Christos Kaikousidis 1, Mohamed Tarek 1, Roxana Aldea 2, João Abrantes 2, Marcelo Boareto 2, Antoine Soubret 2
1 Pumas-ai (, ), 2 Roche Pharma Research and Early Development, Pharmaceutical Sciences (, )
Introduction:
Disease modeling (DM) offers a quantitative framework to characterize and predict disease progression in Alzheimer’s disease, supporting treatment evaluation and clinical trial design. In this work, we leverage a novel DeepNLME-based modeling approach to 1) model the progression of Clinical Dementia Rating (CDR) sum of boxes scores and several longitudinal biomarkers with and without drug and 2) identify promising biomarkers and covariates with a high predictive power of CDR. The DeepNLME disease progression models of CDR were presented in a previous work [1], demonstrating their accuracy.
Methods:
The dataset used in this study was derived from the Gantenerumab trials [2] and comprised 326 subjects in the training set and 352 in the test set. Longitudinal biomarker data were available for plasma, CSF and PET, along with baseline demographic, physiological, and cognitive assessments, including FAQ, MMSE, and ADAS Cog scores. Both placebo and treated subjects were included. Only plasma biomarkers were available for the majority of the subjects and therefore used in this analysis.
To filter the long list of biomarkers, we modeled longitudinal plasma biomarker trajectories using neural network functions of time and random effects:
〖DV〗_Biomarker=NN(t/1000,η_i,ψ^(pk,i))
where time (t) is scaled to avoid vanishing gradients in the neural network, η_i are latent random effects modelling the inter-individual variability (IIV) in the biomarker progression and ψ^(pk,i) is the drug concentration. For tuning the neural network models, Graduate I study was used for training and Graduate II for validation, similar to the CDR model. The individual predictions and visual predictive checks of the biomarkers were found sufficiently good.
After fitting each biomarker independently, the empirical Bayes estimates (EBEs) were extracted for each subject and each biomarker. The EBEs represent embeddings that capture the main information in the biomarker data for a particular subject. To help filter the biomarkers, the biomarkers’ EBEs were incorporated as baseline covariates to augment the CDR model using the DeepPumas sequential covariate modelling workflow. Various combinations of baseline covariates and longitudinal biomarkers were tried. A Jacobian based sensitivity analysis was then performed to identify the most influential biomarker EBEs and/or baseline covariates.
This analysis was used to filter most biomarkers, shortlisting a few potentially influential ones that may enhance the predictive accuracy of the disease progression model. However, because the EBEs are derived from complete biomarker trajectories and subsequently used to inform disease dynamics, the approach violates time causality, where future biomarker values can inform past CDR scores. Consequently, the results are exploratory, highlighting potential associations between specific biomarkers and CDR score trajectories rather than supporting predictive application.
Results:
The individual biomarker models performed well for both the flatter placebo profiles and the more complex trajectories in the active group, with diagnostic plots such as predictions vs observations indicating good fit and no systematic trends. Visual predictive checks (VPCs) also showed the model can generate realistic subjects and had no outliers across all models.
Incorporating EBEs into CDR predictions improved performance, with the base model achieving 〖LL〗_test=2014 and 〖LL〗_train=1932, compared to 〖LL〗_(test,EBE)=2124 and 〖LL〗_(test,EBE)=2032 for the biomarker augmented model in the placebo population. Further inclusion of baseline cognitive scores led to additional improvement 〖LL〗_(test,full)=2459,〖 LL〗_(train,full)=2450 suggesting these measures provide more substantial complementary information for CDRSB modeling. This is somewhat expected since they are directly correlated with the CDR itself. The corresponding values for the PK population were 〖LL〗_test=2688, 〖LL〗_train=2753 for the base model, 〖LL〗_(test,EBE)=2719, 〖LL〗_(test,EBE)=2835 for the biomarker augmented model and 〖LL〗_(test,full)=3090,〖LL〗_(train,full)=3238 for the full model.
Conclusions:
We propose a novel approach for identifying potential associations between biomarker dynamics and clinical endpoints, such as CDR scores, used to characterize Alzheimer’s disease progression. Incorporating biomarker EBEs improved predictive performance; however, this approach violates time causality and can only be used to identify potential links. A more appropriate strategy to leverage biomarker information is joint modeling with the disease modeling endpoint, which we plan to pursue in future work. However, the proposed approach was successful at filtering out most biomarkers, selecting only the few promising ones for the joint model.
References:
[1] PAGE 33 (2025) Abstr 11775 [www.page-meeting.org/?abstract=11775]
[2] Bateman RJ, Smith J, Donohue MC, Delmar P, Abbas R, Salloway S, et al. Two Phase 3 Trials of Gantenerumab in Early Alzheimer’s Disease. N Engl J Med. 2023;389:1862–76.
Reference: PAGE 34 (2026) Abstr 11978 [www.page-meeting.org/?abstract=11978]
Poster: Drug/Disease Modelling - Other Topics