doi
string | pmcid
string | plain_text
string | technical_text
string | full_text
string | journal
string | topics
list | keywords
list | llm_lingua2_text
string | llm_lingua2_rate
float64 | llm_lingua2_preproc_time
float64 |
|---|---|---|---|---|---|---|---|---|---|---|
10.1038/s41467-020-16904-3
|
PMC7308400
|
Single cell RNA-seq is a powerful method to assign cell identity, but the purity of cell clusters arising from this data is not clear. Here the authors present an entropy-based statistic called ROGUE to quantify the purity of cell clusters, and identify subtypes within clusters.
|
Single-cell RNA sequencing (scRNA-seq) is a versatile tool for discovering and annotating cell types and states, but the determination and annotation of cell subtypes is often subjective and arbitrary. Often, it is not even clear whether a given cluster is uniform. Here we present an entropy-based statistic, ROGUE, to accurately quantify the purity of identified cell clusters. We demonstrate that our ROGUE metric is broadly applicable, and enables accurate, sensitive and robust assessment of cluster purity on a wide range of simulated and real datasets. Applying this metric to fibroblast, B cell and brain data, we identify additional subtypes and demonstrate the application of ROGUE-guided analyses to detect precise signals in specific subpopulations. ROGUE can be applied to all tested scRNA-seq datasets, and has important implications for evaluating the quality of putative clusters, discovering pure cell subtypes and constructing comprehensive, detailed and standardized single cell atlas.
|
IntroductionTissues are complex milieus comprising various cell types and states with specialized roles1. Characterizing the property and function of each pure cell type is a long-standing challenge in biological and medical disciplines. The recent advances in scRNA-seq have transformative potential to discover and annotate cell types, providing insights into organ composition2, tumor microenvironment3, cell lineage4, and fundamental cell properties5. However, the identification of cell clusters is often determined by manually checking specific signature genes, which are arbitrary and inherently imprecise. In addition, different methods and even parameters used for normalization, feature selection, batch correction, and clustering can also confound the final identified clusters6, thus motivating the need to accurately assess the purity or quality of identified clusters (Fig. 1a).Fig. 1The expression entropy model.a Identifying pure cell subtypes in unsupervised single-cell data analysis. b The S–E plot of the Tabula Muris (droplet) dataset. Each point represents one gene. The relationship between S and E was fitted with LOESS regression for each gene. c The S–E plot of a T-cell dataset27 obtained by Smart-seq2 protocol. d Accuracy in identifying differentially expressed genes on data simulated from both NB (left) and ZINB (right) distribution, with subpopulation containing 50% of the cells. The center line indicates the median AUC value. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and whiskers denote 1.5 times the interquartile range. Discriminating power of genes selected by S–E model, HVG, Gini, M3Drop, SCTransform, Fano factor, and RaceID3 (“Methods”) estimated by RF with 50 times cross-validation on both droplet-based dataset (e) and full-length-based dataset (f) listed in Supplementary Table 1. The classification accuracy was measured as the percentage of query cells that were assigned the correct label. The center line indicates the median classification accuracy. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and whiskers denote 1.5 times the interquartile range. g Reproducibility of features of brain replicates (Supplementary Table 3). h ARI for the dataset comprising five cell lines6 when different feature selection methods were used.A pure cluster here is defined as a population where all cells have identical function and state without variable genes. The importance of purity assessment is particularly relevant for analyses that aim to discover novel pure subtypes and further detect the true biological signals. For example, signature genes specific to a pure subpopulation maybe mistakenly considered as the common signals of a mixture due to less guided clustering and annotation. The purity evaluation could therefore eliminate such misleading conclusions, potentially aiding our understanding of cellular function, state, and behavior. While pioneering approaches such as silhouette7, DendroSplit8, and distance ratio9 have been devoted to determining the optimal number of identified clusters by calculating the ratio of within-cluster to intercluster dissimilarity, they are not comparable among datasets and have poor interpretability of cluster purity. For example, an average silhouette value of 0.7 indicates a fairly strong consistency for a given cluster, but it is still unknown whether this cluster is a pure population or a mixture of similar subpopulations especially when frequent dropout events occur.The challenges presented by purity evaluation can be broadly addressed by investigating the number of infiltrating nonself cells and variable genes, which are suited to the intended areas of unsupervised variable gene detection. Given its importance, diverse methods10 have been proposed for the quantification and selection of highly variable genes. In particular, scran11 aims to identify variable genes by comparing variance to a local regression trend. However, the over-dispersion, coupled with the high frequency of dropout events, would often result in the selection of many lowly expressed genes12. Alternatively, although Gini coefficient13 could be used to quantify the variation in gene expression, it is specially designed for rare cell-type identification. New probabilistic approaches for variable gene selection using dropout rates have also been recently adapted12, with the advantage of supporting both pseudotime analysis and discrete clustering, but their usage of dropout metric hinders the capturing of the global distribution of gene expression. Although highly informative genes can also be determined by inspecting their weights during multiple iterations of dimensionality reduction14, such ad hoc approaches are computationally intensive, requiring potentially orders of magnitude more time than methods like HVG and M3Drop.Here, we present an entropy-based model to measure the randomness of gene expression in single cells, and demonstrate that this model is scalable across different datasets, capable of identifying variable genes with high sensitivity and precision. Based on this model, we propose the Ratio of Global Unshifted Entropy (ROGUE) statistic to quantify the purity or homogeneity of a given single-cell population while accounting for other technical factors. We demonstrate that the ROGUE metric enables accurate and unbiased assessment of cluster purity, and thus provides an effective measure to evaluate the quality of both published and newly generated cell clusters. Applying ROGUE to B cell, fibroblast, and brain data, we identified additional pure subtypes and demonstrate the application of ROGUE-guided analysis in detecting the precise biological signals. Our approach is broadly applicable for any scRNA-seq datasets, and is implemented in an open-source R package ROGUE (https://github.com/PaulingLiu/ROGUE), which is freely available.ResultsOverview of ROGUEAs scRNA-seq data can be approximated by negative binomial (NB) or zero-inflated NB (ZINB) distribution15,16, we considered the use of the statistic, S (expression entropy—differential entropy of expression distribution, as defined in “Methods”), to capture the degree of disorder or randomness of gene expression. Notably, we observed a strong relationship between S and the mean expression level (E) of genes, thus forming the basis for our expression entropy model (S–E model, Fig. 1b, c). Moreover, S is linearly related to E for the Tabula Muris dataset2 as expected (Fig. 1b), which is characteristic of current droplet experiments, hence demonstrating the NB nature of UMI-based datasets (“Methods”). For a heterogeneous cell population, certain genes would exhibit expression deviation in fractions of cells, leading to constrained randomness of its expression distribution and hence the reduction of S. Accordingly, informative genes can be obtained in an unsupervised fashion by selecting genes with maximal S-reduction (ds) against the null expectation (“Methods”).To provide a direct purity assessment of putative cell clusters or fluorescence-activated cell sorting (FACS)-sorted populations, here we take advantage of the wide applicability of S–E model to scRNA-seq data and introduce the quantitative measure, ROGUE (“Methods”). Intuitively, a cell population with no significant ds for all genes will receive a ROGUE value of 1, indicating it is a completely pure subtype or state. In contrast, a population with maximum summarization of significant ds will yield a purity score of ~0.S–E model accurately identifies informative genesTo illustrate the performance of our model, we benchmarked S–E against other competing feature selection methods (HVG11, Gini13, M3Drop12, SCTransform17, Fano factor18, and RaceID319) on data simulated from both NB and ZINB distribution (“Methods”). For a fair comparison, we generated a total of 1600 evaluation datasets with subpopulations containing 50, 20, 10, or 1% of the cells, and used AUC as a standard to test the performance of each method. Notably, S–E model consistently achieved the highest average AUC and significantly outperformed other gene selection methods in all tested cases with varied subpopulation proportions or gene abundance levels (Fig. 1d and Supplementary Figs. 1 and 2). Although SCTransform is specially designed for UMI-based scRNA-seq data, it exhibited notable performance on ZINB-distributed datasets (Fig. 1d). As a tool to identify genes specific to rare cell types, Gini showed increased performance when there were subpopulations accounting for <20% of the cells. In contrast, HVG performed better in the presence of cell subpopulations with a larger proportion (Supplementary Figs. 1 and 2).To validate our unsupervised feature selection method in real datasets, we performed cross-validation experiments using random forest classifier (RF)20. We randomly sampled 70% cells from the original dataset as reference, and classified the remaining 30% cells, with clusters defined by the original authors (“Methods”). Intuitively, gene sets that enable higher classification accuracy are more biologically meaningful21. Using 14 previously published datasets derived from both droplet-based and full-length protocols (Supplementary Table 1), we demonstrated that our method consistently identified genes with greater ability of classification when different number (30–5000) of genes were selected (Fig. 1e, f and Supplementary Figs. 3 and 4). Specially, our S–E model showed notable superiority when fewer genes (30–100) were used, demonstrating its sensitivity. Taken together, these results suggest that genes identified by our model are more informative and biologically discriminating.Since datasets derived from the same biological system are expected to have reproducible informative genes12, we tested how our expression entropy model behaves using technical replicates from different tissues (Supplementary Table 2). Notably, genes identified by our S–E model