Os dai atlas11/9/2023 ![]() ![]() ![]() ![]() Different variations of convolutional neural networks (CNNs) or fully connected neural networks are effective at solving segmentation and multi-class classification problems 13, 14, 15, 16, 17. For example, class imbalance can be tackled at the data level, for example, by oversampling rare classes or under-sampling common classes 9, or at the classifier level, for example, by applying class weights 10, 11 or using novelty detection 12. Recent advances in artificial intelligence, especially deep neural networks (DNNs), have provided powerful algorithms to undertake these challenges. At the same time, the availability of image-level labels is much greater, which supports the need for weakly supervised models making use of image-level annotations. Third, training supervised models requires ground truth annotations of every single cell in a large set of images, which is time-consuming and costly. Second, the distribution of proteins in these compartments is highly disproportionate by up to five orders of magnitude (HPAv21) 6, causing an extreme class imbalance. First, around 55% of human proteins localize to multiple subcellular compartments, indicating that they might be involved in several biological processes 6, 7, 8, making it a multi-label classification problem. Classifying the subcellular protein locations from these images is currently hampered by various technical challenges. The Human Protein Atlas (HPA) Subcellular Section has generated the first subcellular proteome map of human cells, consisting of a publicly available dataset of 83,762 confocal microscopy images detailing the subcellular localization of 13,041 proteins (HPAv21) 6 across a multitude of cell lines. The increasing amounts of fluorescent image data necessitate better computational models that are capable of classifying the spatial protein distribution of single cells and ultimately enabling the investigation of how protein spatial regulation contributes to cellular function in health and disease. Tremendous technological progress in microscopy enables increasingly data-rich descriptions of cellular properties, including subcellular protein distribution. Therefore, protein subcellular localization needs to be attributed at the single-cell level. The expression and localization of proteins are well known to vary between healthy human cell types 4, 3, even within genetically identical cell populations 5. While ongoing cell mapping efforts 1, 2, 3 are mainly based on single-cell sequencing technologies, protein localization is crucial to the understanding of biological networks. The function of cellular systems is predominantly defined by the structure, amount, spatial location and interactions of individual proteins that collectively make up the proteome. Nature Methods volume 19, pages 1221–1229 ( 2022) Cite this article Analysis of the Human Protein Atlas Weakly Supervised Single-Cell Classification competition ![]()
0 Comments
Leave a Reply.AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |