Quantifying Uncertainty in Hybrid Models for Systems Biology
At CIBB 2025, held at Politecnico di Milano from September 10 to 12, Stefano Giampiccolo, PhD student at COSBI in collaboration with the University of Trento, presented his research on “Uncertainty quantification in out-of-distribution scenarios for data-driven and hybrid neural-network/mechanistic modeling in computational Systems Biology.”
Recent advances such as Neural Ordinary Differential Equations (Neural ODEs) and Universal Differential Equations (UDEs) have expanded the possibilities of combining data-driven and mechanistic approaches in systems biology. However, these models often struggle in out-of-distribution (OOD) scenarios, where predictions must be made beyond the conditions represented in the training data. This makes uncertainty quantification (UQ) a critical component for ensuring the reliability and robustness of hybrid models.
Stefano’s work explores deep ensembles, a widely used method for UQ, and introduces a new strategy called Maximizing OOD Disagreement (MOD). In this approach, ensemble members are trained not only to fit the data but also to maximize their diversity in OOD regions. This helps prevent the overconfidence that often undermines model predictions when data are scarce or novel.
Through in-silico test cases, Stefano demonstrated the potential of MOD to improve the trustworthiness of Neural ODEs and UDEs under challenging predictive conditions. His results contribute to developing more reliable hybrid models in computational biology, ultimately supporting applications where accurate and interpretable predictions are essential.
Congratulation to Stefano for taking COSBI’s research to the CIBB 2025 conference.