were more reproducible when top 500–2000 genes were used (Fig. 1g and Supplementary Fig. 5a–c). In addition, we also considered four pancreatic datasets (Supplementary Table 3) derived from different technologies and labs. These real datasets are more complex than technical replicates as they included systemic nuisance factors such as batch effects. Despite substantial systematic differences, our model consistently achieved high reproducibility scores (Supplementary Fig. 5d).A major task of feature selection is to identify genes that are most relevant for biological heterogeneity, which can be applied to downstream clustering. We therefore evaluate the performance of S–E model in the context of unsupervised clustering with RaceID319, SC322, and Seurat23. Here we considered five publicly available scRNA-seq datasets with high-confidence cell labels6,9,24,25 (“Methods”). These datasets include cells from different lines, FACS-purified populations, or well-characterized types (Supplementary Fig. 6 and “Methods”), and thus can be considered gold standards. To quantify the similarity between the clusters obtained by different clustering methods and the reference cell labels, we calculated the adjusted Rand index (ARI)26, which is restricted to the interval [0, 1]. For the number of features, we considered the top 100, 500, 1000, or 2000 genes. Our results illustrated that S–E model provides the best performance in terms of ARI in these scenarios (Fig. 1h and Supplementary Fig. 7).As some methods were optimized to detect rare cell types, we tested if the genes selected by our S–E model are effective in uncovering such rare subpopulations. To this end, we first simulated a scRNA-seq dataset (“Methods”), which contains three rare clusters (of 10, 30, and 20 cells, respectively) and two common clusters (of 1000 cells each), and clustered these cells with GiniClust218, RaceID3, as well as S–E model-based Seurat (“Methods”). Of note, all these three methods properly recapitulated the five cell clusters (Supplementary Fig. 8), indicating that S–E model-based Seurat is indeed effective for the recovery of both common and rare cell clusters. In addition to simulated data, we wondered how S–E model behaves in detecting real rare cell types. Since no gold standard is available for such cases, we considered four cell lines (A549, H2228, H838, and HCC827) of Tian et al.6, and generated three common cell types (A549, H2228, and H838; of 500 cells for each) and one rare cell type (HCC827; of 20 or 10 cells, respectively) by down-sampling. Similar to the analysis of simulated data, all the three methods effectively identified both common and rare cell clusters when there were 20 cells of the rare cell type (Supplementary Fig. 9a–c). For the dataset with the rare cell type accounting for lower frequency (10 cells, 0.6% of total cells here), RaceID3 and GiniClust2 exhibited their superiority in uncovering the rare cell type as opposed to S–E model-based Seurat (Supplementary Fig. 9d–f). Thus, although S–E model is effective in uncovering rare subpopulations to a certain extent, methods specifically developed for this purpose, such as GiniClust2 and RaceID3, maybe more appropriate.Evaluation of robustness of ROGUETo test how sensitive ROGUE is to the choice of informative genes, here we considered two scRNA-seq datasets: a T-cell dataset sequenced with Smart-seq25 and a droplet-based dataset2 (Tabula Muris). The results illustrated that the heterogeneity score (1-ROGUE) reached saturation when genes with significant ds were selected (Supplementary Fig. 10), thus we used significant ds to calculate ROGUE in the following analyses. We investigated the performance of ROGUE on 1860 cell populations simulated from both NB and ZINB distribution (2000 cells × 20,000 genes each), with 0.1–50% genes varied in a second cell type (“Methods”). A cell population with both fewer infiltrating nonself cells and varied genes would yield a high purity score, while a population with converse situation is expected to yield a low-purity score. It is evident that the ROGUE index decreased monotonically with the heterogeneity of cell populations (Fig. 2a, b and Supplementary Figs. 11 and 12). ROGUE performed well even when cell populations contained few varied genes (<1%) and infiltrating cells (<1%), indicating ROGUE index provides a sensitive and unbiased measure in response to the degree of cell population purity. The usage of different values of the reference factor K (“Methods”) yielded vary similar results (Supplementary Fig. 13), suggesting that ROGUE is robust to the choice of parameter K within a reasonable range.Fig. 2ROGUE use and performance.a The ROGUE index (reference factor K = 45) decreases monotonically with increasing varied genes in each simulated mixture consisting of two cell types (1:1). The center line indicates the median ROGUE value of n = 50 repeated simulations. The lower and upper hinges represent the 25th and 75th percentiles respectively, and whiskers denote 1.5 times the interquartile range. b The ROGUE values (reference factor K = 45) for the simulated mixtures with cell-type sizes ranging from 1:100 to 1:1. In each mixture, the number of varied genes was 1% of the total gene number (n = 20,000). The center line indicates the median ROUGE value of n = 50 repeated simulations. The lower and upper hinges represent the 25th and 75th percentiles respectively, and whiskers denote 1.5 times the interquartile range. c Pearson correlations of S between the randomly down-sampled datasets (n = 50 runs for each) and the entire datasets (2000 cells) simulated from both NB and ZINB distribution. The center line indicates the median correlation value. The lower and upper hinges represent the 25th and 75th percentiles respectively, and whiskers denote 1.5 times the interquartile range. d Sequencing depth distribution (total UMI counts across cells) for two simulated replicates. The replicate 2 has a sequencing depth ten times that of replicate 1. e The S–E plot of the mixture of replicates 1 and 2 is shown in d. f ROGUE values of n = 100 mixtures versus the silhouette values for every two replicates within individual mixtures. A high silhouette value indicates a substantial difference in sequencing depth between two replicates. g, h The S–E plots and corresponding ROGUE values of 10 cell populations from the PBMC dataset24. i Purity assessment of six human T-cell populations. j Purity evaluation of lung-cancer infiltrating DCs, with each point representing a patient. The center line indicates the median ROUGE value. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and whiskers denote 1.5 times the interquartile range.To address the potential concern that the number of cells may represent an intrinsic challenge for S and ROGUE calculation, particularly if only a small number of cells are collected from given samples, we performed down sampling analysis to test how S was impacted by cell numbers. By calculating the Pearson correlations of S between the randomly down-sampled datasets and the entire datasets, we found the similarity values of >0.99 and demonstrated that our S and ROGUE calculation would not be affected by variation in cell number (Fig. 2c).Sequencing depth can vary significantly across cells, with variation potentially spanning orders of magnitude2, and hence contributes to a substantial technical confounder in scRNA-seq data. We sought to investigate whether ROUGE index can accurately assess the purity of single-cell population while accounting for this technical effect. As test cases, we simulated increasing molecular counts (sequencing depth) in a second mock replicate, with the fold change of gene expression means ranging from 2 to 100 (Fig. 2d and “Methods”). Despite the substantial technical effect, the mixture of each two simulated replicates is expected to be a pure cell population. Here we used silhouette to measure the degree of replicate-to-replicate differences. The results revealed ROGUE values of ~0.99–1 for each population consisting of two replicates, with silhouette values ranging from 0.25 to 0.75 (Fig. 2e, f and Supplementary Fig. 14a). Thus, ROGUE not only offers a robust and sensitive way to estimate the purity of single-cell population, but also accounts for the variation in sequencing depth.ROGUE accurately assesses the purity of cell populationsTo illustrate the applicability of ROGUE index to real data, we first considered an External RNA Controls Consortium (ERCC) dataset24, which is a highly controlled experiment dedicated to understanding the technical variability. All 1015 droplets of this dataset received the same ratio of ERCC synthetic spike-in RNAs, hence no varied RNAs should be detected in principle. We referred to this dataset as an ideal case of pure cell population and found only one RNA with significant ds. Accordingly, this ERCC dataset achieved a ROGUE value of ~1 as expected, thus confirming its purity. Further, we investigated the fresh peripheral blood mononuclear cells (PBMCs) enriched from a single healthy donor24. The authors provided multiple cell types purified by FACS, and thus representing a suitable resource for purity assessment. These cell types in Fig. 2g, including CD4/CD8 naïve T cells and CD4 memory T cells, have been shown to be highly homogeneous populations27, and were detected high ROGUE values (0.94–1) as expected. In contrast, both CD14 monocytes and CD34+ cells are mixtures of diverse subtypes24 and received relatively low ROGUE values (~0.8; Fig. 2h), thus confirming their heterogeneity.In addition to highly controlled datasets, it is also instructive to investigate how ROGUE index performs on pure subtypes identified by unsupervised clustering. Here we first considered six well-defined T-cell subtypes from human colorectal cancer5, which were generated via the Smart-seq2 protocol. All these pure subtypes achieved high ROGUE values of >0.9 (Fig. 2i), versus 0.78 for complete data (Supplementary Fig. 14b). We next examined four dendritic cell (DC) subsets collected from human lung cancers28 and sequenced with inDrop platform. Specially, tumor-infiltrating DC2 cells have been proven to be highly heterogeneous populations29,30 and deviated substantially from the other homogeneous cell types including DC1, LAMP3+ DC, and pDC (Fig. 2j). Taken together, these results illustrate that our ROGUE represents an effective and direct quantification of cell population purity without being affected by technical characteristics.ROGUE-guided analysis enhances cell-type identificationWe next evaluated the potential for ROGUE to guide clustering analysis with silhouette, which investigates whether a certain clustering has maximized intercluster dissimilarity and minimized within-cluster dissimilarity. As a test case, we simulated a scRNA-seq dataset consisting of three cell types A, B, and C (see “Methods” for details), where cell types A and B were similar subtypes with 1% varied genes. We clustered this dataset into 2, 3, 4, and 5 subpopulations respectively by adjusting the resolution parameter in Seurat23 (Fig. 3a), then evaluated the results by inspecting corresponding silhouette and average ROGUE values. Proper clustering of this dataset should result in three subpopulations, one for each cell type. However, silhouette received the maximum value when cell-type A co-clustered with B (Fig. 3b), i.e., when only two clusters were identified, suggesting that such measure is poorly interpretable for cluster