Year: 2020
Type: editorial
Abstract: Recent advances in instrumentation have led to a wealth of novel single-cell technologies, many of which generate large and high-dimensional datasets. Examples include high-parameter flow cytometry (Chattopadhyay et al. 2019), mass cytometry (Bandura et al. 2009) and recently also single-cell sequencing techniques (Kolodziejczyk et al. 2015). Depending on the technology, some techniques are able to provide measurements for thousands of cells, while others can easily measure millions of cells per sample. Also imaging techniques have been rapidly evolving, some of them being able to provide single-cell measurements, e.g. using segmentation techniques. Hybridizations between imaging and cytometry techniques have also emerged as novel tools to combine single-cell multi-parameter measurements with spatial information. These include techniques such as imaging flow cytometry (McGrath et al. 2008), imaging mass cytometry (Giesen et al. 2014), and spatial “omics” techniques (Ståhl et al. 2016). All of these techniques have one thing in common: they require statistical and machine learning techniques to interpret the data and extract novel biological knowledge from it (Saeys et al. 2016). Machine learning techniques are data-driven modeling techniques that aim to learn a model for the data without being explicitly told how to do so. Typically, a specific machine learning task is specified as an optimization problem, and a wide array of techniques can typically be used to solve these optimization problems, all with their own advantages, disadvantages and assumptions they make regarding the data. Machine learning tasks are typically divided into two large categories: unsupervised learning techniques, and supervised learning techniques. Unsupervised learning techniques can be used to learn the structure that is present in the data. Common examples include dimensionality reduction and visualization techniques, clustering techniques, and more recently also the novel class of trajectory inference methods. All of these techniques require no additional information than just the data points, which in the case of single-cell data can be usually represented as a matrix of cells and their properties measured. Supervised learning techniques on the other hand aim to learn a function between the single-cell data and an external variable that should be predicted as well as possible from the single-cell data. This external variable can be cell-specific or biological sample specific information. In single-cell analysis, we generally see two broad applications: identification and classification of cell types from cell-level measurements (e.g. protein, RNA, or image features), and identification and classification of biological samples (e.g. predicting disease status, dose of a drug, or patient survival), often from differential abundance of inferred cell types and other features. Both applications are presented in this special issue. An important prerequisite for machine learning techniques, and data mining in general, is that the input data should be of sufficient quality (the “garbage in, garbage out” principle always applies). In the case of single-cell data, good quality control and preprocessing is a cornerstone on which subsequent machine learning models are built. However, since many of the novel single-cell technologies are still in development, adequate methods for quality and preprocessing are also actively developed alongside them. Thus, robust benchmarks(Saelens et al. 2019; Aghaeepour et al. 2016; Weber and Robinson 2016; Tran et al. 2020) are critical to adequately evaluate both the technology and the methods of analysis. This special issue of Cytometry A features some of the most recent research in machine learning techniques for single cell data. A first issue that is tackled by the authors in this issue is quality control and normalization. Van Gassen et al. present CytoNorm, a novel normalization algorithm to correct for batch effects when dealing with clinical samples across various study batches. The authors propose to use shared controls to learn the appropriate transformations for each batch, and subsequently apply these transformations to the other clinical samples in the batch to remove batch effects. An important message in this work is that batch effects could be cell-type specific, and the authors thus propose a solution based on clustering that learns the cell-type specific transformations. One of the first applications of machine learning in cytometry was the development and application of a wide range of clustering algorithms for automated gating techniques (Van Gassen et al. 2015; Finak et al. 2014; Levine et al. 2015; Ge and Sealfon 2012; Finak et al. 2009; Lo et al. 2009; Chen et al. 2016; Shekhar et al. 2014; Becher et al. 2014; Sörensen et al. 2015; Zare et al. 2010; Linderman et al. 2012; Naim et al. 2010), many of which were evaluated and benchmarked by Weber and Robinson (Weber and Robinson 2016). In this issue, Bligaard Pedersen and Olsen provide some cautionary words when blindly trusting the output of automated gating techniques. The authors stress the importance of using multiple tools to get a good picture of what is going on in the data and advocate the use of more semi-supervised approaches, reinforcing the early work of the FlowCAP challenges (Aghaeepour