Presentation Information
[SS06-10]Modeling Cellular Dynamics with Omics nueral ODE model
*Han Wen1,3,2 (1. Beijing AI for Science Insititute (China), 2. DP Technology (China), 3. Peking University (China))
Keywords:
Purterbation,Proteomics,Neural ODE,Drug Discovery,artificial intelligence
Understanding how cellular systems respond over time to perturbations is key to decoding biological dynamics and informing therapeutic strategies. In this talk, we introduce GraphOmicsODE, a neural ODE framework that integrates prior biological networks—such as STRING—with time-series omics data to model gene expression trajectories under diverse perturbations. It handles non-uniform sampling and sparse data, enabling accurate, generalizable, and stable predictions across cell types, time points, and conditions. Applications include drug response profiling and transferring perturbation effects from cell lines to patients.
We also present ProteinTalks, a large-scale dynamic model trained on over 38 million cancer proteomics measurements. It captures protein network dynamics to predict drug effects, resistance mechanisms, and biomarkers. As a precursor to more generalizable dynamic models, ProteinTalks complements GraphOmicsODE by showcasing the value of time-resolved modeling in the proteome space. Together, these models highlight the potential of neural ODEs in learning interpretable, multi-omic cellular dynamics.
We also present ProteinTalks, a large-scale dynamic model trained on over 38 million cancer proteomics measurements. It captures protein network dynamics to predict drug effects, resistance mechanisms, and biomarkers. As a precursor to more generalizable dynamic models, ProteinTalks complements GraphOmicsODE by showcasing the value of time-resolved modeling in the proteome space. Together, these models highlight the potential of neural ODEs in learning interpretable, multi-omic cellular dynamics.