Deep Dive Into X-Cell: Ci Chu And Bo Wang, Xaira Therapeutics

Author: Ci Chu, SVP of AI-Enabled Discovery, Xaira Therapeutics; Bo Wang, SVP & Head of Biomedical AI, Xaira Therapeutics Collection: UNLOCK-2026-Sessions

Summary

Ci Chu and Bo Wang from Xaira Therapeutics present a detailed technical deep dive into X-Cell, Xaira's virtual cell model designed to predict how cells respond to perturbations. The talk covers both the biology-side data generation strategy (led by Chu, head of "hyperbiology" at Xaira) and the AI modeling innovations (led by Wang, head of biomedical AI). Xaira describes itself as an "AI drug discovery company" that simultaneously builds therapeutic programs and the AI models needed to accelerate them — covering protein design, predictive biology, virtual cell, and patient models.

The central hypothesis of the virtual cell effort is that a model trained on massive causal (interventional) data can predict perturbation effects in biological contexts not seen during training, enabling forward modeling (if I knock out gene X, what molecular state will the cell reach?) and backward counterfactual reasoning (to move a cell from disease state to healthy state, what interventions are required?). First-generation virtual cell models like scGPT were trained on descriptive data from resources like CELLxGENE and Human Cell Atlas and perform well on cell-type annotation but fail to outperform simpler methods on counterfactual perturbation questions. Xaira's thesis is that this is a data-type problem: perturbative, causal data was missing.

Xaira generated two large-scale perturbation datasets — X-Atlas Orion (published June 2025, downloaded 80,000 times, favorably reviewed by key opinion leaders including Dr. Emma Loomberg) and X-Atlas Pisces (published March 2026, more than 3x the size of Orion at ~26 million cells, multi-context with 10 cell types and genome-wide perturbations in a single experiment). The data generation method is "Fixed Perturb-seq" — Xaira's industrialized version of the academic Perturb-seq technique combining CRISPR perturbations with single-cell RNA sequencing. The key innovations are chemical cell fixation (preserving cell state to allow long processing windows without stress artifacts) and rigorous QC with pause points. X-Atlas data is available for academic use at no cost and has attracted multiple commercial licenses.

Bo Wang describes X-Cell's architectural innovations: (1) replacing next-token prediction (used in scGPT) with a diffusion language model, which is better suited to gene expression data where token order is meaningless, and (2) using cross-attention to incorporate five diverse prior knowledge sources (ESM2 protein embeddings, STRING protein-protein interaction database, DepMap cancer dependency maps, JUMP-Cell Painting morphological data, and scGPT cell state encodings). X-Cell outperforms Cell2Sentence, STATE, and scGPT on multiple perturbation prediction benchmarks. X-Cell-Ultra has 4.9 billion parameters — a 92x increase over the original scGPT's 53M parameters — and demonstrates clean scaling laws: test loss and perturbation prediction accuracy improve monotonically with parameter count, and out-of-distribution generalization (across cell types and to external datasets from Marson Lab T-cells) improves at larger scales.