#39 · Systems biology

Identifiability and optimiser benchmarks for parameter estimation in dynamic models

The common thread is reliable recovery of model parameters or operators from data in dynamic life-science models. The BINN study varies architecture and optimisation choices, including expressivity, learning rate and loss weighting, to recover advection-diffusion-reaction operators from sparse data. The SEIR preprint addresses structure instead, reparameterising the model for global identifiability to improve estimation convergence and numerical stability. The biorxiv benchmark compares local, global and hybrid optimisers across 990 method-problem pairs for parameter estimation in dynamic life-science models. Together the studies treat identifiability, optimiser selection and network design as linked determinants of estimation reliability. What is new is the systematic, quantitative treatment of these choices, with the SEIR reparameterisation as the only item here that explicitly targets global identifiability.