Overview

Overall survival (OS) is the endpoint that matters most in oncology, and the one that takes longest to read out, delaying critical go/no-go decisions by months. Early tumor growth dynamics (TGD) data, such as sum of longest diameters, become available much sooner, but traditional summary measures such as objective response rate capture only limited features of TGD trajectories. Nonlinear mixed-effects (NLME) models can characterize these trajectories more thoroughly, but traditional models are complex to build and need to be adapted to different studies, even if the disease is the same.

In this webinar we show how a TGD model built with DeepPumas, by combining NLME modeling with scientific machine learning (SciML) in a single data-driven architecture, captures all possible TGD trajectories without the manual work required by traditional approaches. By adding an OS model on top of the TGD one, we can predict each arm's survival curve from interim data and determine whether a trial will be successful much sooner. This joint TGD-OS model is trained on historical non-small cell lung cancer trials and is treatment-agnostic, though specialized to the tumor type, so it applies to new treatments without retraining. To evaluate whether the model can predict success accurately, even in light of the inevitable bias due to being treatment-agnostic, we test it on a large number of realistic synthetic early trials generated by resampling real trial data.

Date & Time: November 12, 2026, 11 AM ET

Duration: 1 hour

Speakers: Lorenzo Contento (PumasAI), Lu Chen (BMS)

The Session Covers

• How to build a SciML model matching or exceeding the predictive performance of a hand-written NLME mixture model
• How a treatment-agnostic model applied to an unseen randomized trial still permits a comparison of survival between arms, even when only early data is available
• How synthetic early trials can be generated from two withheld trials (one successful and the other not) by using inter-arrival time distributions fitted to the real enrollment patterns and a stopping criterion such as a minimum number of subjects with a given follow-up
• How this allows us to compute true and false positive rates for the go/no-go classifier, to calibrate the decision threshold to reach the desired trade-off and to determine how much interim data is sufficient for a reliable decision

This webinar is curated for pharmacometricians, oncology modeling and simulation scientists, clinical pharmacologists, and study teams working on interim decision-making and model-informed drug development. A working familiarity with NLME modeling helps but isn't required.

Register Now!