betaFeedback

WorksMathematics

GrAFT-PDX: Growth and Failure-Time Modeling of Patient-Derived Xenograft Experiments

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.

Introduction

Patient-derived xenografts (PDXs) provide an important method for studying how human tumors may respond to drug treatments in a controlled setting [50]. In these experiments, tumor tissue obtained from a cancer patient is implanted into an immunocompromised mouse. This allows researchers to monitor tumor growth over time in vivo. Different drugs or drug combinations can then be administered to individual mice to determine which therapies would more likely be effective in patients with similar tumors. Effective treatments shrink the size of the tumor, or slow its growth, and extend the overall survival of the mouse. When conducted across many mice engrafted with multiple PDXs, these experiments yield datasets consisting of longitudinal tumor measurements coupled with mouse survival outcomes.

Because each experiment measures the response of a specific tumor to a particular drug, these datasets let statistical models learn how tumors respond to treatment and identify promising therapies directly from the data. A PDX line can be propagated indefinitely, so the same tumor can be screened against many drugs. PDX biobanks aggregate these screens across many tumors and therapies [32, 3] (Matschkos et al. 2021). Even so, exhaustive screening of every PDX-treatment combination remains infeasible due to cost, time, and labor constraints. Computational models that impute and rank the untested combinations can ease this burden by allowing researchers to analyze potential experiments in silico. Such methods support drug repositioning: the identification of new therapeutic uses for existing drugs.

Keywords and phrases. Bayesian statistics, survival analysis, tensor factorization, functional analysis.

We study the problem of computational modeling in the PDX Encyclopedia (PDXe) [18], a large-scale PDX drug screen containing 223 PDX lines, 62 treatments, and 4190 tested PDX-treatment combinations. Figure 1 illustrates both the incompleteness of the screen and the prediction task considered in this paper. Modeling the PDXe data poses three challenges: measurements are sparse both across and within combinations, tumor growth and survival are coupled through the observation process, and treatment response is heterogeneous across PDX lines. We consider each challenge in turn.

Figure 1. The left panel shows the sparsity of recorded experiments in the PDXe dataset relative to all possible PDX-treatment combinations. The right panel zooms in, showing how GrAFT-PDX is capable of predicting responses for unobserved combinations from sparse, irregular measurements of previously conducted experiments.

The sparsity of PDXe arises at two levels. First, many combinations are never tested; only 4190 of a possible 13826 (30.3%) PDX-treatment pairs are tested. Second, even within tested combinations, tumor size is recorded on only 29.8% of days, on average, over a mouse’s lifespan. Together, these leave the overall dataset 91.7% sparse. A model for these data must therefore impute at two levels, predicting both unobserved combinations and unmeasured points along each continuous tumor trajectory.

The coupling between growth and survival arises from the observation process itself. Tumor trajectories are only partially observed because measurements stop when a mouse dies. This missingness is informative. Mice with aggressive tumors tend to die earlier and therefore contribute no later measurements. Modeling tumor growth independently of survival would then bias estimated growth curves downward, since the fastest growing tumors are absent at later times. Capturing the dual nature of effective therapies—shrinking tumors and extending lifespan—requires joint modeling of longitudinal tumor growth and mouse survival.

Finally, treatment response is heterogeneous across PDX lines. The same therapy can shrink tumors in some lines yet appear far less effective in others, producing different trajectories and different lifespans. This variation is hard to predict from observed molecular covariates alone. This is true even in precision oncology, where genomic matching is an imperfect guide to response, with only around 63% of patients responding to targeted therapies selected from tumor genomics [23]. Capturing this heterogeneity therefore requires a model that learns PDX line specific response structure directly from the screening data rather than relying on observed covariates.

Contributions

Motivated by these challenges, we introduce GrAFT-PDX, a Bayesian model that jointly predicts tumor growth and survival in sparse PDX drug screens. GrAFT-

Related Work

A large body of work in computational drug repositioning seeks to identify new therapeutic uses for existing compounds by analyzing molecular, phenotypic, or pharmacological data [26]. Relational prediction methods have approached this problem by imputing unobserved drug-target or drug-disease associations, including probabilistic matrix completion [37] and methods incorporating side information or causal structure [56, 58]. Closely related work in cancer predicts treatment response from molecular characteristics of tumors or cell lines [33]. On the PDXe dataset used here, Kurilov, Haibe-Kains and Brors [28] compare prediction strategies that represent xenograft response through summaries such as tumor volume, tumor growth slope, and integrated response, while Gupta et al. [21] reduce response to a binary outcome predicted by a single classifier across cancer types and treatments. These approaches predict treatment response for unobserved PDX-treatment combinations using scalar summaries of tumor response. This reduction discards the longitudinal structure of the tumor measurements, motivating models that instead represent the entire tumor trajectory over the course of treatment.

GrAFT-PDX builds on the joint modeling framework by linking time-varying tumor sizes to survival. Joint modeling is a statistical approach where longitudinal measurements and event times depend on a shared latent process [25, 53]. Rizopoulos, Molenberghs and Lesaffre [39] model latent subject specific biomarker trajectories using linear mixed models and link them to a proportional hazards sub-model. Li and Luo (2017) extend this framework to functional data by incorporating scalar and functional predictors into both the longitudinal and proportional hazards submodels, with penalized splines representing the functional effects. Alvares and Rubio [4] combine a generalized linear mixed model with a parametric hazard survival model, linking the processes through shared fixed and random effects. In xenograft experiments, Pan et al. [36] model longitudinal log tumor volume using a constrained linear mixed model and link its mouse specific random effect to a Cox proportional hazards model for survival. More recently, Naguib et al. (2026) link patient specific tumor trajectories to clinical event models to forecast overall survival from immature oncology trials. These approaches capture dependence between repeated measurements and event times, but do not naturally extend to sparse designs such as PDX screens, in which entire PDX-treatment combinations are unobserved.

We also utilize matrix and tensor factorization methods for predicting treatment responses. Large cancer drug screens form sparse response arrays indexed by biological samples and treatments, motivating constrained models that share information across these axes. Matrix factorization methods predict scalar drug response summaries across cancer cell lines and treatments by sharing information across observed combinations [5, 45, 38]. Adding experimental dimensions such as dose or drug combinations extends these matrix factorization approaches to tensor models. Wang et al. [54] use latent tensor reconstruction to model responses across drugs, concentrations, and cell lines. Bertsimas and Pawlowski [7] utilize tensor completion across patients, drugs, and doses to predict potential therapies for cancer patients. Tansey, Tosh and Blei [47] use Bayesian tensor filtering to model dose-response curves across cell lines, drugs, and doses. Together, these methods show how low-rank structure can be used to predict missing responses in sparsely observed cancer drug screens, including responses that vary across experimental dimensions such as dose and drug combinations. However, they do not model how treatment response evolves over time nor do they handle the informative censoring issue of in vivo studies.

PDXe Data and Notation

Each observed mouse ii is associated with a PDX line ℓi∈{1,…,L}\ell_i \in\{1,\ldots,L\} and a treatment ji∈{0,1,…,J}j_i \in\{0,1,\ldots,J\}, where ji=0j_i=0 denotes the untreated control and values ji=1,…,Jj_i=1,\ldots,J denote active treatments or treatment combinations. Multiple mice may share the same PDX-treatment pair, though in PDXe only 3 out of the observed 4190 (0.072%) pairs contain replicate mice.

Tumor measurements are taken at irregularly spaced times. For mouse ii, we record both the raw tumor volume and its volume relative to the starting volume. We collect these quantities in the mouse level tumor record

Yi={ℓi,ji,(ri(m),yi(m),τi(m))m=1Mi},yi(m)=ri(m)ri(1),τi(1)<⋯<τi(Mi).(1)\mathcal{Y}_i = \left\{\ell_i,j_i,\left(r_i^{(m)},y_i^{(m)},\tau_i^{(m)}\right)_{m=1}^{M_i}\right\}, \qquad y_i^{(m)}=\frac{r_i^{(m)}}{r_i^{(1)}}, \qquad\tau_i^{(1)}<\cdots<\tau_i^{(M_i)}. \tag*{(1)}

Here, ri(m)r_i^{(m)} is the raw tumor volume measured at time τi(m)\tau_i^{(m)}, yi(m)y_i^{(m)} is the corresponding relative tumor size, and MiM_i is the number of tumor measurements for mouse ii. Times are measured in days from treatment initiation, so τi(1)=0\tau_i^{(1)}=0 for all ii. The remaining measurement times may differ across mice. By construction, yi(1)=1y_i^{(1)}=1, and subsequent measurements describe multiplicative change from the initial observed tumor volume.

This normalization accounts for variability in initial tumor size. In PDXe, treatment began once tumors reached approximately 200 mm³, but tumors were only monitored at most daily so volume is not identical across mice as some tumors may be larger when treatment begins. Modeling relative tumor size removes baseline tumor volume as a nuisance source of variation and places mice on a common response scale. This convention is also consistent with standard PDX efficacy analyses, where treatment response is often summarized through tumor growth inhibition [55], time-to-event endpoints based on tumor volume doubling [42], or response categories defined relative to baseline tumor burden [18].

Each mouse’s observation sequence also encodes survival information. Let ti∗t_i^* denote the true latent death time for mouse ii. In the PDXe data, every mouse is followed until death. The exact death time is not observed. Instead, τi(Mi−1)\tau_i^{(M_i-1)} is the last time mouse ii was observed alive, and τi(Mi)\tau_i^{(M_i)} is the first time mouse ii was observed to have died:

τi(Mi−1)<ti∗≤τi(Mi).(2)\tau_i^{(M_i-1)} < t_i^* \leq\tau_i^{(M_i)}. \tag*{(2)}

The relation above is the observed death interval for mouse ii. Here τi(Mi)\tau_i^{(M_i)} is also the final tumor measurement time, after which no further measurements are observed.

Model

GrAFT-PDX jointly models the tumor measurements and interval observed death time for each mouse. For mouse ii, the associated PDX line ℓi\ell_i and treatment jij_i determine a continuous time median tumor trajectory and survival distribution. The observed measurements are noisy evaluations of this trajectory, and its previous values influence the mouse’s death risk over time. We define the tumor and death submodels below.

Tumor submodel

The tumor submodel describes how relative tumor size evolves over time. Because relative tumor measurements are strictly positive and tumor growth is naturally multiplicative, we use a lognormal likelihood. Mouse ii’s PDX line ℓi\ell_i and treatment jij_i determine its tumor trajectory. We model the observed relative tumor measurements as

yi(m)∼Lognormal⁡(MEDIAN⁡=Gℓi,ji(τi(m)), SCALE⁡=D(τi(m))),m=2,…,Mi.(3)y_i^{(m)} \sim\operatorname{Lognormal}\left(\operatorname{MEDIAN}=G_{\ell_i,j_i}(\tau_i^{(m)}),\ \operatorname{SCALE}=D(\tau_i^{(m)})\right), \qquad m=2,\ldots,M_i. \tag*{(3)}

Here, Gℓi,ji(t)G_{\ell_i,j_i}(t) is the median relative tumor size at time tt for mice assigned to PDX line ℓi\ell_i and treatment jij_i, while D(t)D(t) controls the amount of measurement variation around this median. The likelihood begins at m=2m=2 because yi(1)=1y_i^{(1)}=1 deterministically. The notation X∼Lognormal⁡(MEDIAN⁡=μ,SCALE⁡=d)X\sim\operatorname{Lognormal}(\operatorname{MEDIAN}=\mu,\operatorname{SCALE}=d) denotes that log⁡X∼N(log⁡μ,d2)\log X\sim\mathcal{N}(\log\mu,d^2). We parameterize the median and scale functions as smoothly evolving over time by modeling them with a fixed cubic B-spline basis (De Boor, [16]),

Gℓ,j(t)≔∑c=1Cθℓ,j,cBc(t),D(t)≔∑c=1CηcBc(t).(4)G_{\ell,j}(t) \coloneqq\sum_{c=1}^{C}\theta_{\ell,j,c}B_c(t), \qquad D(t) \coloneqq\sum_{c=1}^{C}\eta_cB_c(t). \tag*{(4)}

We constrain each ηc>0\eta_c>0 to ensure that D(t)>0D(t)>0. The coefficients η1,…,ηC\eta_1,\ldots,\eta_C determine how measurement variation changes over time and these coefficients have the priors,

ηc∼Exponential⁡(1).(5)\eta_c\sim\operatorname{Exponential}(1). \tag*{(5)}

We share D(t)D(t) across all PDX-treatment pairs because pair specific scale functions would be difficult to learn from the sparse observations.

