Reducing cost of fluorescent microscopy through dye exclusion, distillation techniques, and deep learning

Abstract High Content Screening (HCS) via the Cell Painting assay is a powerful drug discovery tool, but the requirement for five fluorescent channels significantly increases the cost and complexity of image acquisition. Current deep learning models for HCS are trained and evaluated on all five channels, and it remains unclear whether robust biological representations can be maintained when channels are removed at inference time. This work introduces Channel-Reduced DINO (CR-DINO), a self-distillation method built on the DINO framework that generates rich HCS image representations from a reduced number of channels. CR-DINO employs a teacher-student architecture in which the teacher retains full five-channel visibility while the student receives progressively fewer channels following a curriculum-inspired schedule. This scheduled reduction, moving from five channels down to two over the course of training, encourages the student to learn cross-channel dependencies rather than relying on information directly available in its input. Experiments on the Bray dataset, validated through Mode of Action prediction, biological activity property prediction, and image–structure retrieval tasks, show that using only two channels (DNA and Mito) provides results on par with those using the full set, and that CR-DINO recovers performance even for the least informative channel pair (ER and AGP). These results highlight the potential of our method to significantly reduce the reagent and acquisition costs of consecutive HCS experiments without sacrificing the depth of biological insights.