GrAFT-PDX: Growth and Failure-Time Modeling of Patient-Derived Xenograft Experiments
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "GrAFT-PDX: Growth and Failure-Time Modeling of Patient-Derived Xenograft Experiments".
Jane: Here's the gist of the paper. GrAFT-PDX introduces a Bayesian model for patient-derived xenograft (PDX) drug screens that jointly predicts tumor growth and survival outcomes using a learned latent function decomposition The model aims to reduce the PDX screening burden by enabling imputation of missing experiments and early stopping of likely unsuccessful drugs.
Tom: Model Overview GrAFT-PDX represents sparse, irregular tumor measurements as noisy observations of an underlying smooth continuous-time trajectory and allows tumor growth to inform death risk It jointly models the tumor measurements and interval observed death time for each mouse, where the PDX line and treatment determine a continuous time median tumor trajectory and survival distribution.
Jane: The model uses a low-rank tensor decomposition over PDX lines, treatments, and time to share information across observed experiments, predict untested PDX-treatment combinations, and capture heterogeneous treatment responses across PDX lines.
Tom: Data Challenges Addressed The modeling problem in the PDXe dataset involves three main challenges "measurements are sparse both across and within combinations" "tumor growth and survival are coupled through the observation process" "treatment response is heterogeneous across PDX lines" The sparsity arises because many combinations are never tested, only 4190 of a possible 13826 (30.3%) PDX-treatment pairs are tested Furthermore, even within tested combinations, tumor size is recorded on only 29.8% of days, on average, over a mouse’s lifespan.
Title and authors: Tom: We just covered the title GrAFT-PDX: Growth and Failure-Time Modeling of Patient-Derived Xenograft Experiments. To put that into simpler terms, it’s a method for using data from mouse tumors to predict both how fast they grow and whether they will survive a treatment.
Jane: It’s essentially building an AI system that looks at the messy, incomplete records of drug tests in mice and uses that information to guess what might happen next for other untested combinations.
Lu: They are solving a problem where you have thousands of potential drug-tumor matchups, but you only actually observe a fraction of them, so they are using advanced math to fill in the gaps.
Meng: Think about it like this: instead of running every single test possible, this model helps you figure out which tests are most worth running next based on the existing data.
Tom: Right, so it’s about reducing the sheer volume of physical experiments needed by making smarter decisions about which ones to pursue.
Jane: And they do that by modeling the tumor as a continuous path over time, and linking that path directly to survival outcomes using this new Bayesian structure.
Lu: The Bayesian framework is important because it lets them create a probability distribution for all their unknown parameters, giving them a way to measure how confident they are in their predictions.
Tom: Which is good, because when you’re making decisions about drug screening, you don't want just a single answer; you want to know the range of possibilities.
Jane: And the paper specifically focuses on solving three major hurdles in this area: sparse measurements both across and within combinations, the coupling between growth and survival through how measurements stop when a mouse dies, and treatment response being different for different PDX lines.
Meng: Those three challenges sound exactly like the headaches we have when we try to build any big screening pipeline; it’s hard enough just to get the data clean.
Tom: But GrAFT-PDX is tackling those headaches by using a low-rank tensor decomposition over PDX lines, treatments, and time to share information across all three challenges simultaneously.
Lu: So the tensor decomposition is the engine here; it’s what allows them to connect all those disparate pieces of data together so they can make holistic predictions.
Jane: It’s like creating a shared language for all those different types of experimental data—the tumor sizes, the treatment names, and the time points.
Tom: That sounds like a very powerful way to handle heterogeneity across PDX lines and treatments that we see in biology.
Lu: And they manage to capture that heterogeneity by letting the model learn it directly from the screening data rather than forcing us to manually input every known genetic feature into the fitting process.
Jane: So, even without those explicit genomic inputs during training, they are still capturing meaningful treatment-gene response patterns in their predictions.
Meng: That’s a big deal for production systems; if we can build a model that learns those complex interactions automatically, it cuts down on the manual effort of feature engineering.
Tom: It seems like they’re showing that you don't always need to pre-label everything perfectly to get good results from deep learning on biological data.
Jane: And what this means for the listener is that we’re moving toward AI tools that can look at complex biological systems and suggest the next best experiment without needing a perfect, exhaustive dataset beforehand.
Lu: The focus is clearly on using these learned representations to go beyond just predicting outcomes and actually interrogating the biology behind those predictions.
Tom: So, this paper isn't just about getting a better prediction; it’s about building a system that can suggest biologically coherent next steps for drug development.
The paper's summary: Jane: Now that we understand the setup, let's talk about what they actually found in the GrAFT-PDX paper regarding its performance and what it’s summarizing.
Tom: They show that GrAFT-PDX outperforms standard survival baselines when predicting the time of death on their PDXe dataset. That’s a solid quantitative result, because it shows the tumor growth component is actually adding value to those predictions.
Lu: The coupling between the longitudinal growth and the survival outcome is modeled through this continuous trajectory, which accounts for that informative termination of measurements at death, which biases standard models downward.
Jane: Because standard models might miss that crucial information—that they stop measuring when the mouse dies—GrAFT-PDX incorporates that termination directly into its modeling.
Tom: And for predicting tumor response specifically, it’s competitive with strong baselines for experiments held out from the main training set, and it can improve on static models by using that partial tumor trajectory information.
Jane: That means if you have a set of experiments you haven't run yet, GrAFT-PDX can give you a better prediction than just looking at a single snapshot of growth data.
Lu: They use the low-rank tensor decomposition to predict those unobserved combinations by sharing information across all the observed data points in a way that’s computationally efficient for this kind of sparse data structure.
Tom: And it’s not just about prediction; they are also trying to capture how treatment response varies across different PDX lines, which is a big biological complexity they are modeling.
Jane: They achieve that by using the decomposition to share information about the PDX lines, the treatments, and the time points simultaneously.
Lu: The model represents tumor burden as this continuous trajectory and links its history to survival through a hazard function, which provides a novel way of describing how tumor burden impacts survival in mouse models.
Tom: So if we boil it down, they are using a shared structure to connect sparse observations of growth with the outcome of death into one coherent model.
Jane: It’s about creating one unified framework where the history you observe helps predict the future outcome, whether that outcome is tumor size or survival time.
Lu: The novelty lies in how they use this structure to go beyond just correlation and actually build a predictive model of the underlying dynamics of tumor evolution.
Tom: It sounds like they’ve built a system that captures dynamic behavior rather than just static correlations, which is what we want when modeling biological processes.
The paper's improvements: Jane: Now let’s talk about the specific improvements they suggest over previous models or baselines. What exactly does GrAFT-PDX improve upon?
Tom: It explicitly improves upon survival baselines because it incorporates the informative termination of tumor measurements at death, which we talked about earlier. That’s a concrete mechanism for improvement.
Lu: And for predicting tumor response, they say it can improve on static models by incorporating partial tumor trajectory information, even when you don't have all the data points available for a specific combination.
Tom: So they are taking the fact that measurements stop at death and using that dynamic history to make better predictions about future growth and survival than a simple average of what we see so far.
Jane: They also improve on static models by incorporating this partial trajectory information, which is something you just can’t get from just looking at a single tumor size measurement.
Lu: And they use the low-rank tensor decomposition to share information across all observed experiments and predict unobserved combinations, which is the mechanism for improving coverage of the screening space.
Tom: That means they are trying to solve that problem of having too many untested drug-treatment pairs by using latent structure to infer missing data points intelligently.
Jane: And on a more biological level, they improve things by capturing heterogeneous treatment responses across PDX lines through this learned response structure.
Lu: They learn this specific structure directly from the screening data instead of relying on external covariates like genomic matching to define those responses upfront.
Tom: That’s a significant point because it means the model is learning how treatments actually work in context, even if our initial guesses about gene interactions were imperfect.
Jane: So, to summarize their improvements: better survival prediction through dynamic history, better tumor response prediction by using partial trajectory data, and better coverage of untested combinations via the tensor decomposition.
Lu: And they show this structure supports both the prediction task and the biological interrogation task by allowing you to look inside the latent components for mechanistic insights.
Conclusion: Tom: So, to wrap up, GrAFT-PDX is a Bayesian model for PDX drug screens that links tumor growth and survival using a learned latent structure. It helps reduce the burden of screening by imputing missing experiments and stopping likely unsuccessful drugs early.