We fix the basis functions before fitting the model; the model learns only their coefficients. A B-spline basis is determined by its degree and knot sequence. The degree controls the smoothness of the basis functions, while the knots determine where the basis functions are locally active. For a cubic B-spline basis with CC basis functions, we write the knot sequence as κ1≤⋯≤κC+4\kappa_1\le\cdots\le\kappa_{C+4}. The basis function Bc(t)B_c(t) is supported on [κc,κc+4][\kappa_c,\kappa_{c+4}], so it is zero outside this interval and can affect a fitted function only over a local region of time. At any time tt, at most four cubic B-spline basis functions are nonzero. Neighboring basis functions overlap, allowing fitted functions to vary smoothly across time while retaining local control over different parts of the curve. Figure 2 illustrates how these basis functions come together to create D(t)D(t).

Figure 2. Sums of basis splines result in the learned time-varying scale. Interior knots denoted by dashed lines are placed at empirical quintiles of the observed measurement times. Each basis function is active only on a local interval, with neighboring basis functions overlapping.

Having specified the shared scale function D(t)D(t), we now turn to the median tumor trajectories Gℓ,j(t)G_{\ell,j}(t). We collect their spline coefficients in the tensor Θ∈R+L×(J+1)×C\boldsymbol{\Theta}\in\mathbb{R}_{+}^{L\times(J+1)\times C}, with entry θℓ,j,c\theta_{\ell,j,c} giving the coefficient on basis function Bc(t)B_c(t) for PDX line ℓ\ell under treatment jj. We write θℓ,j=(θℓ,j,1,…,θℓ,j,C)∈R+C\boldsymbol{\theta}_{\ell,j}=(\theta_{\ell,j,1},\ldots,\theta_{\ell,j,C})\in\mathbb{R}_{+}^{C} for the corresponding fiber of Θ\boldsymbol{\Theta}, which collects the coefficients defining Gℓ,j(t)G_{\ell,j}(t).

Rather than modeling Θ\boldsymbol{\Theta} as an unrestricted L×(J+1)×CL\times(J+1)\times C tensor, GrAFT-PDX represents it using a rank-KK nonnegative canonical polyadic (CP) decomposition ([11, 22, 27]). A CP decomposition expresses a tensor as a sum of rank-one outer products, with each component coupling one loading pattern from each mode. Nonnegative loadings prevent components from canceling one another, so we can examine how each component contributes additively. Historically, nonnegative CP decompositions have been used both to predict unobserved entries in sparse relational tensors ([43]) and to recover interpretable biological patterns (Armingol et al., 2022). These successful applications motivate using a nonnegative CP decomposition to predict unobserved responses and examine biological patterns in PDXe. Applied to Θ\boldsymbol{\Theta}, each of the KK components combines a pattern of loadings across PDX lines, a pattern across treatments, and a pattern across the spline basis. We write

θℓ,j,c≔∑k=1Kuℓ,kvj,kwc,k,Θ≔∑k=1Kuk∘vk∘wk.(6)\theta_{\ell,j,c} \coloneqq\sum_{k=1}^{K} u_{\ell,k}v_{j,k}w_{c,k}, \qquad\boldsymbol{\Theta} \coloneqq\sum_{k=1}^{K} \mathbf{u}_k \circ\mathbf{v}_k \circ\mathbf{w}_k. \tag*{(6)}

We let ∘\circ denote the outer product. Each term uk∘vk∘wk\mathbf{u}_k \circ\mathbf{v}_k \circ\mathbf{w}_k is a rank-one tensor over PDX lines, treatments, and basis functions. The vector uk∈RL\mathbf{u}_k \in\mathbb{R}^{L} contains the PDX line loadings for latent component kk, vk∈RJ+1\mathbf{v}_k \in\mathbb{R}^{J+1} contains the treatment loadings, and wk∈RC\mathbf{w}_k \in\mathbb{R}^{C} contains the basis function loadings. The integer KK is the number of rank-one components in the CP decomposition. When KK is small relative to LL, J+1J+1, and CC, this representation replaces the unrestricted collection of L(J+1)CL(J+1)C tumor trajectory coefficients with K(L+(J+1)+C)K(L+(J+1)+C) loading parameters. In practice, we use K=5K=5; see Section A.2 for details.

The CP decomposition also has a functional interpretation. Combining Eqs. (4) and (6) gives us a view on how GrAFT-PDX is representing latent functions:

Gℓ,j(t)=∑c=1C(∑k=1Kuℓ,kvj,kwc,k)Bc(t)=∑k=1Kuℓ,kvj,k(∑c=1Cwc,kBc(t)).(7)G_{\ell,j}(t) = \sum_{c=1}^{C} \left(\sum_{k=1}^{K}u_{\ell,k}v_{j,k}w_{c,k}\right)B_c(t) = \sum_{k=1}^{K}u_{\ell,k}v_{j,k}\left(\sum_{c=1}^{C}w_{c,k}B_c(t)\right). \tag*{(7)}

This shows that each latent factor kk defines a shared tumor response function Φk(t)\Phi_k(t). The product uℓ,kvj,ku_{\ell,k}v_{j,k} weights the shared function Φk(t)\Phi_k(t) for pair (ℓ,j)(\ell,j). Thus, Gℓ,j(t)G_{\ell,j}(t) is represented as a sum of shared latent functions whose weights depend on PDX line ℓ\ell and treatment jj.

We enforce positivity through nonnegative factor loadings. Specifically, we place independent exponential priors on the factor loadings,

uℓ,k,vj,k,wc,k∼Exponential⁡(1).(8)u_{\ell,k}, v_{j,k}, w_{c,k} \sim\operatorname{Exponential}(1). \tag*{(8)}

The exponential prior restricts the factor loadings in Eq. (6) to be positive because the distribution only has support over R+\mathbb{R}_{+}. Since cubic B-spline basis functions are nonnegative, the shared functions Φk(t)\Phi_k(t) in Eq. (7) are also nonnegative. The products uℓ,kvj,ku_{\ell,k}v_{j,k} are positive as well, so the resulting median tumor trajectory Gℓ,j(t)G_{\ell,j}(t) is positive over the experimental time window. This is required because Gℓ,j(t)G_{\ell,j}(t) represents median relative tumor size, and relative tumor size is strictly positive. Importantly, positivity does not impose monotone tumor growth. The spline coefficients determine the height of local basis function contributions, not the slope of the trajectory. As the active basis functions change over time, Gℓ,j(t)G_{\ell,j}(t) can increase, decrease, or remain approximately stable while remaining positive.

Death submodel

The death submodel assigns probability to the interval of death for each mouse. Let ti∗t_i^* denote the true latent death time for mouse ii. As described in Eq. (2), the observed survival information is τi(Mi−1)<ti∗≤τi(Mi)\tau_i^{(M_i-1)} < t_i^* \le\tau_i^{(M_i)}, where τi(Mi−1)\tau_i^{(M_i-1)} is the last time mouse ii was observed alive and τi(Mi)\tau_i^{(M_i)} is the first time mouse ii was observed to have died. Since mouse ii is associated with PDX line ℓi\ell_i and treatment jij_i, its death distribution is determined by the survival function Sℓi,ji(t)S_{\ell_i,j_i}(t), defined

Sℓ,j(t)=defPr⁡(ti∗>t∣ℓi,ji).(9)S_{\ell,j}(t) \mathrel{\overset{\mathrm{def}}{=}} \Pr(t_i^* > t \mid\ell_i,j_i). \tag*{(9)}

The survival function Sℓ,j(t)S_{\ell,j}(t) is a probability indexed by time. Therefore, for any given tt it must be the case that 0≤Sℓ,j(t)≤10 \le S_{\ell,j}(t) \le1. Sℓ,j(t)S_{\ell,j}(t) must also be monotonic decreasing in tt, as the event of remaining alive beyond a later time is contained in the event of remaining alive beyond any earlier time. Since all mice are alive at treatment initiation, we also have Sℓ,j(0)=1S_{\ell,j}(0)=1. These constraints ensure that survival probabilities assigned to time intervals are valid. In particular, the probability that mouse ii dies in its observed interval becomes

Pr⁡(τi(Mi−1)<ti∗≤τi(Mi)∣ℓi,ji)=Sℓi,ji(τi(Mi−1))−Sℓi,ji(τi(Mi)).(10)\Pr\left(\tau_i^{(M_i-1)} < t_i^* \le\tau_i^{(M_i)} \mid\ell_i,j_i\right) = S_{\ell_i,j_i}\left(\tau_i^{(M_i-1)}\right) - S_{\ell_i,j_i}\left(\tau_i^{(M_i)}\right). \tag*{(10)}

Directly modeling Sℓ,j(t)S_{\ell,j}(t) with an unconstrained spline would not automatically enforce these properties: the fitted curve could exceed one, become negative, or increase locally, producing invalid interval probabilities in Eq. (10). We therefore define the survival function indirectly through a nonnegative hazard function.

The hazard is a standard modeling object in survival analysis. Rather than modeling the probability of being alive beyond time tt directly, the hazard describes the instantaneous death rate among mice that have survived up to time tt. For a fixed PDX line ℓ\ell and treatment jj, let λℓ,j(t)\lambda_{\ell,j}(t) denote this hazard. For differentiable survival functions, this hazard can also be written as the instantaneous rate of decrease in log survival:

λℓ,j(t)=lim⁡Δ→0Pr⁡(t<z∗≤t+Δ∣z∗>t,ℓ,j)Δ=−ddtlog⁡Sℓ,j(t).(11)\lambda_{\ell,j}(t) = \lim_{\Delta\to0} \frac{\Pr(t < z^* \le t+\Delta\mid z^* > t,\ell,j)}{\Delta} = -\frac{d}{dt}\log S_{\ell,j}(t). \tag*{(11)}

Here z∗z^* denotes a generic latent death time for a mouse from PDX line ℓ\ell under treatment jj. Any nonnegative hazard defines an associated survival function

Sℓ,j(t)=exp⁡(−∫0tλℓ,j(u) du),(12)S_{\ell,j}(t) = \exp\left(-\int_0^t \lambda_{\ell,j}(u)\,du\right), \tag*{(12)}

which is automatically positive, bounded above by one, and nonincreasing. Thus, instead of directly modeling a highly constrained Sℓ,j(t)S_{\ell,j}(t), we model a nonnegative hazard and obtain a valid survival function through Eq. (12).

We model the hazard as the sum of two nonnegative components:

λℓ,j(t)=defλℓ,jbase(t)+λℓ,jtumor(t).(13)\lambda_{\ell,j}(t) \mathrel{\overset{\mathrm{def}}{=}} \lambda_{\ell,j}^{\mathrm{base}}(t) + \lambda_{\ell,j}^{\mathrm{tumor}}(t). \tag*{(13)}

The baseline component λℓ,jbase(t)\lambda_{\ell,j}^{\mathrm{base}}(t) captures the underlying death risk for PDX line ℓ\ell under treatment jj, including survival patterns not explicitly attributed to the tumor trajectory. The tumor history component λℓ,jtumor(t)\lambda_{\ell,j}^{\mathrm{tumor}}(t) captures additional risk associated with the latent tumor trajectory up to time tt. Both components are constrained to be nonnegative, so each can only add to the instantaneous death rate in Eq. (13).

