Xaira’s X-Cell: learning what happens when you turn a gene down

Turn down one gene, predict the consequences

Reduce the expression of one gene in a cell. What happens to the other genes, and eventually to the cell’s function? This is the prediction problem behind X-Cell, Xaira’s first virtual-cell model. Genetic perturbation prediction offers a starting point for drug discovery because many therapeutics reduce protein or gene activity. Predicting responses to drugs and chemicals is a proposed extension of this first model.

Cells can share a genome yet behave differently because they express genes at different levels. Expression helps determine both cell type and cell state. A biological pathway is a chain of interactions among genes and proteins; changing part of that chain can alter the cell’s appearance or function. Predicting a perturbation therefore means predicting a response across an interacting system.

In the Latent Space conversation, Xaira frames drug discovery as three difficult matching problems: the right target, the right molecule, and the right patients. Its platforms divide those responsibilities: - Protein design: develop molecules against previously undruggable targets, building on work originating in co-founder David Baker’s University of Washington group. - Virtual-cell modeling: predict how interventions affect cellular biology and help identify therapeutic targets. - Patient representations: connect molecular information to treatment response. This platform was beginning development at the time of recording.

The platforms are intended to work together. A virtual-cell model might identify a target that protein engineering can address, while patient representations could connect cellular predictions to clinical outcomes. Xaira aims to improve drug-development success rates and halve development cycles; these remain goals. Patient modeling also presents a difficult data problem: molecular samples must be matched to clinical annotations that capture treatment response. Collecting samples across disease severities alone does not solve it.

Cell descriptions leave causal relationships unresolved

Virtual cell remains a broad, unsettled term. Earlier efforts fitted differential equations to observations of biological pathways; one guest argues that a few predefined equations could not capture the complexity. The newer approach learns responses from data. Its longer-term ambition includes predicting how cell states develop over time and change in spatial context, with foundation models providing an initial representation of cells.

[scGPT](https://www.nature.com/articles/s41592-024-02201-0), developed in a University of Toronto lab, illustrates the foundation-model approach: one model supports multiple single-cell tasks, including integration across experimental batches and molecular measurements. Batch effects are differences introduced by laboratories, equipment, or measurement conditions. Batch integration removes those differences while preserving biological signals such as cell-type distinctions. Observational datasets such as CELLxGENE support this descriptive work.

Predicting an intervention asks more of the data. Suppose genes A, B, and C rise and fall together. Several explanations fit: - A regulates B and C: changes in A drive the shared response. - B regulates A and C: B drives it instead. - Separate regulation produces correlated measurements: A regulates B, while something else regulates C. Those networks can predict different outcomes when a gene is suppressed. This ambiguity motivates training causal predictions on intervention data.

The guests report that models trained on descriptive biological data had struggled to beat linear models on perturbation tasks. Their response is to invest in suitable experiments. Accumulated structures in the Protein Data Bank and genomic sequence collections helped enable progress in protein modeling; the team argues that virtual-cell and patient-response models need comparable foundations of high-quality data.

Perturb-seq links each intervention to its response

Perturb-seq combines pooled CRISPR perturbations—targeted interventions in gene expression—with single-cell RNA sequencing, which measures expression cell by cell. Genes have two roles in the dataset: intervention targets and measured features. Each known perturbation is associated with expression changes across other genes in the same cell. The companion article calls Xaira’s perturbation dataset X-Atlas.

The workflow makes a large pooled experiment readable: 1. Target a gene. A guide RNA directs Cas9 equipped with a silencing effector to the gene’s promoter, the region preceding transcription, reducing the gene’s expression. 2. Pool many interventions. Individual cells receive single-gene perturbations, while the experiment collectively covers roughly 20,000 gene targets across many cells. 3. Read the intervention identity. Detecting the guide RNA in each cell identifies which gene was targeted. 4. Measure the response. Single-cell RNA sequencing links that intervention to a broad expression profile from the same cell.

The guide supplies both an instruction and an identity. Researchers can run many perturbations together and recover which intervention belongs to which response. The team says pooling eliminates plate-to-plate differences associated with separate experiments, removing unwanted variation that the model would otherwise have to learn around.

Scaling creates another problem: the cells keep changing while they wait. The reported 25 million cells remained after stringent quality filtering; producing that dataset can require harvesting tens or hundreds of millions. A fresh-cell protocol designed for an hour or two can stretch into a 14-hour processing day, long enough for cellular stress signals to appear.

Xaira introduced chemical fixation to preserve cell state early, while maintaining compatibility with downstream molecular measurements. Processing could then happen at different times without living cells enduring the entire operation. The team reports that this preserved data quality and reduced susceptibility to batch effects: laboratory engineering changed the quality of the learning signal.

Expand biological context, then transfer beyond easy experiments

The first datasets used two immortalized cell lines, cells maintained through repeated growth in the laboratory. Their ability to yield millions of cells made large screens practical. Stem-cell experiments came later, alongside a T-cell line and work with induced pluripotent stem cells (iPSCs). In one experiment, iPSCs differentiated into ten cell types while undergoing genome-scale perturbations.

