Analysis of the Human Protein Atlas Image Classification competition

Ouyang, Wei; Winsnes, Casper F.; Hjelmare, Martin; Cesnik, Anthony J.; Åkesson, Lovisa; Xu, Hao; Sullivan, Devin P.; Dai, Shubin; Lan, Jun; Park, Jinmo; Galib, Shaikat M.; Henkel, Christof; Hwang, Kathleen; Poplavskiy, Dmytro; Tunguz, Bojan; Wolfinger, Russel D.; Gu, Yinzheng; Li, Chuanpeng; Xie, Jinbin; Buslov, Dmitry; Fironov, Sergei; Kiselev, Alexander; Panchenko, Dmytro; Cao, Xuan; Wei, Runmin; Wu, Yuanhao; Zhu, Xun; Tseng, Kuan-Lun; Gao, Zhifeng; Ju, Cheng; Yi, Xiaosu; Zheng, Hongdong; Kappel, Constantin; Lundberg, Emma

doi:10.1038/s41592-019-0658-6

Cited by 97 publications

(104 citation statements)

References 42 publications

Supporting

Mentioning

103

Contrasting

Order By: Relevance

“…The algorithm performed almost as accurately as human experts, and with greater speed and reproducibility. Furthermore, it could quantify the spatial information 8 . "When we can quantify it, and not just describe it with a label, we can integrate it with other types of data."…”

Section: The Spatial Dimensionmentioning

confidence: 99%

Deep learning takes on tumours

Landhuis¹

2020

Nature

View full text Add to dashboard Cite

Section: The Spatial Dimensionmentioning

confidence: 99%

Deep learning takes on tumours

Landhuis¹

2020

Nature

View full text Add to dashboard Cite

“…While the number of studied proteins could be expanded by pooling additional IF and AP-MS datasets gathered in disparate cell types and conditions, we considered the importance of a controlled cellular context in prototyping any new approach. Using a deep convolutional neural network trained to recognize patterns in IF images 18 , we embedded each protein as a 1024-dimension feature vector, capturing its spatial distribution relative to counter-stained cellular landmarks (Methods). Similarly, the node2vec deep neural network 20 was used to embed each protein as a second 1024-dimension feature vector based on its interaction neighborhood within the AP-MS data, including directly and indirectly associated proteins (Methods, Extended Data Fig.…”

Section: Protein Position and Distance Two Waysmentioning

confidence: 99%

“…IF images interrogating protein locations in the HEK293 cell line were downloaded from the HPA Cell Atlas 5,18,19 . Physical protein interactions detected by AP-MS in the HEK293T cell line were downloaded from the BioPlex 2.0 protein interaction database 6 .…”

Section: Data Sourcesmentioning

confidence: 99%

“…7a). DenseNet: For each image, a 1024-dimension feature embedding was generated from a DenseNet-121 model 48 optimized for annotating HPA images as previously described 18 . This method uses the protein channel and all three reference channels.…”

Section: Data Embeddingsmentioning

confidence: 99%

“…positioning has become increasingly quantitative in recent years, based on the ability of deep learning systems to recognize complex patterns in data [18][19][20][21][22][23] . Once an IF or AP-MS dataset has been analyzed to determine protein positions (or position distributions), a natural next step is to compute distances between these distributions, raising the possibility that the two types of protein distance might be calibrated and combined within integrated maps of cellular architecture (Fig.…”

