Morphological fingerprints enable machine learning based inference of neuroblastoma cell states without transcriptomics

Inference of cancer cell states is essential for understanding oncogenic mechanisms and predicting clinical outcomes, yet current reliance on transcriptomic profiling limits scalability and real-time monitoring. Here, we show that cell morphology provides a low-dimensional, observable representation of cellular identity and its dynamics. Using neuroblastoma (NB) as a model system, we establish a machine learning- morphology profiling framework that infers adrenergic (ADRN) and mesenchymal (MES) cell states directly from high-dimensional morphological fingerprints without reliance on transcriptomic measurements. By benchmarking against single-cell RNA sequencing (scRNA-seq), we demonstrate that morphology-defined states closely align with transcriptomic profiles at single-cell resolution. We further show that cell state transitions are represented as continuous trajectories within a morphology-defined state space. Perturbations targeting distinct regulatory layers, including ROCK signaling and epigenetic regulation via EZH2, drive convergent trajectories along a shared phenotypic axis. Together, these results establish cell morphology as a scalable and non-destructive readout of cell state with machine learning providing a unified framework for high-throughput phenotyping and real-time tracking of cancer cell state plasticity.