ӳ����ý

David Van Valen
Division of Biology and Biological Engineering, California Institute of Technology
Meeting: Single-cell biology in a software 2.0 world

Multiplexed imaging methods can measure the expression of dozens of proteins while preserving spatial information. While these methods open an exciting new window into the biology of human tissues, interpreting the images they generate with single cell resolution remains a significant challenge. Current approaches to this problem in tissues rely on identifying cell nuclei, which results in inaccurate estimates of cellular phenotype and morphology. In this work, we overcome this limitation by combining multiplexed imaging’s ability to image nuclear and membrane markers with large-scale data annotation and deep learning. We describe the construction of TissueNet, an image dataset containing more than one million paired whole-cell and nuclear annotations across eight tissue types and five imaging platforms. We also present Mesmer, a single model trained on this dataset that can perform nuclear and whole cell segmentation with human-level accuracy – as judged by expert human annotators and a panel of pathologists – across tissue types and imaging platforms. We show that Mesmer accurately measures cell morphology in tissues, opening up a new observable for quantifying cellular phenotypes in vivo. We make this model available to users of all backgrounds with both cloud-native software and on-premise software. Last, we also describe ongoing work to develop a similar resource and models for dynamic live-cell imaging data.

Emily Laubscher
Van Valen Lab, California Institute of Technology
Primer: Integrating heterogeneous measurements in single cells

Recent advances in imaging and machine learning have increased our ability to capture information about biological systems in the form of images. Given that spatial genomics allows us to capture both the “parts list” and spatial variation in living systems, images have the potential to be a universal image type for biology. In this talk, we describe the field’s progress towards such a vision. We describe our experience using live-cell imaging and end-point spatial genomics to integrate heterogeneous measurements in single cells and highlight the common computer vision challenges raised by this approach. We discuss how these problems might be solved using modern deep learning methods. We conclude by describing our latest work on using weakly supervised learning to perform spot detection in multiplexed RNA FISH experiments.