purity as opposed to ROGUE, which reached saturation when there were three clusters (Fig. 3c). Repeating the simulation with varied differences in cell-type A, B, and C yielded equivalent performance for these two methods (Supplementary Fig. 15a–f). Such performance was also observed when different values of the reference factor K were used (Supplementary Fig. 16). Since ROGUE can provide direct purity quantification of a single cluster and is independent of methods used for normalization, dimensionality reduction, and clustering, it could also be applied to guide the splitting (re-clustering) or merging of specific clusters in unsupervised clustering analyses.Fig. 3ROGUE enhances single-cell clustering and cell-type identification.a t-SNE plots of a simulated dataset containing three cell types. Corresponding silhouette values (b) and average ROGUE values (c) when there were 2, 3, 4, and 5 putative clusters, respectively. d UMAP plots of lung-cancer-associated fibroblasts, color-coded by clusters in original paper (left; Supplementary Fig. 17a) and re-clustered labels (right). e ROGUE values of different clusters before (left) and after (right) re-clustering. Each point represents a patient. The center line indicates the median ROUGE value. The lower and upper hinges represent the 25th and 75th percentiles respectively, and whiskers denote 1.5 times the interquartile range. f UMAP plot of expression levels of MYH11 and MEF2C. g Differences in hallmark pathway activities scored using GSVA.To test how ROGUE could help the clustering of real datasets, we examined a previously reported dataset of cancer-associated fibroblasts (CAFs)31 from lung tumors. CAFs have been reported to represent a highly heterogeneous population and may play a tumor-supportive role in the tumor microenvironment32. We found that the seven identified fibroblast clusters received low ROGUE values (Fig. 3d, e and Supplementary Fig. 17a). We therefore performed re-clustering analysis with the goal of exploring the extent of heterogeneity and identified a total of 11 clusters with a higher average ROGUE value (Fig. 3d, e). In addition to the two classical subtypes of CAFs (myofibroblastic CAFs and inflammatory CAFs), we also found the presence of antigen-presenting CAFs (apCAFs) that was characterized by the high expression of CD74 and MHC class-II genes (Supplementary Fig. 17b). The apCAFs were firstly discovered as a fibroblast subtype in mouse pancreatic ductal adenocarcinoma (PDAC), but barely detectable in human PDAC without forming a separate cluster33. The considerable existence of apCAFs in lung cancer thus may indicate potential differences between different cancer types.Furthermore, we noted that the myCAFs (AF_C02_COL4A1, ROGUE = 0.81) identified by original authors could be further segregated into three distinct subpopulations, including BF_C01_RGS5 (ROGUE = 0.84), BF_C02_ACTA2 (ROGUE = 0.87), and BF_C03_GPX3 (ROGUE = 0.94). Interestingly, the signature genes of AF_C02_COL4A1 described by original authors were actually specific to one of these three subpopulations, including MEF2C in BF_C01_RGS5 and MYH11 in BF_C02_ACTA2 (Fig. 3f). Pathway analysis also revealed that the NOTCH signaling was activated in BF_C01_RGS5 (Fig. 3g) rather than a common signal of AF_C02_COL4A131. Despite the considerable increase of overall ROGUE index, BF_C00_AOL10A1, BF_C04_COL1A2, and BF_C05_PLA2G2A still received relatively low ROGUE values, thus deserving further investigation. Overall, ROGUE-guided analysis not only discovered novel cell subtypes, but also enabled the detection of the true signals in specific pure subpopulations.ROGUE-guided analysis identifies pure B cell subtypesB cells are key components in tumor microenvironment but have unclear functions in antitumor humoral response34. Here we investigated previously reported liver- and lung-tumor-infiltrating B cells31,35 and found that they received relatively low ROGUE values (Fig. 4a). Thus, we applied further clustering analysis coupled with ROGUE to these B cells in an attempt to discover pure subtypes. A total of seven clusters were identified, each with its specific marker genes (Fig. 4b–d). Cells from the first B-cell subset, B_C0_JUNB, specifically expressed signature genes including JUNB and FOS, thus representing activated B cells36. The second subset, B_C1_TXNIP, showed high expression of glycolysis pathway genes (Supplementary Table 4), indicating its metabolic differences. ACTB, a gene involved in antigen presenting, was highly expressed in the third subset (B_C2_ACTB). Pathway activity analysis also revealed a strong antigen processing and presentation signal in this subset (Supplementary Table 4). The fourth cluster, B_C3_FCER2, characterized by high expression of HVCN1 and genes involved in B-cell receptor signaling pathway (Supplementary Table 4), was largely composed of pre-activated B cells37. The fifth cluster, B_C4_MX1, predominantly composed of interferon-induced B cells38, expressed high levels of MX1, IFI6, and IFI44L. The sixth cluster, B_C5_CD3D, expressed key markers of both T- and B-cell lineages (Fig. 4d), thus maybe the dual expressers (DEs)-like lymphocytes39 or doublets. The remaining B cells, falling into the seventh cluster, B_C6_LRMP, exhibited high expression of LRMP and RGS13, indicative of the identity of germinal center B cells40.Fig. 4ROGUE-guided analysis in the identification of pure B-cell subtypes.a The S–E plots and ROGUE values of liver- and lung-tumor-infiltrating B cells, respectively. UMAP plots of 4291 B cells, color-coded by their associated clusters (b) and tissues (c). d Gene expression heatmap of seven B-cell clusters. Rows denote marker genes and columns denote different clusters. e ROGUE values of seven identified B-cell subtypes. Each point represents a patient. The center line indicates the median ROUGE value. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and whiskers denote 1.5 times the interquartile range. f Tissue preference of each B-cell subtype in liver cancer estimated by RO/E27, the ratio of observed to expected cell numbers calculated by the chi-square test. g The average fractions of B_C02_ACTB and B_C04_MX1 in each patient across tissues, where error bars representing ±s.e.m. *p < 0.05, **p < 0.005, Student’s t test. The Kaplan–Meier curves of TCGA LUAD (h) and LIHC (i) patients grouped by the 13 markers (Supplementary Table 5) of B_C02_ACTB.Both DEs/doublets-like and germinal center B cells exhibited low ROGUE values (Fig. 4e), but the limited cells did not permit further clustering. For germinal center B cells, we readily detected the high expression of proliferating marker genes, including MKI67 and STMN1 (Supplementary Fig. 18), in a fraction of these cells, thus explaining the heterogeneity to some extent. In contrast to these two clusters, we found ROGUE values of >0.92 for each of the remaining five clusters (Fig. 4e), demonstrating that they were all highly homogeneous B-cell subtypes. By calculating the ratio of observed to expected cell numbers with the chi-square test (RO/E), we noted that both B_C02_ACTB and B_C04_MX1 contained mainly cells from tumor, with RO/E values >1 (Fig. 4f). Similar analyses stratified by patient further confirmed this trend (Fig. 4g). Based on the independent TCGA lung adenocarcinoma (LUAD) cohort dataset, patients with higher expression of the marker genes of B_C02_ACTB (normalized by MS4A1; Supplementary Table 5) showed significantly worse overall survival (Fig. 4h). Such survival difference was also observed in TCGA liver hepatocellular carcinoma (LIHC) cohort dataset (Fig. 4i). Thus, the clinical implication deserves further study to investigate what specific roles B_C02_ACTB cells play in tumor microenvironment. In summary, identifying pure subtypes with ROGUE-guided analysis could enable a deeper biological understanding of cell state and behavior.Application to brain data and batch effect evaluationIn addition to cancer data, we also demonstrated the application of ROGUE in analyzing the brain transcriptome dataset2, which harbors a high degree of heterogeneity for those encapsulated cell classes. This dataset identified seven distinct cell types, of which oligodendrocyte and neuron cell types had low ROGUE values of <0.8, versus ~0.9–1 for the remaining five cell classes (Fig. 5a). We therefore applied further clustering guided by ROGUE to oligodendrocyte which is of enough cells (n = 3401), and identified ten refined cell subtypes, each with its specific marker genes (Fig. 5b, c). Except cluster 6, we found ROGUE values of ~0.9–1 for all the other nine clusters, suggesting their purity (Fig. 5d). To investigate potential functions of these subtypes, we compared pathway activities and found considerable phenotypic diversity. For example, cluster 5 showed a strong signal of axon guidance signaling (Fig. 5e), while neurotrophin signaling pathway was highly activated in cluster 1 (Fig. 5f). This example further illustrates how ROGUE plays a key role in uncovering pure subpopulations.Fig. 5The application of ROGUE in brain data and batch effect evaluation.a ROGUE values of seven distinct brain cell types as defined by the original publication2, with each point representing a sample. The center line indicates the median ROUGE value. The lower and upper hinges represent the 25th and 75th percentiles respectively, and whiskers denote 1.5 times the interquartile range. b UMAP plot of the ten identified clusters of oligodendrocytes (n = 3401), color-coded by their associated clusters. c Expression heatmap of cell-type-specific genes of the ten oligodendrocyte clusters. d ROGUE values of oligodendrocyte clusters. Each point represents a sample. The center line indicates the median ROUGE value. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and whiskers denote 1.5 times the interquartile range. e, f Enriched pathways for cluster 5 (e) and cluster 1 (f), respectively. g ROGUE values were shown for batch 1 (the control group), batch 2 (the stimulation group), and aggregated cell population (batch 1 and batch 2) for each cell type. For fair comparison, we equalized the number of cells in each group by down-sampling. The center line indicates the median ROGUE value. