How to make hybrid models more reliable
In systems biology, hybrid modeling approaches such as Neural Ordinary Differential Equations (NODEs) and Universal Differential Equations (UDEs) are becoming increasingly powerful tools. By combining the interpretability of mechanistic models with the flexibility of machine learning, they allow researchers to describe complex biological systems more accurately.
However, these models are typically trained on limited biological data, raising a crucial question: how reliable are their predictions when exploring conditions beyond the training set?
In a new study, Stefano Giampiccolo, together with Luca Marchetti and Giovanni Iacca, introduces a novel framework to quantify out-of-distribution (OOD) uncertainty in NODEs and UDEs. This approach directly addresses one of the key challenges in model-driven biology — ensuring that predictions remain trustworthy even when data are scarce or extrapolation is required.
Quantifying OOD uncertainty enables researchers to:
- Identify when a model is venturing into unexplored regions of the parameter space
- Improve the robustness and interpretability of biological simulations
- Support safer and more reliable decision-making in computational biology
This contribution marks an important step toward making AI-assisted modeling more transparent, robust, and applicable to real-world biological systems.
Read the full preprint