We choose this additive specification, following the linear hazard models of Aalen [1, 2]. Whereas standard additive hazard models allow the current value of a time-varying covariate to contribute to death risk, our tumor history term depends on the latent tumor trajectory over all previous times. Under the spline representations introduced below, an additive hazard lets us precompute all required basis integrals and convolutions before inference. A proportional hazards formulation would introduce a nonlinear, nested time integral that cannot be computed before inference and would require numerical integration for each likelihood and gradient evaluation, making posterior inference computationally impractical.

The baseline hazard is represented using the same cubic B-spline basis introduced in § 4.1. For PDX line ℓ\ell and treatment jj, we define

λℓ,jbase(t)≔∑c=1Cexp⁡(∑k=1Kuℓ,k(d)vj,k(d)wc,k(d))Bc(t).(14)\lambda_{\ell,j}^{\mathrm{base}}(t) \coloneqq\sum_{c=1}^{C} \exp\left(\sum_{k=1}^{K} u_{\ell,k}^{(d)}v_{j,k}^{(d)}w_{c,k}^{(d)}\right) B_c(t). \tag*{(14)}

The exponential transformation above ensures that each basis coefficient θℓ,j,c(d)\theta_{\ell,j,c}^{(d)} is positive. Since the B-spline basis functions are nonnegative, the resulting baseline hazard λℓ,jbase(t)\lambda_{\ell,j}^{\mathrm{base}}(t) is nonnegative over the modeled time window. This lets the death factors uℓ,k(d),vj,k(d),wc,k(d)u_{\ell,k}^{(d)}, v_{j,k}^{(d)}, w_{c,k}^{(d)} vary freely on the real line, so we place standard normal priors on them,

uℓ,k(d),vj,k(d),wc,k(d)∼N(0,1).(15)u_{\ell,k}^{(d)}, v_{j,k}^{(d)}, w_{c,k}^{(d)} \sim\mathcal{N}(0,1). \tag*{(15)}

The tensor of log-coefficients (log⁡θℓ,j,c(d))(\log\theta_{\ell,j,c}^{(d)}) is represented by a rank-KK CP decomposition separate from the nonnegative CP decomposition of the tumor trajectory coefficient tensor Θ\boldsymbol{\Theta} in Eq. (6). Here uℓ,k(d)u_{\ell,k}^{(d)} is the loading of PDX line ℓ\ell on baseline hazard component kk, vj,k(d)v_{j,k}^{(d)} is the corresponding loading of treatment jj, and wc,k(d)w_{c,k}^{(d)} determines how that component contributes through basis function Bc(t)B_c(t). Thus, we use a nonnegative CP decomposition for the tumor coefficient tensor and an unconstrained CP decomposition for the baseline hazard log-coefficient tensor, then exponentiate the latter elementwise. Section A.3 compares this construction with alternatives that use the same positivity mechanism in both submodels.

The tumor history component links the death process to the latent tumor trajectory. While λℓ,jbase(t)\lambda_{\ell,j}^{\mathrm{base}}(t) captures baseline death risk, λℓ,jtumor(t)\lambda_{\ell,j}^{\mathrm{tumor}}(t) captures additional risk associated with the latent tumor trajectory up to time tt. Inspired by the weighted cumulative exposure model of Sylvestre and Abrahamowicz [46], we represent this dependence using a convolution term in the hazard. For PDX line ℓ\ell under treatment jj, we define

λℓ,jtumor(t)≔∫0tβconv(t−s)Gℓ,j(s) ds,βconv(a)≔∑c=1CβcBc(a).(16)\lambda_{\ell,j}^{\mathrm{tumor}}(t) \coloneqq\int_{0}^{t} \beta_{\mathrm{conv}}(t-s)G_{\ell,j}(s)\,\mathrm{d}s,\qquad\beta_{\mathrm{conv}}(a) \coloneqq\sum_{c=1}^{C}\beta_c B_c(a). \tag*{(16)}

Here, Gℓ,j(s)G_{\ell,j}(s) is the latent median relative tumor size at previous time ss, and βconv\beta_{\mathrm{conv}} is a shared lag response kernel represented in the same cubic B-spline basis as Gℓ,jG_{\ell,j}. The argument t−st-s is the lookback time from the current time tt to the tumor burden at time ss. Equivalently, writing a=t−sa=t-s, larger values of βconv(a)\beta_{\mathrm{conv}}(a) indicate lags at which historical tumor burden contributes more strongly to current death risk. Thus, λℓ,jtumor(t)\lambda_{\ell,j}^{\mathrm{tumor}}(t) is a weighted summary of previous tumor burden, where the shared kernel βconv\beta_{\mathrm{conv}} determines which parts of the past trajectory are most relevant for current death risk. Importantly, we model the full tumor burden history because longitudinal measurements can provide survival information that is lost when tumor response is reduced to a scalar summary ([59]).

Expanding out βconv(a)\beta_{\mathrm{conv}}(a) and Gℓ,j(t)G_{\ell,j}(t) in Eq. (16) gives

λℓ,jtumor(t)=∑c=1C∑d=1Cβcθℓ,j,d∫0tBc(t−s)Bd(s) ds.(17)\lambda_{\ell,j}^{\mathrm{tumor}}(t)=\sum_{c=1}^{C}\sum_{d=1}^{C}\beta_c\theta_{\ell,j,d}\int_{0}^{t}B_c(t-s)B_d(s)\,\mathrm{d}s. \tag*{(17)}

Although λℓ,jtumor(t)\lambda_{\ell,j}^{\mathrm{tumor}}(t) is written as a double sum over all basis function pairs, most terms vanish at a given time tt. The factor Bd(s)B_d(s) is nonzero only on its local support s∈[κd,κd+4]s\in[\kappa_d,\kappa_{d+4}], while Bc(t−s)B_c(t-s) is nonzero only when (t−s)∈[κc,κc+4](t-s)\in[\kappa_c,\kappa_{c+4}]. Thus, the integrand is nonzero only on the portions of [0,t][0,t] where these two local supports overlap. On each overlap, both basis functions have a polynomial form, so when nonzero Bc(t−s)Bd(s)B_c(t-s)B_d(s) is a polynomial in ss. Consequently, each integral in Eq. (17) reduces to a finite sum of piecewise polynomial integrals on regions of overlap.

We place an index biased shrinkage prior on each kernel coefficient,

βc∼Exponential⁡(10c−1∫0TBc(a) da).(18)\beta_c\sim\operatorname{Exponential}\left(10^{c-1}\int_{0}^{T}B_c(a)\,\mathrm{d}a\right). \tag*{(18)}

Under the standard parameterization of the exponential distribution, Exponential⁡(λ)\operatorname{Exponential}(\lambda) has mean 1/λ1/\lambda. Therefore Eq. (18) implies

Eβc[∫0TβcBc(a) da]=101−c.(19)\mathbb{E}_{\beta_c}\left[\int_{0}^{T}\beta_cB_c(a)\,\mathrm{d}a\right]=10^{1-c}. \tag*{(19)}

Thus, the factor 101−c10^{1-c} sets the prior expected total weight of the cc-th kernel component. The normalization by ∫0TBc(a) da\int_{0}^{T}B_c(a)\,\mathrm{d}a 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. Without normalization, basis functions with larger area would receive larger prior expected contribution to βconv\beta_{\mathrm{conv}} solely because of the geometry of the chosen basis.

To understand the induced prior on the full convolution kernel, consider the kernel weight accumulated up to time tt.

Eβ1,…,βC[∫0tβconv(t−s) ds]=∑c=1C101−c∫0tBc(a) da∫0TBc(a) da⏟def=Fc(t)∈[0,1].(20)\mathbb{E}_{\beta_1,\ldots,\beta_C}\left[\int_{0}^{t}\beta_{\mathrm{conv}}(t-s)\,\mathrm{d}s\right]=\sum_{c=1}^{C}10^{1-c}\underbrace{\frac{\int_{0}^{t}B_c(a)\,\mathrm{d}a}{\int_{0}^{T}B_c(a)\,\mathrm{d}a}}_{\mathrel{\mathop{\mathrm{def}}}=F_c(t)\in[0,1]}. \tag*{(20)}

Here Fc(t)F_c(t) is the fraction of the cc-th basis function’s area accumulated by time tt. Because the B-spline basis functions are local and ordered from recent to more distant lags, only a few nearby basis functions contribute to βconv(a)\beta_{\mathrm{conv}}(a) at each lag aa. At later lags, basis functions with larger indices cc contribute most, and the corresponding kernel components have geometrically smaller expected total weights. The prior consequently imposes stronger shrinkage on distant tumor history effects, which are weakly informed because fewer mice survive long enough to contribute information about long lookback windows. The coefficients βc\beta_c are constrained to be nonnegative, so previous tumor burden can only increase the hazard; because these coefficients are shared across all PDX lines and treatments, βconv(a)\beta_{\mathrm{conv}}(a) provides a common summary of how tumor history contributes to survival.

Inference and Synthetic Testing

The GRAFT-PDX model induces a posterior distribution over the latent tumor trajectories, survival functions, measurement error function, and tumor history kernel. We generate posterior samples from this model using Hamiltonian Monte Carlo (HMC), and use those posterior draws to compute summary statistics.

Posterior target and inference

Let Ψ\Psi collect all unknown model parameters: the tumor factor loadings, death factor loadings, measurement scale coefficients, and tumor history kernel coefficients. That is,

Ψ≔{uℓ,k,vj,k,wc,k,uℓ,k(d),vj,k(d),wc,k(d),ηc,βc}.(21)\Psi\coloneqq\left\{u_{\ell,k}, v_{j,k}, w_{c,k}, u_{\ell,k}^{(d)}, v_{j,k}^{(d)}, w_{c,k}^{(d)}, \eta_c, \beta_c\right\}. \tag*{(21)}

Given the mouse observation records defined in Eq. (1) where Yi\mathcal{Y}_i contains the relative tumor sizes and observation times for mouse ii, posterior inference then targets

p(Ψ∣Y1:N)∝p(Y1:N∣Ψ)p(Ψ).(22)p\left(\Psi\mid\mathcal{Y}_{1:N}\right) \propto p\left(\mathcal{Y}_{1:N} \mid\Psi\right)p(\Psi). \tag*{(22)}

The likelihood term combines the lognormal tumor likelihood with the interval censored survival likelihood defined in § 4. The prior p(Ψ)p(\Psi) is the product prior over all parameters.

Although the posterior is parameterized in terms of spline coefficients and low-rank tensor factors, these parameters are not the main inferential targets. Each posterior draw of Ψ\Psi induces tumor quantities such as the median latent tumor trajectory Gℓ,j(t)G_{\ell,j}(t) and scale function D(t)D(t), survival quantities such as the hazard components λℓ,jbase(t)\lambda_{\ell,j}^{\mathrm{base}}(t) and λℓ,jtumor(t)\lambda_{\ell,j}^{\mathrm{tumor}}(t) and the survival function Sℓ,j(t)S_{\ell,j}(t), and the shared convolution kernel βconv(a)\beta_{\mathrm{conv}}(a). For each posterior draw, we evaluate these functions, then summarize their values across draws using means or quantiles. For example, if Ψ(1),…,Ψ(R)\Psi^{(1)}, \ldots, \Psi^{(R)} are posterior draws, the posterior mean of the survival function for PDX line ℓ\ell under treatment jj is estimated by

S^ℓ,j(t)≔1R∑r=1RSℓ,j(t;Ψ(r))≈EΨ∼p(Ψ∣Y)[Sℓ,j(t;Ψ)∣Y].(23)\widehat{S}_{\ell,j}(t) \coloneqq\frac{1}{R}\sum_{r=1}^{R}S_{\ell,j}\left(t;\Psi^{(r)}\right) \approx\mathbb{E}_{\Psi\sim p(\Psi\mid\mathcal{Y})}\left[S_{\ell,j}(t;\Psi)\mid\mathcal{Y}\right]. \tag*{(23)}

We use analogous Monte Carlo summaries for tumor trajectories, hazard functions, posterior predictive distributions, and the convolution kernel.

