From Graphs to States: Interpretable Forecasting of Biological Systems Dynamics
Event details
| Date | 27.08.2025 |
| Hour | 10:00 › 12:00 |
| Speaker | Jérémy Baffou |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Charlotte Bunne
Thesis advisor: Prof. Pascal Frossard
Thesis coadvisor: Dr. Dorina Thanou
Co-examiner: Prof. Pierre Vandergheynst
Abstract
Recent advancements in sequencing and imaging technologies have brought an unprecedented view of biological systems and their dynamics. However, existing analytical methods often inadequately capture the inherent temporal and spatial complexities of such data. Frequently, these datasets are modeled as time series of hand-crafted features, obscuring spatial structure and limiting the interpretability of disease evolution over time. Effective modeling of such systems necessitates methods that are agnostic to data modality and explicitly designed to handle sparsity and asynchronous observation patterns inherent in real-world clinical and biological time series. This work reviews recent methodological developments applicable to dynamic biological systems, evaluating their capabilities and limitations, and highlighting integrative approaches that leverage their complementary strengths. We first introduce Dynamic Graph Neural Networks (GNNs), emphasizing their capability to adapt static graph models to temporal data via hierarchical state representations. Next, we discuss Deep State Space Models (SSMs), highlighting structured latent spaces that efficiently capture long-range dependencies using dual convolutional and recurrent frameworks. Finally, we examine generative modeling approaches for disease progression using medical imaging, focusing on flow-based latent space trajectories that capture continuous pathological changes. In addition to reviewing these methods, we discuss their integration into flexible, interpretable frameworks that capture latent biological dynamics and support modeling of patient trajectories, enabling improved interpretation and clinical decision-making.
Selected papers
Exam president: Prof. Charlotte Bunne
Thesis advisor: Prof. Pascal Frossard
Thesis coadvisor: Dr. Dorina Thanou
Co-examiner: Prof. Pierre Vandergheynst
Abstract
Recent advancements in sequencing and imaging technologies have brought an unprecedented view of biological systems and their dynamics. However, existing analytical methods often inadequately capture the inherent temporal and spatial complexities of such data. Frequently, these datasets are modeled as time series of hand-crafted features, obscuring spatial structure and limiting the interpretability of disease evolution over time. Effective modeling of such systems necessitates methods that are agnostic to data modality and explicitly designed to handle sparsity and asynchronous observation patterns inherent in real-world clinical and biological time series. This work reviews recent methodological developments applicable to dynamic biological systems, evaluating their capabilities and limitations, and highlighting integrative approaches that leverage their complementary strengths. We first introduce Dynamic Graph Neural Networks (GNNs), emphasizing their capability to adapt static graph models to temporal data via hierarchical state representations. Next, we discuss Deep State Space Models (SSMs), highlighting structured latent spaces that efficiently capture long-range dependencies using dual convolutional and recurrent frameworks. Finally, we examine generative modeling approaches for disease progression using medical imaging, focusing on flow-based latent space trajectories that capture continuous pathological changes. In addition to reviewing these methods, we discuss their integration into flexible, interpretable frameworks that capture latent biological dynamics and support modeling of patient trajectories, enabling improved interpretation and clinical decision-making.
Selected papers
- Albert Gu · Karan Goel · Christopher Re, "Efficiently Modeling Long Sequences with Structured State Spaces" (https://iclr.cc/virtual/2022/oral/6960)
- Jiaxuan You, Tianyu Du, and Jure Leskovec, "ROLAND: Graph Learning Framework for Dynamic Graphs" (https://dl.acm.org/doi/abs/10.1145/3534678.3539300)
- C. Liu et al., "ImageFlowNet: Forecasting Multiscale Image-Level Trajectories of Disease Progression with Irregularly-Sampled Longitudinal Medical Images" (https://ieeexplore.ieee.org/abstract/document/10890535)
Practical information
- General public
- Free
Contact
- edic@epfl.ch