DeepCyTOF: Automated Cell Classification of Mass Cytometry Data by Deep Learning and Domain Adaptation

摘要: Mass cytometry or CyTOF is an emerging technology for high-dimensional multiparameter single cell analysis that overcomes many limitations of fluorescence-based flow cytometry. New methods analyzing data attempt to improve automation, scalability, performance, and interpretation generated in large studies. However, most current tools are less suitable routine use where must be standardized, reproducible, interpretable, comparable. Assigning individual cells into discrete groups types (gating) involves time-consuming sequential manual steps untenable larger The subjectivity gating introduces variability the impacts reproducibility comparability results, particularly multi-center FlowCAP consortium was formed address these issues it aims boost user confidence viability automated methods. We introduce DeepCyTOF, a standardization approach based on multi-autoencoder neural network. DeepCyTOF requires labeled from only sample. It domain adaptation principles generalization previous work allows us calibrate between source distribution (reference sample) multiple target distributions (target samples) supervised manner. apply two datasets primary immune blood cells: (i) 14 subjects with history infection West Nile virus (WNV), (ii) 34 healthy different ages. Each sample 42 antibody markers, 12 which were used our analysis, at baseline three stimuli (PMA/ionomycin, tumor line K562, WNV). In each we manually gated reference automatically gate remaining uncalibrated samples. show classification highly concordant obtained by over 99% concordance. Additionally, stacked autoencoder, one building blocks 4th challenge FlowCAP-I competition demonstrate performs relative all introduced this competition. conclude autoencoders combined procedure offers powerful computational semi-automated such sufficient accurately