11Cell Engineering

Up to now you have designed one molecule at a time: a protein that folds (Chapter 8), a binder that sticks (Chapter 9), a small molecule that drugs a target (Chapter 10). Cell engineering asks a bigger question. Given a living cell, which set of deliberate interventions pushes it into a state you want: a T cell that resists exhaustion, a stem cell nudged toward a chosen lineage, a tumor cell coaxed back toward something normal. The molecule is the lever; the cell is the machine you are trying to reprogram. This chapter follows the same spine as the rest of Part II. First the problem of choosing perturbations, then the models trained on Perturb-seq to predict their effects, and finally the sobering distance between predicting a cell in a dish and helping a patient.

11.1The problem: programming cells

Start with two definitions. A cell's state is, for our purposes, its transcriptome: the vector of which genes are transcribed and how much, exactly the readout that single-cell RNA sequencing (scRNA-seq) gives you (Chapter 4). A perturbation is a deliberate change you make to the cell: knock a gene out (CRISPR cuts it so it stops working), turn one on or down (CRISPR activation or interference, CRISPRa/CRISPRi, which steer a gene's expression without cutting), or wire several of these into a small synthetic circuit. Programming a cell means finding the perturbation, or combination of perturbations, that moves the cell from its starting state A to a desired state B.

Two things make this hard, and neither is solved by a bigger dataset alone. The first is combinatorics. With roughly 20,000 human genes there are about 200 million single-and-double combinations and astronomically more triples, so you cannot test them all, especially not in a patient-relevant cell. The second is that genes interact. Knock out two genes and the effect is often not the sum of the two single knockouts, a phenomenon geneticists call epistasis. That non-additivity is exactly the interesting biology, and exactly what makes the search space refuse to factorize into independent choices.

This reframes target discovery (Chapter 6). There the question was which gene is causally tied to disease. Here you assume you have candidate levers and ask which interventions, applied together, produce the cell state you want. The genetics primer (Chapter 5) supplies the vocabulary of gene function and interaction that the state space is built on.

11.2Models that attempt it

The training data is Perturb-seq: a pooled CRISPR screen with a single-cell RNA-seq readout, so that every cell carries a label for which gene was perturbed and a full transcriptome of the result (Replogle et al., 2022). One cell is one (perturbation, response) pair, and a single experiment yields millions of them (Chapter 4). A perturbation-response model learns from these pairs to predict the expression shift a held-out perturbation would cause.

GEARS attacks the unseen-perturbation problem with a graph (Roohani et al., 2024). It places every gene in a gene-gene network built from known pathways and co-expression, represents a perturbation by its neighborhood in that graph, and so predicts the effect of a gene it never saw perturbed by borrowing from its neighbors that it did. It handles single and two-gene perturbations, and its authors report gains on combinations where one or both genes were never perturbed alone.

scGPT-perturb takes a different route: fine-tune a single-cell foundation model. Models like scGPT (Cui et al., 2024) and Geneformer (Theodoris et al., 2023) are transformers pretrained on tens of millions of cells to learn a general representation of cell state, then adapted to the perturbation task. These general-purpose single-cell models are appendix-tier for us; see the single-cell foundation model appendix for how they are pretrained and where they help.

The most ambitious recent effort is Arc Institute's State, a "virtual cell" model (Adduri et al., 2025). It has two modules. State Embedding (SE) maps a raw transcriptome into a smooth latent space that dampens technical noise and aligns cell types. State Transition (ST) is a transformer that predicts how a cell's embedding shifts under a given perturbation, and a decoder maps the shifted embedding back to gene expression. It was trained on one of the largest compendia assembled, roughly 170 million observational and over 100 million perturbational cells across dozens of cell lines (the Tahoe-100M dataset among them), and it targets genetic, chemical, and cytokine perturbations with an explicit goal of transferring across cell types. Arc also launched a Virtual Cell Challenge in 2025 that held out an unseen cell type (H1 human embryonic stem cells) with 300 CRISPRi perturbations, a purpose-built benchmark that drew over a thousand teams and made cross-context generalization the headline metric.

11.3What they do well and what is still hard

Give the field its due. These models recover, for genuinely hard perturbations, which genes respond and in which direction, and they do it across thousands of candidate perturbations fast enough to prioritize wet-lab follow-up. On its own benchmarks State improves the discrimination of perturbation effects and the detection of truly differentially expressed genes over prior models. As a hypothesis generator that ranks what to try next, this is real value.

Now the hard part, and it is genuinely hard. A careful 2025 benchmark tested scGPT, scFoundation, and GEARS against deliberately simple baselines: predicting the mean of the training perturbations, or an additive model for pairs (Ahlmann-Eltze et al., 2025). For perturbations of genes not seen in training, the deep models did not beat predicting the training mean. For two-gene combinations, they did not beat the additive baseline. This is the load-bearing caveat of the whole area: much of the impressive-looking correlation comes from the fact that most perturbations barely move most genes, so a constant baseline already scores high. A more nuanced follow-up finds that deep models do win on the subset of "resistant" perturbations, recovering which genes flip and in which direction, but not the magnitude of the response.

Two more gaps sit between a good benchmark number and a therapy. The first is distribution shift. Perturb-seq is run mostly in a handful of immortalized cell lines, and models trained on them transfer poorly to primary cells, other labs' batches, or a patient's tissue, because batch and cell-type effects can swamp the perturbation signal. That is precisely why the Virtual Cell Challenge scores held-out cell types. The second is the dish-to-patient gap, the same translation gap that haunts every chapter in this part. A predicted transcriptome shift is not a cellular phenotype, a phenotype is not a clinical benefit, and none of these models say anything about delivery, getting the edit into the right cells inside a body. Cell engineering also echoes protein design (Chapter 9): in-silico success rates run far ahead of validated wet-lab wins, so treat rankings as leads, not conclusions.