Home
 /
Papers
 /
Systems Biology
 /
X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models
Preprint
Open access
Top 1% · Validated by QED

X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models

Chloe Wang, Mehran Karimzadeh, Neal G. Ravindra +27 more
March 2026
 ·
bioRxiv
 ·
DOI 
10.64898/2026.03.18.712807
99
%
QED Score
QED Impressions

Establishing that causal perturbation prediction follows power-law scaling and achieving zero-shot generalization to unseen primary human cells, this work overturns the assumption that cellular models cannot extrapolate, opening a scalable path to context-agnostic drug target discovery.

Run full QED analysis
Read paper
Save

QED claims analysis

Summary

X-Cell, a diffusion language model trained on X-Atlas/Pisces—the largest genome-wide CRISPRi Perturb-seq resource comprising 25.6 million perturbed single-cell transcriptomes across 16 diverse cellular contexts—achieves state-of-the-art perturbation prediction through cross-attention to multi-modal biological priors and demonstrates that perturbation prediction follows power-law scaling consistent with large language models, enabling zero-shot generalization to unseen cell types and primary human cells while revealing that perturbations exhibit context-dependent regulatory effects with conserved perturbations targeting core metabolic machinery and context-specific perturbations targeting cell-type identity pathways.

See full analysis
Main Claim 1 –

X-Atlas/Pisces: 25.6M-cell CRISPRi Perturb-seq across 16 contexts

Supporting evidence

  • Multi-context sampling boosts regulatory network recall over single-context depth
Main Claim 2 –

X-Cell diffusion language model with multimodal prior cross-attention

Supporting evidence

  • Iterative diffusion refinement improves perturbation prediction accuracy
  • X-Cell achieves >5× gains over prior perturbation prediction models
  • DepMap and STRING priors most improve X-Cell perturbation prediction
  • X-Cell generalizes zero-shot from genetic to chemical perturbations
Main Claim 3 –

Conserved metabolic perturbations versus context-specific identity pathways

Main Claim 4 –

Zero-shot prediction of T cell inactivation in activated Jurkat

Abstract

Causal models of cellular systems hold the promise to empower broad biological discovery, including the systematic identification of novel targets for drug discovery. Predicting how genetic and pathway perturbations reshape gene expression across diverse cellular contexts is a prerequisite for building generalizable cellular foundation models. However, current methods typically fail to extrapolate beyond their training distributions because they rely predominantly on observational expression atlases rather than interventional perturbation data. We present X-Atlas/Pisces, the largest genome-wide CRISPRi Perturb-seq compendium to date, comprising 25.6 million perturbed single-cell transcriptomes across 16 biologically diverse contexts, including widely used cell lines, induced pluripotent stem cells (iPSCs), resting and CD3/CD28 activated Jurkat T lymphoma cells, and multi-lineage differentiating iPSCs. Leveraging this resource, we develop X-Cell, a diffusion language model that predicts perturbation responses by iteratively refining control-to-perturbed state transitions through cross-attention to multi-modal biological priors derived from natural language, protein language models, interaction networks, genetic dependency maps, and morphological profiles. X-Cell outperforms existing state-of-the-art models by up to five-fold on key metrics such as Pearson Δ (correlation between predicted and observed perturbation-induced log-fold changes), and demonstrates zero-shot prediction of T cell inactivating perturbations in stimulated Jurkat cells. We scale X-Cell to 4.9 billion parameters (X-Cell-Ultra), the largest causal perturbation model to date. We demonstrate for the first time that perturbation prediction follows power-law scaling with an exponent matching large language models. X-Cell-Ultra demonstrates zero-shot generalization to novel biological contexts, including unseen iPSC-derived melanocyte progenitors and primary human CD4+ T cells from multiple donors, and outperforms all baselines after self-supervised test-time adaptation. These results demonstrate that coordinated scaling of causal perturbation data and model capacity yields foundation models capable of generalizable perturbation prediction across cellular contexts, with potential applications for improving computational target identification, validation, and context-specific therapeutic prioritization.

Related papers

How the QED Score works

QED Score is an AI-based quality metric for life science research. It rates each paper on two dimensions: originality (is the finding new and important?) and validity (does the evidence support the claims?). Author names and affiliations are removed before scoring. Each manuscript is broken down into its claims and the experiments behind them, and specialized AI agents then examine those claims for statistical rigor, consistency across figures, and contradictions with the existing literature. A verification layer re-checks every issue they flag. Scores are percentile ranks against all papers in the corpus. In a blinded study of papers where QED Score and journal rank disagreed, domain experts sided with QED Score in 75% of cases.
Read the methodology

Free access for academic researchers

Create your free QED account to validate your research, strengthen grant proposals, and uncover scientific insights.
We've sent you an access link.
Please check your inbox.

Didn't get your email? Check your spam folder or reach out to info@qedscience.com

Oops! Something went wrong while submitting the form.
Looking for QED for pharma, biotech or life science organizations?