🔬Causal Models Need Causal Data - Xaira’s X-Cell model for Drug Discovery (Bo Wang & Ci Chu, Chief Discovery Officer & Chief AI Scientist)

There's a moment in AI research that shows up as a flat line on a chart, and it's more damning than any error message. Bo Wang and Ci Chu at Xaira Therapeutics hit exactly that wall: their model's test loss stopped improving past 1.5 billion parameters, even as training loss kept dropping. Scale up to 3.1 billion parameters and the model actually falls off the trend line. That's the tell. No amount of extra compute or extra parameters was going to fix it, because the problem wasn't the model. It was the data.

The data in question came from CELLxGENE, a massive public resource built by the Chan Zuckerberg Institute that catalogs RNA expression across 168 million cells, tracking somewhere between 20,000 and 30,000 genes per cell. It's a roughly four-trillion-entry matrix, and it has done for virtual cell modeling what the Protein Data Bank did for structural biology. Bo Wang's own scGPT, one of the most cited models trained on this kind of data, became a foundation stone for a lot of what followed. But there's a catch: CELLxGENE tells you what cell states look like, not what causes them. Gene expression changes are tangled together, correlated in ways that make it nearly impossible to untangle cause from effect just by observing.

So Chu and Wang's teams built something new instead of trying to squeeze more out of what existed. X-Atlas is a dataset generated through CRISPR-based perturbation experiments run in parallel, millions of times over, each one nudging a single gene and recording what happens upstream and downstream. That's the causal signal missing from observational data like CELLxGENE. Feed a model enough of these targeted interventions and, in theory, it can start predicting what a drug or gene edit will actually do to a cell, rather than just describing what cells tend to look like. X-Cell is the model trained on top of that, and unlike its predecessor, it scales the way you'd want a model to scale: more parameters and more compute actually buy you more performance.

What's striking is the shape of the investment behind this. Nobody's disclosed exact figures, but data collection and lab infrastructure likely ran into the tens of millions of dollars, while compute, headcount, and modeling research probably cost a few million on top of that. That ratio is backwards from typical AI pretraining economics, where compute dominates. It looks more like an RL rollout budget than a data-hungry language model budget, which says something about where the actual scarcity sits in biology right now. It's not FLOPs. It's wet-lab experiments that generate information no public database currently contains.

The promotions that followed — Chu to Chief Discovery Officer, Wang to Chief AI Scientist — happened after this work was already underway, which tells you how seriously Xaira is treating the causal-data bet as core strategy rather than a side project. Whether X-Cell generalizes cleanly to real human cells in real lab settings, and whether it can meaningfully beat the stubbornly strong linear baselines that have dogged this field, remains the open question the team is still chasing.