CellFluxV2: An Image Generative Foundation Model for Virtual Cell Modeling

Abstract

Building a virtual cell that simulates cellular behavior in silico is a central goal of computational biology. We introduce CellFluxV2, an image-generative model that predicts how cell morphologies change in response to chemical and genetic perturbations. A core innovation of CellFluxV2 is to learn distribution-level transformations from unperturbed to perturbed cells within the same experimental batch using flow matching, enabling it to disentangle true perturbation effects from confounding batch effects. Incorporating three methodological advances, CellFluxV2 achieves up to a 77% improvement in image fidelity over diffusion- and GAN-based baselines, while maintaining biological fidelity comparable to ground-truth images. Scaling up CellFluxV2, we establish the first scaling laws in image-based virtual cell modeling, demonstrating that performance improves consistently with both dataset size and model capacity. Furthermore, the scaled-up model generalizes well to out-of-distribution perturbations and exhibits two novel capabilities: batch-effect correction and cell-state interpolation. Together, these results position CellFluxV2 as a powerful foundation model advancing the vision of a virtual cell, unlocking novel opportunities for in silico drug screening.