In this paper, we explore whether sparse dictionary learning (DL), which has emerged as a powerful method for extracting semantically meaningful concepts from large-scale language models (LLMs) trained on text data, can be applied to scientific data that are difficult to interpret by humans, such as vision-based models trained on cell microscopy images. We propose a novel method that combines a sparse DL algorithm, iterative codebook feature learning (ICFL), with a PCA whitening preprocessing step derived from control data. This successfully retrieves biologically meaningful concepts, such as cell types and genetic variations, and reveals subtle morphological changes caused by human-interpretable interventions, suggesting a promising new direction for scientific discovery through mechanistic interpretation of biological images.
