This project focuses on developing advanced AI methodologies to model and understand dynamic biological systems, particularly plant-pathogen interactions. The primary goal is to leverage Physics-Informed Neural Networks (PINNs) and Variational Autoencoders (VAEs) to improve the integration and interpretation of omics data in these systems. Specifically, the internship will involve building an initial model based on Ordinary Differential Equations (ODEs) to describe the gene expression dynamics of plants, which will be integrated into the existing HIVE platform [1]. HIVE (Horizontal Integration analysis using VAE), developed by Dr. Bottini’s team, is a framework for analyzing transcriptomics data from unpaired experiments. The internship will begin with designing an ODE system to model the temporal changes in gene expression under external stresses such as pathogen attacks or environmental conditions [3]. This model will simulate the evolution of gene counts as longitudinal data, enabling the characterization and inference of key interaction parameters. By identifying molecular signatures and their activation timings during stress responses, the project seeks to uncover the defense mechanisms that plants deploy during infection. Once the ODE model is established, it will be extended into a PINN framework [2] to estimate parameters efficiently. This involves constructing a loss function that incorporates ODE constraints, ensuring the model adheres to biological principles. The neural network will be trained on synthetic data to fit gene expression dynamics and infer the interaction parameters. A well-adapted neural network architecture will be developed to balance predictive accuracy with interpretability, capturing the key dynamics of the biological system. The intern will focus on integrating the ODE model into the HIVE platform, facilitating the identification of critical data features and capturing plant-pathogen interaction dynamics in a reduced-dimensional space. While full integration and application to real-world data may extend beyond the internship, the project will provide a solid foundation in modeling dynamic biological systems using state-of-the-art AI techniques.
[1] G. Calia, S. Marguerit, A. P. Zotta Mota, M. Vidal, M. Seynabou-Fall, H. T. Nguyen, A. Bhat, H. Schuler, C. Gwizdek, A. C. M. Brasileiro, et al. Disentangling plant response to biotic and abiotic stress using hive, a novel tool to perform unpaired multi-transcriptomics integration. bioRxiv, pages 2024–03, 2024.
[2] P. J. Hossie, B. Laroche, T. Malou, L. Perrin, T. Saigre, and L. Sala. Simulating interactions in microbial communities through physics informed neural networks: towards interaction estimation. 2024.
[3] X. Liu, D. Igarashi, R. A. Hillmer, T. Stoddard, Y. Lu, K. Tsuda, C. L. Myers, and F. Katagiri. Decomposition of dynamic transcriptomic responses during effector-triggered immunity reveals conserved responses in two distinct plant cell populations. Plant Communications, 2024. 2