The design objective is information per dollar across biological contexts. More cells or sequencing reads alone do not capture the value of observing an intervention in another cell type. The team describes the differentiation experiment as roughly 10,000 perturbations multiplied across ten cell types, expanding both the interventions and the settings in which their effects could be observed.

Why build a model if exhaustive screening is possible? For questions confined to tractable cell lines, the team concedes that researchers can often run the experiment. The model’s intended value is transfer into settings where exhaustive experiments are impractical, including primary cells, organs, and interacting biological systems. Training on scalable systems, then transferring or fine-tuning the model, could prioritize hypotheses for expensive laboratory validation.

Spatial measurements add another dimension. Bulk RNA measurements pool tissue, losing the association between a measurement and its originating cell. Single-cell measurements recover that association, but not necessarily location. For cancer immunology, location can distinguish a T cell near a tumor from one unable to penetrate it, and help explain differences between attacking and non-attacking cells.

Tissue-section assays preserve coordinates while measuring cellular appearance, proteins, or gene expression. Models must then represent neighboring cells as well as individual cells. scGPT Spatial addresses spatial single-cell measurements; the discussion also describes early work predicting expression from stained tissue images. The X-Cell release discussed here did not handle spatial measurements. Spatially aware representations were ongoing work intended for a later version.

Diffusion refines expression profiles without imposing gene order

Autoregressive language modeling predicts tokens in sequence. Earlier single-cell approaches borrowed that setup and imposed gene orderings, sometimes by ranking expression values. But gene-expression measurements have no inherent sentence order: rearranging the measured genes should not change the biology represented. This concerns expression measurements, not the nucleotide sequence within a gene.

X-Cell uses diffusion language modeling to generate high-dimensional expression profiles through bidirectional, iterative refinement. The analogy is typing versus editing: autoregression constructs a sequence, while diffusion repeatedly improves a rough draft. Here, a transformer refines a noisy expression profile into a predicted response to the supplied perturbation.

The conceptual interface is expression levels plus a perturbation in; new expression levels out, with measurements associated with individual cells. The team reports that additional diffusion steps reduced loss and improved agreement with observed profiles, without giving step counts or numerical improvements in the conversation.

A host proposes a transformer operating on a set of genes with encoded identities instead. The response emphasizes generating the full transcriptomic profile, the cell’s gene-expression measurements, rather than a low-dimensional representation or a small predefined gene subset. This is the guest’s rationale for a largely decoder-only generative approach; the exchange does not establish that the proposed set-based alternative was tested and ruled out.

For X-Cell, data contributes most

X-Cell incorporates biological priors, information about genes and cells supplied alongside the experimental learning signal. Its five categories are: - Gene descriptions: embeddings of ChatGPT-generated gene information, an approach compared with GenePT. - Protein–protein interactions: networks describing relationships among proteins. - Gene essentiality: cancer-related information from DepMap. - Morphology: information about cellular appearance. - Cell-type representations: embeddings describing cellular identity.

The team reports that these priors improve context-specific predictions; inspecting their learned weights can help identify which information matters for a cell type. Training priors need not be resupplied at inference, because they are incorporated into learnable parameters. Additional context can still condition predictions at inference, which the guest likens to prompting or in-context learning.

The reported contribution ranking is data first, architecture second, priors third—for this model. In comparisons that trained alternative models on the same datasets, diffusion reportedly improved harder generalization tasks. Priors produced large gains for some cell types and marginal gains for others. The guest suggests that cross-attention may limit how much metadata the model can use, making better integration an open research question.

The model has 4.9 billion parameters. Its perturbation data spans 25 million quality-filtered cells and 16 cell types or biological contexts, with genome-wide interventions and expression measurements.

The companion article adds a scaling observation: on a single, relatively small dataset, test loss flattened beyond about 1.5 billion parameters while training loss kept improving, and a 3.1-billion-parameter model fell off the scaling trend. The article interprets this as an information bottleneck and reports restored scaling with roughly 30 times as much information. It does not define that information measure or quantify the resulting performance gain.

Withhold the response the model must predict

The clearest generalization test separates knowledge of a cell state from knowledge of its response to intervention. The team screened resting and activated regulatory T cells, activating their receptors to initiate proliferation. Training included resting-cell perturbations but withheld activated-cell perturbation responses. The model received information describing the activated state, then predicted how genetic interventions would affect it.

The linear baseline transferred the perturbational change measured in resting cells directly onto activated cells. It assumed activation and genetic perturbation have additive effects. X-Cell was tested on whether it could capture how an intervention’s effect depends on activation state. The team reports predictions of expected inactivation from perturbing the T-cell receptor complex, plus putative inactivators found in the screen. Follow-up laboratory validation of those hits was still planned.