We implement GrAFT-PDX in Stan ([49]) and fit the model using the No-U-Turn Sampler (NUTS), an adaptive HMC algorithm ([10, 24]). For each fitted model, we run four chains with 1000 warmup iterations and 500 post-warmup sampling iterations per chain, yielding 2000 retained posterior draws. We use no thinning and set max_treedepth to 7.

Synthetic recovery of the tumor history kernel

The convolution kernel βconv(a)\beta_{\mathrm{conv}}(a) summarizes how previous tumor burden contributes to current death risk. Before applying GrAFT-PDX to PDXe, we test whether it can recover a known kernel from synthetic data. This ensures that the model’s structure does not introduce spurious relationships between tumor history and death and that the model can identify a relationship when one is present.

We generate synthetic datasets using the same time scale, spline basis, and knot locations as in our PDXe analysis in § 6. For each synthetic dataset, we first draw one set of tumor trajectory factors and baseline hazard factors from the model prior. This prior draw defines the latent tumor trajectories Gℓ,jtrue(t)G_{\ell,j}^{\mathrm{true}}(t) and synthetic baseline hazards Rℓ,jtrue(t)R_{\ell,j}^{\mathrm{true}}(t) for all observations. The synthetic death hazard is

λℓ,jtrue(t)=Rℓ,jtrue(t)+∫0tβconvtrue(t−s)Gℓ,jtrue(s) ds.(24)\lambda_{\ell,j}^{\mathrm{true}}(t) = R_{\ell,j}^{\mathrm{true}}(t) + \int_{0}^{t}\beta_{\mathrm{conv}}^{\mathrm{true}}(t-s)G_{\ell,j}^{\mathrm{true}}(s)\,\mathrm{d}s. \tag*{(24)}

For each mouse, we draw a latent death time from the survival distribution induced by Eq. (24). Synthetic tumor measurements are generated from the lognormal observation model every three days. We use a three day spacing because it approximately matches the observation frequency in the PDXe data. Let ti∗t_i^* be the exact synthetic death time and let T={0,3,6,…}\mathcal{T}=\{0,3,6,\ldots\} denote the observation grid. We record death through the interval (ai,bi](a_i,b_i], where ai=max⁡{t∈T:t<ti∗}a_i=\max\{t\in\mathcal{T}:t<t_i^*\} and bi=min⁡{t∈T:t≥ti∗}b_i=\min\{t\in\mathcal{T}:t\geq t_i^*\}.

We consider two data generating settings for βconvtrue(a)\beta_{\mathrm{conv}}^{\mathrm{true}}(a), a null kernel and a Gaussian kernel,

βconvtrue(a)∈{βconvNull(a)=0,βconvGauss(a)=5×10−4exp⁡(−(a−20)2102)}.(25)\beta_{\mathrm{conv}}^{\mathrm{true}}(a) \in\left\{ \beta_{\mathrm{conv}}^{\mathrm{Null}}(a)=0,\quad\beta_{\mathrm{conv}}^{\mathrm{Gauss}}(a)=5\times10^{-4}\exp\left(-\frac{(a-20)^2}{10^2}\right) \right\}. \tag*{(25)}

The two settings for βconvtrue(a)\beta_{\mathrm{conv}}^{\mathrm{true}}(a) target complementary recovery questions. The null kernel βconvNull(a)\beta_{\mathrm{conv}}^{\mathrm{Null}}(a) tests whether the fitted model avoids inferring a tumor history effect when no such effect is present. The Gaussian kernel βconvGauss(a)\beta_{\mathrm{conv}}^{\mathrm{Gauss}}(a) has peak height 5×10−45\times10^{-4} at lag a=20a=20 days. Writing its exponent in the usual Gaussian form gives standard deviation 10/2≈7.110/\sqrt{2}\approx7.1 days, so approximately 95% of its mass is concentrated between lags 6 and 34 days. Thus, the Gaussian setting creates a localized effect centered about three weeks before the current hazard time and tests whether GrAFT-PDX can recover both the existence of a tumor history effect and the region where that effect is concentrated. Figure 3 summarizes these recovery tests. Visually, we find that GrAFT-PDX consistently recovers the true effect of historical tumor burden on survival.

Figure 3. Across 10 independently generated synthetic datasets per kernel, GrAFT-PDX recovers the localized tumor history effect in the Gaussian setting while returning estimates near zero under the null.

Quantitatively, we assess how accurately GrAFT-PDX recovers the cumulative tumor effect in the Gaussian case. Let

M=∫0TβconvGauss(a) daandM^=∫0Tβ^conv(a) da(26)M=\int_{0}^{T}\beta_{\mathrm{conv}}^{\mathrm{Gauss}}(a)\,da \qquad\text{and} \qquad\widehat{M}=\int_{0}^{T}\widehat{\beta}_{\mathrm{conv}}(a)\,da \tag*{(26)}

denote the areas under the true and estimated kernels, respectively. We measure recovery of the cumulative effect using the ratio M^/M\widehat{M}/M, for which a value of one indicates that the estimated and true kernels have equal total area. Across the ten datasets, M^/M=0.97\widehat{M}/M=0.97 with a standard deviation of 0.14, indicating that the model recovered the true cumulative effect.

Results

We evaluate GrAFT-PDX on the PDXe dataset of Gao et al. [18]. First, we assess held-out prediction of survival and tumor response under partially observed tumor trajectories, comparing GrAFT-PDX to survival baselines, tumor response baselines, and simpler versions of the GrAFT-PDX model. We then interrogate the learned convolution kernel, βconv\beta_{\mathrm{conv}}, to assess how previous tumor burden contributes to current death risk. To evaluate the scientific usefulness of GrAFT-PDX, we examine whether posterior predictive outcomes preserve drug-gene response patterns and whether GrAFT-PDX suggests biologically coherent candidates for follow-up experiments.

Held-out prediction on PDXe

We evaluate prediction using 20 train-test splits constructed at the PDX-treatment combination level. Within each split, we exclude all mice from each test combination from training. We evaluate predictions for each test mouse at censoring times q∈{0,5,10,15}q \in\{0, 5, 10, 15\}. The censoring times chosen are intentionally small as one use for GrAFT-PDX is to identify unsuccessful PDX experiments early. At q=0q = 0, the fitted model has observed no outcomes from the test combination. At q>0q > 0, prediction conditions on the mouse’s tumor measurements at times τi(m)≤q\tau_i^{(m)} \le q and on survival through qq. Its subsequent tumor measurements and death interval remain held out for evaluation. Figure 4 illustrates the information update induced by censoring. At q=0q = 0, prediction is driven solely by the learned parameters, whereas at q=15q = 15, the prediction additionally reflects the mouse’s early observed tumor response.

Figure 4. Conditioning on the first 15 days of tumor measurements shifts the posterior tumor trajectory toward the observed early response and removes posterior predictive death mass before day 15. Ribbons denote the 50% credible interval. For tumor size, the predictions are posterior median whereas for death density they are the posterior predictive distribution.

For mouse ii, implanted with PDX line ℓi\ell_i and assigned treatment jij_i, let Si(t):=Sℓi,ji(t)S_i(t) \mathrel{\mathop:}= S_{\ell_i,j_i}(t). Survival through the censoring time contributes Si(q)S_i(q) to the conditioning likelihood by Eq. (9). Increasing qq therefore provides additional early tumor measurements while shortening the unobserved portion of the mouse’s trajectory.

For survival prediction, conditioning on survival through qq truncates and renormalizes the posterior predictive death distribution. Let ai=τi(Mi−1)a_i = \tau_i^{(M_i-1)} and bi=τi(Mi)b_i = \tau_i^{(M_i)}, so that the observed death interval is (ai,bi](a_i, b_i]. Conditional on a posterior draw of SiS_i and on survival through qq, the predictive survival function is

Pr⁡(ti∗>t∣ti∗>q,Si)=Si(t)Si(q),t≥q.(27)\Pr(t_i^* > t \mid t_i^* > q, S_i) = \frac{S_i(t)}{S_i(q)}, \qquad t \ge q. \tag*{(27)}

If the observed death interval for mouse ii satisfies bi>qb_i > q, then the conditional probability assigned to the future death interval is

Pr⁡(ai<ti∗≤bi∣ti∗>q,Si)=Si(max⁡{ai,q})−Si(bi)Si(q).(28)\Pr(a_i < t_i^* \le b_i \mid t_i^* > q, S_i) = \frac{S_i(\max\{a_i,q\}) - S_i(b_i)}{S_i(q)}. \tag*{(28)}

Mice with bi≤qb_i \le q are not included in the prediction task because their recorded death interval ends before or at the censoring time. In evaluation, we average the conditional interval probabilities across posterior draws to obtain posterior predictive probabilities.

We evaluate predictive performance using four complementary metrics. The death log probability evaluates the probability assigned to the observed future death interval after conditioning on survival through qq. Given a death interval (ai,bi](a_i,b_i], this score is

log⁡Pr⁡(ai<ti∗≤bi∣ti∗>q).\log\Pr\left(a_i < t_i^* \le b_i \mid t_i^* > q\right).

The log probability rewards models that assign high probability to the observed death interval. The median death mean absolute error (MAE) compares the posterior median predicted death time to the first recorded death time bib_i, giving an error in days. The death concordance index (C-index) evaluates ranking accuracy by measuring whether mice predicted to die earlier do in fact die earlier. For tumor response, we evaluate mRECIST classification accuracy using the clinical categories: complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). This metric evaluates whether the model predicts clinically interpretable tumor response categories.

We compare GrAFT-PDX to separate survival and tumor response baselines. The survival baselines are the Cox proportional hazards model [15] equipped with the Breslow baseline hazard estimator [9] and the Royston-Parmar spline-based survival model [40]. The tumor response baselines are multinomial regression, a multilayer perceptron classifier, and XGBoost [12]. Baseline hyperparameter values and tuning grids are given in Section A.1.

Figure 5 shows that GrAFT-PDX achieves higher death log probabilities and lower death time prediction errors than the survival baselines across all censoring times. Relative to Royston-Parmar, its death log probability is higher by 0.11–0.12 nats, the MAE of its predicted median death time is lower by 0.43–0.89 days, and its C-index matches or exceeds that of Royston-Parmar by up to 0.004 across q∈{0,5,10,15}q \in\{0, 5, 10, 15\}. For tumor response prediction, the mean mRECIST accuracy of GrAFT-PDX is 61.6% at q=0q = 0, compared with 63.6% for the best static classification baseline, multinomial regression. From q≥10q \ge10 it outperforms multinomial regression with a mean accuracy of 64.5% at q=10q = 10 increasing to 66.1% at q=15q = 15. As the mRECIST baselines cannot condition on partial tumor trajectories, the q>0q > 0 results highlight a unique capability of GrAFT-PDX among the evaluated models to update tumor predictions using the partial trajectory observed before censoring.

Box plots of death log probability, median death MAE, death C-index, and mRECIST accuracy

Figure 5. Predictions are evaluated at censoring times q∈{0,5,10,15}q \in\{0, 5, 10, 15\}. The Cox model is omitted from the log-probability panel because the Breslow estimator for the baseline hazard assigns 0 probability to some held-out death intervals. For mRECIST accuracy, the GrAFT-PDX results show how partial tumor trajectory information improves predictive performance. For survival, we see that GrAFT-PDX outperforms baselines across all metrics with an increasing performance gap with increasing qq.

Ablation study

We next evaluate whether the tumor history component improves survival prediction on held-out data. We compare the full model to an ablated model that keeps the same baseline hazard model, spline basis, low-rank factorization, priors, train-test splits, and HMC settings, but removes the convolution term from the hazard. Under this ablation, λℓ,j(t)=λℓ,jbase(t)\lambda_{\ell,j}(t) = \lambda_{\ell,j}^{\mathrm{base}}(t), so survival risk can vary across PDX lines, treatments, and time, but it cannot depend directly on the latent tumor trajectory. This particular ablation tests if the learned convolution kernel βconv(a)\beta_{\mathrm{conv}}(a) is necessary to make accurate predictions. Moreover, in order to use GrAFT-PDX to truncate unsuccessful PDX experiments early, we must confirm that βconv(a)\beta_{\mathrm{conv}}(a) is the mechanism through which tumor size informs death risk.

