From Descriptive Data to Causal Engineering: The Shift in Virtual Cell Modeling

Moving from descriptive to causal data is the next frontier in AI-driven drug discovery. While current foundational models are good at recognizing patterns in static gene expression, they cannot predict how cells respond to interventions. This is a major gap for drug development. The bottleneck is not compute or parameter count, but the quality of information. By shifting from observational datasets to high-throughput causal experiments, teams like Xaira are turning biology into an engineering discipline. For practitioners in AI for science, this means competitive advantage no longer comes from scaling models on public archives, but from proprietary, high-fidelity data loops that enable counterfactual prediction.

The Hidden Cost of Descriptive Success

The field has long relied on observational databases like CELLxGENE, which map how cells look under various conditions. These datasets are large, but they are fundamentally descriptive. As Bo Wang and Ci Chu explain, correlation is not causation. Observational data can be fit by an infinite number of causal regulatory networks. When teams try to predict how a cell will change if a specific gene is inhibited, models trained on descriptive data often fail to outperform simple linear baselines.

The correlation data in the descriptive dataset can be fit with many possible causal structures... We believe our provisional data are underpowered to learn causality truly.

-- Ci Chu

The implication is clear: scaling parameters on smoothie data, where tissue is ground up and averaged, creates a ceiling. You can increase compute and parameters indefinitely, but you will not solve the underlying lack of causal information.

Why Immediate Pain Creates Lasting Moats

Xaira’s approach, X-Atlas and X-Cell, prioritizes the generation of causal data through high-throughput Perturb-seq. This involves CRISPR-based experiments that disrupt genes one at a time and measure the subsequent gene expression across millions of cells.

This process is difficult. It requires handling hundreds of millions of cells, managing chemical fixation to prevent stress signals, and ensuring data quality across a 14-hour workday. While this work is expensive and operationally grueling, it creates a competitive advantage. Most teams avoid this level of wet lab complexity, opting for easier, publicly available data. By absorbing this immediate operational pain, Xaira builds a proprietary data layer that allows their models to generalize to unseen cell types and primary human cells, a feat that models trained on public data cannot replicate.

The System Responds: Why Architecture Matters

The shift from autoregressive architectures like GPT to diffusion models for gene expression is a direct response to the nature of biological data. Autoregressive models force an arbitrary order on genes, treating them like words in a sentence. However, gene expression is a set of simultaneous activations. By switching to a bi-directional diffusion process, the model can iteratively refine predictions without assuming a sequence, which better mimics the regulatory dynamics of a cell.

It is visually very clear to see that X-Cell prediction is much more similar to ground truth than the linear baseline. This is a wild moment... This is the first time that someone can put together not just one perturbation, but seven genome-wide perturbation campaigns together.

-- Bo Wang

This demonstrates that when the architecture is aligned with the underlying system dynamics, the model stops hallucinating patterns and starts capturing actual biological responses.

Key Action Items

• Audit your data source (Immediate): Determine if your training data is descriptive or causal. If it is observational, stop expecting your model to predict counterfactual outcomes accurately.
• Prioritize bits per dollar over raw volume (Next Quarter): Shift focus from simply aggregating public datasets to generating high-fidelity, causal data points. The information density of your data is a better predictor of model performance than parameter count.
• Invest in operational hard work (6-12 months): Identify the most difficult, unglamorous part of your data pipeline. This is where your moat is built; competitors will avoid this work precisely because it is difficult.
• Adopt diffusion-based architectures for set-based data (6-12 months): If you are working with high-dimensional, non-sequential data, move away from autoregressive models. Diffusion processes offer a better way to model simultaneous state changes.
• Build for Out-of-Context generalization (12-18 months): Design your validation benchmarks to specifically test generalization to unseen cell types or donors. If your model only performs well on the data it was trained on, it has not learned the underlying system dynamics; it has merely memorized the training set.