Algorithm to find new identifiable reparametrizations of parametric rational ODE models

TL;DR

This paper introduces a new algorithm to find globally identifiable reparametrizations of parametric rational ODE models, broadening the class of models that can be transformed into identifiable forms, especially for linear models.

Contribution

The paper develops an algorithm that extends existing methods to find globally identifiable reparametrizations for a wider range of models, including explicit formulas for certain linear models.

Findings

A globally identifiable reparametrization always exists for linear models.

Explicit reparametrization formulas are provided for specific linear compartmental models.

The method is demonstrated on multiple examples with detailed analysis available online.

Abstract

Structural identifiability concerns the question of which unknown parameters of a model can be recovered from (perfect) input-output data. If all of the parameters of a model can be recovered from data, the model is said to be identifiable. However, in many models, there are parameters that can take on an infinite number of values but yield the same input-output data. In this case, those parameters and the model are called unidentifiable. The question is then what to do with an unidentifiable model. One can try to add more input-output data or decrease the number of unknown parameters, if experimentally feasible, or try to find a reparametrization to make the model identifiable. In this paper, we take the latter approach. While existing approaches to find identifiable reparametrizations were limited to scaling reparametrizations or were not guaranteed to find a globally identifiable reparametrization even if it exists, we significantly broaden the class of models for which we can find a globally identifiable model with the same input-output behavior as the original one. We also prove that, for linear models, a globally identifiable reparametrization always exists and show that, for a certain class of linear compartmental models, with and without inputs, an explicit reparametrization formula exists. We illustrate our method on several examples and provide detailed analysis in supplementary material on github.