The Bitter Lesson for Biology — Adam Green on Virtual Cells and Scaling Laws

Virtual cell models represent a fundamental shift in biological research from traditional mechanistic reductionism to large-scale, top-down machine learning. Adam Green, founder of Markov Biosciences, argues that training models on observational single-cell RNA-seq data using a generative ranking objective—specifically the geometric Plackett-Luce loss—outperforms models trained on perturbation data. This approach treats cells as "specimens" that encode biological state, allowing for the discovery of regulatory mechanisms and clinical insights without requiring exhaustive experimental simulation. By applying these models to antibody drug conjugates, researchers can identify biomarkers like tetraspanin microdomains that explain differential drug efficacy across cancer subtypes. Ultimately, these unified biological world models offer a path toward controlling complex cellular systems, potentially surpassing human-legible mechanistic models by leveraging high-dimensional data patterns that exceed traditional analytical capacity.