Poster #93: Pascal Kataboh PhD

Identifiability and Experimental Design for Neural–Mechanistic Hybrid Models Using Physics-Informed Neural Networks in Colorectal Cancer

Pascal Kataboh, Amirkhan Aidarkhan

Neural-mechanistic hybrid models replace one term of a differential equation model with a neural network and retain the rest. Such hybrids are known to impair identifiability of the remaining mechanistic parameters, but the cause has not been established. We address this in a reduced model of the WNT–RA–HOX signaling network of colorectal cancer initiation, across four regimes from Normal to Severe APC Loss. We first characterize the fully mechanistic model's identifiability using Fisher information, profile likelihood, and Hamiltonian Monte Carlo posterior inference. All three agree on what time-course data can recover: in high-WNT regimes, β-catenin saturates and APC functionality becomes non-identifiable by any method. A learned term then degrades recovery in a specific way: the network absorbs the basal production of its host equation, under every constraint parameterization tested, including the monotone bounded form recommended in the literature. The cause is not architectural: the constraint that the learned term vanish at zero regulator is imposed where the data never reach, so it holds without constraining the fit. The remedy is experimental. One knockdown condition driving the regulator to zero, at a fixed condition count, reduces the basal-parameter error from 2.2% to 0.0%; an information-matched control that moves regulators away from zero produces no improvement. A training-free reachability analysis identifies the least costly knockdown for each learnable term, and a prospective test confirms its predictions. Identifiability of hybrid models is thus governed by whether the experimental design visits the constrained region.