⚛️ biophysics

Integrating cell-state and enhancer constraints discriminates models of PU.1 regulation during hematopoietic differentiation

By integrating quantitative cell-state measurements with enhancer constraints, this study demonstrates that PU.1 regulation during hematopoietic differentiation cannot be explained by autoregulation and cell-cycle feedback alone, but requires an additional cell-state-dependent input mediated by a RUNX1-extended regulatory architecture.

Original authors: Igoshin, O. A., Zhang, Y.

Published 2026-10-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Igoshin, O. A., Zhang, Y.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Inside the body, a single master switch controls the fate of blood cells. This switch is a protein called PU.1, and its job is to decide whether a stem cell will become a macrophage that fights infection or a B-cell that produces antibodies. The amount of this protein present in a cell acts like a volume knob: low levels guide the cell toward one path, while high levels push it toward another. For years, scientists believed they understood how this volume knob was tuned. They thought the cell simply adjusted its own internal machinery, creating a self-reinforcing loop where more protein led to even more protein, eventually locking the cell into a specific identity. This idea seemed to explain how different blood cell types could coexist stably. However, when researchers tried to test this theory against real measurements of how fast cells divide and how much protein they actually make, the old explanation began to fall apart.

A team of researchers at Rice University set out to fix this broken model by treating the cell not as a theoretical concept, but as a physical system with strict rules. They gathered existing data on three specific things: how much PU.1 protein was in different blood cell types, how fast the cells were making new protein, and how long it took for each cell type to divide. They found that the old theory, which relied on a simple feedback loop between protein levels and cell division speed, could not match the reality of the data. The theory predicted that cells with more protein would always divide more slowly, but the measurements showed a different pattern. In some cell types, the cells with the highest protein levels were actually dividing faster than those with lower levels. Furthermore, the old model could only explain one type of immature cell, whereas the data clearly showed two distinct groups of immature cells with different characteristics.

To solve this puzzle, the scientists built a series of increasingly complex computer simulations, testing each one against the hard numbers they had collected. First, they tried to save the old idea by simply adjusting the speed at which cells diluted their protein during division, using the actual measured division times. This did not work; even with the correct division speeds, a single self-regulating loop could not produce the two different immature cell states observed in nature. The researchers realized that the cell must have an extra input, a second factor helping to turn the PU.1 volume knob up or down depending on the cell's identity. They tested a known partner protein, C/EBPα, which interacts with PU.1. While this pair could create the right number of stable states, the model still failed to place the immature cells in the correct positions relative to the mature ones. The simulation kept predicting that the immature cells would have too little protein compared to what was actually measured.

The solution appeared when the team added a third player to the mix: a protein called RUNX1. This protein binds to the same genetic region that controls PU.1 production. When the researchers included RUNX1 in their model, allowing it to influence the production of PU.1 alongside the other factors, the simulation suddenly matched all the data perfectly. The model could now reproduce the exact protein levels, synthesis rates, and division times for all the different cell types, including the two distinct groups of immature cells. The researchers found that this new architecture required a very sharp, switch-like response from the C/EBPα protein, suggesting that the cell uses a highly sensitive mechanism to make these critical decisions.

This work does more than just correct a model; it reveals that the regulation of blood cell identity is more complex than a simple self-loop. It shows that the cell relies on a combination of its own growth rate, a partnership with C/EBPα, and a crucial boost from RUNX1 to set the correct levels of the master protein. The study suggests that the two groups of immature cells previously observed are likely real, distinct populations that exist because of this specific regulatory architecture. While the researchers used computer simulations to reach these conclusions, the model is tightly constrained by real experimental measurements, making the findings a robust guide for future biological experiments. The next step for scientists will be to test these specific predictions in the lab, confirming whether RUNX1 and the sharp switch-like behavior of C/EBPα are indeed the keys that unlock the correct cell fate.

Problem
Hematopoietic differentiation involves the establishment of distinct, stable molecular states associated with specific lineages, such as B cells, macrophages, and progenitor cells. The transcription factor PU.1 is a central regulator of this process, with its abundance varying quantitatively across these states. A previous influential study proposed an autoregulation-and-growth-feedback (AR&GF) model, suggesting that PU.1 autoregulation coupled with a feedback loop where PU.1 slows cell division (reducing protein dilution) generates three stable steady states corresponding to B cells, progenitors, and macrophages. However, this model was developed primarily to explain qualitative trends and had not been quantitatively evaluated against the specific measurements of PU.1 abundance, synthesis rates, and cell-cycle durations that motivated it. The central challenge addressed in this paper is to determine whether the AR&GF mechanism can quantitatively reproduce these reported cell-state measurements and, if not, to identify the minimal regulatory features required to do so while satisfying independent enhancer-reporter constraints.

Methodology
The authors employed a model-discrimination strategy using a sequence of progressively more mechanistic dynamical systems models. They treated reported measurements of PU.1-GFP abundance, PU.1 synthesis rates, and cell-cycle lengths as quantitative constraints on steady states and removal rates. Additionally, they used enhancer-reporter perturbation data to constrain the PU.1 production function.

