Mouse tumor drug screens need not be run in full: GrAFT-PDX couples growth and death in one model and can stop losing combinations early

GrAFT-PDX puts two things that PDX screens previously handled separately, how a tumor grows or shrinks over time and when the mouse dies, into a single probabilistic model, so untested PDX-treatment combinations also receive predictions. Earlier approaches reduced each combination's response to a scalar summary such as tumor volume, growth slope or a binary outcome, then predicted unobserved combinations from it, discarding the longitudinal structure of the tumor measurements. On PDXe with 20 train-test splits at the PDX-treatment combination level, GrAFT-PDX achieves higher death log probability than the Cox proportional hazards model and the Royston-Parmar spline survival model at every censoring time, and lower mean absolute error for predicted median death time.

The model treats each mouse's tumor measurements as noisy samples of a smooth continuous trajectory, and lets the history of that trajectory feed back into death risk. Tumor measurements stop when a mouse dies, and that missingness is informative: fast-growing tumors die earlier and contribute no later measurements, so fitting growth curves alone biases them downward. Tumor trajectories are represented with a cubic B-spline basis whose coefficient tensor is given a nonnegative CP decomposition over PDX lines, treatments and basis functions; death risk is written as a baseline hazard plus a convolution of tumor history with a shared lag kernel weighting past tumor burden, and all parameters are inferred with Stan's NUTS sampler.

Removing the tumor-history convolution term degrades prediction of held-out death intervals, indicating that this pathway carries information. At a censoring time of zero, when no outcome from that combination has been observed, the confidence interval for the death-prediction difference between full and ablated models includes zero. Once early tumor measurements enter the conditioning, the full model's held-out death log pointwise predictive density exceeds the ablated model at every censoring time with confidence intervals excluding zero, while the confidence intervals for the tumor-measurement difference always include zero.

Evaluation is confined to the single large PDXe drug-screen dataset, with prediction tasks holding out PDX-treatment combinations and conditioning on observed early tumor measurements and survival at several censoring times; the synthetic experiments only test whether the model recovers a preset lag kernel. The authors propose incorporating biological covariates such as mutation status, copy-number alterations and cancer type into the latent representation, and developing prioritization rules that account for posterior uncertainty and validation cost.

The authors note that the fitted kernel's peak location comes from a smooth, regularized kernel and should be read mainly as evidence for a short recent lag window rather than a specific delay; which physiological processes raise death risk with tumor burden is not measured in PDXe and would need further study. Generated data preserve the direction of gene associations but with point estimates of smaller magnitude on average, and four of the ten ranked candidates have 90% posterior predictive intervals for additional lifespan that include zero.