Clustering of high-content screen images to discover off-target phenotypes
summary
Claims that simple, unsupervised feature summaries plus clustering reveal off‑target cellular phenotypes missed by one‑dimensional scores, demonstrating coherent clusters on real screens (a demo subset of 1,700 images within a full screen of 45,000 images and up to half a million in newer sets). Implementation uses scikit-image, pandas, MongoDB and mini-batch k‑means with precomputed nearest neighbors and a notebook/ML‑D3 visualization called husk; parts run and produce useful clusters though some live demo pieces and preprocessing still need polish.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:00What is highcontent screening
- 5:20Map function
- 7:20Feature map
- 9:20Scanning images
- 16:30Empty LD3
- 17:05Conclusions
- 19:50Questions
related talks
25mThe Art of Clustering: The Good, The Bad and The Beautiful
03h 51mMachine Learning Summit
47mBANKSY: A Spatial Omics Algorithm that Unifies Cell Type Clustering and Tissue Domain Segmentation
27mKubernetes + AI Joining Forces in the Battle Against Cancer
22mOne-click analysis: Detecting and visualizing insights automatically
50mBuilding Intelligent Mobile Apps using Machine Learning
30mWords in Space
01h 00mData, visualization, and designing AI