Abstract
Molecular regulation of cell fate decisions underlies health and disease. To identify molecules that are active or regulated during a decision, and not before or after, the decision time point is crucial. However, cell fate markers are usually delayed and the time of decision therefore unknown. Fortunately, dividing cells induce temporal correlations in their progeny, which allow for retrospective inference of the decision time point. We present a computational method to infer decision time points from correlated marker signals in genealogies and apply it to differentiating hematopoietic stem cells. We find that myeloid lineage decisions happen generations before lineage marker onsets. Inferred decision time points are in agreement with data from colony assay experiments. The levels of the myeloid transcription factor PU.1 do not change during, but long after the predicted lineage decision event, indicating that the PU.1/GATA1 toggle switch paradigm cannot explain the initiation of early myeloid lineage choice.
Introduction: The "Black Box" of Cell Fate
In developmental biology, identifying when a cell decides its future identity is notoriously difficult. While we can observe "lineage markers"—specific proteins that turn on once a cell has committed—these signals are often significantly delayed. This creates a "black box" period: if the exact moment of decision is unknown, researchers cannot accurately identify the molecular signals that triggered it.
The authors address this by leveraging the fact that cells divide. If a decision happens in an ancestor cell, its descendants will show synchronized (correlated) marker onsets. Conversely, if decisions happen independently in daughter cells, those signals will be uncorrelated. The challenge lies in creating a mathematical framework to map these correlations back to a specific "hidden" time point within a complex family tree.
Methodology: Inferring Hidden Decisions
The researchers introduced a computational method based on Maximum Likelihood Estimation (MLE) applied to tree-structured data. They represented every possible differentiation scenario within a genealogy as a "hidden tree." Using a probabilistic graphical model, they calculated the likelihood of each scenario, accounting for cell lifetimes and the stochastic nature of gene expression that causes the delay between decision and marker detection.
To handle the computational complexity of analyzing trees up to 12 generations deep, the team employed dynamic programming to efficiently sum over all possible differentiation paths. The model was validated using a mathematical branching process to see if predicted decision times could accurately forecast the results of independent "colony assays."
Results: Decisions are "Old News"
The model revealed that the majority (74%) of lineage decisions in hematopoietic stem and progenitor cells (HSPCs) happen in the first or second generation of the experiment, even though markers do not appear until much later. There is a staggering delay between the decision and the marker onset—averaging 78 hours for the granulocyte-macrophage (GM) lineage and 54 hours for the megakaryocyte-erythroid (MegE) lineage. This represents a lag of roughly 4–6 cell divisions.
The model's predictive accuracy was confirmed by matching the frequencies of pure versus mixed colonies in independent experiments. For example, it correctly predicted that approximately 60% of colonies would be mixed GMMegE. Interestingly, the probability of differentiating over time was almost identical for both GM and MegE lineages, suggesting a common clock or mechanism governs the timing of these choices.
Discussion: Refuting the Toggle Switch Paradigm
For decades, the PU.1/GATA1 "toggle switch"—a mechanism where two proteins inhibit each other until one dominates—was the textbook model for blood cell choice. However, the authors found that PU.1 levels remain stable during the actual inferred decision time ($p = 0.25$). In contrast, simulated toggle switch models show drastic changes in PU.1 production at the decision point ($p < 10^{-16}$).
These results suggest that the PU.1/GATA1 switch is not the "trigger" that initiates the decision. Instead, it likely acts as an "effector" that locks in the decision long after it has been made by an unidentified mechanism. This "retrospective inference" approach could be applied to other biological processes, such as identifying the exact moment a healthy cell turns cancerous or when a somatic cell begins reprogramming into a pluripotent state.
Summary updated