Speaker
Description
Neural ordinary differential equation frameworks, such as Biologically-Informed Neural Networks (BINNs), have shown strong potential for learning mechanistic laws from sparse biological data. However, most existing approaches assume homoscedastic Gaussian noise, overlooking biologically meaningful variability arising from cell-to-cell heterogeneity and experimental measurement processes. In this work, we extend the BINN framework by introducing a learnable noise model that enables the identification of additive, multiplicative, or mixed noise structures directly from data. Using population growth systems motivated by cell proliferation assays, we demonstrate that the approach accurately recovers underlying noise types, captures state-dependent variability, and produces well-calibrated uncertainty estimates. These results highlight the importance of modelling structured noise for interpreting biological dynamics and provide a general framework for integrating data-driven uncertainty into neural ODE models, with applications to developmental and cellular systems.
