Distributional Modeling of Single-Cell Perturbation Responses

This direction is split into other directions: cell-to-cell variability. Its last version stays readable and citable here.

Version 1 of 172 papers82 conceptsfrozen 2026-10-04

In traditional biology, we often describe the effect of a drug or a genetic mutation by its "average" effect. We treat a population of cells as a single unit and measure how the mean expression of a gene shifts. However, as Gefei Wang and colleagues argue in their work on scLAMBDA, the "average cell" is a biological myth.

In reality, a single drug might kill 90% of a tumor while leaving a small, resistant sub-population untouched. A mean-seeking model would see this as a successful "average" reduction in tumor markers, completely missing the rare cells that will eventually cause a relapse. To solve this, a new research direction has emerged: moving away from predicting point estimates and toward modeling the entire probability distribution of cellular states.

The Missing Data Problem

The fundamental challenge in single-cell perturbation modeling is a "missing data" problem. To know exactly how a specific cell responds to a treatment, you would need to measure it, apply the treatment, and measure it again. But single-cell sequencing is destructive; the act of measuring the cell destroys it.

As Charlotte Bunne explains in the CellOT framework, we are forced to compare two different populations: one group of cells that was never treated (the control) and a different group that was (the perturbed). The goal of the researcher is to learn the "coupling" or the "flow" that maps the distribution of healthy cells to the distribution of treated cells.

From Means to Manifolds

Traditional regression models suffer from "mean-seeking" behavior. When faced with uncertainty, they predict the average, which Sichen Zhu and Peng Qiu note in their work on Stem effectively "averages out" the natural stochasticity of biology.

To capture the full diversity of life, researchers are employing three primary mathematical mechanisms:

  1. Optimal Transport (OT): Frameworks like CAROT and CMonge use OT to find the most efficient way to "push" the mass of the control distribution into the perturbed distribution. This allows models to predict how specific sub-clusters—rather than just the whole population—will shift.
  2. Generative Flow Matching: Instead of a static map, models like CellFlow and scDFM learn the continuous "velocity field" that moves cells from one state to another. This is particularly useful for combinatorial perturbations, where a researcher might want to simulate the effect of two drugs that have never been tested together in a lab.
  3. Functional Diffusion: Xinyu Yuan and Jian Tang introduced PerturbDiff, which treats the entire cell distribution as a single random variable in a high-dimensional space. By performing diffusion on the distribution itself, the model can capture complex, multi-modal shifts that simpler models miss.

Why Heterogeneity is a Signal, Not Noise

In many fields, variance is treated as measurement noise. In this research direction, variance is the signal. Jonathan G. Hedley and colleagues demonstrate in Modeling Gene Expression Distributional Shifts that perturbations often increase the variance of a population or create bimodal states (where some cells respond and others don't).

This distributional perspective extends beyond gene expression into physical form. Reed Naidoo and Chris Bakal developed FORM, a 3D generative framework that predicts the distribution of cell shapes. They argue that because identical cells react differently to the same drug, a deterministic model that predicts a single "average" shape is fundamentally unbiological.

Generalizing to the Unseen

The ultimate goal of these models is "zero-shot" prediction: the ability to predict how a cell will respond to a gene knockout or a drug that has never been seen in training data.

By training on massive datasets—such as the 100 million cells used to train STATE—these models learn the underlying "grammar" of cellular response. They recognize that a cell's response is a function of its starting state (its "stemness," its position in the cell cycle, or its epigenetic landscape). As C. Y. He shows with MORPH, by modeling the full statistical distribution, we can finally begin to navigate the nearly infinite space of genetic and chemical interactions that define modern medicine.

Papers in this version

Showing 60 of 72 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_distributional_modeling_of_single_cell_p_eac946f1,
  title = {Distributional Modeling of Single-Cell Perturbation Responses},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributional-modeling-of-single-cell-perturbation-responses/rdr_eac946f1d97c4541aa480a07912b7c87},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_distributional_modeling_of_single_cell_p_eac946f1_v1,
  title = {Distributional Modeling of Single-Cell Perturbation Responses},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/distributional-modeling-of-single-cell-perturbation-responses/rdr_eac946f1d97c4541aa480a07912b7c87/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}