Xaira Therapeutics' X-Cell Model and the Need for Causal Data in Drug Discovery
Xaira Therapeutics' leadership, Bo Wang and Ci Chu, detailed their X-Cell model, a set-level diffusion transformer designed to predict genome-scale transcriptional responses to genetic perturbations. The model is trained explicitly on interventional data to capture causal biological relationships rather than just correlations. Traditional AI models in biology often rely on observational data, which struggles to identify true cause-and-effect relationships. By generating and training on causal, interventional data, X-Cell can improve the accuracy of predicting how cells respond to genetic modifications, accelerating drug discovery. X-Cell was trained on the X-Atlas/Pisces dataset containing 25.6 million perturbed single cells across 7 CRISPRi Perturb-seq screens. It utilizes cross-attention to integrate multi-modal biological priors, allowing it to generalize zero-shot to unseen cell types and perturbations.
## BACKGROUND
In drug discovery, understanding how genetic perturbations (like knocking out a gene) affect cellular behavior is crucial. Traditional machine learning models often rely on observational data, which can identify correlations but cannot prove causality. Causal inference and interventional datasets, such as those generated by CRISPR screens, allow models to predict the direct consequences of genetic changes.