In classical mathematical biology, researchers often rely on the concept of the "average cell." Whether modeling a signaling pathway with ordinary differential equations (ODEs) or analyzing bulk transcriptomic data, the underlying assumption is that a population can be represented by a single, idealized set of parameters.
However, biological populations are inherently heterogeneous. Genetically identical cells exposed to the exact same stimulus frequently exhibit radically different behaviors. Treating a population as a uniform group doesn't just introduce noise—it fundamentally misrepresents the underlying biology. To capture the true dynamics of living systems, a growing body of research shifts the modeling paradigm away from scalar averages and toward distributional modeling, treating entire cell populations as continuous probability densities.
The Flaw of the Average Cell
To understand why mean-based modeling fails, consider the phenomenon of fractional killing in cancer therapeutics, where a drug dose eliminates only a specific percentage of a tumor population. If you model this system using a single "average" cell, a deterministic model must conclude that every cell becomes partially dead—a physical and biological impossibility. As Steffen Waldherr and Frank Allgöwer demonstrate in Estimation of biochemical network parameter distributions in cell populations, capturing the distribution of parameters across cells is the only way to mechanistically explain why identical cells choose different fates like life or death.
Averages also act as a mask that hides critical, low-abundance subpopulations. For instance, in bacterial infections, bulk multi-omic measurements aggregate data across millions of organisms, completely obscuring specialized, slow-growing, or dormant sub-populations that drive antibiotic persistence (Max Conway and Lu Yu in Transcriptome and proteome analysis of Salmonella enterica serovar Typhimurium systemic infection of wild type and immune-deficient mice). Similarly, within a single histologically defined tissue type, individual cells can exhibit massive variations in metabolic RNA turnover that remain entirely invisible to static transcriptomic snapshots (Kun Yin et al. in Dyna-vivo-seq unveils cellular RNA dynamics during acute kidney injury via in vivo metabolic RNA labeling-based scRNA-seq).
Mapping Populations to Probability Densities
Instead of fitting a single parameter vector $\theta$ to an experimental dataset, distributional modeling treats $\theta$ as a random variable governed by a probability density function $p(\theta)$. The goal then becomes reconstructing this density function from experimental snapshots, such as flow cytometry or single-cell sequencing data.
Mechanistic Mixture Models
When a population contains distinct, hidden cell types (e.g., fast-dividing vs. quiescent stem cells), the overall population density can be modeled as a mixture of multiple sub-populations. In ODE Constrained Mixture Modelling: A Method for Unraveling Subpopulation Structures and Dynamics, Jan Hasenauer et al. introduced a framework that embeds deterministic biochemical models directly inside statistical mixture distributions. This allows researchers to use static "snapshot" data to resolve nearly indistinguishable subpopulations and estimate their internal, time-varying reaction rates.
Convex Optimization and Maximum Entropy
Inferring a continuous parameter distribution from noisy, high-throughput data is a severely underdetermined inverse problem. To solve this without making rigid, artificial assumptions about the shape of the distribution, Steffen Waldherr et al. developed a framework combining partial differential equations (PDEs) with convex optimization in Density-based modeling and identification of biochemical networks in cell populations. By pairing this with a Maximum Entropy approach, the framework finds the most unbiased parameter distribution that remains perfectly consistent with the observed population data.
Turning Variability into a Statistical Signature
If parameters vary from cell to cell, that variation itself contains a structural pattern. Rather than treating cell-to-cell diversity as experimental noise, researchers can leverage it as a rich source of information to uncover hidden biological mechanisms.
In Identification of a molecular system that regulates growth cone membrane potential during growth cone guidance, Tatsuya Yamada demonstrates that treating individual model parameters as random variables transforms cellular variability into a statistical signature. By analyzing how these parameters covary across a population, the author successfully decoded interconnected signaling pathways in developing neurons that were completely obscured by traditional averaging.
This concept is formalized in high-content screening by frameworks like the one proposed by Didier Surdez et al. in A Probabilistic Model for Cell Population Phenotyping Using HCS Data. By utilizing copula-based joint distributions, their model captures the intricate, non-linear dependencies between different cellular features across a population, providing a far more sensitive readout for drug effects than simple well-averaging.
