Latent Space

Bo Wang & Ci Chu on why virtual cells need causal data, not atlases

Bo Wang & Ci Chu· SVP, Biomedical AI & SVP, AI-Enabled Discovery, Xaira Therapeutics at Xaira Therapeutics
·~90 min·English·Latent Space
LLMTrainingAI Company
TL;DR

Xaira's virtual-cell leads on why the bottleneck to predicting biology isn't the model but the data — and how industrializing CRISPR perturbation screens into causal datasets let a diffusion model predict how cells respond to genes it has never seen.

01Core Mental Model

AI Models Begin With the Data

<strong>The bottleneck to predicting biology isn't the model, it's the data</strong> — protein design took off because 70 years of curated structures existed, while virtual cells lagged for want of that same scale of high-quality causal data.

we don't have the right biological data really to power the training of a predictive model.

— Ci Chu, Latent Space
Key Insight
The strategic move isn't a cleverer architecture — it's vertical integration of the training set. When every lab can fine-tune the same open models, the durable advantage shifts to whoever owns the causal-data factory upstream of the model.
02The Core Problem

Correlation Can't Reveal Cause

<strong>Observational data is underpowered to learn causality</strong> — when genes rise and fall together, countless wiring diagrams explain the same pattern, and only intervening on a gene reveals which one is actually in charge.

Fundamentally, we believe observational data are underpowered to learn causality truly.

— Ci Chu, Latent Space
Key Insight
This is the quiet reason big single-cell foundation models kept losing to linear baselines on perturbation tasks: trained on descriptive atlases, they mastered the correlational shadow of biology, not the causal levers a drug actually pulls.
03The Data Engine

Manufacturing Causality at Scale

<strong>Perturb-seq pools thousands of single-gene interventions into scalable causal screens</strong> — CRISPR silences one gene per cell, a guide-RNA barcode records the target, and single-cell RNA-seq reads all ~20,000 genes; the released dataset spans 25 million quality-filtered cells across seven genome-wide campaigns.

you can easily detect stress signals from the cells and from your scientists in the lab.

— Ci Chu, Latent Space
Key Insight
The unlock is as much industrial as scientific. Academic protocols work on fresh cells for an hour, but genome-wide screens mean handling hundreds of millions of cells across a 14-hour day. Chemical fixation to 'freeze' cell state is what turns a fragile assay into a reproducible data factory.
04The Architecture

Editing Cells, Not Typing Them

<strong>Gene expression isn't a sentence, so X-Cell doesn't write it left-to-right</strong> — instead of autoregression's forced word order, a diffusion model starts from a rough guess of the whole cell and iteratively refines it, editing rather than typing.

So you kind of iteratively edit the gene expression predictions until it minimize the losses.

— Bo Wang, Latent Space
Key Insight
Choosing diffusion is really a statement about the data's geometry. Language is a sequence; a cell's transcriptome is a set. Matching the generative process to that structure — rather than borrowing GPT's recipe wholesale — is where the accuracy gain on hard, out-of-distribution tasks came from.
05The Payoff

Predicting a Cell It Has Never Seen

<strong>The whole point of a model is to answer questions you never ran the experiment for</strong> — trained only on resting T-cells, X-Cell predicted activated-T-cell perturbations better than the linear baseline, and a version trained on a single T-cell line still predicted accurately in primary cells from multiple donors.

we told the model hey this is how the active T-cell look like now go and predict what all of the perturbation are going to do in this active T-cell.

— Ci Chu, Latent Space
Key Insight
This is the line between an expensive atlas and an actual model. If you can only predict cells you already screened, you didn't need AI — you needed a spreadsheet. Generalizing to contexts too costly or impossible to screen — animals, organoids, patients — is the entire economic case for the virtual cell.
06What Actually Mattered

Data Beats Architecture Beats Priors

<strong>Ablations put data quality and scale first, architecture second, biological priors third</strong> — X-Cell injects five kinds of prior knowledge, from literature embeddings to protein-interaction networks, but the causal data did the heaviest lifting.

we find that the qualities the amount of the data sets matter the most.

— Bo Wang, Latent Space
Key Insight
The priors didn't help uniformly — their gains varied sharply by cell type, which says biological knowledge is better treated as context-dependent conditioning than as a universal performance boost. It's also a check on leaderboard instinct: the long list of priors a model advertises on its architecture diagram is the cheapest part of the win, not the load-bearing one.
07The Human Shift

The Agent Codes, the Human Debugs

<strong>Agentic AI flipped the scientist's day from writing code to debugging it, making taste the scarce input</strong> — when everyone can generate code, knowing which experiments are worth running, and validating models against real wet-lab evidence, is what separates good science from burned tokens.

we used to spend lots of time coding a little bit time just debugging but now we let the agent do most of the coding but we spend most of time debugging

— Bo Wang, Latent Space
Key Insight
There's a second-order point buried here: if code is cheap, the constraint moves to judgment and to ground truth. That's why Xaira pairs open-sourced models and data with a wet lab — the validation signal, not the code, is now the rate-limiter on real discovery.