Jane: The key implication is that we now have a unified way to look at these complex, sparse data problems by treating tumor measurements as noisy observations of an underlying smooth continuous trajectory.
Lu: It provides a foundation for forecasting treatment response and prioritizing follow-up experiments by ranking untested combinations based on predicted additional lifespan.
Meng: From an engineering standpoint, the challenge remains how robust this model is when you try to deploy it at scale and what breaks when you run it against real-world screening data instead of synthetic data.
Tom: It’s a question for future work, but the paper shows that the model's latent structure is really good for both making predictions and actually asking biological questions about why those predictions happen.
Jane: So, we’ve seen how they use tensor decomposition to share information across PDX lines, treatments, and time to tackle sparsity in drug screens.
Lu: And it’s a framework that allows us to understand the latent structure of tumor response and survival more deeply than before.
Meng: I just want to make sure we know if this model is going to be able to handle the complexity of real clinical data, because synthetic recovery experiments are great, but real patient data has its own set of weirdness.
Tom: That’s where future work will focus, integrating external covariates like mutation status into the model to make those predictions even more robust.
Jane: So this GrAFT-PDX paper is a solid step forward in creating tools that can handle the complexity of PDX screens by linking longitudinal tumor response and survival together.
NIHAR MAUSKAR, SEBASTIAN SALAZAR, AARON SCHEIN, WESLEY TANSEY
Department of Computer Science, University of Chicago · Department of Computer Science, Columbia University · Data Science Institute, University of Chicago · Department of Statistics, University of Chicago · Computational Oncology, Memorial Sloan Kettering Cancer Center
stat.AP
Submitted: 2026-11-11
Updated: 2026-11-11
Comments: 27 pages, 12 figures
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: GrAFT-PDX introduces a Bayesian model for patient-derived xenograft (PDX) drug screens that jointly predicts tumor growth and survival outcomes using a learned latent function decomposition The model…
Key concepts
- Latent Function Decomposition
- This technique breaks down complex tumor measurements into simpler, underlying patterns. It helps the model share information across different PDX lines, treatments, and time points. By learning these hidden patterns, the model can make predictions about combinations it hasn't seen before.
- Tumor History Kernel
- This is a learned component that captures how much past tumor growth influences current death risk. The model finds that the most relevant history for predicting death is usually recent tumor burden, suggesting that what happened in the last couple of weeks matters most for survival predictions.
- Low-Rank Tensor Decomposition
- This mathematical method organizes data involving PDX lines, treatments, and time into a compact structure. It allows the model to efficiently share knowledge across all observed experiments, enabling it to predict outcomes for untested combinations by leveraging similar past data points.
Terminology
Summary
GrAFT-PDX introduces a Bayesian model for patient-derived xenograft (PDX) drug screens that jointly predicts tumor growth and survival outcomes using a learned latent function decomposition The model aims to reduce the PDX screening burden by enabling imputation of missing experiments and early stopping of likely unsuccessful drugs. p. 0 Here
we present GrAFT-PDX, a Bayesian model of PDX drug screens that aims to
reduce the PDX screening burden by enabling imputation of missing experiments and early stopping of likely unsuccessful drugs. Unlike conventional
survival or growth models, GrAFT-PDX jointly predicts tumor growth and
survival outcomes using a learned latent function decomposition.Where the paper says it
Model Overview
GrAFT-PDX represents sparse, irregular tumor measurements as noisy observations of an underlying smooth continuous-time trajectory and allows tumor growth to inform death risk It jointly models the tumor measurements and interval observed death time for each mouse, where the PDX line and treatment determine a continuous time median tumor trajectory and survival distribution. The model uses a low-rank tensor decomposition over PDX lines, treatments, and time to share information across observed experiments, predict untested PDX-treatment combinations, and capture heterogeneous treatment responses across PDX lines. p. 2 GrAFT-PDX uses a low-rank tensor decomposition
over PDX lines, treatments, and time to share information across observed experiments, predict untested PDX-treatment combinations, and capture heterogeneous treatment responses
