Author response: Comprehensive and unbiased multiparameter high-throughput screening by compaRe finds effective and subtle drug responses in AML models
Year: 2022
Type: peer-review
Abstract: Article Figures and data Abstract Editor's evaluation eLife digest Introduction Results Discussion Materials and methods Appendix 1 Data availability References Decision letter Author response Article and author information Metrics Abstract Large-scale multiparameter screening has become increasingly feasible and straightforward to perform thanks to developments in technologies such as high-content microscopy and high-throughput flow cytometry. The automated toolkits for analyzing similarities and differences between large numbers of tested conditions have not kept pace with these technological developments. Thus, effective analysis of multiparameter screening datasets becomes a bottleneck and a limiting factor in unbiased interpretation of results. Here we introduce compaRe, a toolkit for large-scale multiparameter data analysis, which integrates quality control, data bias correction, and data visualization methods with a mass-aware gridding algorithm-based similarity analysis providing a much faster and more robust analyses than existing methods. Using mass and flow cytometry data from acute myeloid leukemia and myelodysplastic syndrome patients, we show that compaRe can reveal interpatient heterogeneity and recognizable phenotypic profiles. By applying compaRe to high-throughput flow cytometry drug response data in AML models, we robustly identified multiple types of both deep and subtle phenotypic response patterns, highlighting how this analysis could be used for therapeutic discoveries. In conclusion, compaRe is a toolkit that uniquely allows for automated, rapid, and precise comparisons of large-scale multiparameter datasets, including high-throughput screens. Editor's evaluation This paper aims to address the current gap in the efficient analysis of large-scale multiparameter flow cytometry and other datasets. The authors offer a software toolkit with an efficient algorithm for comparing numerous samples at once. The study is well presented and is relevant to single cell analysis research. https://doi.org/10.7554/eLife.73760.sa0 Decision letter eLife's review process eLife digest Biology has seen huge advances in technology in recent years. This has led to state-of-the-art techniques which can test hundreds of conditions simultaneously, such as how cancer cells respond to different drugs. In addition to this, each of the tens of thousands of cells studied can be screened for multiple variables, such as certain proteins or genes. This generates massive datasets with large numbers of parameters, which researchers can use to find similarities and differences between the tested conditions. Analyzing these ‘high-throughput’ experiments, however, is no easy task, as the data is often contaminated with meaningless information, or ‘background noise’, as well as sources of bias, such as non-biological variations between experiments. As a result, most analysis methods can only probe one parameter at a time, or are unautomated and require manual interpretation of the data. Here, Chalabi Hajkarim et al. have developed a new toolkit that can analyze multiparameter datasets faster and more robustly than current methods. The kit, which was named ‘compaRe’, combines a range of computational tools that automatically ‘clean’ the data of background noise or bias: the different conditions are then compared and any similarities are visually displayed using a graphical interface that is easy to explore. Chalabi Hajkarim et al. used their new method to study data from patients with acute myeloid leukemia (AML) and myelodysplastic syndrome, two forms of cancer that disrupt the production of functional immune cells. The toolkit was able to identify subtle differences between the patients and categorize them into groups based on the proteins present on immune cells. Chalabi Hajkarim et al. also applied compaRe to high-throughput data on cells from patients and mouse models with AML that had been treated with large numbers of specific drugs. This revealed that different cell types in the samples responded to the treatments in distinct ways. These findings suggest that the toolkit created by Chalabi Hajkarim et al. can automatically, rapidly and precisely compare large multiparameter datasets collected using high-throughput screens. In the future, compaRe could be used to identify drugs that illicit a specific response, or to predict how newly developed treatments impact different cell types in the body. Introduction Technological developments have accelerated the generation of large-scale multiparameter screening data through methodologies such as high-content microscopy and high-throughput flow cytometry (Boutros et al., 2015; Saeys et al., 2016; Caraus et al., 2015). These technologies can test hundreds of samples (such as drug treatments) each with tens of thousands of events (e.g. cells) labeled for numerous biomarkers (such as cytoplasmic or membrane markers). However, analyzing this massive multiparameter data to provide an overview of similarities and differences between hundreds of samples is still a challenge (Boutros et al., 2015; Saeys et al., 2016; Caraus et al., 2015). This analytical challenge is further complicated by various sources of bias and noise often existing in the data, such as batch effect and signal drift (a gradual shift in the marker intensity across a multi-well plate) (Boutros et al., 2015; Saeys et al., 2016; Caraus et al., 2015). There have been efforts to cluster samples from large-scale multiparameter (multidimensional) screening data. A simple approach is to use a representative value for each cell marker such as median fluorescence intensity (MFI) for clustering samples (Cossarizza et al., 2019). However, using a single representative value can easily lead to loss of information about biologically relevant variance within and between cell subpopulations. Meta-clustering with single-cell clustering algorithms has been suggested to cluster samples based on the similarity of the centroids of cell subpopulations identified in the individual samples (Qiu et al., 2011; Levine et al., 2015; Van Gassen et al., 2015; Ogishi et al., 2021). While these algorithms are widely used in single-cell data analysis for clustering cells, they are not efficient for clustering of samples. This is because centroid-based analysis can be misleading when subclusters are not sufficiently distinct or the number of sub-clusters varies. Additionally, the heavy computing cost of meta-clustering makes it poorly suited for analyses of large datasets with many samples. Manual gating and machine learning based on prior knowledge have been used to cluster samples (Amir et al., 2019; Bruggner et al., 2014), but using prior knowledge for subpopulation identification can both lead to biased interpretations and failure to make de novo discoveries. Dimension reduction methods (Lvd and Hinton, 2008; Amir et al., 2013; McInnes et al., 2018) coupled with the Jensen-Shannon divergence (JSD) metric have also been used to cluster multidimensional samples (Amir et al., 2013). These algorithms including factor analysis and principal component analysis (PCA) still require excessive computing costs with an inherent information loss. It is also important to note that none of the methodologies developed so far efficiently correct for sources of bias and noise in large-scale multiparameter screening data. Available computational toolkits (BioScience E, 2020; Potdar et al., 2020; Boutros et al., 2006) mostly allow for single-parameter or unautomated analyses of large-scale screening data using the aforementioned methods. In these toolkits, each well should be first represented by a single parameter such as cell counts or centroids or they require manual intervention. To provide a useful toolkit for precise and effective interpretation of small- to large-scale multiparameter screening data, we developed compaRe. This toolkit has several unique modules for quality control, bias correction, pairwise comparisons, clustering, and data visualization. The quality control and bias correction modules can effectively reveal and remove various sources of bias in the screening data. compaRe clusters samples by measuring the similarity between them using a dynamic mass-aware gridding algorithm. This algorithm increases the robustness of the toolkit to the size of data and signal shift (a technical term referring to batch effect and signal drift), while guaranteeing fast clustering, as it does not bear the computing cost of dimension reduction and subsampling. The toolkit is available both as a command-line version and a graphical user interface (GUI) version that provides various visualizations to help with the interpretation of its readouts. compaRe performed robustly in the presence of background noise and batch effects even where these input data artifacts could not be corrected. compaRe analyses of multiparameter mass and flow cytometric data from acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS) patient samples revealed interpatient heterogeneity and recognizable phenotypic profiles. When applied to high-throughput flow cytometry of the dose response of AML samples treated with various drugs, compaRe successfully corrected for various sources of bias and clustered the samples based on their response to treatment, allowing for detection of both drastic and subtle phenotypic responses. Results compaRe is a comprehensive toolkit for multiparameter screening data compaRe is designed to analyze the data from small to large-scale multiparameter screening assays such as high-throughput flow cytometry, high-content microscopy, mass cytometry, and standard flow cytometry. The toolkit comprises several modules for quality control, bias correction, clustering, and visualization. Figure 1 shows the modules for a high-throughput flow cytometry of AML samples taken from a mouse model treated with various drugs. During quality control, several sources of bias such as autofluorescence, bioluminescence, carryover effect, edge effect, signal drift, and cell viability drift (drift in the number of live cells across the plate) were identified. The bias correction module could effectively correct for signal and cell viability drifts (two main sources of bias in high-throughput screening with fluorescent markers) using regression analysis (Figure 