TL;DR
This paper introduces a probabilistic model for continuous labeling assays that accounts for biological variability in cell cycle lengths, a factor previously ignored by deterministic models. This new model provides more accurate and unbiased estimates of cell proliferation rates, even in the presence of significant biological noise, and offers a way to disentangle variability at different biological levels.
What Problem Does This Paper Address?
Accurately measuring how quickly cells divide (proliferation rate) is fundamental for understanding tissue development, maintenance, and regeneration. Continuous labeling assays, where cells are exposed to a label that gets incorporated into newly synthesized DNA, are a common in vivo method for this. However, existing models for analyzing these assays are largely deterministic. They assume all cells have the same cell cycle length and that this length is consistent across different tissue samples. This simplification ignores the inherent biological variability at both the individual cell level and the sample level, which can lead to significant biases (up to 40%) in the estimated proliferation rates. The paper aims to address this by developing a more realistic probabilistic model that captures these sources of variability.
Methods
The core of the paper is the development of a probabilistic model for continuous labeling assays. Instead of deterministic equations, they use random variables to represent key biological parameters:
- Cell Cycle Length ($T_{i,j}$): This is modeled hierarchically. The sample-level mean cell cycle length ($T_i$) is drawn from a log-normal distribution, and then individual cell cycle lengths ($T_{i,j}$) within that sample are also drawn from a log-normal distribution, dependent on the sample's mean. This captures both sample-to-sample variation and within-sample variation.
- Proliferation State ($g_{i,j}$): This is modeled using a Bernoulli distribution, representing the probability that a cell is actively proliferating (the "growth fraction").
- Cell Age ($a_{i,j}$): The initial age of cells at the start of the experiment is modeled using a uniform distribution.
- S-phase Fraction ($f_{i,j}$): This is assumed to be constant across all cells and samples.
The model then derives an analytical expression for the likelihood of observing a certain number of labeled cells in a sample, considering the probabilistic nature of cell cycle lengths and the binomial distribution of labeled cells within a sample. This likelihood function is then used for parameter estimation via maximum likelihood.
Datasets
The paper uses simulated data for its parameter recovery study. They generate synthetic experimental data based on their probabilistic model to test the accuracy and bias of their method compared to the traditional Nowakowski model. No real-world experimental datasets are explicitly mentioned as being used for training or evaluation, beyond the simulated data.
Key Insights and Ideas
- Biological Variability Matters: The study strongly demonstrates that ignoring biological variability in cell cycle lengths (both between cells and between samples) leads to systematically biased estimates of proliferation rates.
- Probabilistic Approach Unlocks Information: By embracing stochasticity, the model can not only provide unbiased estimates but also disentangle the contributions of different sources of variability (cell-level vs. sample-level).
- Deterministic Models Can Be Misleading: The parameter recovery study shows that even with a relatively small number of measurements, the deterministic Nowakowski model can produce significantly inaccurate results, with errors up to 40%.
- Hierarchical Modeling is Key: The hierarchical modeling of cell cycle lengths effectively captures the nested structure of biological variability, where cells within a sample are more similar to each other than to cells in other samples.
Key Results
- The probabilistic model yields unbiased estimates of cell cycle length, growth fraction, and variability parameters ($\sigma$ and $\sigma_T$).
- The deterministic Nowakowski model can produce biased estimates of cell cycle length, with errors up to 40% observed in parameter recovery studies.
- The probabilistic model's accuracy is less sensitive to the number of samples per time point compared to the Nowakowski model, which shows increased bias with fewer measurements.
- The model accurately estimates error intervals, providing a more reliable assessment of parameter uncertainty.
- The authors have developed an open-source Python tool and a web service for easy parameter estimation using their probabilistic model.
Limitations and Future Work
- Asymmetric Cell Division Assumption: The analytical likelihood derivation is currently limited to the case of asymmetric cell division, which is well-justified for homeostatic stem cell populations. The authors note that extending this to symmetric cell divisions (relevant for expanding populations) would be straightforward in principle but finding a tractable likelihood might be challenging. They suggest approximate Bayesian computation or other likelihood-free methods as potential solutions.
- Cell Cycle Length Distribution: The model assumes log-normal distributions for cell cycle lengths. While this is supported by experimental data, other distributions (e.g., gamma, normal) could also be explored. The presence of sub-populations might lead to multimodal distributions, which is not explicitly handled.
- Cell Phase Stretching Model: The model uses a specific "cell phase stretching" model for variability. The influence of other models of cell phase length variability on the estimates remains to be elucidated.
Future work could involve extending the model to symmetric cell divisions, exploring alternative distributions for cell cycle lengths, and investigating the impact of different cell phase variability models.
Summary updated