decide if a treatment is working. These rules, known as RECIST, were designed for older chemotherapy drugs that shrink tumors quickly and steadily. They do not account for the complex, delayed, and sometimes deceptive behavior of immunotherapy. A newer set of guidelines, called iRECIST, was created to handle this confusion by telling doctors to wait and re-scan patients before declaring a treatment a failure. While this waiting period helps avoid stopping effective treatment too early, it also means that patients whose cancer is truly growing must wait weeks longer before their doctors can switch them to a different plan. During this wait, the cancer continues to grow unchecked. Researchers at the University of Minnesota, led by Aaron Li and Jasmine Foo, set out to solve this timing problem by building a mathematical model that simulates what happens inside a patient's body during immunotherapy.

The researchers created a computer model that tracks three main things: the cancer cells, the active immune cells attacking them, and a specific type of genetic material called circulating tumor DNA, or ctDNA, which leaks into the bloodstream when cancer cells die. In their model, the "tumor" seen on a scan is not just cancer cells; it is the sum of the cancer cells plus the immune cells swarming inside it. This distinction is vital because it explains why a tumor might appear to grow even as the cancer cells are being killed. The model also tracks ctDNA, which acts as a molecular fingerprint of the cancer. When cancer cells die, they release more of this DNA into the blood. The researchers used data from hundreds of real patients treated with a drug called atezolizumab to tune their model, ensuring it matched the real-world ups and downs of tumor sizes and DNA levels.

Once the model was calibrated, the team used it to generate thousands of virtual patients to test different ways of predicting the outcome. They wanted to see if combining the scan data with the blood DNA data could tell the difference between a tumor that is truly growing and one that is just inflamed. They trained a computer program, known as a random forest classifier, to look at the patterns of tumor size and ctDNA levels over time. The results showed that this combined approach was far better at making the right call than looking at scans alone. The computer model could distinguish between true progression and pseudoprogression with much higher accuracy than current standard rules, even when the scan measurements were slightly noisy or imperfect.

Perhaps the most significant finding was the speed of the new method. The standard rule, iRECIST, requires a doctor to wait for a second scan, usually four to eight weeks after the first sign of growth, to confirm if the tumor is truly getting worse. In the researchers' simulations, this waiting period allowed true cancer growth to increase by a median of 144 percent before a new treatment could be started. In contrast, the model using both scan and blood data could make the same determination about 37 days earlier. This earlier detection means that patients with truly aggressive cancer could be switched to a different treatment much sooner, potentially preventing the disease from spreading further, while patients with pseudoprogression could be reassured that their current treatment is still working.

The researchers also tested how well their method held up under different conditions, such as when the blood tests were less precise or when the time between tests varied. The model remained robust, maintaining its accuracy even when the data was not perfect. This suggests that the approach is practical for real-world use, where medical tests are never perfectly clean. However, the team was careful to note that these results come from computer simulations based on existing data. To prove that this method works in the clinic, they argue that doctors need to collect paired data from real patients—scans and blood tests taken at the same time—specifically during the period when a tumor looks like it is growing. Without this specific type of data, it is difficult to know for sure if the model's predictions will hold up in every patient.

The study does not claim to have solved the mystery of pseudoprogression once and for all, nor does it suggest that doctors should stop using current guidelines immediately. Instead, it offers a powerful new tool and a clear path forward. By showing how the interplay between immune cells, cancer cells, and blood DNA creates the confusing signals we see on scans, the researchers have provided a way to cut through the noise. Their work suggests that if we can synchronize our imaging and blood testing, we might be able to see the truth of a patient's response much sooner, sparing them from the anxiety of waiting and the danger of unnecessary delays in treatment. The next step, as the authors propose, is to gather the real-world evidence needed to turn this mathematical insight into a standard part of cancer care.

Technical Summary: Explaining and Predicting Pseudoprogression in Immunotherapy through Joint Modeling of Immune Infiltration and ctDNA Dynamics

Problem Statement
Pseudoprogression (PsP) is a clinical phenomenon observed in patients treated with immune checkpoint inhibitors (ICIs), where tumor lesions initially appear to enlarge on radiographic imaging before eventually regressing. This creates a significant diagnostic challenge: misclassifying PsP as true progression may lead to the premature cessation of effective therapy, while waiting to confirm progression in patients with true disease progression delays the switch to salvage therapies. Current guidelines, such as RECIST 1.1, often misclassify PsP as progression, while immunotherapy-specific guidelines like iRECIST attempt to mitigate this by requiring a follow-up assessment (4–8 weeks later) to confirm progression. However, this delay can result in substantial additional tumor growth for patients with true progression. While biomarkers like circulating tumor DNA (ctDNA) have shown potential in distinguishing these responses, robust statistical analysis is hindered by the scarcity of longitudinal datasets containing both synchronized tumor imaging and ctDNA measurements.