et al. 2013; Scheuermann et al. 2013). In another article on automated population identification, Ji et al. propose a new algorithm to learn interpretable gates in a supervised fashion when class label information is available. They develop new methods that enable smooth gradient-based optimization of discrete single-cell classification cost functions, and, complementary to existing, unsupervised automated gating techniques, their algorithm is able to learn hyper-rectangular regions in marker space that are optimized for classification of cytometry samples with regard to the label information available at the patient level. In a subsequent article by Lucchesi et al., the authors discuss the use of computational flow cytometry methods for vaccine studies. The authors combine classical, existing automated gating techniques based on clustering (FlowSOM) with tSNE visualization of cells to interpret undefined phenotypes of B cell subsets elicited upon vaccination. Using this pipeline, they identified different B cell populations, including plasmablasts, plasma cells and germinal center B cells. They find that in particular plasma cells were more abundant in mice primed with the adjuvant, and also highlight intermediate stages of cell differentiation elicited by immunization that could be more clearly described using the machine learning based approaches. A subsequent series of articles in this special issue discusses novel machine learning challenges for image-based single-cell data. Since deep learning methods have been particularly successful for image analysis. Deep learning methods have been particularly successful for image analysis, and this application has also impacted the single-cell field. Su et al. provide an overview of deep learning based techniques for single-cell optical image analysis. They cover novel learning paradigms such as transfer learning, multimodal learning, multitask learning, and end-to-end learning, and describe applications such as image segmentation, super-resolution image reconstruction, cell tracking, cell counting, cross-modal image reconstruction, and design and control of cell imaging systems. Several articles in this issue also develop and apply deep learning and other machine learning techniques, in particular to imaging flow cytometry data. Mochalova et al. describe a novel application of using convolutional neural networks (CNNs) to quantitatively measure particle targeting of cells. They conclude that CNNs in this case allow high-throughput object detection with close-to-human precision. In addition, these methods allow more objective image analysis as they do not depend on subjective choices of parameter settings when using classical image processing techniques. In another article, Lippeveld et al. compare classical machine learning approaches to deep learning approaches for the problem of white blood cell classification. In this article, the authors also work on imaging flow cytometry data, and evaluate whether machine learning approaches can distinguish different cell types using only the label-free information available. They conclude that for some cell types it is possible to distinguish them using label-free information, but deep learning techniques in this case did not outperform classical machine learning approaches based on manually engineered features. A different application of machine learning for imaging flow cytometry is presented by Piasecka et al. In this article, the authors explore the potential of trajectory inference methods to map the activation state of eosinophils. Using features from dark-field images they apply diffusion maps to order the cells based on similar image features. They validate their approach by comparing to exogenous cell activation using eotaxin and endogenous activation, and show that the diffusion map correlates very well with the measured activation level. Trajectory inference methods are also the topic of the work by Verrou et al., where the authors aim to learn the signalling dynamics present in mass cytometry data. They propose a number of new evaluation metrics to quantify the performance of learning signalling dynamics, and show that the inferred trajectories provide more insights. They also present a new method based on Diffusion Maps and Principal Curves that showed great promise to model the progression of signal transduction. To support all machine learning related research for spatial analysis of cell images, Rose et al. present PySpacell, a Python package that aims to measure spatial effects of adherent cells in microscopy images. This toolbox can be used for light microscopy images of cells as well as other types of single-cell data such as in situ transcriptomics or metabolomics. The authors took care to make their tool compatible with other common analysis tools such as CellProfiler, Fiji, and Icy. Such efforts are critical to make complex algorithmic tools available to a wide audience of researchers. A final note on deep learning is provided by Michal Kozubek, who reflects on the potential of deep learning, in particular the use of generative models, for cell image synthesis, the goal of which is to render realistic, but synthetic cell images.
Source: Cytometry Part A
DOI: 10.1002/cyto.a.23986
Authors: Yvan Saeys, Greg Finak
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: bronze
OpenAlex ID: W3010601552