across PDX lines. The model represents tumor burden as a continuous trajectory and links
its history to the survival process providing a novel description of how tumor burden impacts
survival in mouse models.Where the paper says it
Data Challenges Addressed
The modeling problem in the PDXe dataset involves three main challenges
measurements are sparse both across and within combinations
tumor growth and survival are coupled through the observation process
treatment response is heterogeneous across PDX lines
The sparsity arises because many combinations are never tested, only 4190 of a possible 13826 (30.3%) PDX-treatment pairs are tested Furthermore, even within tested combinations, tumor size is recorded on only 29. p. 1 3%) PDX-treatment pairs are tested. Second, even within tested
combinations, tumor size is recorded on only 29. Second, even within tested
combinations, tumor size is recorded on only 29. 8% of days, on average, over a mouse’s
lifespan.Where the paper says it
Where the paper says it
Model Components and Structure
The tumor submodel uses a lognormal likelihood where the median relative tumor size is determined by the PDX line and treatment, modeled as a smoothly evolving function using a fixed cubic B-spline basis. The median tumor trajectory Gl,j (t) is represented using a rank-K nonnegative canonical polyadic (CP) decomposition to share information across PDX lines, treatments, and the spline basis. p. 2 GrAFT-PDX uses a low-rank tensor decomposition
over PDX lines, treatments, and time to share information across observed experiments, predict untested PDX-treatment combinations, and capture heterogeneous treatment responses
across PDX lines. The model represents tumor burden as a continuous trajectory and links
its history to the survival process providing a novel description of how tumor burden impacts
survival in mouse models. We model the hazard as the sum of two nonnegative components:
(13) λℓ,j (t)
def = λ
base
ℓ,j (t) + λ
tumor
ℓ,j (t). The baseline component λ
base
ℓ,j (t) captures the underlying death risk for PDX line ℓ under
treatment j, including survival patterns not explicitly attributed to the tumor trajectory.Where the paper says it
Where the paper says it
Inference and Validation
GrAFT-PDX uses Hamiltonian Monte Carlo (HMC) to induce a posterior distribution over all unknown model parameters. The model is validated through synthetic recovery experiments, where it recovers a known kernel from synthetic data, showing that the tumor history component contributes to death risk. p. 9 Synthetic recovery of the tumor history kernel. The convolution kernel βconv(a)
summarizes how previous tumor burden contributes to current death risk.Where the paper says it
Key Findings and Interpretations
The learned tumor history kernel βconv(a) places most of its mass on recent tumor history, with the posterior mean attaining its maximum at a lag of approximately two weeks. This suggests that current death risk is more strongly associated with tumor burden accumulated during the preceding weeks than with either instantaneous tumor size alone or more remote tumor history. p. 7 The argument t−s is
the lookback time from the current time t to the tumor burden at time s. Equivalently, writing
a = t−s, larger values of βconv(a) indicate lags at which historical tumor burden contributes
more strongly to current death risk. These successful applications motivate using a nonnegative CP decomposition to predict unobserved responses and examine biological patterns in PDXe. Applied to Θ, each of
the K components combines a pattern of loadings across PDX lines, a pattern across treatments, and a pattern across the spline basis.Where the paper says it
Where the paper says it
Limitations and Future Directions
A limitation is that GrAFT-PDX currently learns response structure from tumor and survival outcomes alone, without incorporating external covariates such as mutation status or treatment targets Future work should integrate these sources of information to improve predictions for PDX lines and treatments with few observations. Additionally, a full experimental design strategy requires decision rules that account for posterior uncertainty, validation cost, and the relative consequences of false positive and false negative recommendations. The model's latent structure supports both prediction and biological interrogation.
Candidate Experiments
Table 1 presents the ten highest ranked untested PDX-treatment combinations identified as candidate future experiments. The model predicts that the top three candidates, X-3483/CM, X-3746/CM, and X4644/CM, will respond to a combination therapy of LEE011 and binimetinib. These top candidates are consistent with a biological hypothesis involving concurrent MAPK-pathway activation and CDK4/6-axis dysregulation. The model's predictions provide biologically coherent hypotheses for future experimentation.