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and whiskers denote 1.5 times the interquartile range. *p < 0.05, **p < 0.005, Student’s t test. h ROGUE values for individual-specific cell populations and aggregated populations (all individuals). All cells used here were from the control group. Subsampling was performed to equalize the number of cells in each group. The center line indicates the median ROUGE value. The lower and upper hinges represent the 25th and 75th percentiles, respectively, and whiskers denote 1.5 times the interquartile range. *p < 0.05, **p < 0.005, Student’s t test.To investigate if ROGUE is effective in evaluating the impact of batch effect, we studied a dataset of human PBMCs containing multiple distinct cell types38. Cells of this dataset were previously split into two groups—the interferon-beta (IFN-β)-stimulated group and the culture-matched control group, thus could be considered as two batches. Then we applied ROGUE to assess the purity of each cell type (as defined by the original authors) in individual bathes as well as the aggregated cell population (batch 1 and batch 2), and found that ROGUE detected considerable purity reduction in the aggregated group (Fig. 5g).As cells of this dataset were collected from eight unrelated individuals, we also tested how ROGUE behaves in estimating the variability (i.e., batch effect) among patients. Here we only used cells from the control group so that the evaluation would not be influenced by IFN-β perturbation. As expected, the aggregated cell populations of all individuals received significantly lower ROGUE values as opposed to patient-specific populations for each cell type (Fig. 5h). Thus, ROGUE offers a reasonable method for estimating the impact of batch effect.DiscussionPurity assessment of identified cell clusters is paramount to the interpretation of scRNA-seq data. This assessment is especially pertinent as increasingly rare and subtle cell subtypes are being uncovered. To address this computational challenge, we present the S–E model and demonstrate that this model is capable of identifying variable genes with high sensitivity and precision, and thus could be applied to both clustering and potentially pseudotime analyses. By taking advantage of the wide applicability of S–E model, we develop the statistic ROGUE to quantify the purity of single-cell populations. Through a wide range of tests, we demonstrate that our entropy-based measure, ROGUE, is broadly applicable for datasets from different platforms, protocols and operators, and able to successfully quantify the purity of singl-cell populations regardless of uncontrollable cell-to-cell variation.When using ROGUE to assess the purity of four DC subtypes from human lung tumors, we found that DC2 was a heterogeneous population, which is consistent with previous findings30. Such heterogeneous populations like DC2 may have different properties and specialized roles in the cancer microenvironment, and could be assessed in a similar fashion with ROGUE. Accordingly, future studies could focus on these cell populations and hence may deepen our understanding of cellular origins of cancer. In addition, ROGUE addresses an important need in unsupervised single-cell data analyses, i.e., to effectively assess the quality of published or newly generated clusters. Often, unsupervised clustering may lead to under- or over-clustering of cells due to the lack of universal stands for clustering quality. By quantifying cluster purity with ROGUE before and after clustering or re-clustering, we were able to detect low-purity clusters and perform further analysis to discover pure subtypes. Improving the purity and credibility of the ever-increasing number of cell types is a mounting challenge with explosive efforts toward single-cell sequencing, and ROGUE could become a potential standard for judging the quality of cell clusters.Our ROGUE-guided analysis on fibroblasts identified a novel subpopulation in lung cancer, apCAFs, which highly expressed CD74 as well as MHC class-II genes and had a strong antigen-presenting signal. These cells have been speculated to deactivate CD4 T cells and decrease the CD8+ to Treg ratio in mouse PDAC33, but have unclear role in the lung-cancer microenvironment, hence requiring further investigation. Moreover, when applying ROUGE to B-cell analysis, we found an interesting pure cluster B_C02_ACTB that displayed high expression of genes involved in antigen processing and presentation. Cells from this cluster were preferentially enriched in tumors and were associated with poor prognostic outcomes in both lung and liver cancer. We therefore hypothesize that these cells may contribute to immune suppression in the cancer microenvironment and hence curtail antitumor immunity, although further studies are required to define the roles of these cells. Such approaches for discovering novel or additional pure subtypes can also be extended to other published or newly generated scRNA-seq datasets.When determining the purity of cell clusters, we recommend a ROGUE value of 0.9 as a suitable threshold, at which the number of infiltrating cells and varied genes is well constrained. But for low-quality data or continuous data, the threshold could be determined by considering the global ROGUE values. Although ROGUE can be very efficient and effective, we anticipate that additional extensions could enable enhanced performance, for example, assessing the purity of integrated cell populations from different protocols and platforms. Overall, our ROGUE metric provides a robust and direct measure for cluster purity in the presence of substantial technical confounders. We expect the ROGUE metric to be broadly applicable to any scRNA-seq datasets, and anticipate that our strategy will improve the rigor and quality of unsupervised single-cell data analysis.MethodsExpression entropy modelFor droplet datasets, the observed UMI count can be modeled as a NB random variable, which also arises as a Poisson–Gamma mixture411\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{array}{l}X_{ij}\sim {\mathrm{Poisson}} \,( {s_j\lambda _{ij}})\\ \, \,\,\,\,\,\lambda _{ij}\sim {\mathrm{Gamma}}\,( {\alpha _{ij},\beta _{ij}})\end{array},$$\end{document}Xij~Poisson(sjλij)λij~Gamma(αij,βij),where λij represents the true expression value that underlies the observed UMI count Xij of gene i in cell j, and sj denotes the size normalization factor in cell j. The αij and βij are shape parameter and rate parameter respectively. Given the assumption that the shape parameter α is a constant across cells and genes, and that the rate parameter β is a constant of gene i across cells41,42, αij and βij can be expressed as α and βi, respectively. Then the distributions can be recognized as: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _i\sim {\mathrm{Gamma}}\left( {\alpha ,\beta _i} \right)$$\end{document}λi~Gammaα,βi and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{ij}\sim {\mathrm{Poisson}}\,( {s_j\lambda _i})$$\end{document}Xij~Poisson(sjλi). We denote2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_{ij}^\prime = \frac{{X_{ij}}}{{s_j}},$$\end{document}Xij′=Xijsj,as the normalized expression of gene i in cell j, and use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Bbb E}\left( {X_i^\prime } \right)$$\end{document}EXi′ (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_i^\prime$$\end{document}Xi′ is the normalized expression assigned to gene i and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Bbb E}\left( {X_i^\prime } \right)$$\end{document}EXi′ is the expectation across cells) as the moment estimation of λi. For the Gamma distribution, the rate parameter could therefore be calculated based on the maximum likelihood estimation3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _i = \frac{\alpha }{{\lambda _i}} = \frac{\alpha }{{{\Bbb E}\left( {X_i^\prime } \right)}}.$$\end{document}βi=αλi=αEXi′.To capture the degree of disorder or randomness of gene expression, here we considered the use of differential entropy defined as434\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H\left( X \right) = - \int_{ - \infty }^{ + \infty } {p\left( x \right) \cdot {\mathrm{ln}}\,p\left( x \right)dx},$$\end{document}HX=−∫−∞+∞px⋅lnpxdx,where X is a continuous random variable and p(x) is the probability density function. Differential entropy is an extension of Shannon entropy, which is used to measure the average surprisal of a continuous probability distribution, and has shown notable performance in our supervised gene selection method E-test44. Specially, for the gamma distributed random variable λi, its differential entropy can be computed as5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_i = \alpha - {\mathrm{ln}}\beta _i + {\mathrm{ln}}\Gamma \left( \alpha \right) + \left( {1 - \alpha } \right) \cdot \varphi \left( \alpha \right) = {\mathrm{ln}}\frac{\alpha }{{\beta _i}} + a = {\mathrm{ln}}{\Bbb E}\left( {X_i^\prime } \right) + a,$$\end{document}Si=α−lnβi+lnΓα+1−α⋅φα=lnαβi+a=lnEXi′+a,where φ is the digamma function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a = \alpha - {\mathrm{ln}}\alpha + {\mathrm{ln}}\Gamma \left( \alpha \right) + \left( {1 - \alpha } \right) \cdot {\upvarphi}\left( \alpha \right)$$\end{document}a=α−lnα+lnΓα+1−α⋅φα is a constant. Although other pioneering methods such as Scnorm45, scran46, and BASiCS47 can be used to calculate size factors, we considered the library size normalization of each cell defined as the total UMI counts divided by the mean total UMI counts across cells41. Accordingly, the expectation of library size factor across cells is equal to 1. Given Eq. (2) and that the gene expression and library size are two independent random variables42, for a given gene i, we have6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\Bbb E}\left( {X_i} \right) = {\Bbb E}\left( {X_i^\prime \times s} \right) = {\Bbb E}\left( {X_i^\prime } \right) \times {\Bbb E}\left( s \right) = {\Bbb E}\left( {X_i^\prime } \right) \times 1 = {\Bbb E}\left( {X_i^\prime } \right),$$\end{document}EXi=EXi′×s=EXi′×Es=EXi′×1=EXi′,where Xi is the observed expression assigned to gene i and s is the library size assigned to cells. Thus, for each cell type, the differential entropy of λi could be computed as7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_i = {\mathrm{ln}}{\Bbb E}\left( {X_i} \right) + a.$$\end{document}Si=lnEXi+a.We formulate the null hypothesis that there is only one Poisson–Gamma component for each gene in a given population (H0) and thus the corresponding differential entropy can be calculated with Eq. (7). Then we assume that each cell represents its own cluster and use Xij as a moment estimation of the mean expression of such cluster. In this way, we define the entropy reduction of gene i across n cells as8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ds_i = {\mathrm{differential}}\,{\mathrm{entropy}}\,{\mathrm{under}}\,H_0 - {\mathrm{average}}\,{\mathrm{actual}}\,{\mathrm{differential}}\,{\mathrm{entropy}}\\ = {\mathrm{ln}}{\Bbb E}\left( {X_i} \right) - \frac{{\mathop {\sum }\nolimits_{j = 1}^n \,( {{\mathrm{ln}}X_{ij}} )}}{n},$$\end{document}dsi=differentialentropyunderH0−averageactualdifferentialentropy=lnEXi−∑j=1n(lnXij)n,which captures the degree of disorder or randomness of gene expression44. Given that genes under H0 (non-variable genes) account for the major proportion, we fit the relationship between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{ln}}\,{\Bbb E}\left( {X_i} \right)$$\end{document}lnEXi and average actual differential entropy, and calculate corresponding residual as dsi to improve the performance (Fig. 1b, c). The significance of ds is estimated based on a normal distribution approximation and is adjusted using Benjamini–Hochberg method. We also extended such procedure to full-length datasets and found that our approach consistently outperformed other gene selection methods (Fig. 1f and Supplementary Fig. 4).Data simulationWe simulated droplet datasets with NB distribution. Mean gene abundance levels E were sampled from the log-normal distribution\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ln\left( E \right)\sim {N}\left( {\mu ,\sigma ^2} \right),$$\end{document}lnE~Nμ,σ2,with parameters μ = 0 and σ = 2. The number of transcripts for each gene were drawn from\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N_{ij} \sim {\mathrm{NB}}\left( {E_i,r} \right).