For each censoring time q∈{0,5,10,15}q \in\{0, 5, 10, 15\}, we evaluate both models using five-fold cross-validation under the held-out prediction task defined above. We partition the observed PDX-treatment combinations into five disjoint folds, assigning all mice from the same combination to a single fold. Each model is fit to four folds and evaluated on all mice belonging to combinations in the remaining fold, so every mouse contributes one out-of-sample prediction. For each test mouse ii, the death contribution is the log pointwise predictive density (LPPD) of its future death interval conditional on survival through qq, as defined in Eq. (28). The tumor contribution is the log joint posterior predictive density of all tumor measurements observed after qq. To isolate the predictive contribution of tumor history, we define the per mouse differences

ΔLPPD⁡io(q)=defLPPD⁡i,fullo(q)−LPPD⁡i,no historyo(q),o∈{death,tumor}.(29)\Delta\operatorname{LPPD}_{i}^{o}(q) \mathrel{\overset{\mathrm{def}}{=}} \operatorname{LPPD}_{i,\mathrm{full}}^{o}(q) - \operatorname{LPPD}_{i,\mathrm{no\ history}}^{o}(q), \qquad o \in\{\mathrm{death}, \mathrm{tumor}\}. \tag*{(29)}

The death difference measures whether the tumor history component improves held-out survival prediction, while the tumor difference checks whether including this component changes the quality of the fitted tumor trajectories. We summarize these differences by their mean and standard error across held-out mice.

Figure 6 summarizes the mouse level differences in death and tumor LPPD. At q=0q = 0, the mean death ΔLPPD\Delta\mathrm{LPPD} is positive, but its 95% confidence interval includes zero. At q∈{5,10,15}q \in\{5, 10, 15\}, the mean death ΔLPPD\Delta\mathrm{LPPD} is positive and each 95% confidence interval excludes zero, indicating that the full model assigns greater probability to held-out death intervals than the ablated model once early tumor measurements are available. By contrast, the 95% confidence intervals for the mean tumor ΔLPPD\Delta\mathrm{LPPD} include zero at every censoring time, providing no evidence of a systematic difference between the two models. Together with the synthetic recovery experiment, these results provide strong evidence that tumor history contributes to death risk in the PDXe data and that GrAFT-PDX can detect and characterize such effects when they are present.

Per mouse differences in held-out predictive accuracy for death and tumor size

Figure 6. Per mouse differences in held-out predictive accuracy between the full model and the ablation at censoring times q∈{0,5,10,15}q \in\{0, 5, 10, 15\}. Positive values favor the full model and error bars represent the 95% confidence interval for the mean per mouse difference over a five-fold cross validation. Once early tumor measurements are available (q≥5q \ge5), the tumor history component improves death interval prediction without materially changing prediction of subsequent tumor measurements.

Interpreting the learned tumor history kernel

The kernel βconv(a)\beta_{\mathrm{conv}}(a) weights the latent median tumor size Gℓ,j(t−a)G_{\ell,j}(t-a) when computing the hazard at time tt. Its shape describes how the weight assigned to a given tumor burden varies with the time elapsed since that burden occurred. Applied to PDXe, βconv(a)\beta_{\mathrm{conv}}(a) provides new insight into how important historical tumor burden is to current death risk. This construction follows weighted cumulative exposure and distributed lag models, which allow past covariate values to contribute differently according to their lag [46, 19]. Related joint models use weighted biomarker histories to link longitudinal measurements to survival [31]. We estimate the kernel from PDXe to identify which periods of tumor history are most strongly associated with current death risk.

As shown in Figure 7, the fitted kernel places most of its mass on recent tumor history. Its posterior mean is small at lag zero, rises to its maximum at a lag of 11 days, and then declines at longer lags. Thus, the fitted model associates current death risk more strongly with tumor burden accumulated during the preceding weeks than with either instantaneous tumor size alone or more remote tumor history. Because the location of the maximum is inferred from a smooth, regularized kernel, we interpret the figure primarily as evidence for a short recent lag window rather than any specific delay.

Figure 7. Posterior estimate of the tumor history kernel βconv(a)\beta_{\mathrm{conv}}(a), where aa denotes the lag in days. Larger values assign greater weight to tumor burden observed at the corresponding lag when constructing the current tumor history hazard. The fitted kernel is concentrated at recent lags, with the posterior mean attaining its maximum at approximately two weeks. Crucially, we find that it does not decrease monotonically as we had initially expected.

This association is broadly consistent with earlier analyses, although those analyses impose different association structures. In xenograft experiments, Pan et al. [36] linked longitudinal tumor volume and survival through shared random effects and found a positive association between tumor volume and death risk. In clinical oncology, Claret et al. [13, 14] identified early change in tumor size and time to tumor growth as predictors of overall survival, while Tardivon et al. [48] found that both time to tumor growth and the instantaneous rate of tumor size change were informative.

The posterior mean’s peak away from zero indicates that elevated tumor burden may increase death risk through physiological deterioration that develops over time. Mouse experiments have documented metabolic disruption, muscle loss, and weight loss emerging over successive days [35]. These observations provide a plausible explanation for why earlier tumor burden could be associated with current death risk, although they do not establish which processes underlie the fitted kernel. Because PDXe does not measure all of these changes, evaluating this explanation would require further studies and analyses that track them alongside tumor volume and survival.

Qualitative analysis of latent components

Nonnegative CP decompositions give us the ability to look inside of our model and examine its inferential components to ensure they admit plausible interpretations. In a good model, such qualitative analysis should reveal the biological mechanisms that underpin the statistical inference. We examine the K=5K = 5 decomposition from the final posterior draw and visualize selected components in Figure 8. For completeness, Section A.5 visualizes all latent components.

Figure 8. Latent tumor components from a single draw of the K=5K = 5 model. For each component, the panels display the ten PDX lines with the largest loadings, the ten treatments with the largest loadings, and its latent function.

Among the five latent components, component 4 exhibits the clearest correspondence between the treatments and PDX lines. Two of its highly weighted treatments contain binimetinib, an inhibitor of MEK1 and MEK2, while three of its highly weighted PDX lines—X-2613, X-3483, and X-4668—are melanomas harboring the activating BRAF V600E mutation. Because BRAF signals through MEK in the MAPK pathway, this grouping agrees with experimental evidence that BRAF-mutant tumors are particularly sensitive to MEK inhibition [44]. The component also includes LEE011, an inhibitor of CDK4 and CDK6, together with four PDX lines—X-3671, X-2613, X-3483, and X-4668—having deletions of CDKN2A, CDKN2B, or both. These genes encode negative regulators of CDK4/6 activity, and loss of CDKN2A has previously been associated with sensitivity to CDK4/6 inhibition in melanoma [57]. The latent function decreases over the first 60 days, consistent with a shrinking tumor in a group containing established genomic relationships.

Component 1 provides a complementary example in which the treatments are negatively related to the top PDX loadings. Six of its ten highly weighted treatments contain encorafenib, a selective RAF inhibitor whose activity is concentrated in tumors harboring activating BRAF V600 mutations [17]. Among the ten highly weighted PDX lines, however, only X-2182 carries BRAF V600E. Consistent with this genomic mismatch, the temporal factor increases early and remains substantial rather than shrinking. Component 1 may therefore represent sustained tumor growth under BRAF-directed treatment in a collection of PDX lines that largely lacks the relevant genomic mutation.

Biological association preservation

We next ask whether posterior predictive outcomes from GrAFT-PDX retain genomic response patterns observed in PDXe. The model is fit using tumor trajectories and survival outcomes alone; genomic covariates are not included. We opt not to use genomic information as there is no clear method to incorporate it and the quantity of genomic information would make posterior inference computationally intractable. Therefore, we can use genomic data after fitting as an external check on the model-generated outcomes. To evaluate this, we compare association analyses fit to the observed data with the same analyses fit to a model-generated dataset.

The generated dataset is constructed out of sample. First, we partition the treated PDX-treatment pairs into five folds. For each fold, we train GrAFT-PDX on the other four folds and use it to generate posterior predictive outcomes for the held-out treated pairs. We retain all untreated mice in every training set. Aggregating the held-out predictions across folds gives one generated dataset whose treated outcomes are predicted out of sample, though with weak dependence due to being constructed from the jackknife-style cross-validation procedure.

We analyze the observed and generated datasets in the same way. We first define the treatment effect using lifespan as the response. Let sˉℓ,j\bar{s}_{\ell,j} denote the average lifespan among mice from PDX line ℓ\ell assigned treatment jj, and let sˉℓ,0\bar{s}_{\ell,0} denote the average lifespan among untreated mice from the same PDX line. We define

Δℓ,j=defsˉℓ,j−sˉℓ,0.(30)\Delta_{\ell,j} \overset{\mathrm{def}}{=} \bar{s}_{\ell,j} - \bar{s}_{\ell,0}. \tag*{(30)}

Thus, Δℓ,j\Delta_{\ell,j} compares lifespan under treatment jj to no treatment for the same PDX line.

To relate the treatment effect to tumor genomic information, we fix a treatment jj, a gene gg, and a gene alteration type m∈{amplification,deletion,mutation}m \in\{\text{amplification},\text{deletion},\text{mutation}\}. Let xℓ,g(m)∈{0,1}x_{\ell,g}^{(m)} \in\{0,1\} indicate whether PDX line ℓ\ell has alteration type mm in gene gg. For each combination of treatment, gene, and gene alteration, we model the mean treatment effect as

Eℓ[Δℓ,j∣xℓ,g(m)=0]=defγj,Eℓ[Δℓ,j∣xℓ,g(m)=1]=defγj+δj,g,m.(31)\mathbb{E}_{\ell}\left[\Delta_{\ell,j} \mid x_{\ell,g}^{(m)} = 0\right] \overset{\mathrm{def}}{=} \gamma_j,\qquad\mathbb{E}_{\ell}\left[\Delta_{\ell,j} \mid x_{\ell,g}^{(m)} = 1\right] \overset{\mathrm{def}}{=} \gamma_j + \delta_{j,g,m}. \tag*{(31)}

Here, γj\gamma_j is the average treatment effect among PDX lines without alteration type mm in gene gg, while δj,g,m\delta_{j,g,m} encodes the association between treatment response and the external genomic covariate xℓ,g(m)x_{\ell,g}^{(m)}. We estimate δj,g,m\delta_{j,g,m} using the linear regression

Δℓ,j=γj+δj,g,mxℓ,g(m)+ϵℓ.(32)\Delta_{\ell,j} = \gamma_j + \delta_{j,g,m}x_{\ell,g}^{(m)} + \epsilon_\ell. \tag*{(32)}

The regression is fit by ordinary least squares over PDX lines ℓ\ell for which treatment jj, untreated control outcomes, and alteration status for gene gg are available.

We also compute treatment effects for both the minimum tumor size and the mean tumor size over the mouse’s observed tumor trajectory. In these analyses, sˉℓ,j\bar{s}_{\ell,j} and sˉℓ,0\bar{s}_{\ell,0} denote the corresponding average tumor size response among treated and untreated mice from PDX line ℓ\ell. For lifespan, larger values of Δℓ,j\Delta_{\ell,j} indicate longer survival under treatment. For minimum and mean tumor size, smaller values indicate lower tumor burden under treatment.

Agreement across genomic associations

We compare observed and generated association coefficients across the full collection of fitted regressions. For a given treatment jj, gene gg, alteration type mm, and response rr, let δ^j,g,m,robs\widehat{\delta}_{j,g,m,r}^{\mathrm{obs}} denote the coefficient estimated from the observed dataset, and let δ^j,g,m,rgen\widehat{\delta}_{j,g,m,r}^{\mathrm{gen}} denote the corresponding coefficient estimated from the generated dataset. This comparison assesses whether the generated outcomes preserve all the gene association patterns estimated from the observed PDX screen.

