🔬Causal Models Need Causal Data - Xaira’s X-Cell model (Bo Wang & Ci Chu)
Summary
Xaira’s thesis is to make drug discovery more engineering-like by connecting three AI systems end to end. Protein design supplies molecules for previously undruggable targets, X-Cell predicts how interventions alter cell biology, and patient-representation models aim to identify responders. The ambition is not merely faster target discovery but turning “trial and error and guesswork” across the path from target identification through Phase III into a more systematic process.
The episode’s central technical claim is that causal models require causal training data, not ever-larger collections of observational cells. Static expression profiles can show genes A, B, and C moving together but do not by themselves establish which regulates which; accordingly, existing foundation models often fail to beat linear baselines on perturbation tasks. Xaira’s answer is seven genome-wide Perturb-seq campaigns spanning 16 biological contexts and 25 million cells after stringent filtering: “Observational data are underpowered to learn causality truly.”
Xaira’s wet-lab system may be at least as important as its 4.9-billion-parameter model. Pooled CRISPR perturbations silence genes across hundreds of millions of cells, while single-cell RNA sequencing records the downstream effects on the transcriptome; barcoding keeps perturbation identity recoverable without plate-to-plate batch effects. The operational moat lies in industrializing academic assays, including fixing cells early enough to avoid introducing handling stress.
X-Cell replaces an unnatural autoregressive ordering of gene-expression values with iterative diffusion and conditions training on five kinds of biological prior knowledge. The model repeatedly edits a noisy transcriptomic prediction toward a refined response, while literature-derived embeddings, protein-protein interactions, DepMap, morphology, and scGPT embeddings provide context. The discussion ranked data quality, amount, and scale first; architecture second; and prior knowledge third.
The strongest evidence is out-of-context generalization rather than an incremental win on an in-distribution benchmark. Trained only on perturbations in resting T cells, X-Cell predicted perturbation effects after activation; it also predicted a held-out iPSC-derived cell type and transferred from a T-cell line to primary T cells from multiple donors. That supports the intended use case: learn where exhaustive screening is feasible, then transfer predictions into biological contexts where experimentation is expensive or impossible.
The linear-baseline dispute turns on whether benchmarks measure average expression or the intervention-induced change that matters biologically. Sparse single-cell data can make a mean profile look artificially strong under mean absolute error, sometimes even stronger than technical replicates; Xaira instead emphasizes similarity between predicted and observed expression deltas. Chu’s “wow moment” was simply aligning the raw-data heat maps: X-Cell looked substantially more like ground truth than the linear prediction.
The platform remains an early cellular scaffold, not a demonstrated solution to clinical attrition. Wang cited a world where 90% of diseases have no cure and Phase III success can be as low as 5%–10%, but X-Cell has not yet modeled patients, organs, or longitudinal cellular dynamics. The next areas the guests identified are multimodal and combinatorial perturbations, spatial context, protein-scale measurement, and repeated measurement of the same living cell—the latter being Wang’s definition of a “real virtual cell.”
Deep dive
1. Xaira is linking three AI platforms into one drug-development loop
Chu described Xaira as an AI-enabled drug-discovery company whose output must ultimately be therapeutics, not model benchmarks. Its three platforms are protein design, a virtual cell for predicting biological response, and patient-representation models intended to match therapeutics to responders.
Protein design grew from co-founder David Baker’s group and targets the molecule-making problem; X-Cell addresses what genetic interventions—and potentially drugs—will do to cells; the emerging patient layer addresses translation. Together, Chu hopes they can turn an “artisanal trial-and-error” process “more and more into an engineering discipline.”
Wang’s differentiator was the connective tissue: target predictions should favor opportunities compatible with protein engineering, while cellular models should connect to patient data carrying clinical utility. Xaira is therefore trying to apply AI from target identification and molecule design through Phase I–III trials, reducing both failure rates and cycle time.
2. A representation of a cell is not yet a virtual cell
Wang traced “virtual cell 1.0” to attempts roughly 20 years ago to encode pathways with differential equations fitted to observations. He called that effort largely unsuccessful because biology was “way too complicated to write in a few predefined set of differential equations.”
The data-driven revival followed language models. Wang’s Toronto lab released scGPT roughly four months after ChatGPT, using one foundation model for tasks such as batch integration and multi-omic alignment. The governing virtual-cell 2.0 intuition became: “If we cannot describe, let’s learn it.”
Yet Wang resisted equating foundation models with virtual cells. Foundation models mainly provide useful semantic representations; a true virtual cell must model interventions and dynamics, potentially including development over time and spatial change. “The foundation model is really just the starting point.”
