Fig. 6: Zero-knowledge pooled optical screening analysis.
From: NIS-Seq enables cell-type-agnostic optical perturbation screening
a, p65–mNeonGreen channel images of 712,146 HeLa cells stimulated with IL-1β were embedded using the SwinV2-T computer vision model pre-trained on a general image dataset, k-means clustered based on 768 penultimate layer activations and visualized using UMAP dimensionality reduction. b, Cluster representation among top-ranking perturbed genes, identified by zero-knowledge image clustering. Perturbed genes covered by at least 15 cells were ranked by the maximum over representation of any cluster as compared to non-targeting control cells. c, Single-cell localizations of top-ranking outlier genes visualized on the same UMAP embedding as used in a.