Because both coordinates are estimated association coefficients, we fit the calibration line using weighted orthogonal distance regression, an errors-in-variables method that accounts for uncertainty in both covariates and responses [8]. For each fixed alteration type mm and response rr, let Km,r\mathcal{K}_{m,r} denote the collection of treatment-gene pairs included in the corresponding panel. We use these coefficient pairs to fit

δ^j,g,m,rgen=ρm,rδ^j,g,m,robs+εj,g,m,r,(j,g)∈Km,r.(33)\widehat{\delta}^{\mathrm{gen}}_{j,g,m,r} = \rho_{m,r}\widehat{\delta}^{\mathrm{obs}}_{j,g,m,r} + \varepsilon_{j,g,m,r}, \qquad(j,g) \in\mathcal{K}_{m,r}. \tag*{(33)}

Here ρm,r\rho_{m,r} measures the multiplicative calibration of generated association coefficients relative to observed association coefficients for alteration type mm and response rr. The ideal value is ρm,r=1\rho_{m,r}=1 as that would mean on average that the model preserves the associations in the original data. One interpretation of ρm,r\rho_{m,r} is as an extension to a Pearson correlation coefficient. Pearson correlation is equivalent to an ordinary least-squares slope after centering and standardizing both variables, whereas ρm,r\rho_{m,r} is an uncentered weighted orthogonal-regression slope that incorporates uncertainty in both coefficient estimates and need not lie in [−1,1][-1,1].

Figure 9 plots δ^j,g,m,rgen\widehat{\delta}^{\mathrm{gen}}_{j,g,m,r} against δ^j,g,m,robs\widehat{\delta}^{\mathrm{obs}}_{j,g,m,r} separately for each alteration type and response. All nine fitted calibration slopes ρ^m,r\widehat{\rho}_{m,r} are positive, ranging from 0.657 to 0.837, with a mean of 0.750. Because every slope is below one, the generated coefficients tend to be closer to zero than the observed coefficients, with a mean fitted magnitude equal to 75.0% of the observed magnitude. Nevertheless, the 95% confidence intervals for the observed and generated coefficients overlap for 99.6% of all gene associations in the dataset, and the generated point estimate lies within the observed 95% confidence interval in 97.4% of cases. Thus, the generated data generally reproduce the association patterns estimated from the observed data within their reported uncertainty, while yielding point estimates of smaller magnitude.

Figure 9. Each panel plot δ^j,g,m,rgen\widehat{\delta}^{\mathrm{gen}}_{j,g,m,r} against δ^j,g,m,robs\widehat{\delta}^{\mathrm{obs}}_{j,g,m,r} for one alteration type and response. The positive ρ^\widehat{\rho} shows that the generated data preserve association directions, while slopes around 0.75 indicate shrinkage of generated effects toward zero. In all plots the dashed line is the ideal y=xy=x line and the red line shows the regression fit.

Repositioning candidates for untested PDX-treatment combinations

A key motivation for GrAFT-PDX is its ability to predict outcomes for PDX-treatment combinations not observed in the original screen. We use this capability to prioritize candidate future experiments. Eligibility is restricted to PDX lines empirically classified as having progressive disease under no treatment. For each eligible line, an untested treatment is considered a candidate if complete response has the highest posterior predictive probability among the mRECIST categories. Candidates are ranked by their predicted additional lifespan, measured in days relative to the untreated outcome for the same PDX line.

Table 1 presents the ten highest ranked candidates. Predicted death time under each condition is the median of its posterior predictive distribution. The accompanying 90% posterior predictive interval (PPI) is computed from the posterior predictive distribution of the difference in death times. The untreated mRECIST classification is computed from the observed data, whereas the treated classification is inferred from posterior predictive tumor trajectories. For PDX lines empirically classified as having progressive disease when untreated, Pr⁡(PD→CR)\Pr(\mathrm{PD} \to\mathrm{CR}) denotes the posterior predictive probability of complete response under the candidate treatment. All ten candidates have positive predicted additional lifespan, although the 90% PPIs include zero for four candidates. We note that the treatments LEE011, binimetinib, and encorafenib each appear more than once.

RankPDX lineTreatmentAdditional lifespan [90% PPI]Pr⁡(PD→CR)\Pr(\mathrm{PD} \to\mathrm{CR})
1X-3483 / CMLEE011 + binimetinib215 [15, 388]1.00
2X-3746 / CMLEE011 + binimetinib176 [4, 352]1.00
3X-4644 / CMLEE011 + binimetinib166 [3, 355]0.76
4X-1228 / NSCLCencorafenib + binimetinib92 [-3, 232]0.99
5X-3483 / CMBKM120 + binimetinib88 [5, 204]0.51
6X-1228 / NSCLCLEE011 + encorafenib87 [-3, 198]0.53
7X-4567 / BRCABKM120 + encorafenib81 [-3, 203]0.75
8X-3483 / CMgemcitabine-50mpk79 [7, 152]0.95
9X-4567 / BRCALEE011 + encorafenib78 [-2, 189]0.75
10X-4567 / BRCALEE011 + binimetinib78 [1, 179]0.50

Table 1. The ten highest ranked untested PDX-treatment combinations identified as candidate future experiments. Each candidate pairs a PDX line empirically classified as having progressive disease under no treatment with a treatment for which complete response receives the greatest posterior predictive probability. Candidates are ordered by predicted additional lifespan.

GrAFT-PDX predicts that the top three candidates, X-3483/CM, X-3746/CM, and X-4644/CM, will respond to a combinations therapy of LEE011 and binimetinib. In our data, all three PDX lines carry a BRAF V600E mutation. Although binimetinib does not directly inhibit BRAF, it inhibits MEK1/2 downstream of BRAF in the MAPK pathway, making MEK inhibition biologically plausible in BRAF-mutant tumors [44, 51]. These same three candidates also show alterations in the CDK4/6 regulatory axis. X-3483/CM has a partial deletion of CDKN2B, and X-3746/CM has a mutation in CDKN2A. Both genes encode INK4-family tumor suppressors that regulate CDK4/6 activity [41]. X-4644/CM has an amplification of CDK6, which LEE011 directly targets [52]. Thus, the highest-ranked candidates are consistent with a biological hypothesis involving concurrent MAPK-pathway activation and CDK4/6-axis dysregulation. This shows that the model’s predictions provide biologically coherent hypotheses for future experimentation.

Conclusion

This paper introduced GrAFT-PDX, a hierarchical Bayesian model for analyzing longitudinal PDX data that jointly models tumor progression and survival under treatment. GrAFT-PDX combines spline-based tumor trajectories, low-rank sharing across PDX lines and treatments, and a survival model in which latent tumor history contributes to death risk through a learned convolution kernel. This construction is designed for sparse PDX screens, where many PDX-treatment combinations are unobserved and where tumor measurements are informatively truncated by death. Across held-out and partially observed prediction tasks, GrAFT-PDX improves survival prediction relative to standard survival baselines and produces competitive tumor response predictions that can update as early tumor measurements are incorporated.

GrAFT-PDX links treatment response to survival through the latent tumor trajectory. The learned convolution kernel identifies the lag window over which previous tumor burden is most associated with current death risk, while the latent representations capture heterogeneity in response across tumors and treatments. Although genomic covariates are not used in fitting the model, posterior predictive outcomes preserve biologically meaningful treatment-gene response patterns and yield coherent hypotheses for untested PDX-treatment combinations. Thus, the model’s latent structure supports both prediction and biological interrogation.

Several limitations point to natural extensions. GrAFT-PDX currently learns response structure from tumor and survival outcomes alone, without incorporating external covariates such as mutation status, copy-number alterations, cancer type, treatment targets, or known pathway relationships. Adding these sources of information could improve prediction for PDX lines and treatments with few observations and would be essential for settings involving new PDX lines or new therapies. The current framework also ranks candidate experiments using posterior predictive summaries, but a full experimental design strategy would require decision rules that account for posterior uncertainty, validation cost, and the relative consequences of false positive and false negative recommendations. Future work should therefore integrate biological covariates into the latent representation, develop uncertainty aware criteria for experimental prioritization and early stopping, and validate the resulting candidates in independent PDX screens or prospective experiments.

Taken together, 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.

Acknowledgments

We are grateful to Olivia Tsang, Ankur Garg, Amber Lee, and Christopher Wolfram for their thoughtful comments on figures and prose. WT is supported by the NIH/NCI (R37 CA271186, U54 CA274492, P30 CA008748), Break Through Cancer, the Fund for Innovation in Cancer Informatics, the Cancer AI Alliance, the Tow Center for Developmental Oncology, and the Maurice Campbell Initiative at Memorial Sloan Kettering Cancer Center.

Additional Model and Evaluation Details

Baseline and hyperparameter tuning information

The implementations for multinomial regression and the multilayer perceptron come from the Scikit-Learn Python library. The implementation of XGBoost comes from the XGBoost Python library and the implementations for the Royston-Parmar model and Cox proportional hazards model come from the Lifelines Python library. Cross-validation was done on a held-out test set and details about hyperparameter grids chosen for tuning can be found in Table 2. We select the mRECIST baseline hyperparameters to maximize validation accuracy on the tuning split. For the survival baselines the selection criterion was highest mean log probability.

BaselineTuned hyperparametersSearch grid / candidatesBest selected setting
Multinomial RegressionCC, max_iterC∈{0.01,0.1,1.0,10.0,100.0}C \in\{0.01, 0.1, 1.0, 10.0, 100.0\}; max_iter = 2000CC = 1.0; max_iter = 2000
Multilayer Perceptronhidden_layer_sizes; alpha; learning_rate_init; early_stopping; max_iterhidden_layer_sizes ∈{(32,),(64,),(64,32)}\in\{(32,), (64,), (64, 32)\}; alpha ∈{10−4,10−3,10−2}\in\{10^{-4}, 10^{-3}, 10^{-2}\}; learning_rate_init ∈{10−3,10−2}\in\{10^{-3}, 10^{-2}\}; early_stopping = True; max_iter = 1000hidden_layer_sizes = (64, ); alpha = 10−410^{-4}; learning_rate_init = 10−310^{-3}; early_stopping = True; max_iter = 1000
XGBoostn_estimators; max_depth; learning_rate; subsample; colsample_bytreen_estimators ∈{100,500,900,1400}\in\{100, 500, 900, 1400\}; max_depth ∈{2,4,6,8}\in\{2, 4, 6, 8\}; learning_rate ∈{0.03,0.05,0.1}\in\{0.03, 0.05, 0.1\}; subsample ∈{0.8,1.0}\in\{0.8, 1.0\}; colsample_bytree ∈{0.8,1.0}\in\{0.8, 1.0\}n_estimators = 1400; max_depth = 8; learning_rate = 0.03; subsample = 1.0; colsample_bytree = 0.8
Cox PHcox_step_sizecox_step_size ∈{0.03,0.1,0.3}\in\{0.03, 0.1, 0.3\}cox_step_size = 0.03
Royston-Parmarrp_df; rp_time_scalerp_df ∈{2,3,4,5}\in\{2, 3, 4, 5\}; rp_time_scale ∈{30.0,100.0,365.0}\in\{30.0, 100.0, 365.0\}rp_df = 3; rp_time_scale = 30.0

Table 2. Hyperparameter tuning grids for the mRECIST and survival baselines.

Choosing Latent Dimension Size

We quantify performance of GrAFT-PDX for different values of latent dimension KK on fully held-out data, q=0q = 0. We consider K∈{3,5,7}K \in\{3, 5, 7\} and hold out 20% of the drug-PDX combinations as a test set. We repeat this procedure across 10 different train-test splits and calculate the relevant metrics on each split. We report the mean absolute error (MAE) and the average log pointwise predictive density (LPPD) [20]. We complement LPPD with three held-out diagnostics: time-of-death absolute error for the posterior median, concordance index (C-index), and mRECIST accuracy. Figure 10 shows the results of this testing. C-index, LPPD for survival, median MAE, and mRECIST accuracy are all approximately the same across our various values of KK. Tumor size LPPD seems to slightly decrease beyond K=5K = 5. Ultimately, we choose K=5K = 5 because it gives comparable survival performance to other choices while avoiding the apparent decline in tumor predictive density seen at K=7K = 7.