$$\end{document}Nij~NBEi,r.For each simulated dataset, the dispersion parameter r (r = α)48 was set to a fixed value, ranging from 5 to 20 (Supplementary Fig. 1). In addition, we simulated full-transcript datasets with ZINB distribution. The dropout rates for each gene was modeled with the sigmoid function49\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_i \sim {\mathrm{sigm}}\left( { - \left( {\gamma _0 + \gamma _1E_i} \right)} \right),$$\end{document}Pi~sigm−γ0+γ1Ei,with parameters γ0 = −1.5 and γ1 = 1/median(E). Each simulated scRNA-seq dataset contained 20,000 genes and 2000 cells (Supplementary Fig. 2).Differentially expressed genes were added in a fraction of cells (1–50%, Supplementary Fig. 1 and 2), with fold changes sampled from the log-normal distribution (μ = 0 and σ = 2). Genes with a >1.5-fold decrease or increase in mean expression were considered as ground truth DE genes.Feature selection methodsThe HVG method11 identifies variable genes by comparing the coefficient of variation squared to a local regression trend, and was implemented with the BrenneckeGetVariableGenes function in the M3Drop12 package. In the Gini index model proposed in GiniClust13, a gene is considered as informative if its Gini is higher than expected from the maximum observed expression. We copied the source code of original GiniClust (GiniClust_Preprocess.R, GiniClust_Filtering.R and GiniClust_Fitting.R) (https://github.com/lanjiangboston/GiniClust/tree/master/Rfunction), and defined the Gini_fun function in our scripts to select genes. M3Drop uses dropout rates for variable gene selection and was implemented with the M3DropFeatureSelection function in the M3Drop package. The SCTransform method17 selects genes with Pearson residuals from the regularized negative binominal regression and was implemented with the SCTransform function in Seurat package. In addition, we implemented the Fano factor method as used in the script GiniClust2_Fano_clustering.R from GiniClust218. The feature selection step in RaceID319, which selects genes with a second-order polynomial fit between the expression variance and log-transformed mean, was implemented according to the fitBackVar function in RaceID3.Datasets used for clustering-based evaluationTo evaluate the performance of different feature selection methods in the context of unsupervised clustering, here we considered five publicly available scRNA-seq datasets. The first dataset6 consists of five cell lines (A549, H1975, H2228, H838, and HCC827) and was sequenced with 10X Genomics protocol, with a total of 3918 cells. The second dataset was generated by the same study6. This dataset comprises three cell lines (H1975, H2228, and HCC827) and was sequenced with CEL-seq2 protocol. The third dataset24 was created by processing multiple FACS-purified cell populations and was sequenced with 10X Genomics protocol. Considering that some populations such as CD8+ cytotoxic T cells were relatively heterogenous24, here we only used CD19 B cells, CD4 naïve T cells, CD56 NK cells, and CD14 monocytes, which were readily distinguishable (Supplementary Fig. 6a). The fourth dataset contains cells from human pancreatic islet and was generated by Smart-seq protocol25. These pancreatic cell types including alpha, beta, delta, and gamma cells are well-characterized and have been shown to be distinct23,44, thus were used for benchmarking (Supplementary Fig. 6b). The remaining dataset comprises multiple immune cell types9, with cells sequenced by Smart-seq2 protocol. Although the cell labels in original publication were assigned using unsupervised clustering, cross-validation experiments revealed that the major cell types (macrophages, DCs, lymphocytes, and exhausted CD8 T cells) were readily distinguishable (Supplementary Fig. 6c). We therefore also consider this dataset for benchmarking.Cross-validation experiments and gene reproducibilityTo illustrate the performance of S–E model in real datasets (Supplementary Table 1), we performed cross-validation experiments using the procedure as implemented in scmap: (i) we randomly selected 70% of the cells as the reference set, (ii) we then identified informative genes (based on the reference set) with different feature selection methods respectively, (iii) we further trained the RF classifier50 using the reference set with only informative genes selected by different methods (cell labels were defined with unsupervised clustering by the original authors), (iv) the remaining 30% cells were considered as query set, and corresponding cell types were predicted with the trained classifier, (v) the classification accuracy was then quantified with the accuracy score50, which is the similarity between the predicted cell types and the original cell types of the query set, (vi) finally, we repeated this entire procedure for n = 50 times for each dataset.We calculated the reproducibility by intersecting the corresponding sets of variable genes as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{Reproducibility}} = \frac{{{\mathrm{Geneset}}_{{\mathrm{replicate}} - 1}^{m - n} \cap {\mathrm{Geneset}}_{{\mathrm{replicate}} - 2}^{m - n}}}{n},$$\end{document}Reproducibility=Genesetreplicate−1m−n∩Genesetreplicate−2m−nn,where m denotes the adapted gene selection method and n is the number of top-ranked variable genes.Rare cell-type simulationWe simulated the synthetic scRNA-seq data following the same approach in GiniClust2 (https://github.com/dtsoucas/GiniClust2/blob/master/Rfunction/Generate_Simulated_Data.R), specifying two large 1000 cell clusters, and three rare clusters of 10, 20, and 30 cells, respectively. To test the performance of our method, we applied our S–E model to the raw count data to select informative genes and performed follow-up clustering with standard clustering procedure in Seurat. The R scripts of RaceID3 and GiniClust2 were accessed through https://github.com/dgrun/RaceID3_StemID2_package and https://github.com/dtsoucas/GiniClust2, respectively.ROGUE calculationBy taking advantage of the wide applicability of S–E model to scRNA-seq data, we introduce the statistic ROGUE to measure the purity of a cell population as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{ROGUE}} = 1 - \frac{{\mathop {\sum }\nolimits_{{\mathrm{sig}}} ds}}{{\mathop {\sum }\nolimits_{{\mathrm{sig}}} ds + K}},$$\end{document}ROGUE=1−∑sigds∑sigds+K,where the parameter K is used for two purposes: (i) constrain the ROGUE value between 0 and 1, (ii) serve as a reference factor to provide the purity evaluation. Consider a reference dataset with maximum summarization of significant ds. We set the value of K to one-half of the maximum. In this way, ROGUE will receive a value of 0.5 when summarized significant ds is equivalent to one-half of the maximum. A cell population with no significant ds for all genes will receive a ROGUE value of 1, while a population with large summarization of significant ds is supposed to yield a small purity score. We reasoned that Tabula Muris can be considered as such a plausible reference dataset because it comprises cells from 20 organs, which represents a highly heterogeneous population and was sequenced with both 10X Genomics and Smart-seq2 protocols2. As the technical variation associated with PCR, which is present in full-length-based but not droplet-based technology, will affect the value of ds, we calculated the summarization of significant ds of Tabula Muris for both 10X Genomics and Smart-seq2 datasets (Supplementary Fig. 19). Accordingly, we set the default value of K to one-half of the summarization, i.e., 45 for droplet-based data and 500 for full-length-based data, receptively. The K value can also be determined in a similar way by specifying a different reference dataset in particular scRNA-seq data analyses. Users should be careful when using the default K value on datasets of different species, and we recommend the user to determine the K value by specifying a highly heterogeneous dataset of that species with the DetermineK function in ROGUE package.Silhouette coefficientTo assess the differences of simulated replicates and the separation of different cell clusters, we calculated the silhouette width7, which is the ratio of within-cluster to intercluster dissimilarity. Let a(i) denote the average dissimilarity of cell i to all other cells of its cluster A, and let b(i) denote the average dissimilarity of cell i to all data points assigned to the neighboring cluster, whose dissimilarity with cluster A is minimal. The silhouette width for a given cell i is defined as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s\left( i \right) = \frac{{b\left( i \right) - a(i)}}{{{\mathrm{max}}\left( {a\left( i \right),b\left( i \right)} \right)}}.$$\end{document}si=bi−a(i)maxai,bi.A high s(i) value suggests that the cell i is well assigned to its own cluster but poorly assigned to neighboring clusters.Sequencing depth simulationSequencing depth can vary significantly across cells and thus contributes to a substantial technical confounder in scRNA-seq data analysis. To illustrate that ROGUE is robust to sequencing depth, we generated simulated populations, each consisting of two replicates with only differences in sequencing depth (Fig. 4d and Supplementary Fig. 7a). In each simulation, we varied the sequencing depth of the two replicates as\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _{{\mathrm{replicate}} - 2,i} = \mu _{{\mathrm{replicate}} - 1,i} \cdot \delta ,\,i \in \left\{ {1, \ldots ,n} \right\},$$\end{document}μreplicate−2,i=μreplicate−1,i⋅δ,i∈1,…,n,where n is the number of genes, μ is the mean expression level, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\delta \in \left\{ {2,\,5,\,10,\,20,\,50,\,70,\,100} \right\}$$\end{document}δ∈2,5,10,20,50,70,100.Generation of simulated cell typesTo demonstrate the potential for ROGUE to guide single-cell clustering, we used NB model as aforementioned to simulate different scRNA-seq datasets, each consisting of three cell types A, B, and C (1000 cells × 10,000 genes each), where A and B were similar subtypes. For the three scenarios shown in Fig. 3a and Supplementary Fig. 15a, d, we introduced 500, 1000, and 800 varied genes between cell-type A and cell-type B/C, respectively, with fold changes drawn from the log-normal distribution (μ = 0 and σ = 2). In addition, we simulated 100, 100, and 120 highly variable genes between cell-type B and C respectively, with fold changes sampled from a log-normal distribution with μ = 0 and σ = 1. The results were visualized using t-distributed stochastic neighbor embedding (t-SNE) implemented in R package Rtsne.Analysis of the fibroblast and B-cell datasetsTo demonstrate the application of ROGUE-guided analysis in identifying pure subpopulations and detecting precise biological signals, we performed re-clustering analysis of the fibroblast and B-cell datasets31,35. We filtered out low-quality cells with either <600 expressed genes, over 25,000 or below 600 UMIs. After filtration, a total of 4291 B cells and 1465 fibroblasts were remained. We further applied our S–E model to the raw count data to select informative genes. Although other pioneering methods could be used to calculate size factors39,40,45, we normalized the gene expression matrices using regularized NB regression in Seurat23. The top 3000 genes with maximal ds were used for PCA analysis. To remove batch effects between donors, we performed batch correction using BBKNN51 with the first 50 PCs. Using the leiden clustering approach implemented in scanpy52, each cell cluster was identified by its principle components. This yielded 11 fibroblast subtypes and 7 B-cell subtypes as shown in Figs. 3d and 4b, which were visualized in 2D projection of UMAP53 with default parameters. Accordingly, the purity score of each cluster was calculated with the rogue function in our R package. The calculation of ROGUE is based on raw count data and is independent of methods used for normalization, dimensionality reduction, and clustering.Pathway and TCGA data analysisTo characterize and detect the pathway signals in specific fibroblast subtypes, we performed pathway analyses using hallmark pathways from the molecular signature database54 with GSVA55. The TCGA LUAD and LIHC data were used to investigate the prognostic effect of 13 signature genes (Supplementary Table 5) derived from B_C2_ACTB. To eliminate the effects of different B-cell proportions, we normalized the mean abundance level of these 13 marker genes by the expression of MS4A1 gene, and performed subsequent statistical analyses using GEPIA256 with default parameters.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article.Supplementary information