3. Observational expression cannot identify the direction of biology
Chu’s causal-data argument began with three correlated genes: if A, B, and C rise together, A might regulate B and C, B might regulate A and C, or an unobserved factor might drive them all. Many incompatible regulatory structures can fit the same descriptive data.
That ambiguity explains why models trained on large observational atlases excel at descriptive tasks such as batch harmonization and technology correction, yet frequently fail to beat linear methods on counterfactual questions: “If I did this to the cell, then what would happen?”
Protein modeling advanced on more than 70 years of curated structures and sequences, Chu argued, whereas virtual-cell and patient-response modeling lack comparable high-quality data. Patient samples are especially scarce when molecular profiles must be matched to clinical annotations and treatment response, so Xaira began by manufacturing causal cellular data itself.
4. Perturb-seq turns a pooled experiment into a genome-scale causal matrix
Perturb-seq combines pooled CRISPR-Cas9 perturbation with single-cell RNA sequencing. One axis records which of roughly 20,000 genes was suppressed; the other records how expression across the transcriptome changed, producing the rich two-dimensional training data Chu considers analogous in shape to protein-structure datasets.
Each cell receives Cas9 plus a guide RNA whose sequence directs the machinery to a gene’s promoter. The guide acts as both address and recoverable barcode: researchers can mix thousands of perturbations in one pool, then infer from sequencing which gene was silenced in every cell.
On the readout axis, single-cell RNA sequencing measures expression across approximately 20,000 genes per cell. Combining both technologies makes it possible to knock down genes across many cells and observe their effects on every other measured gene without running 20,000 separate plates.
Scale created a second scientific problem. Xaira may process tens or hundreds of millions of cells to obtain 25 million passing its strictest filters; fresh-cell workflows stretched across a 14-hour day and could introduce stress signals in the cells. Chemical fixation now locks cellular state early and enables time-shifted processing with fewer batch effects.
5. Biological diversity matters more than accumulating raw cell counts
Chu corrected the premise that Xaira began with stem cells. Its first datasets used two readily scalable immortalized lines, one cancer-derived, which captured characteristics of cell types derived from colorectal cancer and hematopoietic cells; later campaigns expanded into T-cell lines, primary cells, and induced pluripotent stem cells.
The most ambitious screen differentiated iPSCs into 10 cell types in one unrestricted experiment, then performed genome-scale perturbation across them. The experiment effectively multiplied a roughly 10,000-experiment library across 10 cell types—“a library-on-library experiment.” The objective was maximizing “bits per dollar” rather than sequencing reads alone.
The host pressed on what isolated cells omit: neighbors, tissue organization, and multicellular physiology. Chu agreed that exhaustive screens cannot be run in organs, animals, or every translational model; the purpose of a model is to learn from scalable systems, then produce higher-quality hypotheses for expensive physiological experiments.
Spatial assays preserve the x-y position of cells while measuring morphology, proteins, or genome-wide expression. The current X-Cell release does not model that information, but Wang said the next version will infer spatially aware representations; longer term, he expects virtual cells to combine RNA, proteomics, other regulatory omics, and imaging.
6. Diffusion fits unordered expression data better than next-token prediction
Early single-cell foundation models borrowed autoregressive language-model training, which requires ordering genes like words. Wang sees a category error: DNA bases have a meaningful sequence, but an expression profile is essentially a matrix of gene counts whose biology does not change when its columns are shuffled.
X-Cell instead uses a diffusion language model. Wang’s analogy contrasted typing “I like coffee” from left to right with editing an entire rough sentence repeatedly; the model begins with a noisy expression profile and iteratively refines it, with prediction loss continuing to fall across diffusion steps.
The host reduced the interface to its essentials: the input includes cell context and a perturbation, and the output is the resulting cell-wide expression profile. Wang emphasized generative, decoder-only modeling because Xaira wants the full high-dimensional transcriptome, not predictions for a small predefined gene panel.
7. Biological priors help, but the causal dataset supplies most of the lift
X-Cell conditions training on five kinds of prior information: literature-derived gene descriptions generated through ChatGPT and embedded in GenePT, protein-protein interaction networks, DepMap cancer-essentiality information, morphology, and scGPT embeddings encoding cell-type context.
These priors become learned parameters, so users need not supply them at inference. The system can also accept additional conditions in a prompt-like or in-context fashion, while the weights assigned to different knowledge sources offer some interpretability into which prior mattered for a particular cell type.
Asked where the “big alpha” resides, the discussion gave a clear ordering: data quality, volume, and scale matter most; moving from autoregression to diffusion gives significant improvements, especially on unseen contexts; priors rank third and vary by condition. Some cell types benefit greatly, while others show only marginal gains.
