Perturbation modelling aims to predict how cell state changes in response to chemical or biological interventions. As the field has grown, this work has moved toward general-purpose virtual cell models (VCs), which can act as digital twins of cellular systems. In 2025, the Arc Institute hosted the inaugural Virtual Cell Challenge (VCC), focused on predicting the effects of unseen CRISPRi interventions in H1 human embryonic stem cells.
Our team, Mean Predictors, ranked 3rd on the generalist leaderboard by predicting pseudo-bulked differential expression profiles from single-cell data and perturbation descriptors from external datasets. A key contributor to our performance was including the number of cells per perturbation as an explicit model feature.