The modeling workflow proceeded as follows:

Evaluation of the AR&GF Model: The authors first tested the original AR&GF model against the data. They found the model failed to reproduce the two distinct progenitor groups (pro-low and pro-high) and the non-monotonic relationship between PU.1 levels and cell-cycle length.
Autoregulation with Discrete Growth Rates (AR&DGR): They replaced the universal growth-feedback function with cell-type-specific dilution rates derived directly from measured cell-cycle lengths. They tested if PU.1 autoregulation alone, with shared production parameters but distinct removal rates, could fit the data.
Cell-Type-Specific Autoregulation: To test if an additional production-side input was necessary, they introduced a cell-state-dependent scaling factor ( $X$ ) to the PU.1 production function.
Positive Feedback with C/EBP $\alpha$ : They replaced the abstract factor $X$ with a specific biological candidate, C/EBP $\alpha$ , creating a mutual positive feedback loop (PU.1 $\leftrightarrow$ C/EBP $\alpha$ ). This model was constrained by both cell-state data and PU.1 reporter activity data from enhancer mutations.
RUNX1-Extended Architecture: Finally, they incorporated RUNX1 as an additional input to the PU.1 enhancer, based on evidence that RUNX1 binding sites significantly affect PU.1 expression. This "RUNX1 model" was fitted to the expanded dataset including RUNX1 perturbation data.

Throughout, the models were implemented as ordinary differential equations (ODEs) and fitted using particle-swarm optimization to minimize the error between model predictions and the joint ranges of experimental data.

Key Contributions and Results

Quantitative Inconsistency of the AR&GF Model: The study demonstrates that the original AR&GF model is quantitatively inconsistent with the motivating data. It predicts only one progenitor attractor rather than the two observed groups (pro-low and pro-high) and imposes a monotonic relationship between PU.1 levels and cell-cycle length that contradicts the data (where pro-high cells have high PU.1 but short cell cycles).
Insufficiency of Dilution Differences: Replacing the growth-feedback assumption with measured, cell-type-specific dilution rates (AR&DGR model) does not resolve the discrepancy. A single shared PU.1 autoregulatory production function cannot intersect the distinct removal functions of B cells, pro-low, and pro-high cells at the required abundance and synthesis rate ranges.
Necessity of Cell-State-Dependent Production: The results indicate that PU.1 production must depend on an additional cell-state-associated input beyond autoregulation and dilution. A model with a cell-type-specific scaling factor successfully reproduces the data, suggesting a second regulatory factor.
Failure of C/EBP $\alpha$ Feedback Alone: When C/EBP $\alpha$ is modeled as the second factor in a closed-loop feedback system constrained by enhancer-reporter data, the model fails to position the pro-low attractor correctly. While an open-loop fit works, the closed-loop dynamics with a Hill coefficient of 2 (based on known binding sites) are too gradual. Even increasing cooperativity leads to bistability where neither state matches the measured pro-low range.
Success of the RUNX1-Extended Model: The inclusion of RUNX1 as an additional input to the PU.1 enhancer allows the model to satisfy all constraints. With a high effective cooperativity (Hill coefficient $\approx$ 10) for PU.1-dependent C/EBP $\alpha$ activation, the RUNX1-extended model jointly reproduces the cell-state measurements (abundance and synthesis rates) and the enhancer-reporter perturbation data. Nullcline analysis shows that RUNX1 shifts the PU.1 nullcline upward, correctly positioning the pro-low attractor within the measured range.

Significance
The paper claims that integrating dynamical systems constraints with cis-regulatory (enhancer) constraints is necessary to distinguish between models that merely generate multistability and those that are quantitatively consistent with biological reality. The study supports a regulatory architecture for PU.1 that combines cell-type-specific dilution, PU.1-C/EBP $\alpha$ feedback, and an additional enhancer-level input represented by RUNX1.

The authors state that this framework generates specific, experimentally testable hypotheses:

The pro-low system may exhibit bistability, suggesting the existence of two progenitor subpopulations with different PU.1 abundances but similar growth rates.
The high effective cooperativity required for C/EBP $\alpha$ activation suggests mechanisms beyond simple binding site occupancy, such as C/EBP $\alpha$ autoregulation or co-condensation of transcription factors.
The model predicts that basal promoter-driven PU.1 expression is higher in B cells than in macrophages, despite lower overall production rates in B cells, implying that cell-type differences in production are shaped by enhancer-level regulation rather than basal promoter activity alone.

The authors modestly note that their conclusions are specific to the integration of these particular datasets under their chosen assumptions and do not imply a unique biological requirement for C/EBP $\alpha$ and RUNX1, acknowledging limitations such as the use of RNA-seq as a proxy for protein production and the aggregation of progenitor data that may obscure distinct subpopulations.

More like this