Xaira's X-Cell Bets on Lab-Made Data, Not Bigger Models
Xaira Therapeutics traded a bigger model for a 30x richer dataset, training X-Cell to predict how a drug or a gene edit reshapes a human cell.
Xaira Therapeutics' Chief AI Scientist Bo Wang and Chief Discovery Officer Ci Chu built X-Cell, a model meant to answer one narrow but valuable question: what happens inside a cell when you change a single gene? Their path there was not a bigger network. It was a much bigger, more targeted pile of lab-generated data, as the pair described on the Latent Space podcast.
The wall that more parameters could not fix
CELLxGENE, a 168M-cell database built by the Chan Zuckerberg Institute, maps each cell to how many times 20K-30K genes were detected in it, forming a roughly 4 trillion-entry matrix. That scale made CELLxGENE the reference dataset for so-called Virtual Cell models, much the way the Protein Data Bank became the reference dataset for structural biology models. It also inspired a wave of RNA-expression models, including Wang's own scGPT, which became the starting point for Xaira's new model.
The catch: models trained only on CELLxGENE describe which cell types and states tend to go together, but changes in gene expression are highly correlated, so those models struggle to say what actually causes what. Xaira's team ran into this directly: training on a single, smallish dataset, a 3.1B parameter model flatlined on test loss while training loss kept dropping, a classic sign the model had run out of new information to learn from, not new capacity to hold it. Neither more parameters nor more compute could push past that wall.
Building X-Atlas: data built for cause and effect
Xaira's answer was X-Atlas, a dataset built specifically to expose cause and effect. Using CRISPR-based experiments that run millions of tests in parallel, the team could turn down one gene at a time and watch what happened upstream and downstream of it, cell by cell. Repeating that across many genes teaches a model what changes when a given gene is switched off, instead of only which genes tend to move together.
X-Atlas gave the X-Cell model roughly 30x more information-rich training data than the earlier CELLxGENE-only approach. With that richer data in place, scaling resumed: more parameters and more training compute again improved performance, instead of hitting the same wall the 3.1B model hit. Along the way, the team also dropped autoregression in favor of diffusion models for X-Cell.
None of this came cheap. Latent Space's own estimate puts data collection and infrastructure alone at a few tens of millions of dollars, with compute, headcount, and research adding a few million more on top, a budget shaped more like a reinforcement-learning rollout than a typical data-rich pretraining run.
Why this matters for drug discovery
A model that can predict how a cell responds to a drug or a gene edit before anyone touches a real cell could narrow down which lab experiments are worth running at all. That is the pitch behind X-Cell, and it is why Chu and Wang's bet on causal, CRISPR-generated data is being watched closely across the field. Both were promoted after recording the podcast episode: Chu to Chief Discovery Officer, Wang to Chief AI Scientist, underscoring how central this bet has become to Xaira's strategy.
Key facts
- X-Cell is Xaira Therapeutics' model for predicting gene-expression changes caused by drugs or gene edits, built by Bo Wang and Ci Chu.
- A 3.1B parameter model trained on CELLxGENE (168M cells, 20K-30K genes tracked per cell, a roughly 4 trillion-entry matrix) flatlined on test loss.
- X-Atlas, a CRISPR-based dataset, supplied roughly 30x more information-rich training data and let scaling resume.
- The team switched from autoregression to diffusion models; Latent Space estimates data collection alone cost a few tens of millions of dollars.
Check your understanding
- Why did the 3.1B parameter model trained on CELLxGENE stop improving even as compute and parameters increased?
- How do the CRISPR-based experiments behind X-Atlas help X-Cell learn cause and effect instead of just correlation?