8. Out-of-context transfer is the result Xaira built the system to test
The episode rejected the idea that a virtual cell should replace experiments already easy to perform. Its purpose is to transfer causal knowledge toward primary cells, tissues, animals, and eventually humans: “generate data where it is possible” and predict where exhaustive intervention studies are infeasible.
For the T-cell test, Xaira trained on perturbations in resting regulatory T cells, then supplied the unperturbed profile of activated cells and requested all activated-state perturbation effects. X-Cell recovered expected TCR-complex biology and putative inactivation hits without seeing perturbation data in the activated context.
In the differentiated-iPSC experiment, Xaira withheld an entire cell type from training. The model still produced strong predictions across thousands of perturbations and genes, indicating transfer across cell type rather than interpolation among perturbations observed in the same context.
A third test moved from a T-cell line to recently published primary T-cell Perturb-seq data from Alex Marson’s lab. X-Cell, trained on the line rather than donor cells, made accurate predictions across primary cells from multiple donors—the most direct validation offered for training on scalable but imperfect proxies.
9. Linear baselines win weak benchmarks but fail the transfer problem
The host invoked the field’s “famous or infamous” perturbation challenge: complex foundation models often do not beat simple linear baselines. Wang conceded that scGPT and similar models trained on static expression can lose under mean absolute error, especially on the small Replogle datasets.
Single-cell matrices are sparse, so predicting an average profile can minimize MAE without capturing intervention biology; Wang noted that an average may even score better than technical replicates treated as ground truth. Chu emphasized metrics comparing predicted and observed expression changes, including Pearson correlation on perturbational deltas.
Chu’s decisive check was visual: line up ground truth, the linear extrapolation, and X-Cell’s heat maps. “It’s visually very clear” that X-Cell more closely matches the real changes, making the result legible to biologists rather than dependent on one aggregate score.
Seven genome-wide screens also expose why nonlinearity matters. Housekeeping-gene perturbations can be context-universal, while developmental or neuronal-differentiation genes light up specifically in iPSCs. A useful model must learn both invariant effects and cell-specific regulation; simply transplanting a perturbational delta cannot do both.
10. Single-gene screens are a scaffold for combinatorial and clinical biology
The host challenged whether single-gene knockdowns miss redundant networks. Chu agreed, illustrating one-gene leverage with X-chromosome dosage compensation: a non-coding RNA coats one female X chromosome, suppresses most of it, and consigns it to a Barr body—yet other systems may require disabling several redundant genes before any phenotype appears.
Xaira is expanding along three axes: beyond transcriptomics into multimodal readouts; beyond single-gene loss of function into pathway activation and inhibition; and beyond monoculture cell lines into primary cells, organoids, more complex translational systems, and direct in-vivo perturbation screens.
X-Cell can simulate combinatorial interventions by perturbing two gene tokens together. Wang kept the caveat intact: without actual combinatorial training data, “the accuracy may not be there,” so those in-silico results currently generate hypotheses rather than validated predictions.
Wang framed the eventual target with stark numbers: 90% of diseases have no cure, and he cited Phase III success as low as 5%–10%. The “holy grail” is not replacing biologists but generalizing into otherwise inaccessible contexts, eventually selecting the patients most likely to respond to a therapeutic.
11. Open science shifts the advantage toward execution, taste, and measurement
Wang described an unsustainable personal split—officially 80% Xaira and 20% university, “in reality 100% on Xaira, 100% on this.” Agents help monitor papers and write code, but the pace still creates anxiety; increasingly, students spend less time coding and more time finding AI-generated bugs.
His prescription for researchers was “taste”: select consequential questions, understand implementations deeply, and avoid letting agents “just burn tokens without purpose.” Industry has GPUs and scale; academia retains freedom, niche expertise, teaching-driven renewal, and access to some healthcare datasets that companies cannot readily obtain.
Wang defended open sourcing because virtual-cell science is too early for isolated datasets to maximize progress. He wants shared data, common quality standards, and model iteration resembling the protein ecosystem built around PDB, AlphaFold, and RoseTTAFold; academic labs can invent assays, while industry robustifies and scales them.
The wet-lab discussion emphasized that biology extends beyond cell-based Perturb-seq: researchers want to measure proteins, metabolites, lipids, protein–protein interactions, spatial cell-cell interactions, and whole-animal perturbations at scale. Their magic-wand bottlenecks were measurements, not larger models. The wet-lab answer prioritized protein sequencing at genomic scale—including abundance, modification, conformation, and localization—while Wang wants repeated transcriptomic measurements from the same living cells. Even a small longitudinal gene panel would help; a full, non-destructive transcriptome would enable temporal dynamics as “the real virtual cell.”