High-content imaging (HCI) enables the characterization of cellular states through the extraction of quantitative features from fluorescence microscopy images. Despite the widespread availability of HCI data, the development of generalizable feature extraction models remains challenging due to the heterogeneity of microscopy images, as experiments often differ in channel count, cell type, and assay conditions. To address these challenges, we introduce uniDINO, a generalist feature extraction model capable of handling images with an arbitrary number of channels. We train uniDINO on a dataset of over 900,000 single-channel images from diverse experimental contexts and concatenate single-channel features to generate embeddings for multi-channel images. Our extensive validation across varied datasets demonstrates that uniDINO outperforms traditional computer vision methods and transfer learning from natural images, while also providing interpretability through channel attribution. uniDINO offers an out-of-the-box, computationally efficient solution for feature extraction in fluorescence microscopy, with the potential to significantly accelerate the analysis of HCI datasets.
