Diego Valderrama1, Ana Victoria Ponce-Bobadilla2, Corinna Maier2, Insa Winzenborg2, Sven Mensing2, Holger Fröhlich1,3, Sven Stodtmann2
1Department of Bioinformatics, Fraunhofer Institute for Algorithms and Scientific Computing (SCAI), 2AbbVie Deutschland GmbH & Co. KG, 3Bonn-Aachen International Center for Information Technology (B-IT), University of Bonn
Introduction Inflammatory Bowel Disease (IBD) affects over 10 million people worldwide. Current advanced treatments may still fail due to side effects or loss of efficacy, underscoring the need for better solutions [1]. While machine learning (ML) and quantitative systems pharmacology (QSP) have enhanced treatment insights, both approaches struggle to integrate diverse data types [2-4]. We propose a multimodal neural ode modeling method that estimates latent representations to clarify the roles of biomarkers in predicting clinical outcomes and enables counterfactual simulations, offering a robust framework for advancing IBD research and treatment strategies. Methodology We utilized multimodal data from various clinical trials involving multiple drugs, with longitudinal data collection for continuous, categorical, and ordinal variables. The dataset comprised baseline measurements, clinical endpoints, and biomarker information presented as log2 changes from the baseline measurement. We employed three distinct biomarker sets: i) the most predictive biomarkers identified through prior SHAP analysis [5], ii) biomarkers highly correlated with this predictive set, and iii) common biomarkers shared across the drugs. For model evaluation, we established validation scenarios that focused on reconstruction accuracy, prediction performance on test patients, and simulation capabilities. Clinical data was divided into 80% for training and 20% for testing, while maintaining similar patient distributions in both splits. The training set was augmented to approximately 2,000 subjects, with a balanced number of patients per drug. We utilized standard pharmacokinetic graphical evaluation methods alongside performance metrics such as correlation coefficients, accuracy, F1 score. We extended the previously published MultiNODEs synthetic generation model [6]. This model comprises two branches: one for static variables and another for longitudinal data through a neural ODE (NODE) [7]. We transformed the longitudinal branch into a modular encoder with separate modules for clinical endpoints and biomarker data. Latent representations were concatenated and processed through a linear layer to learn the latent initial conditions (LIC) for the NODE, which also integrated dose information to capture drug effects. To reconstruct both ordinal/categorical and continuous longitudinal data, we introduced a singular decoder for the NODE’s dynamics. Ordinal variables were decoded using an ordinal regression model. Leveraging the MultiNODEs framework, we adopted a variational inference (VI) approach with a mixture of Gaussian priors for static representation and a Gaussian prior for each modular encoder. The loss function was tailored for each data type, employing mean squared error (MSE) for continuous variables and log-likelihood for ordinal variables [8,9]. Results Model reconstruction on the test dataset closely aligns with actual data, demonstrating reliable reconstructions of biomarker and endpoint time profiles across both training and test subsets. This performance is consistent across various drugs, underscoring the model’s robustness in handling diverse treatments. Its ability to extrapolate in time effectively, even with missing data or limited patient information, further confirms its capacity for reliable predictions. Preselecting the biomarkers improved the model performance. These results indicate that the model generalizes well to new test patients, exhibiting similar behavior across all drugs analyzed. While correlation values between the reconstruction and real test data exceed 0.65 for different variable groups. For clinical endpoints, accuracy and F1 scores reach 0.83 and 0.76 during induction and 0.82 and 0.78 in maintenance when leveraging full patient data. When relying solely on induction data for extrapolation, performance slightly declines to (0.75, 0.71) in induction and (0.77, 0.72) in maintenance. Conclusions The proposed model tackles challenges in clinical data, including integrating patient information from various clinical studies over time. By utilizing a modular multimodal architecture, the model effectively reconstructs data and predicts patient-specific clinical endpoints across multiple data types. This architecture may also facilitate the simulation of counterfactual scenarios, serving as a valuable tool for advancing research, guiding the development of new treatments, and enhancing future clinical trial designs. References 1.M’koma, A. E. (2013). Inflammatory bowel disease: an expanding global health problem. Clinical Medicine Insights: Gastroenterology, 6, CGast-S12731. 2.Rogers, K. V., Martin, S. W., Bhattacharya, I., Singh, R. S. P., & Nayak, S. (2021). A dynamic quantitative systems pharmacology model of inflammatory bowel disease: Part 1–model framework. Clinical and Translational Science, 14(1), 239-248. 3.Rogers, K. V., Martin, S. W., Bhattacharya, I., Singh, R. S. P., & Nayak, S. (2021). A dynamic quantitative systems pharmacology model of inflammatory bowel disease: Part 2–application to current therapies in Crohn’s disease. Clinical and Translational Science, 14(1), 249-259. 4.Stalgis, C., Deepak, P., Mehandru, S., & Colombel, J. F. (2021). Rational combination therapy to overcome the plateau of drug efficacy in inflammatory bowel disease. Gastroenterology, 161(2), 394-399. 5.Lundberg, S. M., & Lee, S.-I. (2017). A unified approach to interpreting model predictions. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1705.07874 6.Wendland, P., Birkenbihl, C., Gomez-Freixa, M., Sood, M., Kschischo, M., & Fröhlich, H. (2022). Generation of realistic synthetic data using multimodal neural ordinary differential equations. NPJ Digital Medicine, 5(1), 122. 7.Chen, R. T., Rubanova, Y., Bettencourt, J., & Duvenaud, D. K. (2018). Neural ordinary differential equations. Advances in neural information processing systems, 31. 8.McCullagh, P. (1980). Regression models for ordinal data. Journal of the Royal Statistical Society: Series B (Methodological), 42(2), 109-127. 9.Rosenthal, E. (2019). spacecutter: Ordinal regression in PyTorch [Computer software]. GitHub. https://github.com/EthanRosenthal/spacecutter Funding and Abbvie role disclosure: AbbVie funded this study and participated in the study design, research, analysis, data collection, interpretation of data, reviewing, and approval of the publication. All authors had access to relevant data and participated in the drafting, review, and approval of this publication. No honoraria or payments were made for authorship.
Reference: PAGE 33 (2025) Abstr 11717 [www.page-meeting.org/?abstract=11717]
Poster: Methodology – AI/Machine Learning