Mouse cancer trials often place the tumor in its organ of origin, such as the brain, adrenal gland or pancreas, instead of under the skin of the flank. These "orthotopic" tumors behave more like human disease. They grow differently, spread to the organs human tumors spread to, and express different driver genes. They also break the statistics used to analyze such trials.
With a flank tumor, the animal is euthanized when a caliper reading crosses a size limit. Because the reading that triggers the dropout has been recorded, standard mixed-effects growth models give unbiased treatment effects. An internal tumor cannot be measured with calipers. Animals are removed when they show symptoms such as neurological signs, ascites or weight loss. Those symptoms track tumor burden that was never recorded, so the data go missing in a way that depends on what was not seen.
We propose to measure how badly this biases treatment-effect estimates in small orthotopic cohorts. We then propose to fix it by design and analysis. The design step is to record a cheap, dense secondary signal, such as bioluminescence imaging every few days, alongside sparse volumetric MRI. The analysis step is to model the two signals jointly in a low-dimensional bivariate mixed model, so that dropout becomes approximately ignorable given everything observed.
Core Claim
Our central claim concerns small cohorts of roughly 6 to 15 animals per arm. In that setting the analysis recommended for subcutaneous trials, a linear mixed model on tumor volume, becomes measurably biased when it is applied to orthotopic trials with symptom-triggered removal. A bivariate mixed model that adds a dense auxiliary burden signal restores near-nominal bias and confidence-interval coverage. More complex joint survival-longitudinal models do not achieve this at these sample sizes.
The point is not a new estimator. It is a practical rule for when an orthotopic trial's missing data can be safely ignored, plus a sampling-density guideline for the auxiliary signal that makes it so.
Why This Gap, and Not the Obvious One
The obvious idea at this intersection is "delta-radiomics or deep temporal models on orthotopic imaging, pretrained on BraTS." Delta-radiomics here means tracking how many image-derived texture and shape features change between scans; BraTS is a large public collection of annotated brain-tumor MRI.
We reject that idea for three reasons, all drawn from the literature.
First, the clinical temporal-imaging work operates at sample sizes one to two orders of magnitude larger than any orthotopic cohort. Cho et al. classify radiation necrosis versus recurrence with a recurrent network over four MRI scans in 369 lesions. Zheng et al. predict pathologic response from pre/post CT changes in 263 patients, filtered through LASSO.
Second, Cho et al. themselves found that the higher-capacity 3D version overfit even at their sample size.
Third, the human-to-mouse anatomical and stromal gap is large. Human stroma is replaced by murine stroma, and the host immune system is absent (Golebiewska et al.; Preclinical Experimental Models for Human Glioma).
A radiomic transfer pipeline evaluated on 10 mice per arm would be uninterpretable. The more basic problem, which gets the treatment-effect estimate itself right, has not been addressed.
The closest existing work is On the Choice of Longitudinal Models for the Analysis of Antitumor Efficacy in Mouse Clinical Trials of Patient-derived Xenograft Models. That paper analyzed 225 mice carrying subcutaneous breast-cancer patient-derived tumors with caliper readings every 3 days. It found that 88% of deaths were planned sacrifices triggered by observed size. Under that condition, a linear mixed model is the right tool, and a joint shared-random-effects model is badly biased: it estimated a growth slope of −0.02 versus 0.24 log(mm³)/week.
This paper motivates our study and also constrains it. It tells us not to reach for joint models reflexively. But its conclusion rests on the dropout trigger being observed, which is exactly what fails in orthotopic placement.
Han et al. supply the experimental template. They implanted SH-SY5Y-Luc neuroblastoma cells into the adrenal gland and the flank of NCG mice and tracked bioluminescence every 4 days. Orthotopic tumors grew faster, metastasized to liver, spleen and intestines, and shortened median survival to about 30 days versus about 40 days for flank tumors. In other words, the orthotopic arm loses animals earlier and for reasons less tied to any recorded measurement. Han et al. compared arms with survival curves and imaging snapshots, not with a missing-data-aware trajectory model.
Research Program
The model treats each animal's tumor history as two linked measurement streams. The first is log tumor volume from MRI, which is sparse (for example weekly) but anatomically accurate. The second is log bioluminescent flux, which is dense (every 2 to 4 days) and cheap, but distorted by tissue depth and by loss of signal from necrotic tumor cores.