Supplementary Information
Peer Review File
Reporting Summary
|
nature communications
|
[
"Article"
] |
[
"RNA sequencing",
"Computational models",
"Software"
] |
complex cell types specialized roles1 Characterizing function cell type challenge advances scRNA-seq discover annotate cell types insights organ tumor microenvironment3 cell cell clusters checking signature genes arbitrary imprecise methods parameters normalization feature selection batch correction clustering confound clusters6 assess purity clusters (Fig. 1a).Fig. expression entropy model Identifying pure cell subtypes single-cell data analysis S–E plot Tabula Muris dataset Each point represents one gene relationship E fitted with LOESS regression S–E plot T-cell dataset27 Smart-seq2 protocol Accuracy identifying differentially expressed genes data NB ZINB distribution subpopulation 50% cells center line median AUC value lower upper hinges represent 25th 75th percentiles whiskers 1.5 times interquartile range Discriminating power of genes S–E model HVG Gini M3Drop SCTransform Fano factor RaceID3 estimated by RF with 50 times cross-validation Supplementary Table 1. classification accuracy measured as percentage query cells assigned correct label center line median classification accuracylower upper hinges represent 25th 75th percentiles whiskers 1.5 interquartile range Reproducibility features brain replicates (Supplementary Table 3) ARI dataset five cell lines6 different feature selection methods used pure cluster defined population cells identical function state without variable genes purity assessment relevant for analyses novel pure subtypes detect true biological signals signature genes pure subpopulation mistakenly common signals mixture guided clustering annotation purity evaluation could eliminate misleading conclusions understanding cellular function state behavior approaches silhouette7 DendroSplit8 distance ratio9 optimal number clusters not comparable poor interpretability cluster purity average silhouette value 0.7 indicates strong consistency cluster unknown whether cluster pure or mixture subpopulations frequent dropout events challenges purity evaluation addressed investigating infiltrating nonself cells variable genes unsupervised variable gene detection diverse proposed for quantification selection highly variable genes scran11 identify variable genes comparing variance local regression trend over-dispersion high frequency dropout events selection lowly expressed genes12 Gini coefficient13 quantify variation gene expression designed for rare cell-typeprobabilistic approaches for variable gene selection using dropout rates adapted12 pseudotime analysis discrete clustering dropout metric hinders global distribution gene expression informative genes determined by inspecting weights during dimensionality approaches computationally intensive more time than HVG M3Drop present entropy-based model randomness gene expression in single cells scalable across datasets identifying variable genes with high precision propose Ratio of Global Unshifted Entropy (ROGUE) statistic to quantify purity single-cell population technical factors ROGUE metric enables unbiased assessment cluster purity effective measure quality of published generated cell clusters Applying ROGUE to B cell fibroblast brain data identified pure subtypes application ROGUE-guided analysis detecting precise biological signals approach applicable for scRNA-seq datasets implemented in open-source R package ROGUE freely available scRNA-seq data approximated by negative binomial considered use statistic S (expression to capture disorder randomness gene expression observed strong relationship between S mean expression level (E of genes basis for expression entropy modelS linearly related to E Tabula Muris dataset2 (Fig. characteristic of current droplet experiments nature UMI-based datasets heterogeneous cell population genes exhibit expression deviation constrained randomness expression distribution reduction S informative genes obtained selecting genes with maximal S-reduction (ds) against null expectation purity assessment cell populations S–E model scRNA-seq data quantitative measure ROGUE cell population with no significant ds genes ROGUE value 1 pure subtype population with maximum summarization significant ds purity score ~0.S–E model informative benchmarked S–E against competing feature selection methods (HVG11 Gini13 M3Drop12 SCTransform17 Fano factor18 RaceID319) data NB ZINB distribution generated 1600 evaluation datasets subpopulations 50 20 10 1% cells used AUC standard S–E model achieved highest average AUC outperformed other gene selection methods cases varied subpopulation proportions gene abundance levels (Fig. 1d Supplementary Figs. 1 and 2)SCTransform designed for UMI-based scRNA-seq data exhibited performance on ZINB-distributed datasets (Fig. 1d). genes rare cell types Gini increased performance subpopulations <20% cells HVG performed better subpopulations larger proportion Figs. 1 2) unsupervised feature selection method performed cross-validation experiments using random forest classifier (RF randomly sampled 70% cells original dataset classified remaining 30% defined authors gene sets higher classification accuracy biologically meaningful21 14 datasets method identified genes greater classification when (30–5000) genes selected (Fig. 1e f 3 4) S–E model showed superiority when fewer genes (30–100) used results suggest genes identified model more informative biologically discriminating datasets same system reproducible tested expression entropy model using technical replicates from different tissues Table 2) genes identified S–E model more reproducible when top 500–2000 genes used (Fig. 1g considered four pancreatic datasets datasets more complex included systemic nuisance factorsdifferences model achieved high reproducibility scores Fig. task feature selection genes relevant for biological heterogeneity downstream clustering performance S–E model unsupervised clustering with RaceID319 SC322 Seurat23 considered five scRNA-seq datasets with high-confidence cell labels6,9,24,25 datasets include cells from different lines FACS-purified populations well-characterized types Fig. 6 gold standards similarity between clusters calculated adjusted Rand index (ARI)26 restricted to interval [0, 1] considered top 100 500 1000 2000 genes results S–E model best performance scenarios (Fig. 1h Fig. 7) methods rare cell types tested if genes S–E model effective uncovering rare subpopulations simulated scRNA-seq dataset three rare clusters two common clusters 1000 clustered with GiniClust218 RaceID3 S–E model-based Seurat methods recapitulated five cell clusters Fig. 8) S–E model-based Seurat effective for recovery of common rare cell clusters wondered S–E model detecting real rare cell typesno gold standard considered four cell lines (A549 H2228 H838 HCC827) Tian et al.6 generated three common cell types (A549 H2228 H838 500 cells one rare type (HCC827 20 or 10 cells by down-sampling three methods identified common rare cell clusters 20 cells rare type Fig. 9a–c). dataset rare cell type lower frequency (10 cells 0.6% total cells RaceID3 GiniClust2 rare cell type S–E model Seurat 9d–f). S–E model effective rare subpopulations methods GiniClust2 RaceID3 appropriate.Evaluation robustness ROGUETo ROGUE genes considered two scRNA-seq datasets T-cell dataset Smart-seq25 droplet-based dataset2 heterogeneity score (1-ROGUE) saturation when genes significant ds selected used significant ds calculate ROGUE investigated performance ROGUE on 1860 cell populations NB ZINB distribution 0.1–50% genes varied second cell type population fewer nonself cells varied genes high purity score converse low-purity score ROGUE index decreased with heterogeneity cell populationsFigs. 11 ROGUE performed populations varied genes (<1%) infiltrating cells<1%) index sensitive unbiased measure cell population purity different values reference factor K similar results Fig. ROGUE robust parameter K.Fig. 2ROGUE use performance ROGUE index K = 45 decreases increasing varied genes mixture cell types (1:1) center line median ROGUE value n = 50 simulations hinges 25th 75th percentiles whiskers 1.5 times interquartile range ROGUE values 45 mixtures cell-type sizes 1:100 to 1:1 varied genes 1% total gene number (n = center line median ROUGE value n = 50 simulations 25th 75th percentiles whiskers 1.5 times interquartile range Pearson correlations randomly down-sampled datasets = 50 entire datasets (2000 cells) center line median correlation value hinges 25th 75th percentiles whiskers 1.5 times interquartile range Sequencing depth distribution two replicates replicate 2 sequencing depth ten times replicate 1. S–E plot mixture replicates 1 2 shown d.ROGUE values n = 100 mixtures versus silhouette values two replicates high silhouette value indicates difference sequencing depth replicates S–E plots ROGUE values 10 cell populations PBMC dataset24. Purity assessment six human T-cell populations Purity evaluation lung-cancer infiltrating DCs point patient center line median ROUGE value lower upper hinges represent 25th 75th percentiles whiskers 1.5 times interquartile range number cells challenge S ROGUE calculation down sampling analysis S cell numbers Pearson correlations S down-sampled datasets datasets similarity values >0.99 S ROGUE calculation affected variation cell number (Fig. 2c).Sequencing depth across cells orders contributes technical confounder scRNA-seq data ROUGE index assess purity single-cell population technical effect simulated increasing molecular counts (sequencing second mock replicate fold change gene expression means 2 to 100 (Fig. 2d two replicates expected pure cell population used silhouette measure replicate-to-replicate differences results revealed ROGUE values ~0.99–1 population two replicates silhouette values 0.25 to 0.75 (Figf Supplementary Fig. 14a). ROGUE purity single-cell population accounts variation sequencing.ROGUE assesses purity cell considered External RNA Controls Consortium (ERCC) dataset24 controlled experiment technical variability 1015 droplets received same ratio ERCC synthetic spike-in RNAs no varied RNAs ideal pure cell population found one RNA with significant ds. dataset achieved ROGUE value ~1 confirming purity investigated fresh peripheral blood mononuclear cells) from single healthy provided multiple cell types purified by FACS suitable resource purity assessment cell types Fig. 2g CD4/CD8 naïve T CD4 memory T cells homogeneous detected high ROGUE values (0.94–1) CD14 monocytes CD34+ cells diverse received low ROGUE values (~0.8 confirming