🔬 Causal Models Need Causal Data - Xaira’s X-Cell model for Drug Discovery (Bo Wang & Ci Chu, Chief Discovery Officer & Chief AI Scientist)

This section discusses Zera Therapeutics' mission to leverage AI for drug discovery, focusing on three main platforms: protein design, visual cell or foundation models of biology, and patient representation models. The goal is to create an end-to-end AI-driven drug development process that accelerates timelines and improves success rates by integrating data, models, and experimental feedback. The approach emphasizes the importance of high-quality, causal data and the potential of virtual cell models (like Excel) to predict cellular responses and inform clinical development.

This part explains Zera's development of the Excel virtual cell model, trained on large-scale perturbation data, including genome-wide CRISPR experiments combined with single-cell RNA sequencing. Excel predicts how gene perturbations affect cell states, functions, and pathways, aiming to simulate cellular responses accurately. The model employs a diffusion language approach, integrating diverse biological priors for better interpretability and generalization. It is designed to predict outcomes of genetic and chemical perturbations, even in unseen contexts, moving toward a comprehensive virtual cell.

The discussion covers the modern methods used in high-throughput biological experiments, especially CRISPR-based perturb-seq technology. Datasets include millions of cells across various conditions, cell types, and sources (e.g., cell lines, primary cells, stem cells). Engineering innovations enable the handling of massive, high-quality data with minimal batch effects. The goal is to generate causal, rich datasets that can train AI models to understand cell biology dynamically, including multi-omics and spatial data integration.

The conversation details the architectures behind Excel, emphasizing the shift from autoregressive models to diffusion models inspired by language modeling. Using attention mechanisms and a focus on generating full transcriptomic data, the model captures gene regulatory networks and dynamic cellular responses. Furthermore, integrating prior biological knowledge—such as protein interaction networks and literature embeddings—enhances prediction accuracy, interpretability, and the ability to generalize to novel contexts.

This section highlights the difficulties in benchmarking models against linear baselines, especially on sparse single-cell data. It underscores the importance of training on high-quality, causal datasets and demonstrates Excel’s superior ability to predict gene expression changes and cellular responses compared to previous models. The value of large, diverse datasets and biological validation through experiments is emphasized as essential for building trust and advancing virtual cell concepts.

Looking ahead, the discussion touches on integrating spatial transcriptomics, multimodal data, and longitudinal cell measurements for richer models. The ultimate goal is to develop virtual cells capable of simulating entire biological systems, including complex in vivo environments, to predict drug responses in humans. The importance of open science, data sharing, and collaboration across academia and industry to accelerate progress is strongly advocated.