Two further tests broadened the withheld context: - An unseen differentiated cell type: one type from the iPSC experiment was excluded from training. After training on the remaining types and other datasets, the model reportedly predicted responses across thousands of genes and perturbations in the withheld type. - Primary cells from multiple donors: a model trained on a T-cell line was evaluated against a primary T-cell perturbation dataset from Alex Marson’s lab. The team reports transfer across donors, extending prediction into cells that are harder to screen at scale.

These results are reported qualitatively, without accuracy scores, uncertainty estimates, or donor-cohort sizes in the conversation. Patient-level drug response remains a long-term goal, requiring substantial further validation and fine-tuning beyond these cellular experiments.

Measure expression changes, then test what single genes miss

Evaluation must distinguish a plausible expression profile from a correct intervention response. Mean absolute error (MAE) measures discrepancies in expression values, but sparse single-cell data can reward an average profile. A guest reports cases where an average achieved lower MAE than technical replicates used as ground truth. Low overall error can therefore obscure whether the model predicts what the perturbation changed.

The team emphasizes Pearson delta, described as similarity between predicted and observed expression changes following perturbation. In a qualitative heat-map comparison, X-Cell’s predicted changes reportedly looked closer to observations than the linear baseline’s. No numerical comparison accompanies that account. The guests also acknowledge the need for consistently accepted benchmarks and more biological validation.

Across seven genome-wide campaigns, the team observed both shared and context-dependent effects. Housekeeping genes illustrated broadly shared responses across the tested contexts; developmentally relevant genes associated with neuronal differentiation stood out in the iPSC experiments. The biological case for nonlinear models is that they must learn which effects transfer and which depend on cell context.

Single-gene effects can range from extensive to hidden. In X-chromosome dosage compensation, a noncoding RNA helps suppress much of one X chromosome’s expression, forming a Barr body. Redundancy produces the opposite case: if four genes perform the same function, suppressing one may cause no observable change. No visible single-gene effect does not establish that the function is unimportant.

Xaira describes expanding the experimental platform along three axes: - Richer measurements: add molecular measurements beyond RNA expression. - Broader interventions: activate or inhibit whole pathways, extending beyond individual genes. - More complex systems: move beyond cell-line monocultures into primary cells, organoids, and direct screens in living organisms.

X-Cell can already simulate combined interventions by perturbing two gene tokens simultaneously. But generating a combinatorial prediction does not establish its accuracy. Without training on actual combined interventions, the team treats those outputs as hypotheses for testing.

Open science connects discovery, scale, and research judgment

Keeping up with AI research is itself a workflow problem. One guest describes using agentic AI to monitor papers daily and holding weekly lab discussions, while still struggling with the publication pace. Industry’s computing resources provide an advantage; the proposed academic strengths are deep expertise in specific areas and freedom to explore questions outside industrial priorities.

Teaching sustains research expertise. Updating a University of Toronto deep-learning course for roughly 600 students annually forces one guest to keep learning and find clearer explanations. Discussions with students generate research ideas. Academic hospital collaborations can also provide access to healthcare datasets that are difficult for industry to obtain; the guest cites Canadian ultrasound-model work while noting that this advantage may be specific to that setting.

For laboratory methods, the proposed division of labor is academic discovery followed by industrial scaling and refinement. Xaira’s workflow builds on academic advances in CRISPR and single-cell sequencing. Slow, sometimes accidental discoveries can become robust processes producing large amounts of high-quality training data once industry supplies resources and engineering.

Open science supports that handoff. The guests point to open protein-structure data and models such as RoseTTAFold and AlphaFold 2 as precedents for rapid iteration. They advocate shared virtual-cell datasets and standardized experimental protocols, allowing an early field to combine resources and exchange ideas.

As agents make coding more accessible, choosing a worthwhile question becomes more consequential. Implementing a method exposes details hidden in a paper’s equations, even as agents shift human effort toward debugging. Research judgment develops through implementation and real-world feedback: peers and professors help shape questions, while wet-lab and clinical collaborators test predictions against biology.

The next bottlenecks: proteins and living cells over time

Asked which bottleneck they would remove, the guests return to measurement. One wish is protein measurement at genomics scale. RNA indicates which proteins may be produced, but proteins perform much of the cell’s functional work. Useful measurements would capture abundance, modifications after production, location within the cell, and conformational state, ideally at single-cell or spatial resolution.

Another wish is repeated measurements of the same living cells. Destructive sequencing provides snapshots. Following the same cells over time could reveal their trajectories and responses directly, supplying data for the fuller virtual-cell ambition: predicting how cellular states develop.

Small gene panels could be a starting point. Early approaches discussed include sampling a little cytoplasm, like taking a biopsy of a cell, or collecting secreted vesicles from culture medium to measure production over time. Partial sampling is distinct from measuring an entire transcriptome while keeping the cell alive, a capability one guest says remained unavailable.

The episode closes with an invitation to comment on the released X-Cell paper and join the work. At the time of recording, Xaira was recruiting engineers, technologists, biologists, drug-discovery researchers, AI scientists, and computational biologists, with open roles listed on Xaira.com.