heterogeneity ROGUE index on pure subtypes unsupervised clustering considered six T-cell subtypes from human colorectal cancer5 generated Smart-seq2 protocol pure subtypes achieved high ROGUE values >0.9 (Fig. versus 0.78 complete dataexamined four dendritic cell (DC) subsets from human lung sequenced with inDrop tumor-infiltrating DC2 cells heterogeneous deviated from cell types DC1 LAMP3+ DC pDC (Fig. results illustrate ROGUE effective cell population purity without affected technical characteristics.ROGUE-guided analysis enhances cell-type evaluated potential ROGUE guide clustering analysis with silhouette intercluster dissimilarity within-cluster dissimilarity simulated scRNA-seq dataset three cell types A B C similar 1% varied genes clustered dataset into 2 3 4 5 subpopulations resolution parameter Seurat23 evaluated results silhouette ROGUE values Proper clustering should three subpopulations one each cell type silhouette maximum value when cell-type A co-clustered with B two clusters poorly interpretable for cluster purity ROGUE saturation when three clusters Repeating simulation with varied differences cell-type A B C yielded equivalent performance performance observed when different values reference factor K usedROGUE purity quantification cluster independent normalization reduction clustering splitting merging clusters unsupervised clustering analyses.Fig. 3ROGUE enhances single-cell clustering cell-type identification t-SNE plots simulated dataset three cell types silhouette values average ROGUE values 2 3 4 5 clusters UMAP plots lung-cancer-associated fibroblasts color-coded clusters original paper Supplementary Fig. re-clustered labels ROGUE values clusters before after re-clustering point represents patient center line median ROUGE value lower upper hinges represent 25th 75th percentiles whiskers 1.5 times interquartile range UMAP plot expression levels MYH11 MEF2C Differences hallmark pathway activities scored GSVA ROGUE clustering examined dataset cancer-associated fibroblasts lung tumors heterogeneous population tumor-supportive seven fibroblast clusters low ROGUE values (Fig. 3d Fig. performed re-clustering analysis heterogeneity identified 11 clusters higher average ROGUE value(myofibroblastic inflammatory antigen-presenting (apCAFs high expression CD74 MHC class-II genes Fig. 17b). apCAFs discovered fibroblast subtype mouse pancreatic ductal adenocarcinoma barely detectable in human PDAC apCAFs lung cancer differences cancer types myCAFs (AF_C02_COL4A1 ROGUE = 0.81) segregated into three subpopulations BF_C01_RGS5 (ROGUE = 0.84), BF_C02_ACTA2 0.87) BF_C03_GPX3 = 0.94) signature genes AF_C02_COL4A1 specific to subpopulations MEF2C BF_C01_RGS5 MYH11 BF_C02_ACTA2 NOTCH signaling activated in BF_C01_RGS5 AF_C02_COL4A131 increase ROGUE index BF_C00_AOL10A1 BF_C04_COL1A2 BF_C05_PLA2G2A low ROGUE values further investigation ROGUE-guided analysis discovered novel cell subtypes enabled detection true signals in subpopulationsROGUE-guided analysis B cell subtypesB key tumor microenvironment unclear functions antitumor investigated liver- lung-tumor-infiltrating B low ROGUE values applied clustering analysis ROGUE cells discover pure subtypes seven clusters identified specific marker genes (Fig. 4b–d). first B-cell subset B_C0_JUNB expressed signature genes JUNB FOS activated B second subset B_C1_TXNIP high expression glycolysis pathway genes metabolic differences ACTB gene antigen presenting expressed third subset (B_C2_ACTB). analysis strong antigen processing presentation signal fourth cluster B_C3_FCER2 high expression HVCN1 genes B-cell receptor signaling pathway pre-activated B cells37 fifth cluster B_C4_MX1 interferon-induced B expressed high levels MX1 IFI6 IFI44L sixth cluster B_C5_CD3D expressed markers T- B-cell lineages remaining cells seventh cluster B_C6_LRMP high expression LRMP RGS13 germinal center B cells40-guided analysis B-cell subtypes S–E plots ROGUE values liver lung-tumor-infiltrating B cells UMAP plots 4291 B cells color-coded clusters tissues Gene expression heatmap seven B-cell clusters Rows genes columns clusters ROGUE values seven B-cell subtypes point represents patient center line median ROUGE value lower upper hinges 25th 75th percentiles whiskers 1.5 interquartile range Tissue preference B-cell subtype cancer RO/E27 ratio observed expected cell numbers chi-square test average fractions B_C02_ACTB B_C04_MX1 patient error bars < 0.05 < 0.005 Kaplan–Meier curves TCGA LUAD LIHC patients grouped 13 markers B_C02_ACTB DEs/doublets-like germinal center B cells low ROGUE values limited cells permit clustering germinal center B cells high expression proliferating marker genes MKI67 STMN1 explaining heterogeneity ROGUE values >0.92 remaining five clusters homogeneous B-cell subtypesratio observed expected cell numbers chi-square test B_C02_ACTB B_C04_MX1 contained cells tumor RO/E values >1 (Fig. analyses patient confirmed trend (Fig. TCGA lung adenocarcinoma patients higher expression genes B_C02_ACTB showed worse survival (Fig. survival difference observed TCGA hepatocellular carcinoma) cohort dataset. clinical implication deserves further study roles B_C02_ACTB cells tumor microenvironment identifying pure subtypes ROGUE-guided analysis deeper understanding cell state behavior.Application brain data batch effect application ROGUE brain transcriptome dataset2 high heterogeneity identified seven cell types oligodendrocyte neuron cell types low ROGUE values <0.8 versus ~0.9–1 remaining five (Fig. applied clustering ROGUE oligodendrocyte identified ten refined cell subtypes specific marker genes (Fig. 5b Except cluster 6 found ROGUE values ~0.9–1 other nine clusters suggesting purity (Fig compared pathway activities found diversity cluster 5 showed strong axon guidance signalingneurotrophin activated cluster 1 (Fig. illustrates ROGUE uncovering subpopulations. application ROGUE brain data batch effect evaluation ROGUE values seven brain cell types each point sample center line median ROUGE value lower upper hinges represent 25th 75th percentiles whiskers 1.5 times interquartile range UMAP plot ten clusters oligodendrocytes (n = 3401) color-coded heatmap cell-type-specific genes ten oligodendrocyte clusters ROGUE values clusters Each point represents sample center line median ROUGE value lower upper hinges 25th 75th percentiles whiskers 1.5 times interquartile range Enriched pathways cluster 5 1 ROGUE values batch 1 2 cell population equalized cells down-sampling center line median ROGUE value lower upper hinges 25th 75th percentiles whiskers denote 1.5 times interquartile range *p < 0.05 **p < 0.005 test ROGUE values individual-specific cell populations populations cells control group Subsampling equalize cells center line median ROUGE valuelower upper hinges represent 25th 75th percentiles whiskers denote 1.5 times interquartile range. *p < 0.05 **p < 0.005 Student’s t test ROGUE impact batch effect studied dataset human PBMCs multiple cell Cells split into two interferon-beta (IFN-β)-stimulated group culture-matched control group two batches applied ROGUE purity each cell type individual bathes cell population (batch 1 batch 2) ROGUE detected purity reduction in group (Fig. cells collected from eight unrelated individuals tested ROGUE variability batch effect among patients used cells control group IFN-β perturbation aggregated cell populations received lower ROGUE values patient-specific populations (Fig. 5h). ROGUE offers reasonable method for estimating impact batch effect.DiscussionPurity assessment of cell clusters paramount interpretation scRNA-seq data pertinent rare subtle cell subtypes uncovered S–E model variable genes high sensitivity precision applied to clustering pseudotime analyses statistic ROGUE quantify purity of single-cell populationsentropy-based measure ROGUE applicable for datasets different platforms protocols operators purity of-cell populations cell-to-cell variation ROGUE purity four DC subtypes human lung tumors DC2 heterogeneous consistent with previous heterogeneous populations different properties roles in cancer microenvironment assessed with ROGUE future studies could focus deepen understanding cellular origins cancer ROGUE addresses need unsupervised single-cell data analyses quality of clusters unsupervised clustering under- over-clustering universal clustering quality quantifying cluster purity with ROGUE low-purity clusters analysis pure subtypes Improving purity credibility of cell types challenge single-cell sequencing ROGUE could potential standard for quality cell clusters ROGUE analysis on fibroblasts identified novel subpopulation in lung cancer apCAFs expressed CD74 MHC class-II genes strong antigen-presenting signal cells deactivate CD4 T cells decrease CD8+ to Treg ratio in mouse PDAC33 unclear role in lung-cancer microenvironment further investigation ROUGE to B-cell analysis found pure cluster B_C02_ACTB high expression of genes antigen processing presentationCells from cluster enriched in tumors associated with poor outcomes in lung liver cancer hypothesize cells may contribute to immune suppression cancer curtail antitumor immunity further studies required approaches for discovering pure subtypes to other scRNA-seq datasets determining purity cell recommend ROGUE value 0.9 threshold infiltrating cells varied genes constrained for low-quality threshold determined global ROGUE values ROGUE efficient effective anticipate additional extensions enhanced performance assessing purity integrated cell populations protocols ROGUE metric robust measure for cluster purity technical confounders expect applicable to scRNA-seq datasets strategy improve rigor quality of unsupervised single-cell data analysisMethodsExpression entropy droplet datasets observed UMI count modeled NB random variable Poisson–Gamma mixture411[12pt{minimal}\usepackage{amsmath{wasysym{upgreek\oddsidemargin-69pt}{document}{array}}X_{ij}\mathrm{Poisson}}\lambda _{ij}}{ij{Gamma}}\alpha _{ij\beta _{ij}}{array}{document}Xij~Poisson(sjλij)λij~Gamma(αij λij expression value observed UMI count Xij gene i cell j sj size normalization factor cell j αij βij shape parameter rate parameter shape parameter α constant across cells genes rate parameter β constant gene i αij βij expressed as α and βidistributions recognized as[12pt]{minimal{amsmath{wasysym\oddsidemargin}{-69pt}{document}$$\lambda _i\sim\mathrm{Gamma}}\left\alpha \beta _i\end{document}λi~Gammaα,βi[12pt]{minimal}{amsmath{wasysym{upgreek}{\oddsidemargin}{-69pt}{document}$$X_{ij}\sim\mathrm{Poisson}}\lambda _i\end{document}Xij~Poisson(sjλi).