Figure 10. Held-out performance for K∈{3,5,7}K \in\{3, 5, 7\}. Survival metrics are similar across the three choices, while held-out tumor LPPD decreases for K=7K = 7. We therefore use K=5K = 5 in the main experiments.

Choice of positivity parameterization

Both the tumor trajectory and baseline hazard require positive spline coefficients. Because the B-spline basis functions are non-negative, positivity can be imposed either by constraining the latent factors directly or by applying an exponential link after reconstructing the log coefficient tensor. Let

Tℓ,j,c(u,v,w)=def∑k=1Kωℓ,kωj,kωc,k.\mathcal{T}_{\ell,j,c}(u,v,w) \overset{\mathrm{def}}{=} \sum_{k=1}^{K} \omega_{\ell,k}\omega_{j,k}\omega_{c,k}.

We compare three constructions. The constrained construction sets both the tumor and baseline hazard coefficients equal to Tℓ,j,c\mathcal{T}_{\ell,j,c} with nonnegative factors. The unconstrained construction exponentiates Tℓ,j,c\mathcal{T}_{\ell,j,c} in both submodels, allowing the underlying factors to be unconstrained. The mixed construction used by GrAFT-PDX directly constrains the tumor factors but exponentiates an unconstrained factorization for the baseline hazard coefficients. Positive factors retain their Exponential(1) priors, unconstrained factors receive N(0,1)\mathcal{N}(0,1) priors, and all remaining model components, including the nonnegative convolution coefficients βc\beta_c, are unchanged. We fit each construction with K∈{3,5,7}K \in\{3,5,7\} using the same 10 train-test splits, spline basis, and fitting procedure. Figure 11 reports held-out tumor and death LPPD. The upper panels compare the constructions at each fixed KK, while the lower panels show the value of KK attaining the highest mean LPPD for each construction and outcome.

Held-out tumor and death LPPD under mixed, constrained, and unconstrained positivity constructions

Figure 11. Held-out tumor and death LPPD under the mixed, constrained, and log-linked positivity constructions across 10 common train-test splits. The mixed construction yields the highest death LPPD at every fixed KK and after allowing KK to vary, while retaining tumor LPPD comparable to alternative constructions.

The mixed construction consistently provides the highest held-out death LPPD. Tumor prediction is broadly comparable between the mixed and constrained constructions. We find that the fully unconstrained construction deteriorates as KK increases. Therefore, we use the mixed construction in the main analysis as it improves survival prediction without any loss in tumor prediction.

Full Generative Model

A.4. Full Generative Model. The following priors are placed on the model parameters:

uℓ,k(d),vj,k(d),wc,k(d)∼N(0,1),u_{\ell,k}^{(d)}, v_{j,k}^{(d)}, w_{c,k}^{(d)} \sim\mathcal{N}(0,1),
uℓ,k,vj,k,wc,k,ηc∼Exponential⁡(1),u_{\ell,k}, v_{j,k}, w_{c,k}, \eta_c \sim\operatorname{Exponential}(1),
βc∼Exponential⁡(10c−1∫0TBc(a) da).\beta_c \sim\operatorname{Exponential}\left(10^{c-1}\int_{0}^{T} B_c(a)\,\mathrm{d}a\right).

for ℓ=1,…,L\ell= 1,\ldots,L, j=0,…,Jj = 0,\ldots,J, k=1,…,Kk = 1,\ldots,K, and c=1,…,Cc = 1,\ldots,C. Here B1,…,BCB_1,\ldots,B_C denote the fixed spline basis functions. Together with the tumor and death submodels, these priors define the full generative model. The interval-observed likelihood is used for the PDXe death records, while the right-censored form is included because the held-out prediction task conditions on survival through a censoring time.

Death submodel

θℓ,j,c(d)=exp⁡(∑k=1Kwℓ,k(d)vj,k(d)wc,k(d))\theta_{\ell,j,c}^{(d)} = \exp\left(\sum_{k=1}^{K} w_{\ell,k}^{(d)} v_{j,k}^{(d)} w_{c,k}^{(d)}\right)
βconv(a)=∑c=1CβcBc(a)\beta_{\mathrm{conv}}(a) = \sum_{c=1}^{C} \beta_c B_c(a)
λℓ,jbase(t)=∑c=1Cθℓ,j,c(d)Bc(t)\lambda_{\ell,j}^{\mathrm{base}}(t) = \sum_{c=1}^{C} \theta_{\ell,j,c}^{(d)} B_c(t)
λℓ,jtumor(t)=∫0tβconv(t−s)Gℓ,j(s) ds\lambda_{\ell,j}^{\mathrm{tumor}}(t) = \int_{0}^{t} \beta_{\mathrm{conv}}(t-s)G_{\ell,j}(s)\,\mathrm{d}s
λℓ,j(t)=λℓ,jbase(t)+λℓ,jtumor(t)\lambda_{\ell,j}(t) = \lambda_{\ell,j}^{\mathrm{base}}(t) + \lambda_{\ell,j}^{\mathrm{tumor}}(t)
Sℓ,j(t)=exp⁡(−∫0tλℓ,j(s) ds)S_{\ell,j}(t) = \exp\left(-\int_{0}^{t} \lambda_{\ell,j}(s)\,\mathrm{d}s\right)
Pr⁡(ai<ti∗≤bi)=Sℓi,ji(ai)−Sℓi,ji(bi)\Pr(a_i < t_i^* \le b_i) = S_{\ell_i,j_i}(a_i) - S_{\ell_i,j_i}(b_i)
Pr⁡(ti∗>qi)=Sℓi,ji(qi)\Pr(t_i^* > q_i) = S_{\ell_i,j_i}(q_i)

Tumor submodel

θℓ,j,c=∑k=1Kuℓ,kvj,kwc,k\theta_{\ell,j,c} = \sum_{k=1}^{K} u_{\ell,k}v_{j,k}w_{c,k}
Gℓ,j(t)=∑c=1Cθℓ,j,cBc(t)G_{\ell,j}(t) = \sum_{c=1}^{C} \theta_{\ell,j,c}B_c(t)
D(t)=∑c=1CηcBc(t)D(t) = \sum_{c=1}^{C} \eta_c B_c(t)
yi(m)∼Lognormal⁡(MEDIAN⁡=Gℓi,ji(τi(m)), SCALE⁡=D(τi(m)))y_i^{(m)} \sim\operatorname{Lognormal}\left(\operatorname{MEDIAN}=G_{\ell_i,j_i}\left(\tau_i^{(m)}\right),\ \operatorname{SCALE}=D\left(\tau_i^{(m)}\right)\right)

All Tumor Size Latent Components.

All Tumor Size Latent Components

Figure 12. Latent tumor components from a single draw of the K=5K = 5 model. For each component, the panels display the ten PDX lines with the largest loadings, the ten treatments with the largest loadings, and the corresponding temporal factor.