We fit a bivariate linear mixed model with these components:
- each stream has its own animal-level random intercept and slope;
- the random effects of the two streams are correlated;
- each stream has a fixed treatment-by-time effect.
Formally, for animal $i$ at time $t$, the two measurements are modeled as
$$ y^{(k)}_{it} = \beta^{(k)}_0 + \beta^{(k)}_1 t + \gamma^{(k)} z_i t + b^{(k)}_{0i} + b^{(k)}_{1i} t + \varepsilon^{(k)}_{it}. $$
Here $k \in \{\text{MRI}, \text{BLI}\}$ indexes the stream, and $z_i$ is the treatment indicator. The quantity $\gamma^{(\text{MRI})}$ is the treatment effect on volumetric growth rate, which is our target. The four random effects $b$ are jointly Gaussian, and $\varepsilon$ is measurement noise.
The model has about a dozen variance parameters, which is fit for single-digit-per-arm cohorts in a way a 3D network is not. The mechanism works as follows. If the symptom-driven removal hazard depends on true burden, and true burden is well predicted by the last bioluminescence reading together with the shared random effects, then dropout is missing at random given the bivariate history. A likelihood-based fit is then unbiased without any explicit dropout model.
The same structure can borrow strength across irregular, animal-specific visit schedules. That is the population-plus-individual idea of Lu et al., who learned a shared projection plus per-subject projections for irregular Alzheimer's imaging histories. Here it is implemented as random effects rather than through optimization of learned projections.
The program has three parts.
The first is a simulation study. Its structure reproduces the reported mouse-clinical-trial design: vehicle and two dose arms, measurements every 3 days, and sacrifice rates near 88%. We change one thing: removal is triggered by a latent burden variable rather than by the recorded volume. We vary four factors:
- the correlation between bioluminescence and latent burden;
- the bioluminescence sampling interval;
- the MRI interval;
- the cohort size, from 6 to 30 per arm. This extends the paper's simulated sample sizes of 300, 200 and 100 mice down to orthotopic scale.
The output is a map of where volume-only analysis fails and how dense the auxiliary signal must be to repair it.
The second is a new orthotopic study with a built-in ground truth. It repeats Han et al.'s adrenal neuroblastoma model with cyclophosphamide versus vehicle. It adds a matched flank arm measured by calipers, weekly MRI, and bioluminescence every 3 days, with symptom-based humane endpoints. A pre-randomized subset of animals is necropsied at fixed days regardless of symptoms, and tumor weight and volume are recorded.
These scheduled sacrifices are missing completely at random by construction, so they give unbiased cross-sectional burden at known times. Each analysis method's predicted mean burden at those days can then be checked against direct measurement.
The third is a second organ to test generality: intracranial glioma xenografts. We would use one model with a methylated MGMT promoter and one without, treated with temozolomide. Golebiewska et al. report that MGMT methylation tracks temozolomide sensitivity in their glioma avatars. That gives an expected direction for the treatment contrast. Neurological-symptom removal in brain tumors is the extreme case of unobserved triggers.
Baselines
The naive reference point is what orthotopic papers typically report: a log-rank comparison of survival plus a t-test on the last available imaging value per animal. This is how Han et al. summarize their arms. It ignores trajectories entirely, and it conflates growth with the timing of symptoms. Beating it means producing a treatment-effect estimate with smaller error against the scheduled-necropsy truth. It also means detecting the cyclophosphamide effect with fewer animals at equal power.
The strongest existing method is the volume-only linear mixed model. The mouse-clinical-trial paper shows it is reliable and unbiased when dropout is triggered by an observed measurement. It is also what a careful statistician would apply today.
All methods are fit to identical data with identical random-effect structure for the MRI stream, using standard restricted-maximum-likelihood software and no method-specific tuning. Beating it means two things. In simulation, bias in $\gamma^{(\text{MRI})}$ must be reduced by at least half, with 95% interval coverage within 92–97%, in regimes where volume-only coverage falls below 90%. In vivo, the predicted burden must agree with the scheduled necropsies where the volume-only model does not.