Traditional Modeling:
[Experimental Population] ───(Average)───> [Single Scalar Value] ───> [Deterministic Model]
│
(Inaccurate Prediction)
Distributional Modeling:
[Experimental Population] ───(Density Estimation)───> [Probability Density p(θ)] ───> [Mechanistic/Generative Model]
│
(True Population Dynamics)
Overcoming the Mean-Seeking Trap with Generative Models
Traditional machine learning and regression techniques suffer from an inherent "mean-seeking" behavior. When trained using standard loss functions like Mean Squared Error (MSE), a model minimizes its loss by outputting the expected value (the mean) of the target distribution. In genomics and spatial biology, this flattens natural stochasticity, erasing multi-modal states and variance.
To preserve biological heterogeneity, modern frameworks deploy deep generative models designed to sample directly from complex, high-dimensional probability distributions:
- Conditional Diffusion Models: In Diffusion Generative Modeling for Spatially Resolved Gene Expression Inference from Histology Images, Sichen Zhu et al. utilize diffusion models to predict spatial gene expression profiles directly from H&E histology images. Because the model learns to reverse a noise process rather than minimizing a point-wise distance, it successfully captures the localized, stochastic variations in gene expression that standard regression methods smooth away.
- Distributional Flow Matching: For predicting how cell populations respond to external perturbations like genetic edits or drugs, Chenglei Yu et al. developed scDFM in scDFM: Distributional Flow Matching Model for Robust Single-Cell Perturbation Prediction. By modeling the entire distribution shift over time, the framework accurately captures multi-modal state transitions and variance changes triggered by combinatorial drug treatments.
- Generative Adversarial Networks (GANs): To bridge the gap between highly variable experimental data and rigid biophysical equations, Jaimit Parikh and Timothy Rumbell introduced a GAN-based framework in Generative adversarial networks for construction of virtual populations of mechanistic models: simulations to study Omecamtiv Mecarbil action. The GAN learns to generate "virtual populations" of mechanistic heart cell models whose collective parameter distributions perfectly mirror real-world physiological diversity.
Why It Works: Heterogeneity as a Functional Driver
The ultimate justification for distributional modeling is that heterogeneity is not merely an ornamental detail—it is a fundamental driver of macroscopic biological behavior.
First, heterogeneity acts as a critical stabilizing mechanism. As Monica Schliemann demonstrates in Heterogeneity reduces sensitivity of cell death for TNF-Stimuli, cell-to-cell variability serves as a biological buffer. A heterogeneous population exhibits significantly lower sensitivity to molecular perturbations than a nominal "average" cell, protecting the tissue as a whole from erratic, localized fluctuations in signaling proteins.
Second, population variance completely alters system-level dynamics. In continuous cell cultures, Jorge FERNANDEZ-DE-COSSIO-DIAZ et al. showed in Maximum entropy and population heterogeneity in continuous cell cultures that the "dynamical landscape" of a culture—including its multi-stable regimes and metabolic transitions—is highly sensitive to the degree of internal variation. Homogeneous models completely miss critical phenomena like unexpected byproduct accumulation and shifts in overall population stability.
Finally, accounting for the entire distribution is essential for effective therapeutic design. Traditional drug discovery looks for compounds that optimize an average response, which frequently leaves substantial sub-populations of cells entirely dysfunctional or vulnerable to toxic side effects. By simulating virtual populations of diseased neurons, Sushmita L. Allam et al. demonstrated in Neuronal Population Models Reveal Specific Linear Conductance Controllers Sufficient to Rescue Preclinical Disease Phenotypes that multi-target therapies can be engineered to rescue an entire heterogeneous population. These distribution-aware interventions successfully restore functional consistency across a diverse cellular landscape, vastly outperforming single-target drugs designed for the "average" cell.
Key Foundational Work
Density-based modeling and identification of biochemical networks in cell populations (2010) — contributed 1 concept elements to this direction.
Estimation of biochemical network parameter distributions in cell populations (2009) — contributed 1 concept elements to this direction.