High‐dimensional flow cytometry data: goldmine or fool's gold?

Cytometry Part A, vol. 105, pp. 425–427

Abstract

Since its inception in the 1970s flow cytometry has been a valuable tool to study the characteristics of cells. Technical advancements lead to a rapid increase in the number of lasers and detectors, enabling assessment of an expanding number of parameters. As every flow cytometry user has experienced the frustration of not being able to fit all markers of interest into their panel, the prevailing mindset seemingly became the bigger the better. So, when spectral flow cytometers became commercially available, their ability to combine spectrally similar fluorophores, measure cellular autofluorescence and utilize the full light spectrum more effectively was received with enthusiasm. These attributes make spectral cytometry particularly applicable for limited or precious (clinical) material. The recent paper from Dott et al. describes the development of a standardized protocol for sample handling, staining and acquisition, for the application of two large spectral panels in a cohort study using the ID7000™ Spectral Cell Analyzer (Sony Biotechnology).1 The authors specifically address the repeatability and reproducibility of staining over time, which is relevant in the context of cohort studies. By combining a 34- and 35-color spectral panel, the authors were able to quantify and identify 182 cell phenotypes in whole blood samples. Our expanding immunological knowledge of rare and unique subsets through deep profiling has necessitated the increasing number of cellular parameters required in a single panel for accurate cell identification. However, a new set of challenges arise with increasing panel size. Although the principle of panel design and analysis remain similar, assessing data quality, defining cell populations and translating phenotypic changes into biological insight become increasingly difficult. While many papers have addressed one of these challenges, Dott et al. touch upon all three of these concerns.1 It quickly became apparent that full spectral profiling has strict requirements for accurate single stain reference controls and fluorophore combinations. As a result, a large (spectral) cytometry panel is not something that can be improvised but requires rigorous testing and optimization, and an associated substantial financial and time investment. Great efforts have been made to develop publicly available resources to improve technology accessibility and data quality. For instance, assignment of chosen markers to appropriate fluorophores, supported by knowledge of fluorophore characteristics and spectral spreading of a particular instrument, is discussed by Nguyen et al.2 Tools such as spillover spreading matrix and complexity index can also be used to determine optimal fluorophore combinations. Their resolution is assessed following titration and acquisition of appropriate single-color controls. Ferrer-Font et al. describe all necessary steps to go from a theoretically designed panel to an appropriately unmixed multicolor sample.3 Imperative to this, was the appropriate use of single-color cell controls rather than beads. This is because specific fluorophores undergo spectral shifting when bound to beads, resulting in unmixing errors in the multicolour cell samples. This detailed protocol was important to the development of the first 40 color full spectrum panel.4 Additional challenges arise for tissue-derived samples with heterogenous autofluorescent signatures across experimental groups and within a single sample type. Spiteri et al., demonstrates best practices for unmixing with multiple and highly autofluorescent populations.5 When the different signatures are considered and appropriately included in data unmixing, this provides an opportunity to gain extra information on these cell subsets as well as substantially improve population resolution.5, 6 Underlying these studies, is the continued importance of manually assessing unmixing errors and data quality on clean data (i.e., non-debris, single and live cells) to help assess best fluorophore combinations, use of beads vs cells for single color reference controls and appropriate autofluorescence extractions. Large panels are ideally suited for clinical cohorts, as they increase the amount of information acquired from a single run and limited material. However, such studies often have large sample sizes with several timepoints requiring multiple staining runs over weeks or months, introducing batch error. Thus, detailed technical planning is required to ensure consistent sample handling and staining procedures for improved reproducibility. Dott et al. developed a standardized protocol for sample handling, staining and acquisition, for the application of two large spectral panels in a cohort study.1 They describe extensive testing of staining and acquisition reproducibility at five distinct timepoints using both fresh and frozen whole blood and PBMCs, to map the differences resulting from practical choices being made at the point of sample collection. If all preparation still fails to yield comparable data batches, several steps can be taken remove data of insufficient quality and perform batch normalization to allow data combination for analysis. These issues are also applicable to (large) mass cytometry studies making approaches developed for mass cytometry data applicable to (spectral) flow cytometry data.7 A full protocol including steps for data transformation, data quality control and batch effect correction can be found in Den Braanker et al.8 Thus, technical