1, Materials and methods). Figure 1 Download asset Open asset compaRe is a comprehensive suite for multiparameter screening data. High-throughput flow cytometry generates massive multidimensional data from hundreds of samples. compaRe’s quality control (QC) module reveals several sources of bias in the assay such as signal (intensity difference between the top left and bottom right corners) and cell viability drifts. These two are corrected for in the bias correction modules within and between the plates. compaRe performs a pairwise similarity calculation between the samples using dynamic gridding and forming hypercubes (represented by distinct colors). The portions of the data within individual hypercubes are used to calculate similarity. Clustering is performed based on similarity. The graphical user interface (GUI) provides several ways to thoroughly explore and visualize the read-outs. At the core of the compaRe toolkit is a module for pairwise comparisons of samples. It measures the similarity between two samples using a dynamic mass-aware gridding algorithm (Figure 1, Materials and methods). Given two samples, the algorithm divides the higher dimensional space (formed by, for example, cell surface markers) of the samples individually into several spatial units called hypercubes. The average difference between proportions of data points present in corresponding hypercubes across the samples is used to represent similarity. In this setting, the module becomes robust to signal shift and data size difference between the two samples (Appendix 1). This module generates a similarity (affinity) matrix for the clustering module. The clustering module uses a graphical algorithm (Figure 1, Materials and methods). Initially, all nodes (samples) are connected forming a complete weighted graph wherein weights represent similarity values. The graph is then pruned to remove potential false positive edges using a threshold inferred from negative controls (untreated samples). After constructing a linked graph, clustering is tantamount to finding maximal cliques (complete subgraphs that cannot be extended), each containing samples with similar responses. compaRe benefits from parallel computing and modular design. Its modular design allows the modules to run independently; thus, the similarity and clustering modules of compaRe can be potentially applied to any problem space. compaRe is ultra-fast and robust to background noise and batch effect To evaluate the robustness of compaRe’s comparison module to noise and batch effect, we benchmarked it against JSD with UMAP (for simplicity just JSD) and meta-clustering with PhenoGraph (for simplicity just meta-clustering) (Levine et al., 2015). We analyzed the publicly available mass cytometry data of a total of 21 bone marrow aspirate samples collected from 16 pediatric AML patients and five healthy adult donors labeled for detection of 16 cell surface markers (Levine et al., 2015). We introduced random noise with Gaussian distribution to the 16 parameters of each sample to simulate a batch effect. In this setting, although the added noise undermines similarity, the overall cell population configuration remains intact, and consequently the simulated samples will still have the highest similarity with their original samples. Even with the added noise, the comparison module correctly identified similar samples (Figure 2a). Conversely, the batch effect seriously compromised the performance of both meta-clustering and JSD, showing several maximum similarities other than the originals (Figure 2b and C). In additional comparison with FlowSOM and SPADE, other commonly used tools for flow cytometry, compaRe’s performance far exceeded their performance (Appendix 1—figure 1). This result demonstrates the advantage of using dynamic gridding for comparison of samples in the presence of noise or batch effect. Figure 2 Download asset Open asset compaRe robustly measures the similarity between samples in the presence of batch effect. Similarity matrix generated by compaRe is shown in (a). Size and color of dots represent the level of similarity. Self-comparisons were removed. Noise was added (marked with *) to the original 21 mass cytometry samples of bone marrow aspirates from 16 pediatric AML patients (S) and five healthy adult donors (H). Similarity matrices using JSD with UMAP and meta-clustering with PhenoGraph are shown in (b) and (c), respectively. The run time of comparing each sample to itself is shown in (d). Samples were sorted based on their size. Notably, compaRe took only 25 min to analyze the 21 samples (210 pairwise comparisons), without subsampling or dimension reduction. Meanwhile, meta-clustering and JSD took 39 hr and 10 hr respectively. For the feasibility of JSD, we subsampled each sample to 100,000 cells (default value suggested in Amir et al., 2013). When we fixed this limit to 60% of each sample, the computing time of JSD increased to 3 days. To investigate the relation between run time and sample size, we compared each sample to itself and sorted measured times based on sample size (Figure 2d). The run time increased steeply for both meta-clustering and JSD as the sample size increased, while the increase for compaRe was almost unnoticeable. To further