Estimating the division kernel of a size-structured population

Van Hà HoangESAIM 2017 · DOI

TL;DR

This paper introduces a novel adaptive kernel estimator for the "division kernel" (how cell contents are split during division) in a size-structured cell population model. The method achieves optimal convergence rates, even with random observation times and population sizes, offering a robust way to understand cell division dynamics.

What Problem Does This Paper Address?

Understanding how cells divide and distribute their internal components (like toxicity or genetic material) is crucial for modeling population dynamics, aging, and disease. This paper focuses on estimating the "division kernel," which describes the probability distribution of how a parent cell's "size" (e.g., toxicity) is split between its two daughter cells. Existing methods often struggle with data-driven bandwidth selection and achieving optimal rates, especially when dealing with the inherent randomness of biological processes like cell division. The authors aim to develop an estimator that is adaptive to the unknown smoothness of the division kernel and achieves optimal statistical performance.

Methods

The core of the paper is the development of an adaptive kernel estimator for the division kernel, denoted by $h$. The authors first model the cell population using a stochastic differential equation driven by a Poisson point measure, capturing both the growth of cell size (toxicity) and the division process.

The estimation strategy involves:

  • Kernel Estimation: A standard approach where a kernel function $K$ is used to smooth the observed data. The estimator $\hat{h}_l$ depends on a bandwidth parameter $l$.
  • Data-Driven Bandwidth Selection: The crucial innovation is the use of the Goldenshluger-Lepski (GL) method for selecting the bandwidth $l$. This method is designed to adapt to the unknown smoothness of the function being estimated, aiming to minimize the mean integrated squared error (MISE). It involves comparing estimators with different bandwidths to estimate the bias-variance trade-off.
  • Oracle Inequality: The authors derive an oracle inequality, which bounds the MISE of their estimator by comparing it to an "oracle" estimator that would have access to the true, unknown division kernel. This inequality is key to proving the optimality of their method.
  • Stochastic Modeling: The cell population dynamics are described by a stochastic differential equation, and properties of the number of cells and their ages are analyzed.

Datasets

Not discussed in the paper. The authors focus on theoretical development and simulations using synthetic data generated from specified division kernel distributions.

Key Insights and Ideas

  • Adaptive Estimation is Crucial: The paper highlights that an adaptive estimator, which can automatically adjust to the smoothness of the division kernel, is essential for achieving optimal statistical performance.
  • GL Method's Power: The Goldenshluger-Lepski method proves effective for data-driven bandwidth selection in this complex, size-structured population model, even with random observation counts.
  • Exponential Convergence Rates: The estimator achieves exponential convergence rates, which are often superior to traditional polynomial rates in nonparametric statistics, especially for smooth functions.
  • Optimality: The derived upper bounds (from the estimator) and lower bounds (from information theory) match, demonstrating that the proposed estimator achieves the optimal rate of convergence in the minimax sense.
  • Impact of Initial Population Size: The convergence rate is shown to depend on the initial number of cells ($N_0$), with a slightly slower rate for $N_0=1$ due to logarithmic factors.

Key Results

  • Adaptive Kernel Estimator: A new adaptive kernel estimator $\hat{h}$ for the division kernel $h$ is proposed.
  • Oracle Inequality: An oracle inequality is derived for the estimator, bounding its Mean Integrated Squared Error (MISE).
  • Exponential Convergence Rates: The estimator achieves exponential convergence rates of the form $e^{-c T^{2\beta/(2\beta+1)}}$ for smooth kernels, where $\beta$ is the smoothness of the kernel and $T$ is the observation time.
  • Optimality: The estimator's rate of convergence is shown to be optimal in the minimax sense for Hölder classes of functions.
  • Numerical Simulations: Simulations with Beta distributions (representing symmetric and asymmetric divisions) demonstrate that the GL method performs well, often comparable to or better than cross-validation and rule-of-thumb methods, and achieves good reconstructions of the true division kernels.

Limitations and Future Work

  • Complete Data Assumption: The current work assumes complete observation of the division tree. Future work could explore scenarios with partial observations or censored data.
  • Constant Rates: The model assumes constant division and growth rates. Extending the model to time-varying or size-dependent rates would be a natural next step.
  • Asymptotic Behavior: While the paper provides theoretical guarantees for large $T$, practical performance for smaller $T$ might be further investigated.
  • Approximation by PDEs: The authors mention that the stochastic model can be approximated by growth-fragmentation partial differential equations, and this problem is a work in progress. This suggests a direction for connecting microscopic cell dynamics to macroscopic population models.

Summary updated

Abstract

We consider a size-structured model describing a population of cells proliferating by division. Each cell contain a quantity of toxicity which grows linearly according to a constant growth rate α . At division, the cells divide at a constant rate R and share their content between the two daughter cells into fractions Γ and 1 − Γ where Γ has a symmetric density h on [ 0,1 ], since the daughter cells are exchangeable. We describe the cell population by a random measure and observe the cells on the time interval [ 0 ,T ] with fixed T . We address here the problem of estimating the division kernel h (or fragmentation kernel) when the division tree is completely observed. An adaptive estimator of h is constructed based on a kernel function K with a fully data-driven bandwidth selection method. We obtain an oracle inequality and an exponential convergence rate, for which optimality is considered.
BibTeX
@inproceedings{hoang2017estimating,
  author = {Van Hà Hoang},
  title = {{Estimating the division kernel of a size-structured population}},
  booktitle = {ESAIM},
  year = {2017},
  doi = {10.1051/ps/2017011},
}
Figure