mentioning

confidence: 99%

See 2 more Smart Citations

Mapping cell structure across scales by fusing protein images and interactions

Qin

Winsnes

Huttlin

et al. 2020

Preprint

Self Cite

View full text Add to dashboard Cite

The eukaryotic cell is a multi-scale structure with modular organization across at least four orders of magnitude 1,2 . Two central approaches for mapping this structure -protein fluorescent imaging and protein biophysical association -each generate extensive datasets but of distinct qualities and resolutions that are typically treated separately 3,4 . Here, we integrate immunofluorescent images in the Human Protein Atlas 5 with ongoing affinity purification experiments from the BioPlex resource 6 to create a unified hierarchical map of eukaryotic cell architecture. Integration involves configuring each approach to produce a general measure of protein distance, then calibrating the two measures using machine learning. The evolving map, called the Multi-Scale Integrated Cell (MuSIC 1.0), currently resolves 69 subcellular systems of which approximately half are undocumented. Based on these findings we perform 134 additional affinity purifications, validating close subunit associations for the majority of systems. The map elucidates roles for poorly characterized proteins, such as the appearance of FAM120C in chromatin; identifies new protein assemblies in ribosomal biogenesis, RNA splicing, nuclear speckles, and ion transport; and reveals crosstalk between cytoplasmic and mitochondrial ribosomal proteins. By integration across scales, MuSIC substantially increases the mapping resolution obtained from imaging while giving protein interactions a spatial dimension, paving the way to incorporate many molecular data types in proteome-wide maps of cells.Advances in confocal microscopy and immunofluorescence (IF) imaging have created systematic pipelines for mapping the spatial distribution of proteins and other molecules within single cells 3,7,8 . Based on these techniques, the Human Protein Atlas (HPA) has launched an extensive effort to map protein subcellular locations using a library of specific fluorescent antibodies targeting more than 13,000 human proteins 5,9 . The use of multiple dyes with separate emission spectra enables locations to be determined relative to known landmarks such as the nucleus, cytoskeleton and endoplasmic reticulum, with the result that most human proteins can be assigned relative positions at sub-micron resolution.In parallel, advances in mass spectrometry (MS) have provided a complementary means of mapping protein coordinates through their biophysical associations with other proteins 4,10 . MS is now routinely combined with affinity purification (AP-MS) 11 or proximity-dependent labeling 12-16 to enumerate physical protein-protein interactions in vitro or in vivo. Combining AP-MS with epitope tagging, the BioPlex project has generated a systematic map of physical interactions covering approximately 7,500 epitope-tagged human proteins (BioPlex 2.0) 6 , including approximately 56,000 candidate interactions organized into more than 400 multimeric protein complexes.Given that protein imaging and biophysical association are leading approaches for mapping cell structure, with growing data resou...

show abstract

High‐Throughput Automated Subcellular Localisation

Chalkley

Kelly

Mysior

et al. 2020

Encyclopedia of Life Sciences

View full text Add to dashboard Cite

Defining the subcellular localisation of the proteome for an organism of interest is a critical next step following genome sequencing. Knowledge of protein subcellular localisation provides insight into the functionality of the normal cell, as well during disease states. However, the presence of gene isoforms, alternative splicing and posttranslational modifications significantly increase the number of protein variants encoded by a single gene, making this a complex task. In the last 20 years, parallel approaches using fractionation and mass spectrometry, synthesis of large libraries of open reading frames fused to genes encoding fluorescent proteins, as well as production of thousands of antibodies have all contributed to the systematic analysis of protein localisation. Alongside these methods, improved bioinformatic predictors, machine learning and deep learning algorithms have also evolved as essential tools. A combinatorial approach of these methods now brings us close to systematically defining the subcellular proteome for many organisms. Key Concepts Subcellular localisation is a critical determinant in understanding protein function. Data from genome sequencing projects provide the fundamental information from which approaches to understand protein localisation can be initiated. Parallel approaches using fluorescence microscopy are being applied in a high‐throughput manner to systematically reveal the subcellular localisation of large numbers of proteins in different cells. The primary techniques to determine protein localisation are mass spectrometry‐based proteomics, production of antibodies and expression of fluorescently tagged proteins. There is increasing use of computational biology tools to aid the automated classification of subcellular localisation. Large image datasets can be interrogated by machine learning software algorithms to automatically classify proteins to specific localisations. Deep learning methods, which can work independently of training datasets, have become the newest tool to automatically assign protein localisation from image sets. Automated approaches combining both experimental and computational methods are likely to become the primary means by which subcellular localisation is determined from new cell systems.

show abstract

Analysis of the Human Protein Atlas Image Classification competition

Cited by 97 publications

References 42 publications

Deep learning takes on tumours

Deep learning takes on tumours

Mapping cell structure across scales by fusing protein images and interactions

High‐Throughput Automated Subcellular Localisation

Contact Info

Product

Resources

About