show that compaRe can identify phenotypic changes from a high-dimensional dataset, we used a subset of the data with three healthy and two AML samples stained with 29 (15 membrane and 14 intracellular signaling) markers (Appendix 1—figure 2). Taking H1 as reference, we gradually removed 25%, 50%, 75%, and 100% of cells from a target cluster identified by PhenoGraph. The gradual removal can be regarded as a phenotypic change and the 75% reduction can potentially resemble a rare cell population (a small cluster of cells). As shown in the UMAP projections, the similarity decreased concurrently and more drastically after 100% reduction when phenotypic changes were detected, indicating compaRe is sensitive to phenotypic changes and the existence of rare cell populations. compaRe reveals interpatient similarity Non-AML myeloid neoplasias such as MDS can evolve to become AML. Over time, about one-third of all MDS cases develop into AML (DeVita and Lawrence, 2015; Niederhuber et al., 2020). The risk of developing AML largely depends on the MDS subtype at the time of diagnosis, with high-risk MDS developing into AML more often than the lower-risk MDS subtypes (Greenberg et al., 2012). As many immunophenotypic abnormalities are not unique to MDS, several diagnostic flow cytometric antibody panels have been proposed (van Dongen et al., 2012; Alhan et al., 2016). The EuroFlow AML/MDS antibody panel (van Dongen et al., 2012) aims at the parallel identification and categorization of AML and MDS. Both diseases are heterogeneous, affecting multiple cell lineages and multiple maturation stages. Therefore, this panel concerns major myeloid lineages (neutrophilic, monocytic and erythroid) and the detection of abnormal lymphoid maturation profiles in four tubes. The panel uses four backbone markers to identify myeloblasts and an additional set of 15 markers devoted to the characterization of myeloid lineages (Supplementary files 1 and 2). Unlike the backbone markers, the characterization markers are divided into each tube exclusively. This design was made so that characterization markers from different tubes can be inferred on the same backbone marker subpopulations, but the design makes it impossible to form a multiparameter dataset which is required for clustering methods. However, as compaRe’s comparison module can compare cell population morphologies even in subspaces, we were able to use it to measure similarities between patient samples. We analyzed 25 bone marrow mononuclear cell samples collected from 16 MDS patients and 9 AML patients (Supplementary file 3). The comparison module provided a detailed overview of similarities of samples. As expected, the AML samples exhibited a great amount of interpatient heterogeneity compared to the MDS samples (Figure 3a and b) with all MDS samples clustered together, and the AML samples spread over three clusters. To verify the performance of the module, we visualized the pairwise comparisons using UMAP projection (Figure 3c and Appendix 1—figures 3–26). The measured similarities perfectly matched the projections so that from top left to bottom right, as the similarity decreases, the degree of overlap decreases, and the number of exclusive cell populations increases. Figure 3 Download asset Open asset compaRe highlights immunophenotypic similarities. (a) The similarity band plot visualizes the similarity between a sample specified by its row (band) and other samples measured by compaRe (H: higher-risk MDS, L: lower-risk MDS and A: AML). Each band was independently transformed by an exponential function to emphasize the highest and the lowest similarity values. (b) A graphical representation of the similarities. The graph nodes (samples) were clustered by a random walk. (c) The UMAP projection of A1 sample against the other patient samples is provided as an example. The other projections are given in Appendix 1—figures 3–26. The projections were sorted based on similarity. We further investigated how different the three groups of the AML samples were (Figure 4 and Appendix 1—figure 27). AML samples 1 and 9 of the blue cluster were confirmed to have a high degree of monocytic differentiation with marked expression of the monocytic maturation markers CD14, CD35, CD64, and CD300e. The AML samples of the green cluster, on the other hand, represented a cluster of poorly differentiated AML cases with low expression of differentiation markers and high expression of the stem cell/progenitor markers CD34 and CD117. Unlike the blue cluster with high monocytic differentiation, and the green cluster with poor monocytic differentiation, the AML samples 2 and 5 of the red cluster included both positive and negative populations of CD11b which is a common granulocytic and monocytic maturity marker, a feature observed in all MDS samples as well (Appendix 1—figure 27). Figure 4 Download asset Open asset Immunophenotypic profiles of two different groups of AML patients. Each row shows the UMAP projection of AML samples 1 and 9 (red and orange) vs AML samples 3, 4, 6-8 (blue) of the green cluster of Figure 3b stained by the markers available in each tube. In conclusion, compaRe’s comparison module can be used to optimize true cytometric n-dimensional immunophenotypic characterization of