Biological Association Preservation
The comparison between observed and generated association coefficients shows that the generated data generally reproduce the association patterns estimated from the observed data within their reported uncertainty All nine fitted calibration slopes ρbm,r are positive, ranging from 0.657 to 0.837, with a mean of 0. p. 8 We place an index biased shrinkage prior on each kernel coefficient,
(18) βc ∼ Exponential
10c−1
Z T
0
Bc(a) da
. Under the standard parameterization of the exponential distribution, Exponential(λ) has
mean 1/λ. The
normalization by R T
0 Bc(a) da makes the prior act on integrated kernel weight rather than
on raw coefficient magnitude. Since interior knots are placed at empirical quintiles of the
observed measurement times, basis functions differ in support width, magnitude, and total
area as shown in Figure 2.Where the paper says it
Where the paper says it
Conclusion
GrAFT-PDX provides a unified probabilistic framework for sparse PDX screens, linking longitudinal tumor response, survival, and treatment heterogeneity within a single model Using latent representations of tumor response and survival, the model provides a foundation for forecasting treatment response, prioritizing follow-up experiments, and developing future tools for PDX study design. The model's latent structure supports both prediction and biological interrogation.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Bayesian model for PDX drug screens that enables imputation of missing experiments and early stopping of likely unsuccessful drugs. This system can perform in silico analysis of untested drug-treatment combinations to reduce the burden of physical experiments.
-
Joint modeling framework that links longitudinal tumor growth measurements with mouse survival outcomes using a learned latent function decomposition. The improved AI can predict time of death by incorporating the informative termination of tumor measurements at death, which biases standard models downward.
-
Low-rank tensor decomposition over PDX lines, treatments, and time to share information across observed experiments and predict unobserved combinations. This allows for the prediction of untested PDX-treatment combinations by sharing information across sparse data points.
-
Model that captures heterogeneous treatment responses across PDX lines by learning a PDX line specific response structure directly from the screening data rather than relying on observed covariates. This improves prediction accuracy even when genomic matching is an imperfect guide to response.
-
A mechanism to rank candidate experiments using posterior predictive summaries, such as predicted additional lifespan for untested combinations. This allows for the prioritization of follow-up experiments based on predicted efficacy and safety margins.
-
Ability to detect and characterize tumor history effects in survival prediction through a learned convolution kernel that identifies the lag window over which previous tumor burden is most associated with current death risk. This provides insight into how important historical tumor burden is to current death risk.
-
Capability to preserve biologically meaningful treatment-gene response patterns in posterior predictive outcomes when genomic covariates are not included in fitting the model. This allows for external checking of model-generated outcomes against known biological associations without requiring computationally intractable genomic information during inference.
Abstract
Cancer drugs are often evaluated for efficacy in mice before proceeding to human clinical trials. To do this, human tumors are engrafted onto immune-incompetent mice to create a patient-derived xenograft (PDX). Engrafted tumors can be propagated to evaluate multiple drugs, but the time and costs involved in PDX studies limit high throughput drug screens in mice. Here we present GrAFT-PDX, a Bayesian model of PDX drug screens that aims to reduce the PDX screening burden by enabling imputation of missing experiments and early stopping of likely unsuccessful drugs. Unlike conventional survival or growth models, GrAFT-PDX jointly predicts tumor growth and survival outcomes using a learned latent function decomposition. In experiments on a large-scale PDX dataset, GrAFT-PDX outperforms Cox proportional hazards models at predicting time of death, outperforms machine learning baselines at predicting tumor growth, and recovers known gene-treatment relationships without genomic input.
Sources
- XGBoost: A Scalable Tree Boosting System
- PIONEER: Bayesian Joint Modelling of Mechanistic Tumour Growth and Time-to-Event Endpoints for Dynamic Prediction of Ongoing Oncology Trials
Related papers
- Unraveling time-varying causal effects of multiple exposures: integrating Functional Data Analysis with Multivariable Mendelian Randomization
- Strategic Causal Policy Learning with Manipulable Eligibility
- BayesJudge: Uncertainty-Aware Bayesian Meta-Evaluation of Human and LLM Judgments
- Automating structural reliability analysis with a multi-agent large language model framework
- Improving Minority Population Sampling with BISG Probabilities: Evidence from a Survey of Jewish Americans
- When Do Sports Leagues Stop Being Competitive? Bayesian Detection of Competitive-Balance Eras from Match-Level Data