Does the random nature of cell-virus interactions during in vitro infections affect TCID$_{50}$ measurements and parameter estimation by mathematical models?

TL;DR

This study investigates how the randomness in cell-virus interactions during in vitro infections influences TCID50 measurements and parameter estimation, proposing a new likelihood-based method and comparing stochastic and deterministic models for better accuracy.

Contribution

The paper introduces a novel likelihood approach for parameter estimation from TCID50 data, incorporating stochastic effects and using infectious virus units for improved biological relevance.

Findings

The new likelihood handles measurements beyond detection limits effectively.

Expressing virus in IV units yields more biologically meaningful parameters.

Stochastic models better capture variability in infection assays.

Abstract

Endpoint dilution (TCID50) assays cannot count the number of infectious virions (IVs), and instead are limited to counting the number of Specific INfections caused by the sample (SIN). The latter depends not only on whether virions are infectious, but also on the cells and the experimental conditions under which they interact. These interactions are random and controlled by parameters such as the rates at which IVs lose infectivity, enter cells, or fail to replicate following cell entry. Here, stochastic TCID50 assays are simulated to determine how the random number of infected wells relates to the parameters and the number of IVs in a sample. We introduce a new parameter estimation method based on the likelihood of observing a given TCID50 assay outcome given the model-predicted number of IVs in the sample. We then successively evaluate how parameter estimates are affected by the use of: 1) the new likelihood function vs the typical assumption of Gaussian-distributed measurement errors; 2) IV vs SIN units to express virus in the model; and 3) a stochastic vs an ODE model to simulate the course of a virus infection. Unlike previous methods, the new likelihood correctly handles measurements beyond the detection limits, and results in non-Gaussian distributions. Expressing virus using IV units makes it possible to impose physical constraints (e.g. one IV cannot infect more than one cell), and yields more biologically useful parameters (e.g. mutation emergence likelihood depends on the number of IVs, not SIN, produced). Using a stochastic rather than an ODE model we show that the variability observed between replicate in vitro virus infections is consistent with the level of stochasticity introduced by the TCID50 assay, which can be reduced through better assay design. The framework introduced herein offers several important improvements over current methods and should be widely adopted.