Perturbation modeling

Motivation¶

Perturbation screens measure how cells respond when a gene or a pathway is disrupted. Pooled CRISPR screens with single-cell readouts such as Perturb-seq and CROP-seq introduce one guide RNA per cell, sequence the guide together with the transcriptome, and thereby measure the effect of hundreds to thousands of genetic perturbations in a single experiment Dixit et al., 2016Adamson et al., 2016Datlinger et al., 2017. Genome-scale screens now cover every expressed gene Replogle et al., 2022, and similar designs measure the effects of chemical compounds at several doses Srivatsan et al., 2020.

The analysis of such screens follows a common path: assign guides to cells, identify the cells in which the perturbation worked, quantify and test the effect of each perturbation, find the genes it affects, and group perturbations with similar effects. We follow this path with pertpy Heumos et al., 2026 on two screens, a multimodal screen of immune checkpoint regulators and a genome-scale screen of essential genes.

Dataset¶

Papalexi et al. targeted 26 genes that regulate the immune checkpoint PD-L1 in THP-1 cells with 111 guide RNAs and three biological replicates Papalexi et al., 2021. Their ECCITE-seq protocol measures the transcriptome, four surface proteins including PD-L1, hashtags for sample multiplexing and the guide RNAs of every cell.

Assigning guides¶

Every downstream result depends on knowing which guide a cell received. The guide counts are noisy, since ambient guide RNA is captured in every droplet, so we assign a cell the guide with the most counts only if it has at least five of them Braunger & Velten, 2024.

Each guide is carried by a few hundred cells, and cells whose most abundant guide has fewer than five counts remain unassigned.

Our assignment agrees with the one published by the authors for every cell.

Identifying perturbed cells¶

CRISPR edits do not succeed in every cell, so many cells that received a targeting guide still express their target normally, and their inclusion dilutes the measured effects. Mixscape first removes technical and biological variation that is shared with the control cells of the same replicate by subtracting the mean expression of each cell’s nearest control neighbors. It then classifies every targeted cell as perturbed (KO) or not perturbed (NP) with a mixture model on this perturbation signature Papalexi et al., 2021.

Mixscape detects a perturbation in about three quarters of the cells targeting JAK2, SMAD4, STAT1, IFNGR1 and IFNGR2, whereas most cells targeting other genes are classified as not perturbed. For those genes, either the knockout failed or it does not change the transcriptome of THP-1 cells, which the classification cannot distinguish.

The surface protein measurements provide an independent check. Knocking out the interferon-γ receptor or its downstream signaling should reduce PD-L1, whose expression THP-1 cells induce in response to interferon-γ.

PD-L1 is lower in the knockout cells of IFNGR2, JAK2 and STAT1 than in the control cells and in the cells classified as not perturbed, which confirms the classification with an independent measurement.

Mixscale extends this approach with a continuous score of perturbation strength per cell, which retains cells with partial effects Jiang et al., 2025.

Differential expression per perturbation¶

To find the genes affected by a perturbation, we compare the perturbed cells with the control cells. Cells of the same replicate are not independent samples, so we sum the counts of each group per replicate and test the resulting pseudobulks with PyDESeq2, as described in the differential gene expression chapter Squair et al., 2021. We compare the cells in which IFNGR2 was knocked out with the control cells.

Knocking out IFNGR2 reduces interferon-γ response genes such as IRF1, SOCS1, CXCL9 and IL18BP, as expected from the loss of interferon-γ signaling. With three replicates per group, the dispersion estimates of PyDESeq2 rest on few degrees of freedom, so more replicates would make the results more robust.

Comparing perturbations at scale¶

Genome-scale screens raise questions about the perturbations themselves: which of them change the transcriptome at all, and which act alike. Replogle et al. knocked down every essential gene in K562 cells with CRISPR interference Replogle et al., 2022. We use at most 50 cells per perturbation and all control cells.

Effect size¶

The energy distance (E-distance) compares the distribution of perturbed cells with that of control cells in a low-dimensional space, relative to the spread within each group Peidli et al., 2024. It is zero if both groups are indistinguishable and grows with the strength of the perturbation.

The strongest effects come from knocking down GATA1, the master regulator of the erythroid program of K562 cells, and core components of protein folding, translation, splicing and the Mediator complex.

The E-test assesses whether a perturbation differs from the controls by comparing its E-distance with those obtained after permuting the labels. Since each test requires many permutations, we test a selection of the strongest and of random perturbations against a subsample of the controls.

All ten strong perturbations and seven of the ten random ones differ significantly from the controls after correcting for multiple testing. Many knockdowns of essential genes change the transcriptome only subtly, which the E-test detects even for small E-distances.

Perturbations with similar effects¶

Perturbations of genes in the same complex or pathway cause similar transcriptional changes. We summarize each of the strongest perturbations by the mean of its cells in the PCA space, cluster these profiles, and list the genes in each cluster.

The clusters recover known complexes and pathways, such as the large and small ribosomal subunits, the spliceosome, the RNA exosome, the proteasome, the Mediator, Integrator and PAF1 complexes, and aminoacyl-tRNA synthetases together with eIF2B, whose loss activates the integrated stress response Replogle et al., 2022. Grouping perturbations by their effects therefore assigns functions to poorly characterized genes that cluster with known complexes.

Predicting unseen perturbations¶

Many models aim to predict the transcriptional response to perturbations that were never measured, from variational autoencoders such as scGen Lotfollahi et al., 2019 to foundation models trained on millions of cells. Independent benchmarks found that none of these models consistently outperformed simple baselines such as the mean response of all training perturbations or linear models Ahlmann-Eltze et al., 2025Csendes et al., 2025Kernfeld et al., 2025Wei et al., 2026. Common evaluation metrics also reward models for reproducing the systematic difference between perturbed and control cells rather than perturbation-specific effects Viñas Torné et al., 2026. Predictions should therefore be compared with such baselines before they are used to prioritize experiments.

Quiz¶

Contributors¶

We gratefully acknowledge the contributions of:

Authors¶

• Lukas Heumos

• Mohammad Lotfollahi

Reviewers¶

• Yuge Ji