Speaker
Description
Parameter estimation for ODE models of biological systems is commonly based on continuous or semi-quantitative measurements. Many relevant observations are however inherently binary — such as gene essentiality classifications from CRISPR knockout screens. Treating such data as continuous introduces statistical misspecification, biasing parameter estimates and undermining uncertainty quantification.
We propose a hierarchical gradient-based framework for integrating binary observables into parameter estimation of mechanistic ODE models. The estimation problem is decomposed into an outer optimisation over mechanistic parameters and an inner optimisation over observable-level parameters. Via a KKT-based envelope argument, we show that exact outer gradients are obtained without tracking the dependence of inner on outer parameters — a result that holds even under box constraints. The fully likelihood-based formulation enables standard uncertainty quantification of mechanistic parameters.
We demonstrate the method on joint inference from drug-response viability measurements and binary gene-essentiality data across 233 cancer cell lines using a pan-cancer ERK-RAS-AKT signalling model, and investigate the effect of appropriate binary likelihood specification on parameter identifiability and predictive performance. The method is implemented in pyPESTO.
