Salem, O. | Search bioRxiv

bioRxiv2023

Scalable querying of human cell atlases via a foundational model reveals commonalities across fibrosis-associated macrophages

Single-cell RNA-seq (scRNA-seq) studies have profiled over 100 million human cells across diseases, developmental stages, and perturbations to date. A singular view of this vast and growing expression landscape could help reveal novel associations between cell states and diseases, discover cell states in unexpected tissue contexts, and relate in vivo cells to in vitro models. However, these require a common, scalable representation of cell profiles from across the body, a general measure of their similarity, and an efficient way to query these data. Here, we present SCimilarity, a metric learning framework to learn and search a unified and interpretable representation that annotates cell types and instantaneously queries for a cell state across tens of millions of profiles. We demonstrate SCimilarity on a 22.7 million cell corpus assembled across 399 published scRNA-seq studies, showing accurate integration, annotation and querying. We experimentally validated SCimilarity by querying across tissues for a macrophage subset originally identified in interstitial lung disease, and showing that cells with similar profiles are found in other fibrotic diseases, tissues, and a 3D hydrogel system, which we then repurposed to yield this cell state in vitro. SCimilarity serves as a foundational model for single cell gene expression data and enables researchers to query for similar cellular states across the entire human body, providing a powerful tool for generating novel biological insights from the growing Human Cell Atlas.

bioinformatics↗

bioRxiv2020

Label2label: Using deep learning and dual-labelling to retrieve cellular structures in fluorescence images

Fluorescence microscopy is an essential tool in cell biology to visualise the spatial distribution of proteins that dictates their role in cellular homeostasis, dynamic cellular processes, and dysfunction during disease. However, unspecific binding of the antibodies that are used to label a cellular target often leads to high background signals in the images, decreasing the contrast of a cellular structure of interest. Recently, convolutional neural networks (CNNs) have been successfully employed for denoising and upsampling in fluorescence microscopy, but current image restoration methods cannot correct for background signals originating from the label. Here, we report a new method to train a CNN as content filter for non-specific signals in fluorescence images that does not require a clean benchmark, using dual-labelling to generate the training data. We name this method label2label (L2L). In L2L, a CNN is trained with image pairs of two non-identical labels that target the same cellular structure of interest. We show that after L2L training a network restores images not only with reduced image noise but also label-induced unspecific fluorescence signal in images of a variety of cellular structures, resulting in images with enhanced structural contrast. By implementing a multi-scale structural similarity loss function, the performance of the CNN as a content filter is further enhanced, for example, in STED images of caveolae. We show evidence that, for this loss function, sample differences in the training data significantly decrease so-called hallucination effects in the restorations that we otherwise observe when training the CNN with images of the same label. We also assess the performance of a cycle generative adversarial network as a content filter after L2L training with unpaired image data. Lastly, we show that a CNN can be trained to separate structures in superposed fluorescence images of two different cellular targets, allowing multiplex imaging with microscopy setups where the number of excitation sources or detectors is limited.

bioinformatics↗