References

  1. [1]Aalen, O. (1980). A Model for Nonparametric Regression Analysis of Counting Processes. In Mathematical Statistics and Probability Theory (W. Klonecki, A. Kozek and J. Rosiński, eds.) 1-25. Springer New York, New York, NY.DOI
  2. [2]Aalen, O. O. (1989). A linear regression model for the analysis of life times. Statistics in Medicine 8 907-925. https://doi.org/10.1002/sim.4780080803
  3. [3]Akemi Kido, L., Rodrigues Marusco, M., Aparecida da Silva, E., Do Carmo, L., Beatriz Teodoro Borges, A., Luiz Torres Silva, F., Silveira Ruas, J., Giomo de Lima, D., de Abreu Fernandes, L., Maia Martin Daiggi, C. et al. (2025). Establishing a pediatric solid tumor PDX biobank for precision oncology research. Cancer Biology & Therapy 26 2541974.DOI
  4. [4]Alvares, D. and Rubio, F. J. (2021). A tractable Bayesian joint model for longitudinal and survival data. Statistics in Medicine 40 4213-4229. https://doi.org/10.1002/sim.9024
  5. [5]Ammad-ud Din, M., Georgii, E., Gönen, M., Laitinen, T., Kallioniemi, O., Wennerberg, K., Poso, A. and Kaski, S. (2014). Integrative and Personalized QSAR Analysis in Cancer by Kernelized Bayesian Matrix Factorization. Journal of Chemical Information and Modeling 54 2347-2359. https://doi.org/10.1021/ci500152b
  6. [7]Bertsimas, D. and Pawlowski, C. (2023). Tensor completion with noisy side information. Machine Learning 112 3945-3976. https://doi.org/10.1007/s10994-023-06338-5
  7. [8]Boggs, P. T. and Rogers, J. E. (1989). Orthogonal distance regression.
  8. [9]Breslow, N. (1972). Discussion of the paper by D. R. Cox. Journal of the Royal Statistical Society. Series B (Methodological) 34 216–217.
  9. [10]Brooks, S., Gelman, A., Jones, G. and Meng, X.-L. (2011). Handbook of Markov Chain Monte Carlo. Chapman and Hall/CRC. https://doi.org/10.1201/b10905
  10. [11]Carroll, J. D. and Chang, J.-J. (1970). Analysis of individual differences in multidimensional scaling via an n-way generalization of “Eckart-Young” decomposition. Psychometrika 35 283-319. https://doi.org/10.1007/BF02310791
  11. [12]Chen, T. and Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting System. CoRR abs/1603.02754.DOI
  12. [13]Claret, L., Girard, P., Hoff, P. M., Van Cutsem, E., Zuideveld, K. P., Jorga, K., Fagerberg, J. and Bruno, R. (2009). Model-Based Prediction of Phase III Overall Survival in Colorectal Cancer on the Basis of Phase II Tumor Dynamics. Journal of Clinical Oncology 27 4103-4108. PMID: 19636014. https://doi.org/10.1200/JCO.2008.21.0807
  13. [14]Claret, L., Gupta, M., Han, K., Joshi, A., Sarapa, N., He, J., Powell, B. and Bruno, R. (2013). Evaluation of Tumor-Size Response Metrics to Predict Overall Survival in Western and Chinese Patients With First-Line Metastatic Colorectal Cancer. Journal of Clinical Oncology 31 2110-2114. PMID: 23650411. https://doi.org/10.1200/JCO.2012.45.0973
  14. [15]Cox, D. R. (1972). Regression Models and Life-Tables. Journal of the Royal Statistical Society. Series B (Methodological) 34 187–220.
  15. [16]De Boor, C. (1978). A Practical Guide to Splines. Springer Verlag, New York.
  16. [17]Delord, J.-P., Robert, C., Nyakas, M., McArthur, G. A., Kudchakar, R., Mahipal, A., Yamada, Y., Sullivan, R., Arance, A., Kefford, R. F., Carlino, M. S., Hidalgo, M., Gomez-Roca, C., Michel, D., Seroutou, A., Aslanis, V., Caponigro, G., Stuart, D. D., Moutouh-de Parseval, L., Demuth, T. and Dummer, R. (2017). Phase I Dose-Escalation and -Expansion Study of the BRAF Inhibitor Encorafenib (LGX818) in Metastatic BRAF-Mutant Melanoma. Clinical Cancer Research 23 5339-5348. https://doi.org/10.1158/1078-0432.CCR-16-2923
  17. [18]Gao, H., Korn, J. M., Ferretti, S., Monahan, J. E., Wang, Y., Singh, M., Zhang, C., Schnell, C., Yang, G., Zhang, Y., Balbin, O. A., Barbe, S., Cai, H., Casey, F., Chatterjee, S., Chiang, D. Y., Chuai, S., Cogan, S. M., Collins, S. D., Dammassa, E., Ebel, N., Embry, M., Green, J., Kauffmann, A., Kowalczyk, C., Leary, R. J., Lehar, J., Liang, Y., Loo, A., Lorenzana, E., Robert McDonald, A., McLaughlin, M. E., Merkin, J., Meyer, R., Naylor, T. L., Patrawan, M., Reddy, A., Röelli, C., Ruddy, D. A., Salangsang, F., Santacroce, F., Singh, A. P., Tang, Y., Tinetto, W., Tobler, S., Velazquez, R., Venkatesan, K., Von Arx, F., Wang, H. Q., Wang, Z., Wiesmann, M., Wyss, D., Xu, F., Bitter, H., Atadja, P., Lees, E., Hofmann, F., Li, E., Keen, N., Cozens, R., Jensen, M. R., Pryer, N. K., Williams, J. A. and Sellers, W. R. (2015). High-throughput screening using patient-derived tumor xenografts to predict clinical trial drug response. Nature Medicine 21 1318-1325. https://doi.org/10.1038/nm.3954
  18. [19]Gasparrini, A., Armstrong, B. and Kenward, M. G. (2010). Distributed lag non-linear models. Stat. Med. 29 2224–2234.DOI
  19. [20]Gelman, A., Hwang, J. and Vehtari, A. (2014). Understanding predictive information criteria for Bayesian models. Statistics and Computing 24 997-1016. https://doi.org/10.1007/s11222-013-9416-2
  20. [21]Gupta, S., Mohani, V., Ghisl̄at, G., Ballester, P. and Ahmad, S. (2025). A pan-cancer, pan-treatment model for predicting drug responses from patient-derived xenografts. NAR Genomics and Bioinformatics 7 lqaf111. https://doi.org/10.1093/nargab/lqaf111
  21. [22]Harshman, R. A. (1970). Foundations of the PARAFAC procedure: Models and conditions for an "explanatory" multi-modal factor analysis. UCLA Working Papers in Phonetics 16 1–84.
  22. [23]Haslam, A., Kim, M. and Prasad, V. (2021). Updated estimates of eligibility for and response to genome-targeted oncology drugs among US cancer patients, 2006-2020. Annals of Oncology 32 926–932.DOI
  23. [24]Hoffman, M. D. and Gelman, A. (2014). The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo. Journal of Machine Learning Research 15 1593–1623.DOI
  24. [25]Ibrahim, J. G., Chu, H. and Chen, L. M. (2010). Basic Concepts and Methods for Joint Models of Longitudinal and Survival Data. Journal of Clinical Oncology 28 2796–2801. PMID: 20439643. https://doi.org/10.1200/JCO.2009.25.0654
  25. [26]J̄arada, T. N., Rokne, J. G. and Alhajj, R. (2020). A review of computational drug repositioning: strategies, approaches, opportunities, challenges, and directions. Journal of Cheminformatics 12 46. https://doi.org/10.1186/s13321-020-00450-7
  26. [27]Kolda, T. G. and Bader, B. W. (2009). Tensor Decompositions and Applications. SIAM Review 51 455–500. https://doi.org/10.1137/07070111X
  27. [28]Kurilov, R., Haibe-Kains, B. and Brors, B. (2020). Assessment of modelling strategies for drug response prediction in cell lines and xenografts. Scientific Reports 10 2849. https://doi.org/10.1038/s41598-020-59656-2
  28. [31]Mauff, K., Steyerberg, E. W., Nijpels, G., van der Heijden, A. A. W. A. and Rizopoulos, D. (2017). Extension of the association structure in joint models to include weighted cumulative effects. Stat. Med. 36 3746–3759.DOI
  29. [32]Meric-Bernstam, F., Lloyd, M. W., Koc, S., Evrard, Y. A., McShane, L. M., Lewis, M. T., Evans, K. W., Li, D., Rubinstein, L., Welm, A. et al. (2024). Assessment of patient-derived xenograft growth and antitumor activity: the NCI PDXNet consensus recommendations. Molecular Cancer Therapeutics 23 924–938.
  30. [33]Mourragui, S. M. C., Loog, M., Vis, D. J., Moore, K., Manjon, A. G., van de Wiel, M. A., Reinders, M. J. T. and Wessels, L. F. A. (2021). Predicting patient response with models trained on cell lines and patient-derived xenografts by nonlinear transfer learning. Proceedings of the National Academy of Sciences 118 e2106682118. https://doi.org/10.1073/pnas.2106682118
  31. [35]O’Connell, T. M., Golz̄arri-Arroyo, L., Pin, F., Barreto, R., Dickinson, S. L., Couch, M. E. and Bonetto, A. (2021). Metabolic biomarkers for the early detection of cancer cachexia. Front. Cell Dev. Biol. 9 720096.DOI
  32. [36]Pan, J., Bao, Y., Dai, H. and Fang, H.-B. (2014). Joint longitudinal and survival-cure models in tumour xenograft experiments. Statistics in Medicine 33 3229–3240. https://doi.org/10.1002/sim.6175
  33. [37]Peska, L., Buza, K. and Koller, J. (2017). Drug-target interaction prediction: A Bayesian ranking approach. Computer Methods and Programs in Biomedicine 152 15–21. https://doi.org/10.1016/j.cmpb.2017.09.003
  34. [38]Ren, S., Tao, Y., Yu, K., Xue, Y., Schwartz, R. and Lu, X. (2022). De novo prediction of cell-drug sensitivities using deep learning-based graph regularized matrix factorization. Pac. Symp. Biocomput. 27 278–289.
  35. [39]Rizopoulos, D., Molenberghs, G. and Lesaffre, E. M. E. H. (2017). Dynamic predictions with time-dependent covariates in survival analysis using joint modeling and landmarking. Biometrical Journal 59 1261–1276. https://doi.org/10.1002/bimj.201600238
  36. [40]Royston, P. and Parmar, M. K. B. (2002). Flexible parametric proportional-hazards and proportional-odds models for censored survival data, with application to prognostic modelling and estimation of treatment effects. Stat. Med. 21 2175–2197.pubmed.ncbi.nlm.nih.gov/12210632
  37. [41]Ruas, M. and Peters, G. (1998). The p16INK4a/CDKN2A tumor suppressor and its relatives. Biochimica et Biophysica Acta (BBA) - Reviews on Cancer 1378 F115-F177. https://doi.org/10.1016/S0304-419X(98)00017-1
  38. [42]Saridogan, T., Akcakanat, A., Zhao, M., Evans, K. W., Yuca, E., Scott, S., Kirby, B. P., Zheng, X., Ha, M. J., Chen, H., Ng, P. K. S., DiPeri, T. P., Mills, G. B., Rodon Ahnert, J., DAMODARAN, S. and MERIC-BERNSTAM, F. (2023). Efficacy of futibatinib, an irreversible fibroblast growth factor receptor inhibitor, in FGFR-altered breast cancer. Scientific Reports 13 20223. https://doi.org/10.1038/s41598-023-46586-y
  39. [43]SCHEIN, A., PAISLEY, J., BLEI, D. M. and WALLACH, H. (2015). Bayesian Poisson Tensor Factorization for Inferring Multilateral Relations from Sparse Dyadic Event Counts. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD ’15 1045–1054. Association for Computing Machinery, New York, NY, USA. https://doi.org/10.1145/2783258.2783414
  40. [44]SOLIT, D. B., GARRAWAY, L. A., PRATILAS, C. A., SHAW, A. S., GETZ, G., BASSO, A., YE, Q., LOBO, J. M., SHE, Y., OSMAN, I., GOLUB, T. R., SEBOLT-LEOPOLD, J., SELLERS, W. R. and ROSEN, N. (2006). BRAF mutation predicts sensitivity to MEK inhibition. Nature 439 358–362.pubmed.ncbi.nlm.nih.gov/16273091
  41. [45]SUPHAVILAI, C., BERTRAND, D. and NAGARAJAN, N. (2018). Predicting Cancer Drug Response using a Recommender System. Bioinformatics 34 3907-3914. https://doi.org/10.1093/bioinformatics/bty452
  42. [46]SYLVESTRE, M.-P. and ABRAHAMOWICZ, M. (2009). Flexible modeling of the cumulative effects of time-dependent exposures on the hazard. Statistics in Medicine 28 3437-3453. https://doi.org/10.1002/sim.3701
  43. [47]TANSEY, W., TOSH, C. and BLEI, D. M. (2022). A Bayesian Model of Dose–Response for Cancer Drug Studies. The Annals of Applied Statistics 16 680–705.arxiv.org/abs/1906.04072
  44. [48]TARDIVON, C., DESMÉE, S., KERIOUI, M., BRUNO, R., WU, B., MENTRÉ, F., MERCIER, F. and GUEDJ, J. (2019). Association between tumor size kinetics and survival in patients with urothelial carcinoma treated with atezolizumab: Implication for patient follow-up. Clin. Pharmacol. Ther. 106 810–820.DOI
  45. [49]STAN DEVELOPMENT TEAM (2026). Stan Reference Manual, version 2.39.
  46. [50]TENTLER, J. J., TAN, A. C., WEEKES, C. D., JIMENO, A., LEONG, S., PITTS, T. M., ARCAROLI, J. J., MESSERSMITH, W. A. and ECKHARDT, S. G. (2012). Patient-derived tumour xenografts as models for oncology drug development. Nature Reviews Clinical oncology 9 338–350.DOI
  47. [51]TRAN, B. and COHEN, M. S. (2020). The discovery and development of binimetinib for the treatment of melanoma. Expert Opin. Drug Discov. 15 745–754.DOI
  48. [52]TRIPATHY, D., BARDIA, A. and SELLERS, W. R. (2017). Ribociclib (LEE011): Mechanism of action and clinical impact of this selective cyclin-dependent kinase 4/6 inhibitor in various solid tumors. Clin. Cancer Res. 23 3251–3262.
  49. [53]WANG, J.-L. and ZHONG, Q. (2025). Joint Modeling of Longitudinal and Survival Data. Annual Review of Statistics and its Application 12 449-476. https://doi.org/10.1146/annurev-statistics-112723-034334
  50. [54]WANG, T., SZEDMAK, S., WANG, H., AITTOKALLIO, T., PAHIKKALA, T., CICHONSKA, A. and ROUSU, J. (2021). Modeling drug combination effects via latent tensor reconstruction. Bioinformatics 37 193-i101. https://doi.org/10.1093/bioinformatics/btab308
  51. [55]WU, J. (2010). Statistical inference for tumor growth inhibition T/C ratio. J. Biopharm. Stat. 20 954–964.DOI
  52. [56]YANG, J., LI, Z., FAN, X. and CHENG, Y. (2014). Drug–Disease Association and Drug-Repositioning Predictions in Complex Diseases Using Causal Inference–Probabilistic Matrix Factorization. Journal of Chemical Information and Modeling 54 2562-2569. https://doi.org/10.1021/ci500340n
  53. [57]YOUNG, R. J., WALDECK, K., MARTIN, C., FOO, J. H., CAMERON, D. P., KIRBY, L., DO, H., MITCHELL, C., CULLINANE, C., LIU, W., FOX, S. B., DUTTON-REGESTER, K., HAYWARD, N. K., JENE, N., DOBROVIC, A., PEARSON, R. B., CHRISTENSEN, J. G., RANDOLPH, S., MCTHUR, G. A. and SHEPPARD, K. E. (2014). Loss of CDKN2A expression is a frequent event in primary invasive melanoma and correlates with sensitivity to the CDK4/6 inhibitor PD0332991 in melanoma cell lines. Pigment Cell Melanoma Res. 27 590–600.DOI
  54. [58]ZHANG, W., XU, H., LI, X., GAO, Q. and WANG, L. (2020). DRIMC: an improved drug repositioning approach using Bayesian inductive matrix completion. Bioinformatics 36 2839-2847. https://doi.org/10.1093/bioinformatics/btaa062
  55. [59]ZHOU, J., JIANG, X., XIA, H. A., HOBBS, B. P. and WEI, P. (2023). Landmark mediation survival analysis using longitudinal surrogate. Frontiers in Oncology Volume 12 - 2022. https://doi.org/10.3389/fonc.2022.999324

Paper details

vlnc:W_s74926qa
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents