Boosting Functional Response Models for Location, Scale and Shape with\n an Application to Bacterial Competition

Almond Stöcker, Sarah Brockhaus, Sophia Anna Schaffer, Benedikt von Bronk, Madeleine Opitz, Sonja Greven

Abstract

We extend Generalized Additive Models for Location, Scale, and Shape (GAMLSS) to regression with functional response. This allows us to simultaneously model point-wise mean curves, variances and other distributional parameters of the response in dependence of various scalar and functional covariate effects. In addition, the scope of distributions is extended beyond exponential families. The model is fitted via gradient boosting, which offers inherent model selection and is shown to be suitable for both complex model structures and highly auto-correlated response curves. This enables us to analyze bacterial growth in Escherichia coli in a complex interaction scenario, fruitfully extending usual growth models.

Introduction

In many scientific domains, data is increasingly collected as functions or curves over time or space. Traditional functional regression models typically focus on estimating the conditional mean of these functional responses. However, focusing solely on the mean can obscure critical insights, particularly in biological systems where variability and extreme behaviors—such as extinction events—are as significant as the average trend.

This paper introduces a framework for Functional Generalized Additive Models for Location, Scale, and Shape (Functional GAMLSS). By extending the GAMLSS framework to functional data, researchers can simultaneously model multiple distributional parameters (e.g., location, scale, and shape) as functions of various covariates. This approach provides a comprehensive distributional perspective on functional responses, moving beyond simple mean-only models.

Methodology

The core of the proposed framework is the extension of GAMLSS to functional responses. Instead of predicting a single scalar value, the model estimates a vector of time-varying distributional parameters $\vartheta(t) = (\mu(t), \sigma(t), \nu(t), \dots)$. Each parameter is modeled through an additive predictor that can incorporate a wide array of effect types.

Overview of possible effect types supported by the Functional GAMLSS framework.

As shown in the table above, the framework supports scalar covariates, functional covariates, and grouping variables, allowing for linear effects, smooth interactions, and concurrent or historical functional effects.

To estimate these complex models, the authors employ component-wise gradient boosting. This optimization technique is particularly advantageous for functional data for several reasons:

  1. Automatic Model Selection: It identifies the most relevant covariates and effect types, which is crucial when dealing with high-dimensional potential predictors.
  2. Regularization: Through "early stopping," boosting prevents overfitting, a common issue in functional data where observations along a curve are highly auto-correlated.
  3. Flexibility: It allows for the use of non-standard distributions, such as zero-adapted distributions, which are essential for modeling biological data with zero-inflation (e.g., colony extinction).

Application: Bacterial Competition

The utility of Functional GAMLSS is demonstrated through an analysis of competition between two Escherichia coli strains: a toxin-producing "C-strain" and a toxin-sensitive "S-strain." The study observed 334 sites over 48 hours under varying levels of antibiotic stress (Mitomycin C), which triggers toxin production.

A key methodological choice was the use of a zero-adapted Gamma distribution. This allowed the model to simultaneously capture the continuous growth of surviving colonies and the discrete probability of extinction. Interestingly, the model revealed that the rate of change (derivative) of the C-strain's growth was a more potent predictor of its impact on the S-strain than its absolute size, suggesting that the metabolic activity associated with growth drives toxin production.

Results and Discussion

The analysis yielded several key biological and statistical insights:

  • Temporal Interaction Phases: The model identified two distinct phases in the competition. An initial 10-hour period dominated by toxin-mediated interference was followed by a phase characterized by spatial competition, where physical obstruction became the primary interaction mechanism.
  • Distributional Sensitivity to Stress: While antibiotic stress had a relatively minor impact on the average growth of the S-strain, it significantly increased both the variability of growth and the probability of extinction. These critical effects would have been invisible to standard mean-regression models.
  • Robustness to Auto-correlation: Simulation studies confirmed that the boosting-based approach (implemented in the FDboost R package) outperformed traditional penalized likelihood methods, particularly in the presence of highly auto-correlated functional data.

Conclusion

Functional GAMLSS represents a significant advancement in functional data analysis, offering a flexible and robust framework for modeling entire distributions of functional responses. By leveraging gradient boosting, the method provides inherent model selection and regularization, making it suitable for complex biological and environmental datasets. Future work will explore the inclusion of functional random effects to explicitly model correlation structures and the extension to multivariate functional responses.

Summary updated

Abstract

We extend Generalized Additive Models for Location, Scale, and Shape (GAMLSS)\nto regression with functional response. This allows us to simultaneously model\npoint-wise mean curves, variances and other distributional parameters of the\nresponse in dependence of various scalar and functional covariate effects. In\naddition, the scope of distributions is extended beyond exponential families.\nThe model is fitted via gradient boosting, which offers inherent model\nselection and is shown to be suitable for both complex model structures and\nhighly auto-correlated response curves. This enables us to analyze bacterial\ngrowth in \\textit{Escherichia coli} in a complex interaction scenario,\nfruitfully extending usual growth models.\n
BibTeX
@inproceedings{stocker2020boosting,
  author = {Almond Stöcker and Sarah Brockhaus and Sophia Anna Schaffer and Benedikt von Bronk and Madeleine Opitz and Sonja Greven},
  title = {{Boosting Functional Response Models for Location, Scale and Shape with\textbackslash\{\}n an Application to Bacterial Competition}},
  booktitle = {Open},
  year = {2020},
  doi = {10.1177/1471082x20917586},
}
Figure