Unambiguous signatures of malignancies extracted from images of growing cells

Abstract As malignant transformation arises from dysregulated cellular programming with characteristic morphological changes, we hypothesized that cancer cells exhibit a unique morphological signature detectable in microscopy images of cells in vitro and in situ across modalities. To test this, we developed CellSign , an AI-based framework to generate Cell Dynamics Fingerprints, which (a) reconstruct morphological progression, (b) remove physiological variation, and (c) support automated malignancy assessment. To arrive there, cell morphology is represented by embeddings from vision foundation models, and a process we term Healthy-Component Reduction is used to refine these by subtracting normal physiological variation, thereby exposing residual disease-specific cues. Embeddings from healthy and malignant cells are organized with manifold learning and summarized with kernel density estimation. We show that unambiguous malignant signatures exist and that our method is robust across diverse datasets spanning breast cancer, lung cancer, and leukemia, transferring reliably between populations of single-cell images and multi-cell patches. Graphical Abstract