Assessing Reliability Beyond the Training Domain
A new study co-authored by COSBI researchers has been published in npj Systems Biology and Applications, addressing one of the key challenges in the integration of mechanistic modeling and machine learning for Systems Biology.
The paper, entitled “A novel approach to quantify out-of-distribution uncertainty in Neural and Universal Differential Equations”, is authored by Stefano Giampiccolo, Luca Marchetti and Giovanni Iacca.
Hybrid approaches that combine traditional mathematical modeling with machine learning are becoming increasingly important in computational biology, offering the possibility to integrate mechanistic knowledge with the flexibility of data-driven methods. However, these models often face a critical limitation: their performance can deteriorate when applied to experimental conditions that differ from those encountered during training.
This challenge is particularly relevant in Systems Biology, where predictive models are frequently expected to provide reliable insights under novel biological or experimental scenarios.
In this work, the authors propose a novel framework to quantify the reliability of Neural and Universal Differential Equations in out-of-distribution settings. The approach provides a principled methodology for evaluating the trustworthiness of model predictions beyond the training domain, contributing to the broader effort of making AI-enabled modeling more robust and interpretable.
The study represents an important contribution at the intersection of mechanistic modeling, machine learning and Systems Biology, highlighting the growing importance of uncertainty quantification for the development of reliable predictive models in biomedical research.
Congratulations to all authors of the study for this achievement.
Read the paper here
Explore COSBI publications and scientific outputs here