No reported result found on Lacuna for treatment-effect bias or coverage of any method in orthotopic trials. The only reported reference figures are the subcutaneous growth slopes from the mouse-clinical-trial paper: 0.24 log(mm³)/week for the mixed model and −0.02 for the joint model. Those figures are estimates, not accuracies. We use them to check that our simulator reproduces the subcutaneous regime.
The joint shared-random-effects model links the volume trajectory to a removal hazard. It is the theoretically motivated competitor for informative dropout and is run with the same MRI data plus removal times. That paper reports non-convergence and bias for this model at small N, so we record convergence rates along with bias.
Multiple imputation of missing MRI volumes from bioluminescence is a further competitor. It uses the same auxiliary information but in two stages rather than through a joint likelihood.
The key ablation drops the bioluminescence stream from the bivariate model while keeping the denser MRI schedule that bioluminescence allows us to afford. This isolates whether the gain comes from the auxiliary signal predicting dropout or simply from more measurements.
Why not just…?
Why not just run subcutaneous xenografts, where the safe analysis already works? Because the biology being measured is different. In Han et al.'s neuroblastoma comparison, flank tumors failed to upregulate PHOX2B and GAB2, metastasized mainly to lung, and survived about 10 days longer. A cleanly estimated effect in the wrong biology is not the quantity of interest. Our matched flank arm makes this concrete: if the flank and orthotopic treatment-effect estimates agree once dropout is handled, flank models are vindicated for this question. That would be a useful negative result.
Why not just apply the volume-only mixed model, which the subcutaneous evidence favors? This is the real threat. If symptom-driven removal correlates strongly enough with the last MRI volume, the auxiliary stream adds nothing and our proposal collapses into "use the existing method." The simulation states in advance the burden-to-volume correlation above which this happens. The scheduled-necropsy check settles it empirically. If volume-only matches necropsy truth, we will report that.
Why not just fit a joint survival-longitudinal model to handle informative dropout directly? The mouse-clinical-trial analysis shows such models misattribute dropout and fail to converge at small N. Our cohorts are smaller than its smallest simulated size of 100. We include the joint model as a baseline rather than dismissing it. If it converges and is unbiased in the orthotopic regime, that overturns the subcutaneous lesson, and we would report that too.
Why not just extract delta-radiomic or deep temporal features from the MRI? Even with hundreds of lesions, Cho et al. found their 3D recurrent model overfit relative to the 2D one. The 2D one reached an AUC of 0.878 against 0.822 for maximum diameter alone, a modest margin over a single size measure. At 10 animals per arm, a feature-rich model cannot be validated. It also does not address the missing-data bias, which affects any downstream estimate.
Decisive Checks
The proposal succeeds if three things hold. First, simulation shows a clear regime with these properties:
- symptom-triggered removal pushes volume-only coverage below 90%;
- bioluminescence every 3–4 days restores coverage;
- the joint model is unstable.
Second, in the neuroblastoma study, the bivariate model's predicted burden at scheduled necropsy days falls within the necropsy confidence interval while the volume-only model's does not. Third, the same pattern appears in the intracranial glioma arm, with the temozolomide contrast in the expected MGMT direction.
The proposal fails informatively if bioluminescence carries no information about removal beyond MRI volume. That can happen when necrotic cores erase luciferase signal, which is itself worth knowing for orthotopic trial design. The deliverable in either case is a sampling and analysis rule for small orthotopic trials, grounded in a ground-truth comparison that the field currently lacks.
References
- Prediction of treatment response after stereotactic radiosurgery of brain metastasis using deep learning and radiomics on longitudinal MRI data
- Temporal radiomics for non-invasive preoperative prediction of pathologic complete response to neoadjuvant chemoimmunotherapy in non-small cell lung cancer
- Patient-derived organoids and orthotopic xenografts of primary and recurrent gliomas represent relevant patient avatars for precision oncology
- Preclinical Experimental Models for Human Glioma
- On the Choice of Longitudinal Models for the Analysis of Antitumor Efficacy in Mouse Clinical Trials of Patient-derived Xenograft Models
- Different tumorigenicity and distinct metastasis and gene signature between orthotopic and subcutaneous neuroblastoma xenografted mice
- Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments