Tumor Models of Nontarget and New Lesions

A literature review for the 3 RECIST Lesion Model

Pharmacometrics
references
Which published tumor dynamic models describe nontarget lesion progression and new lesion appearance, how each one does it, and where the RECORD-1 three-lesion model sits among them.
Last edited

October 14, 2026

Verdict: models of nontarget and new lesions are rare, and models that keep the two apart are rarer. Of about 40 tumor modeling papers screened on 2026-10-13, six give new lesions a submodel of their own, two treat new and nontarget lesions as separate processes, and neither of those two is a journal paper: one is the 2011 RECORD-1 poster this project is publishing, and the other is a 2022 conference abstract. No published model found here gives nontarget lesions a size, latent or otherwise, and none other than the poster uses a detection probability for new lesions.

Out of scope: models of a single primary tumor, preclinical models, and models of target lesions alone, except where Section 6 lists the lesion-level ones to say what they leave out.

The project’s other documents are on the project index, and the sources below are also listed in the references.

2. Summary Table

The table puts every model found in one row, ordered by how much of the nontarget and new lesion process it represents.

Models that represent nontarget or new lesions. ✅ modeled as its own process; ⚠️ unclear; ❌ absent
# Paper Nontarget lesions New lesions Drug effect on them Linked to
3.1 Stein 2011, everolimus, RCC ✅ Latent size, Hill function to probability ✅ Latent size from 0.1 cm, detection probability ✅ Same growth equation as targets, shared 5 and 10 mg effect PFS
3.2 Krishnan 2022 (PAGE), afatinib and nintedanib, NSCLC and ovarian ✅ Logistic or time-to-event ✅ Logistic or time-to-event ⚠️ Not stated in the abstract PFS, ORR, DoR
3.3 Zecchin 2016, carboplatin ± gemcitabine, ovarian ❌ ✅ Weibull time-to-event Through week-12 tumor size ratio OS
3.4 Krishnan 2021, docetaxel, breast ❌ ✅ Logistic per scan Through largest lesion time course OS
3.5 Bonate 2013, pazopanib, RCC ❌ ✅ Logistic, once per patient Through fitted shrinkage and regrowth rates None
3.6 Król 2016, colorectal ❌ ✅ Recurrent events Treatment arm on the event rate OS
4.1 Yu 2020, ovarian Lumped Lumped Through dSLD/dt PFS
4.2 Baaz 2023, FOLFOX ± panitumumab, colorectal Lumped Lumped Through dSLD/dt PFS
4.3 Kumar 2021 (QSP), pembrolizumab ± ipilimumab, melanoma Lumped Lumped Through target tumor burden RECIST response
4.4 Qi 2023 (QSP), pembrolizumab, NSCLC Lumped Lumped Kumar’s model, refit PFS
4.5 Cerou 2024, amcenestrant, breast In the PFS hazard In the PFS hazard Through target size slope PFS
4.6 Chigutsa 2026, abemaciclib, breast In the PFS hazard In the PFS hazard A direct drug effect on the hazard PFS
5.1 Vera-Yunca 2022, cetuximab, colorectal ❌ Observed covariate ❌ OS
5.2 Krishnan 2021 multistate, docetaxel, breast ❌ Inside a progression state ❌ OS

3. Separate Submodels for Nontarget or New Lesions

Each entry gives the data, what is modeled for nontarget and for new lesions, how the drug enters, what it is linked to, and what it was read from.

3.1 Stein 2011: Everolimus in Renal Cell Carcinoma

  • Data. RECORD-1, everolimus 10 or 5 mg against placebo, 407 patients with a baseline assessment.
  • Nontarget lesions. A latent percentage change in size grows by the target lesion equation and maps to a probability of progression, \((100\% + \Delta y\%)^8 / \big((150\%)^8 + (100\% + \Delta y\%)^8\big)\).
  • New lesions. A latent lesion starts at 0.1 cm, grows by the same equation, and is detected with probability \(y^3/\big((0.5\,\text{cm})^3 + y^3\big)\).
  • Drug effect. Separate 5 and 10 mg effects on target lesions; one shared effect on nontarget and new lesions, since the data did not support two. Placebo growth rates correlated across the three lesion types.
  • Linked to. PFS, by applying the RECIST rules to the three predictions.
  • Read from. The poster. The companion paper (Stein 2012, BMC Cancer) models target lesions only, and cites the poster as its reference 22 for the nontarget and new lesion work.

3.2 Krishnan 2022: Sub-Endpoints for PFS

  • Data. Three phase 3 trials: afatinib against chemotherapy in non-small cell lung cancer (NSCLC), nintedanib with docetaxel in NSCLC, and nintedanib with carboplatin and paclitaxel in ovarian cancer.
  • Nontarget lesions. Nontarget progression at each scan, by logistic regression, or by a time-to-event model where the logistic fit had estimation problems.
  • New lesions. The same, as a separate model.
  • Drug effect. The abstract does not say whether drug exposure or SLD enters the nontarget and new lesion models.
  • Linked to. PFS, objective response rate (ORR), duration of response (DoR) and best overall response, simulated from the sub-endpoint events.
  • Read from. The PAGE 2022 abstract (number 10071). Krishnan, Friberg, Nyberg, Sarr, Niebecker, Wang and Perez Pitarch, Pharmetheus and Boehringer Ingelheim. No journal version was found. Céline Sarr is an author here and on the 2011 poster.

3.3 Zecchin 2016: Carboplatin in Ovarian Cancer

  • Data. One phase 3 trial, carboplatin against carboplatin plus gemcitabine, 336 patients.
  • Nontarget lesions. Not modeled; the authors list them as future work.
  • New lesions. Time to first observed new lesion, with a Weibull hazard \(h(t) = \lambda \gamma t^{\gamma-1} \exp(0.545\,\text{LIVER} + 0.767\,\text{TSR})\), where LIVER marks liver metastasis at enrollment and TSR is the model’s tumor size ratio, held at its week-12 value after week 12.
  • Drug effect. Indirect, through TSR.
  • Linked to. Overall survival (OS), with new lesion appearance as a time-varying covariate raising the hazard of death 3.4-fold.
  • Read from. The full text, through a summarizing fetch; the coefficient values are ⚠️ until checked against Table 3 of the paper.

3.4 Krishnan 2021: Docetaxel in Breast Cancer

  • Data. The docetaxel arm of AVADO, 183 HER2-negative patients, 586 target lesions in seven organs. New lesions appeared in 121 patients.
  • Nontarget lesions. Not modeled.
  • New lesions. Probability of a new lesion at each scan, \(P = e^{\text{DR}} / (1 + e^{\text{DR}})\), with the linear predictor DR driven by the predicted time course of the largest baseline lesion, three or more liver lesions, and time since enrollment.
  • Drug effect. Indirect, through the lesion-level docetaxel model.
  • Linked to. OS and dropout.
  • Read from. The full text, through a summarizing fetch. That summary flagged hazard ratios in the text that do not match the Table 3 coefficients; check before citing a number.

3.5 Bonate 2013: Pazopanib in Renal Cell Carcinoma

  • Data. Phase 2 (220 patients) and phase 3 (423 patients) trials of pazopanib against placebo.
  • Nontarget lesions. Not modeled.
  • New lesions. Logistic regression of whether a patient ever had a new lesion, on individual tumor model parameters and covariates. Prior radiotherapy, baseline tumor size, shrinkage rate, regrowth rate, study and treatment were significant; placebo patients were about twice as likely to have a new lesion.
  • Drug effect. Treatment arm as a covariate, and indirectly through the shrinkage rate.
  • Linked to. Nothing downstream; the model has no time course for new lesions.
  • Read from. The abstract. The one other RCC paper here besides the poster.

3.6 Król 2016: A Statistical Joint Model in Colorectal Cancer

  • Data. FFCD 2000-05, a phase 3 trial of sequential against combination chemotherapy, 410 patients.
  • Nontarget lesions. Not modeled.
  • New lesions. A recurrent-event process, jointly with left-censored SLD and death, linked by shared random effects.
  • Drug effect. Treatment arm on each of the three processes, to ask which component the treatment acts on.
  • Linked to. OS.
  • Read from. The abstract. This is the biostatistics joint-model literature rather than pharmacometrics, and it does not cite the pharmacometric papers.

4. Nontarget and New Lesions Lumped Into One Hazard

These models give PFS a second route besides target progression, and the second route stands for nontarget progression, new lesions and, in some, death, without separating them. This is now the common pharmacometric approach to PFS.

4.1 Yu 2020

Pooled platinum-resistant ovarian cancer data, four treatments. The nontarget hazard is tied to the rate of change of target tumor size, and PFS is the earlier of target progression, derived from the predicted SLD by the RECIST rule, and nontarget progression. Read from the abstract; Baaz 2023 reports the hazard form as \(h(t) = \max(\alpha \, dS\!L\!D/dt + \beta, 0)\).

4.2 Baaz 2023

Yu’s model fitted to PRIME, FOLFOX with or without panitumumab in RAS wild-type colorectal cancer, with ASPECCT as external validation: \(\alpha = 0.0041\) per mm and \(\beta = 0.035\) per month. “Nontarget progression” here covers new lesions, unequivocal nontarget progression and death. Read from the full text, through a summarizing fetch.

4.3 Kumar 2021

A QSP model of pembrolizumab and ipilimumab in melanoma, with several target lesions per virtual patient and a probabilistic model for progression from “nontarget growth or appearance of metastases”, evaluated every 12 weeks and dependent on the sum of target diameters. Calibrated to RECIST response categories. The paper’s stated finding is that nontarget growth and new lesions drove much of the progression in patients whose target lesions shrank, and it says few QSP or clinical models address them. Read from the full text.

4.4 Qi 2023

Kumar’s nontarget model refitted to pembrolizumab in NSCLC, inside a lesion-level virtual trial used to ask who benefits from treatment beyond progression. Read from the full text.

4.5 Cerou 2024

Amcenestrant in breast cancer. The authors tried Yu’s approach and rejected it on the visual predictive check, and instead linked a single PFS hazard to the slope of target tumor size. Nontarget and new lesions enter only as part of the observed PFS events. Read from the full text.

4.6 Chigutsa 2026

Abemaciclib in breast cancer. The PFS hazard depends on change in target size plus a second drug effect “outside that mediated through” target lesions, justified by nontarget and new lesion progression in 36% of patients with measurable disease. Read from the full text.

5. New Lesions as an Observed Covariate

  • Vera-Yunca 2022. Cetuximab with chemotherapy in colorectal cancer, four trials, 1716 patients. Early appearance of a new lesion predicts OS (HR 4.14), entered as observed data, not modeled.
  • Krishnan 2021, multistate. The same AVADO data as 3.4, with states for stable disease, response, progression, second-line treatment and death. Progression is one state, whatever its cause.

6. Lesion-Level Models of Target Lesions

A growing set of papers model each target lesion separately, by organ. They describe within-patient heterogeneity, which is one reason nontarget and new lesions behave differently from the SLD, but none of them models nontarget or new lesions. Kerioui 2022 says so directly, listing new lesions and nontarget dynamics as information the models leave out.

  • Terranova 2018, CICIL clustering of lesion trajectories.
  • Kerioui 2022 (ESMO Open) and 2023 (Biometrics), atezolizumab in urothelial cancer.
  • Chen 2024, organ-specific growth rates across six colorectal trials.
  • Liu 2026, AI-measured lesions from CROWN (lorlatinib), linked to PFS.
  • Dilley 2026, 55,220 lesions from 20 trials. The preprint version models new-lesion progression with Cox regression, according to a search summary not checked against the paper.

The most cited tumor growth inhibition and OS frameworks (Claret 2009, Chan 2021, Struemper 2025) use target lesions only.

7. Evidence That Each Component Carries Information

These are statistical analyses without a dynamic model. They are the case for modeling nontarget and new lesions at all.

