Why AI Needed Perturbation Data to Model Cells

William Palmer

So... picture you are- you are hurtling down a racetrack at a hundred and forty miles an hour. The tires are screaming, you're right on the limit of adhesion, and suddenly... the rear end loses traction. You're in a high-speed spin. Now, how do you learn to recover from that? Do you do it by, uh, by watching millions of hours of highway driving footage? Just observing cars cruising in their lanes?

William Palmer

No, of course not. Highway driving shows you stable states. It-it-it doesn't show you the physics of recovery. To learn how to catch a spin, you have to actually perturb the system. You have to lose traction on purpose, counter-steer, feel the weight shift, and measure the correction. You need causal feedback.

William Palmer

And the reason I'm talking about race cars is that AI researchers in biology just hit this exact same wall. We've been trying to build a "Virtual Cell"—an AI that understands how human cells work—by feeding it massive, passive datasets. Specifically, CELLxGENE, which is this mind-bogglingly huge database of 168 million cells, essentially a four-trillion-entry matrix of gene expressions. It's an incredible repository, but... it's purely observational. It's just highway driving footage.

William Palmer

And here's the shocker. When researchers scaled up their models on this observational data, they hit a hard, flat wall. The training loss kept dropping, sure, but the test loss? Flatlined. It completely stalled out after about 1.5 billion parameters. It didn't matter if you threw ten times the compute or massive parameters at it; the model's performance just wouldn't budge.

William Palmer

This is what we call the "information gap." The model is limited by the amount of real, causal information in the data. You can't figure out what actually causes what just by looking at correlations. If gene A and gene B always go up together, is A causing B? Is B causing A? Or is there some hidden gene C pulling the strings? You can't tell.

William Palmer

To break through, you have to do what Ci Chu and Bo Wang at Xaira Therapeutics did. They realized that instead of spending their millions on more pre-training GPUs... they needed to invest in physical rollouts. They spent tens of millions of dollars on automated wet labs to run massive, parallel CRISPR experiments. They literally went into human cells and turned the dial down on genes, one by one, to observe the downstream effects.

William Palmer

The result of this massive bet is the X-Atlas compendium, or Pisces, which contains... let me get the exact detail here... "comprising 25.6 million perturbed single-cell transcriptomes across 16 biologically diverse contexts." That is roughly thirty times the causal information we had before. And by feeding *that* into their models, they didn't just crawl over the scaling wall—they blew right through it.

William Palmer

But, okay, once you have this incredible causal data, how do you actually model it? For a while, the state-of-the-art approach was this project called Cell2Sentence. The idea was to treat the cell like a book. You take a cell's profile, sort the genes by how highly expressed they are, write those gene names down in a linear list, and feed that "sentence" into a standard autoregressive language model—like a Gemma-2 27B.

William Palmer

But think about that for a second. A cell is not a book. It doesn't read left-to-right. A cell is a highly parallel, deeply interconnected web of regulatory networks. Forcing that three-dimensional biological logic into a flat, sequential sentence is... well, it's a massive structural mismatch.

William Palmer

So Bo Wang's team at Xaira abandoned autoregression entirely. Instead, they built a model called X-Cell, and they based it on diffusion.

William Palmer

Here is how it works. During training, they take a control cell's profile and randomly mask out, say, 25%, 50%, or even 75% of the gene expression values. Then they replace those masked values with the ground-truth perturbed values from the CRISPR experiments. The model's job is to "denoise" the rest of the cell—to look at those active edits and figure out how they ripple across the remaining twenty to thirty thousand genes all at once, globally and simultaneously, rather than predicting them one by one like words in a sentence.

William Palmer

And the results? They are absolutely wild. "We scale X-Cell to 4.9 billion parameters (X-Cell-Ultra), the largest causal perturbation model to date." And because they had that 30x increase in causal data from X-Atlas, scaling actually worked! The model showed a train scaling exponent of 0.32, proving that biological perturbation follows the same mathematical power-law scaling we see in massive LLMs.

William Palmer

But the real test is in the lab. How does this actually translate to real human cells? Well, X-Cell-Ultra completely beats standard linear baselines—which, honestly, used to outperform previous deep learning models—by a factor of five on Pearson Delta, which measures the correlation of predicted versus observed fold changes.

William Palmer

Even more impressive, it showed zero-shot generalization. It successfully predicted T-cell inactivating perturbations in stimulated Jurkat cells and even in primary human CD4+ T-cells. We are talking about simulating highly complex genetic edits in silico, on a computer, before a scientist ever has to touch a pipette in a physical lab.

William Palmer

It turns out, if you want a model to navigate the chaotic, high-speed physics of a living cell... you can't just teach it to watch the road. You have to teach it how to steer through the skid.

William Palmer

Alright, that's it for me today. Talk to you soon.