patient samples. Interpretation can then be performed in a conventional manner assisted by lower-dimensional projection tools such as PCA and UMAP that promptly provide a phenotypic profile of the patient samples. Identifying cell-subtype-specific drug responses in mouse AML cells We applied compaRe to high-throughput flow cytometry data to identify cell subtype-specific responses evoked by antineoplastic agents in leukemic spleen cells from an AML mouse model. Splenic cells were sorted for c-Kit cell surface expression, allowing for the enrichment of stem/progenitor-type leukemic cells. On ex vivo expansion, these cells continuously expand and differentiate in a similar way as in vivo with a clear stem cell/progenitor population and partial differentiation towards CD11b/Gr-1 or CD16/CD32-expressing myeloid cells. After ex vivo expansion, the leukemic cells were plated onto multi-well plates containing a library of 116 antineoplastic agents including surface and nuclear receptor inhibitors and activators, inhibitors in a as well as negative control (Supplementary file After hr of drug we stained the cells with labeled against three cell surface markers and and cell surface marker expression using a high-throughput flow compaRe corrected the signal drift, sources of bias in cell as well as sources of bias (Appendix 1—figure After clustering and analysis, we each similar drug responses (Supplementary file To an overview of the we generated a of the clusters (Figure and and Materials and methods). We identified a distinct response with decreased and increase of as compared to control one in Figure of the cliques included in this response of drugs in high with However, drugs in this had a effect on live cell and these were for inhibitors (Figure file For in a marked of and increase of further the of compaRe (Figure The is a of such as differentiation, and et al., findings suggest that controls the differentiation towards cells. Figure 5 Download asset Open asset compaRe analysis several distinct cell subtype-specific responses in a high-throughput flow cytometry screening of mouse AML cells. (a) A UMAP plot of cliques identified by compaRe. are by and one is with and increased as compared with two has increased expression compared with (b) of marker are between and 1 marker to make containing control, and and and are (c) plot of to response 1, showing live cell as a of control for and In high and inhibitors of proteins and a reduction in live cell counts but also a reduction of in all the measured markers, which to the loss of differentiation marker positive cells (Figure Figure The proteins of on and of other and to et al., 2016). The enrichment of cells could be to an in differentiation or that of has led to a of cell surface In this cell the leukemic cells are to be present within the differentiation marker negative These cells are potential for treatments against We observed response 2 (Figure had a higher in marker as compared to control, the increase was and to be linked to drug However, three drugs, and the of differentiation marker negative cells (Figure these three drugs have distinct of is a the and is a receptor (Supplementary file together, compaRe analysis of the high-throughput flow cytometry screening data identification of several distinct phenotypic responses in this mouse AML as well as the that of different of can still cluster the they in a specific model. response in with can be one of the of compaRe. The could be visualized in the clusters that compaRe Identifying signal inhibitors in AML cells We further applied compaRe to the drug screening data from an AML patient AML bone marrow mononuclear cells were into a containing a library of drugs and drug in (Supplementary file After hr of drug the cells were stained with labeled against a panel of cell surface markers and a of myeloid markers). A high-throughput flow was used to cell surface marker compaRe analysis identified several distinct response groups (Figure file one had higher in the CD34 and compared to the increase in was to a drug of a cell population that was in the control samples (Figure The of this population was also to a increase in live cell (Figure different drugs had the same effect (Figure most of them signal inhibitors such as and Figure Download asset Open asset of drugs that of cells in an AML patient (a) UMAP of cliques identified by compaRe. are by CD34 and groups of are using a (b) of response plot of markers CD34 and in different of (c) of live cells after hr to different of of drugs that of the cell 2 of two and in different These a in the of and (Appendix 1—figure In the of this response was to cell (Appendix 1—figure the response was without the loss of live cell (Appendix 1—figure but with a in the in the cell differentiation marker (Appendix 1—figure compaRe also response three as distinct from the This with and the phenotypic response in three is subtle but with a distinct increase in cells (Appendix 1—figure This result highlights compaRe analysis is sensitive to identify these subtle Discussion Technological in multiparameter high-throughput screening have
Source: eLife
Language: English
Subfield: Hematology
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $2,500
OpenAlex ID: W4224251511