Xaira builds 'causal data' for drug discovery, scaling AI cell model 30×

Xaira Therapeutics, led by Ci Chu (Chief Discovery Officer) and Bo Wang (Chief AI Scientist), has introduced X-Cell, an AI model designed to predict what happens to human cells when genes are edited or targeted by drugs. The model was trained on X-Atlas, a dataset of CRISPR-based experiments conducted by Chu and Bo's teams. These experiments run millions of tests in parallel, each isolating the effect of changing a single gene, allowing researchers to map upstream and downstream gene interactions and determine causality—a major leap beyond previous observational approaches.

The project emerged from a concrete problem: earlier AI models trained on CELLxGENE, a database of 168M cells built by Chan Zuckerberg Institute, could describe which genes are expressed in which cell types and states but could not predict outcomes of genetic perturbations. CELLxGENE maps gene expression counts across roughly 20K–30K genes per cell and includes detailed metadata for each of the 168M cells—a ~4 trillion-entry matrix that has inspired a zoo of RNA expression models. Bo Wang built scGPT, one of the most influential of these models, which became the starting point for X-Cell. The problem: gene expression changes are highly correlated, making it nearly impossible to determine causation from observational data alone.

The scaling limitation was stark: Xaira's initial 3.1B-parameter model trained on a single, smaller dataset hit a plateau where test loss flatlined while training loss continued to drop—a signature of information starvation. "Neither parameters nor compute will improve performance past this wall," the team found. "For predicting changes to gene expression, you need more information-rich data." By collecting causal experimental data through CRISPR perturbations—generating roughly 30× more information—the team broke through that wall: the model could now continue to scale with both parameters and training compute. The effort likely cost a few tens of millions for data collection and infrastructure, plus a few million more for compute, headcount, and research—a budget structure resembling reinforcement-learning rollout rather than conventional pre-training.

In validation, X-Cell abandoned autoregressive training in favor of diffusion and demonstrated it beats a linear baseline that had outperformed all previous RNA expression models. The team reports generalization to real lab experiments in real human cells, though full details of that validation remain to be seen in the full podcast episode. This causal-data approach represents a shift in strategy for AI-driven drug discovery: rather than scaling model size alone, Xaira is betting that causal information—data that captures what actually causes changes in gene expression—is the limiting factor for predicting drug efficacy and toxicity.

The core insight driving Xaira's work is the distinction between correlation and causation in gene expression modeling. Traditional AI models trained on CELLxGENE can capture which genes are expressed in which cells, but they cannot reliably predict what will happen when a gene is perturbed—because gene expression changes are highly interdependent, making it nearly impossible to reverse-engineer the causal direction from observational data alone. Xaira's solution is to collect experimental data where individual genes are perturbed one at a time using CRISPR, running millions of parallel tests to build a dataset that explicitly encodes causal relationships. This causal data proved to be the bottleneck: the team's earlier model plateaued at 3.1B parameters because the small, correlational dataset had exhausted its information content, but scaling the data ~30× unlocked continued improvement as parameters and compute increased—a classic signal that the model was previously starved for information, not capacity. The strategic promotion of Chu to Chief Discovery Officer and Wang to Chief AI Scientist signals that Xaira views this data-centric, causal-modeling approach as central to its long-term drug-discovery strategy.

No comments yet. Be the first to share your thoughts!