Methodology
The authors developed a mechanistic mathematical model based on ordinary differential equations (ODEs) to simulate tumor-immune dynamics and ctDNA shedding under ICI treatment.

Model Structure: The system tracks four state variables: tumor cells ( $T$ $T$ ), active immune cells ( $I_A$ $I_{A}$ ), exhausted immune cells ( $I_E$ $I_{E}$ ), and circulating tumor DNA ( $C$ $C$ ).
- Tumor-Immune Dynamics: The model incorporates tumor growth, immune cell recruitment, proliferation, and exhaustion. Crucially, it models the "ramp-up" phase where immune cells infiltrate the tumor.
- Observed Volume Definition: Unlike models assuming measured volume reflects only tumor burden, this model defines observed tumor volume ( $V$ ) as the sum of tumor cells and both active and exhausted immune cells ( $V = T + I_A + I_E$ ). This explicitly accounts for immune infiltration contributing to apparent tumor growth.
- ctDNA Dynamics: The model assumes ctDNA is shed upon cell death (both natural turnover and immune-mediated killing) and eliminated from the bloodstream at a fixed rate.
Model Fitting: The model was fitted to longitudinal tumor data from 632 non-small cell lung cancer (NSCLC) patients treated with atezolizumab (POPLAR, BIRCH, OAK, FIR trials) and to a combined tumor-ctDNA dataset from 23 patients treated with pembrolizumab (Bratman et al.). The fitting demonstrated high accuracy (median $R^2 = 0.998$ for tumor size).
Analytic and Clinical Definitions: The authors derived an analytic definition of PsP based on the position of initial conditions relative to a nullcline (where total volume change is zero) and a separatrix (determining long-term tumor escape). They compared these analytic definitions with clinical classifications (RECIST 1.1 and iRECIST) using a simulated cohort of 10,000 patients.
Classification Strategy: Using the model-generated cohorts, the authors trained Random Forest (RF) classifiers to distinguish between three outcomes: progression, pseudoprogression, and response/stable disease. They compared the performance of classifiers using:
1. Tumor data only.
2. ctDNA data only.
3. Coupled tumor and ctDNA data.
  These were benchmarked against RECIST 1.1 and iRECIST under various levels of measurement noise and sampling intervals.

Key Results

Mechanistic Origin of PsP: The model analytically demonstrates that pseudoprogression arises naturally from two mechanisms: continued tumor growth during the initial "ramp-up" phase and the physical volume contribution of infiltrating immune cells.
Model Fit: The ODE model successfully reproduced clinical trajectories for progression, pseudoprogression, and response, achieving high goodness-of-fit metrics on real-world data.
Classification Performance:
- The Random Forest classifier utilizing both tumor and ctDNA data outperformed classifiers using either data source alone.
- In the presence of tumor measurement noise (10% coefficient of variation), the coupled RF classifier significantly outperformed RECIST 1.1 and showed improved robustness over iRECIST.
- The RF classifier achieved a Cohen's Kappa of 0.735 for 3-class classification (vs. 0.656 for RECIST 1.1) and 0.435 for distinguishing progression vs. pseudoprogression (vs. 0.055 for RECIST 1.1) in noisy conditions.
Timing of Classification: The RF classifier provided a median classification time of 45.7 days, which is 37.5 days earlier than iRECIST (82.1 days). This earlier classification potentially limits the median additional tumor volume growth of true progressors by 144% compared to the waiting period required by iRECIST.
Robustness: The classifier's performance remained stable across varying levels of ctDNA measurement noise (up to 40% coefficient of variation) and did not require more frequent sampling than standard tumor imaging schedules.

Significance and Claims
The paper claims that the integration of mechanistic modeling with longitudinal data analysis provides a framework to understand the biological underpinnings of pseudoprogression. Specifically, the authors assert that:

Biological Mechanism: Pseudoprogression is a natural outcome of immune infiltration and delayed immune response, which can be mathematically modeled by including immune cell volume in tumor size assessments.
Predictive Utility: Paired longitudinal measurements of tumor burden and ctDNA dynamics can enable earlier and more accurate discrimination between true progression and pseudoprogression compared to imaging-only criteria.
Clinical Implication: The study motivates the prospective collection of synchronous imaging and ctDNA data during ICI therapy, particularly in cases of suspected disease progression, to facilitate earlier treatment decisions.

Limitations Noted by Authors
The authors explicitly state that their validation on clinical data was limited by the lack of public datasets containing sufficient longitudinal tumor and ctDNA pairs with confirmed pseudoprogression cases. Consequently, the clinical validation of the RF classifier for distinguishing PsP specifically was not possible with available data; the model's performance improvements are currently demonstrated on in silico cohorts generated by their own model. They emphasize that definitive evaluation requires prospective cohorts with synchronized serial imaging and ctDNA measurements.

Technical Summary: Explaining and Predicting Pseudoprogression in Immunotherapy through Joint Modeling of Immune Infiltration and ctDNA Dynamics

More like this