Modeling the inter-individual variability of single-stemmed plant development

Antonin Della Noce, Véronique Letort, Amandine Hansart, Charlotte Baey, G Viaud, Sébastien Barot, Jean‐Christophe Lata, Xavier Raynaud, Paul-Henry Cournède, Jacques Gignoux2016 · DOI

TL;DR

This paper introduces a robust statistical framework for modeling the significant inter-individual variability observed in plant development, specifically focusing on the number of phytomers (plant modules) over time. By employing hierarchical mixed-effects models and an iterative Expectation-Maximization (EM) algorithm, the authors provide a method to capture these individual differences, leading to more accurate predictions of plant growth.

What Problem Does This Paper Address?

Plant development, particularly the rate at which new modules (phytomers) are formed along the main stem, is often modeled using thermal time (cumulative degree-days). However, real-world observations consistently reveal substantial variation in this development rate among individual plants, even under identical environmental conditions. This inter-individual variability is not easily explained by simple environmental factors and, if ignored, can lead to inaccurate model predictions and interpretations. The paper aims to address this by developing a statistical methodology that explicitly accounts for and models this heterogeneity within a plant population.

Methods

The core of the paper lies in its proposed methodological framework for modeling inter-individual variability using hierarchical or mixed-effects models. The key idea is to treat some regression parameters, which describe individual plant responses, as random variables that are identically distributed within the population.

The authors detail an iterative Expectation-Maximization (EM) algorithm for parameter estimation. This algorithm is particularly suited for models with latent variables (like individual parameters) and incomplete data. They present two candidate models for organogenesis:

  • Model 1 (Linear Mixed-Effects Model): This model assumes that the relationship between thermal time and phytomer number is piecewise linear, and importantly, the parameters describing the slopes of these linear segments are treated as random effects. This linearity with respect to individual parameters allows for analytical solutions in the E-step of the EM algorithm.
  • Model 2 (Non-linear Mixed-Effects Model): This model relaxes the linearity assumption for certain parameters, specifically the thermal times associated with developmental transitions (initiation and change points). This introduces non-linearity, making analytical solutions intractable. For this model, they explore two approaches for the E-step:
  • Markov Chain Monte Carlo (MCMC) with Metropolis-Hastings: This is a simulation-based approach to estimate the posterior distribution of individual parameters.
  • Laplace Approximation: This method approximates the posterior distribution with a Gaussian distribution, offering a more computationally efficient alternative when the distribution is unimodal and sharp.

The paper also discusses the Maximization step of the EM algorithm, detailing how to update population parameters and fixed effects. For the linear model, some updates are analytic, while for the non-linear model, numerical optimization methods are employed.

Datasets

The study utilizes experimental data collected on young Acacia erioloba plants grown in the CEREEP-Ecotron IleDeFrance greenhouse. This dataset comprises weekly measurements of phytomer number and daily temperature recordings, covering the first five months of development for 122 plants. The data is used to analyze organogenesis dynamics and to illustrate the proposed modeling framework.

Key Insights and Ideas

  • Inter-individual variability is a crucial characteristic: The paper strongly emphasizes that ignoring the inherent variability in plant development leads to incomplete and potentially misleading models.
  • Hierarchical models offer a powerful solution: By treating individual parameters as random variables drawn from a population distribution, hierarchical models can effectively capture and quantify this variability.
  • EM algorithm is well-suited for this problem: The EM algorithm provides a principled way to estimate parameters in models with latent variables, making it ideal for mixed-effects models.
  • Linearity simplifies the E-step: For models linear in individual parameters, the E-step of the EM algorithm can be solved analytically, leading to efficient parameter estimation.
  • Non-linearity requires advanced techniques: When models become non-linear in individual parameters, MCMC or Laplace approximation become necessary for the E-step, offering trade-offs between accuracy and computational cost.
  • Model 1 provides a good baseline: The linear mixed-effects model (Model 1) demonstrated good adequacy with the observed data, suggesting that for this specific aspect of Acacia development, a simpler linear representation might suffice.
  • Model 2 offers more generality but limited improvement: The non-linear model (Model 2) is more flexible but yielded only minor improvements in this case, suggesting that the thermal times of transition might be well-approximated by fixed effects in this specific dataset.

Key Results

  • The linear mixed-effects model (Model 1) was successfully identified using the EM algorithm. The estimated parameters, including population means and variances for phyllochrons, provided a good fit to the observed data.
  • Figure 2 shows that the predicted population distribution of observations from Model 1 aligns well with the data, with confidence intervals capturing the observed variability.
  • Figure 3 illustrates that the empirical distribution of individual means is close to a Gaussian distribution, supporting the assumptions of the mixed-effects model.
  • Figure 4 demonstrates the individual-level predictions from Model 1, showing how the model can capture the growth trajectory of specific plants within the population's variability.
  • The non-linear mixed-effects model (Model 2), when explored, showed that the parameters related to thermal times of initiation and change points had very low variances, suggesting they could be treated as fixed effects in this context. This led to only marginal improvements over the linear model.

Limitations and Future Work

  • Non-smoothness in Model 2: The paper acknowledges that Model 2's non-smoothness (due to indicator functions) can make numerical optimization of derivatives challenging. They considered a C∞ approximation to address this.
  • Initialization for Model 2: The authors note that the initialization of the EM algorithm for Model 2, using parameters from the linear model, might have constrained the random effects for thermal times to be very small. They suggest that a new initialization procedure might be needed to explore different configurations.
  • Future Work: The authors propose extending this methodology to model other aspects of plant development, such as branching and reiteration, and to integrate functional parameters. The ultimate goal is to develop a full population model for Acacia erioloba that accounts for both organogenesis and functioning, especially considering interactions between them. They also suggest that the linearized version of a non-linear model can provide valuable initialization values for EM algorithms applied to the full non-linear model.

Summary updated

Abstract

International audience
BibTeX
@misc{noce2016modeling,
  author = {Antonin Della Noce and Véronique Letort and Amandine Hansart and Charlotte Baey and G Viaud and Sébastien Barot and Jean‐Christophe Lata and Xavier Raynaud and Paul-Henry Cournède and Jacques Gignoux},
  title = {{Modeling the inter-individual variability of single-stemmed plant development}},
  year = {2016},
  doi = {10.1109/fspma.2016.7818287},
}
Figure