  • Litière 2014 (RECIST Working Group). 13 trials, 3758 patients with breast, lung or colorectal cancer. New lesions and nontarget progression were each independently associated with OS, and a search summary reports that they made up 36% and 28% of progression events. The year is given as 2014 in one source and 2015 in another.
  • Stein 2013 (RECORD-1). In 246 everolimus patients, nontarget growth and new lesions at weeks 2–14 each predicted OS (p < 0.001, univariate).
  • Twelves 2015. Two eribulin trials in breast cancer. Progression by new metastases carried worse OS than growth of existing lesions.
  • Osorio 2019 and Topp 2021. Lesion-level progression under PD-1 blockade. Mixed progression, with some lesions shrinking while others grow, was common.
  • Coy 2019 (Radiology). Components of RECIST progression in clear cell RCC. Not read here; listed because it is the indication.

8. Metastasis Models From Mathematical Oncology

A separate literature models how metastases seed and grow, using the Iwata–Kawasaki–Shigesada size-structured equation (2000) and its successors from Benzekry’s group, including brain metastases in NSCLC (Bilous 2019) and relapse in early breast cancer (Nicolò 2020). These models predict the number and size of new metastases from seeding and growth rates, which is the mechanism the poster’s new lesion submodel approximates with one latent lesion. They are fitted to untreated or post-surgical data, not to RECIST time courses under treatment, and none was found applied to a drug trial’s new lesion data.

9. Where the Poster Sits

Three features of the 2011 poster are not in any published model found here:

  1. Nontarget lesions with a size. Every other model treats nontarget progression as an event with a hazard or a probability; the poster gives it a latent size that grows and shrinks with the target lesion equation.
  2. A detection probability for new lesions. The poster’s latent lesion starts below imaging resolution and becomes visible with a probability set by its size. Bilous 2019 models brain metastases too small to see on imaging, but outside a drug trial.
  3. A drug effect estimated separately on each lesion type. That is what found 5 mg weaker than 10 mg on target lesions but not distinguishable on the other two. The lumped models in Section 4 cannot ask this question.

For the preprint, this section is the case for novelty, and Sections 3 and 4 are the related work. It should cite Krishnan 2022 as the closest later approach, and Yu 2020 as the approach most later work adopted.

10. Where to Start Reading

In order, with the question each one answers.

  1. Yu 2020. The lumped nontarget hazard that most later PFS models extend. Read the hazard definition and the simulation algorithm.
  2. Zecchin 2016. The fullest published new lesion model, and its effect on OS. Read the time-to-event section and Table 3.
  3. Krishnan 2022 (PAGE abstract and poster PDF). The closest approach to the 2011 poster, from a group that includes one of its authors. Read for whether drug or SLD enters the nontarget and new lesion models.
  4. Kumar 2021. The QSP version, and the argument that nontarget and new lesions drive progression under immunotherapy.
  5. Litière 2014. The evidence that each component predicts survival independently, for the preprint’s introduction.
  6. Bonate 2013. The other RCC paper, for comparison in the discussion.
  7. Liu H 2025 review (Liu, Ibrahim, Centanni, Sarr, Venkatakrishnan, Friberg; Adv Drug Deliv Rev). A review of integrated oncology models; check its reference list for anything Sections 3 and 4 missed.

11. Search Coverage and Gaps

  • Europe PMC indexes PubMed and open-access full text. Embase, Scopus and Web of Science were not searched, and paywalled full text was searched only where a web search reached it.
  • PAGE abstracts were reached through web search, not by searching the PAGE archive directly. ACoP, ASCO and AACR abstracts were not searched systematically. Conference work is where this topic seems to live, so this is the likeliest gap.
  • PIONEER (Naguib 2026, arXiv 2607.17908) adds a hazard channel for progression the SLD does not explain, according to a search summary; the abstract does not say, and the full text was not read.
  • Industry QSP platforms that simulate RECIST may model new lesions without publishing it.
  • Details marked “through a summarizing fetch” were read by an automated summary of the full text, and every number from them is unverified.
Back to top