considerations pertinent to developing large panels are being addressed. Once samples have been accurately unmixed, data analysis and interpretation are the next frontier. While manual analysis is still possible for large panels, it becomes increasingly time consuming and difficult to identify novel subsets with unique expression patterns. To aid data exploration, many algorithms have been developed. Generally, a combination of dimensionality reduction and clustering algorithms such as UMAP and FlowSOM are used to visualize the different cell types present in a sample.9, 10 Expression patterns can then be shown in single marker overlay plots or heatmaps for identified clusters. While this is very helpful to explore data and gain an understanding of the cell types that can be found in the data, human interpretation remains necessary. Confirmation through manual gating is required to ensure populations are not a result of unmixing error, technical error or inappropriate settings of dimensionality reduction or clustering algorithms. For example, an elbow plot can determine the number of theoretically ideal clusters within a dataset based on the number of cells and markers. However, it does not take into consideration whether a cluster is functionally or biologically distinct enough to be deemed a separate population. While describing the identifying markers of a cellular phenotype is relevant and necessary, this should not be the only description provided. Splitting samples into small enough clusters will allow identification of differential abundance between groups, regardless of whether these are functionally distinct populations. A minimum population size of 50–100 cells was an unofficial standard often adhered to by flow cytometrists. With increasing panel size and rare population identification, new standards for what constitutes a legitimate population should be implemented. The number of individuals/samples that a cluster can be found in may be a factor to consider, to avoid identification of unique clusters in small studies, that cannot be reproduced in other studies or cohorts or even upon reanalysis of the same dataset. Furthermore, clusters identified as different between groups of interest should always be checked by manual gating to exclude the possibility that they are a result of artifacts in the data, or such minor differences in marker expression that it is unlikely that they are functionally different. As shown by Dott et al. this manual gating can also help relate these clusters to known populations and thereby aid interpretation. A discussion on appropriate statistical analysis of clustered data may be necessary to separate signal from noise. Is it appropriate to perform a t-test on abundance of a cluster of interest between two groups, which understandably is commonly done, when this data is part of a much larger dataset? With the rapid development of analysis methods, the field may need to determine whether incorporation of multiple test corrections or weighted analysis is desirable to contribute to the validity and repeatability of data. With all the challenges and sometimes overwhelming amounts of data that come with high-dimensional flow cytometry data, it is easy to lose sight of the intended goal, namely understanding cellular function. Often, limited information is provided on what a specific phenotype or cluster identified in a study may mean for the outcome of disease. Of course, when describing the discovery of a novel population that is different for a specific group, it is not always possible to include functional data on its likely role. However, at the very least a hypothesis based on literature should be offered for how these cells may contribute to phenotype of the study population. This will help keep the focus on increasing our understanding of the link between the abovementioned data and relevant biological processes, rather than the production of big data for the sake of big data. In summary, the possibilities of big panels and what this can mean for our understanding of the role of cell populations is thrilling. Building and applying panels of this size comes with challenges at all stages but also with many opportunities to gain much more insight into cellular function. The recent paper by Dott et al. addresses many of the technical challenges to apply spectral flow cytometry to large cohorts and obtain reproducible data. They also identify their 182 phenotypes by manual gating but indicate their approach to unsupervised analysis is still under development. This paper shows the growing awareness of the challenges of big flow cytometry data and the ongoing efforts to address them as well as those that still need to be addressed. With great power comes great responsibility, and more emphasis must be put on ensuring that the populations identified are functionally distinct. This will ensure that the data obtained contributes to our understanding of biological processes instead of merely becoming a race to identify the rarest population using the biggest panel. Paula Niewold: Conceptualization; writing – original draft. The peer review history for this article is available at https://www.webofscience.com/api/gateway/wos/peer-review/10.1002/cyto.a.24849.

Authors 1

1. Paula Niewold corresponding Aachen

   Leiden University · Leiden University Medical Center

   Affiliation as printed

   Leiden University Center for Infectious Diseases Leiden University Medical Centre Leiden The Netherlands

   Leiden University Center for Infectious Diseases, Leiden University Medical Centre, Leiden, The Netherlands