[12pt]{minimal}{amsmath}{wasysym{amsfonts{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}$$X_{ij} =\frac{{X_{ij}}}{{s_j}}\end{document}Xij′=Xijsj normalized expression of gene i in cell j use[12pt]{minimal}\usepackage{amsmath}{wasysym{amsfonts{mathrsfs{upgreek\oddsidemargin{-69pt}{document}\Bbb E}\left( {X_i^\prime }\right\end{document}EXi′[12pt]{minimal{amsmath{wasysym{mathrsfs{upgreek\oddsidemargin{-69pt}$X_i^\prime\end{document}Xi′ normalized expression gene i[12pt]{minimal}{amsmath{wasysym{amsfonts{mathrsfs{upgreek}\setlength\oddsidemargin{-69pt}{document}{\Bbb E}\left( {X_i^\prime }\right\end{document}EXi′ expectation across cells moment estimation of λi.Gamma distribution rate parameter calculated maximum likelihood estimation3\documentclass[12pt]{minimal{amsmath\oddsidemargin-69pt}{document}\beta _i = \frac{\alpha\lambda\frac{\alpha\Bbb E}\left( {X_i^\prime \right\end{document}βi=αλi=αEXi′ disorder randomness gene expression differential entropy\documentclass[12pt]{minimal}{amsmath\oddsidemargin-69pt}{document}$H\left( X \right) = - \int{ - \infty + \infty }\left( x \right)\mathrm{ln}}\left( x \right)dx\end{document}HX=−∫−∞+∞px⋅lnpxdx X continuous random variable p(x) probability density function Differential entropy extension Shannon entropy average surprisal continuous probability distribution performance supervised gene selection method E-test44.gamma distributed random variable λi differential entropy computed[12pt]\usepackage{amsmath-69pt$S_i =\alpha -\mathrm{ln}}\beta _i +\mathrm{ln}}\Gamma\left\alpha \right) + {1 - \alpha \right) =\mathrm{ln}}\alpha\beta _i}} + a =\mathrm{ln}}\Bbb\left(_i\prime \right) + a{document}Si=α−lnβi+lnΓα+1−α⋅φα=lnαβi+a=lnEXi′+a φ digamma function[12pt]{minimal\usepackage{amsmath\oddsidemargin{-69pt}}$$a = \alpha -\mathrm{ln}}\alpha +\mathrm{ln}}\Gamma \left( \alpha \right) +\left( {1 - \alpha \right)\upvarphi\left(alpha \right)$\end{document}a=α−lnα+lnΓα+1−α⋅φα constant methods Scnorm45 scran46 BASiCS47 calculate size factors considered library size normalization cell total UMI counts divided by mean total UMI counts across cells41 expectation library size factor across cells equal to 1. Eq. (2) gene expression library size independent random variables42 gene i\documentclass[12pt]{minimal}\usepackage{amsmath}{wasysym}{amsfonts}{mathrsfs}{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}\Bbb E}\left( {X_i \right) = s\right 1\end{document}EXi=EXi′×s=EXi′×Es=EXi′×1=EXi′ Xi observed expression gene i s library size cells.each cell type differential entropy λi computed[12pt]{minimal{amsmath{wasysym{upgreek}\oddsidemargin-69pt}{document}$S_i =\mathrm{ln}} E}\left {X_i}\right) + a\end{document}Si=lnEXi+a formulate null hypothesis one Poisson–Gamma component each gene population (H0) differential entropy calculated Eq. (7) each cell represents cluster use Xij moment estimation mean expression clusterdefine entropy reduction gene i across n cells\documentclass[12pt{minimal{amsmath{upgreek\oddsidemargin-69pt}{document}_i =\mathrm{differential}}{entropy}}{average}}{actual}}{differential}}{entropy}}\\{ln}}\Bbb E\left {X_i}\right) -\frac{{\mathop\sum\nolimits_{j = 1}^n\mathrm{ln}}X_{ij}}\end{document}dsi=differentialentropyunderH0−averageactualdifferentialentropy=lnEXi−∑j=1n(lnXij)n captures disorder randomness gene expression44genes under H0 (non-variable major proportion relationship between[12pt{minimal{amsmath{wasysym\oddsidemargin-69pt}{document\mathrm{ln}}_i{document}lnEXi average differential entropy calculate residual as dsi improve performance (Fig. 1b, c). significance of ds estimated normal distribution adjusted Benjamini–Hochberg method extended procedure to full-length datasets approach outperformed gene selection methods (Fig. 1f Supplementary Fig. 4) simulated droplet datasets with NB distributiongene abundance levels E sampled from log-normal distribution[12pt]{minimal{amsmath{wasysym\oddsidemargin-69pt}}$$ln\left E\right\sim {N} ^2{document}lnE~Nμ,σ2 parameters μ = 0 σ = 2. transcripts each gene drawn[12pt]{minimal}{amsmath{wasysym\oddsidemargin}{-69pt}}$$N_{ij}\mathrm{NB}}\left {E_i,r} \end{document}Nij~NBEi,r simulated dataset dispersion parameter r = α)48 fixed value 5 to 20 (Supplementary Fig. 1) simulated full-transcript datasets with ZINB distributiondropout rates gene modeled sigmoid function49\documentclass[12pt{minimal{amsmath{wasysym\oddsidemargin-69pt\mathrm{sigm}}\left\gamma _0 _1E_i~sigm−γ0+γ1Ei parameters γ0 = −1.5 γ1 = 1/median(E). simulated scRNA-seq dataset 20,000 genes 2000 cells.Differentially expressed genes added fraction cells (1–50% changes sampled log-normal distribution (μ = 0 σ = 2) Genes >1.5-fold decrease increase mean expression ground truth DE genes.Feature selection HVG method11 identifies variable genes coefficient variation squared local regression trend implemented BrenneckeGetVariableGenes function M3Drop12 package Gini index model GiniClust13 gene informative if Gini higher than expected maximum observed expression copied source code original GiniClust defined Gini_fun function scripts select genesM3Drop uses dropout rates variable gene selection implemented M3DropFeatureSelection function SCTransform selects genes with Pearson residuals from negative binominal regression implemented SCTransform function Seurat package implemented Fano factor method GiniClust2_Fano_clustering.R feature selection step RaceID319 selects genes with second-order polynomial fit between expression variance log-transformed mean implemented fitBackVar function RaceID3.Datasets clustering feature selection methods unsupervised clustering considered five scRNA-seq datasets first five cell lines (A549 sequenced with 10X Genomics protocol 3918 cells second dataset three cell lines (H1975 H2228 HCC827) sequenced-seq2 protocol third multiple FACS-purified cell populations sequenced with 10X Genomics protocol used CD19 B cells CD4 naïve T cells CD56 NK cells CD14 monocytes distinguishable fourth dataset cells human pancreatic islet generated Smart-seq pancreatic cell types alpha beta delta gamma cells well-characterized used for benchmarkingremaining dataset multiple immune cell types9 cells sequenced by Smart-seq2 protocol cell labels assigned unsupervised clustering cross-validation experiments revealed major cell types (macrophages DCs lymphocytes exhausted CD8 T cells distinguishable Fig. 6c). consider dataset for benchmarking.Cross-validation experiments gene reproducibilityTo performance S–E model datasets performed cross-validation experiments procedure scmap randomly selected 70% cells reference set identified informative genes different selection methods trained RF classifier50 genes labels defined unsupervised clustering remaining 30% cells query set cell types predicted with trained classifier classification accuracy quantified with accuracy score50 similarity between predicted cell types original cell types repeated procedure = 50 times each datasetcalculated reproducibility intersecting variable genes[12pt{amsmath{wasysym{upgreek\oddsidemargin-69pt\mathrm{Reproducibility}} ={Geneset}} - 1}{m - n} - 2 - n}Reproducibility=Genesetreplicate−1m−n∩Genesetreplicate−2m−nn m adapted gene selection method n top-ranked variable genes.Rare cell-type simulated synthetic scRNA-seq data GiniClust2 specifying two large 1000 cell clusters three rare clusters 10, 20 30 cells test performance applied S–E model raw count data genes performed follow-up clustering standard procedure Seurat R scripts RaceID3 GiniClust2 accessed https://github/dgrun/RaceID3_StemID2_package/dtsoucas/GiniClust2ROGUE calculationBy S–E model scRNA-seq data introduce statistic ROGUE purity cell population[12pt]{minimal}\usepackage{amsmath}{wasysym{amsfonts{mathrsfs{upgreek}\oddsidemargin{-69pt}{document\mathrm{ROGUE}} = 1 - ds + K}}}ROGUE=1−∑sigds∑sigds+K parameter K constrain ROGUE value between 0 and 1 reference factor purity evaluation reference dataset with maximum summarization significant ds set value K to one-half of maximum ROGUE value 0.5 when ds equivalent to one-half maximum cell population with no significant ds for all genes ROGUE value 1 population large summarization significant ds small purity score Tabula Muris plausible reference dataset cells from 20 organs heterogeneous population sequenced with 10X Genomics and Smart-seq2 protocols2. technical variation PCR value ds calculated summarization significant ds Tabula Muris for 10X Genomics and Smart-seq2 datasets Fig.set default value K to one-half summarization 45 for droplet-based 500 full-length-based data K value determined specifying different reference dataset scRNA-seq data analyses careful using default K value datasets different species recommend determine K value specifying heterogeneous dataset with DetermineK function ROGUE package.Silhouette coefficientTo assess differences simulated replicates separation cell clusters calculated silhouette width7 ratio within-cluster to intercluster dissimilarity a(i) average dissimilarity cell i to other cells cluster A b(i) average dissimilarity cell i to all data points neighboring cluster dissimilarity minimal silhouette width for cell i defined as\documentclass[12pt]{minimal}\usepackage{amsmath\left\right =\right - a(i}si=bi−a(i)maxai,bi high s(i) value suggests cell i well assigned to own cluster poorly assigned to neighboring clustersSequencing depth across cells technical confounder scRNA-seq data analysis ROGUE robust generated simulated populations two replicates differences sequencing depth (Fig. 4d Supplementary Fig.7a). each simulation varied sequencing depth two replicates[12pt]{minimal}{amsmath{wasysym-69pt}{document}$\mu\mathrm{replicate}} - 2,i} ={replicate}} - 1,i} \delta\left {1 ,n}\right\end{document}μreplicate−2,i=μreplicate−1,i⋅δ,i∈1,...,n n number genes μ mean expression level[12pt]{minimal}{amsmath{upgreek\oddsidemargin{-69pt}{document}$\delta\left {2,,5,10,20,50,70,,100}\right\end{document}δ∈2,5,10,20,50,70,100.Generation simulated cell potential ROGUE guide single-cell clustering used NB model simulate scRNA-seq datasets three cell types A B C (1000 cells × 10,000 genes A B similar subtypesscenarios Fig. 3a 15a introduced 500 1000 800 varied genes cell A B/C changes log-normal distribution (μ 0 σ = 2) simulated 100 120 variable genes between B C fold changes log-normal distribution μ = 0 σ = 1. results visualized t-distributed stochastic neighbor embedding-SNE) R package fibroblast B-cell ROGUE-guided analysis subpopulations biological signals re-clustering analysis fibroblast-cell filtered low-quality cells <600 expressed genes over 25,000 below 600 UMIs 4291 B cells 1465 fibroblasts remained applied S–E model raw count data informative genes normalized gene expression matrices NB regression Seurat23 top 3000 genes maximal ds used PCA analysis batch effects batch correction BBKNN51 first 50 PCs clustering approach cell cluster identified principle components yielded 11 fibroblast subtypes 7 B-cell subtypes Figs. 3d 4b visualized 2D projection UMAP53 default parameters purity score cluster calculated rogue function R package calculation based raw count data independent normalization reduction clusteringPathway TCGA data pathway signals fibroblast subtypes performed pathway analyses hallmark pathways molecular signature database54 GSVA55 TCGA LUAD LIHC data prognostic effect 13 signature genes Table 5) B_C2_ACTB normalized abundance 13 genes expression MS4A1 gene performed statistical analyses GEPIA256 default parameters.Reporting Nature Research Reporting Summary.Supplementary information Peer Review File Reporting Summary
| 50.4
| 1.137954
|
10.1038/s41467-020-20466-9
|
PMC7794343
|
Dual oxidases (DUOXs), assembled from the catalytic DUOX and the auxiliary DUOXA subunits, produce hydrogen peroxide by transferring electrons from intracellular NADPH to extracellular oxygen in a calcium-activated manner. Here authors report the cryo-EM structures of human DUOX1-DUOXA1 complex in both high-calcium and low-calcium states.
|
Dual oxidases (DUOXs) produce hydrogen peroxide by transferring electrons from intracellular NADPH to extracellular oxygen. They are involved in many crucial biological processes and human diseases, especially in thyroid diseases. DUOXs are protein complexes co-assembled from the catalytic DUOX subunits and the auxiliary DUOXA subunits and their activities are regulated by intracellular calcium concentrations. Here, we report the cryo-EM structures of human DUOX1-DUOXA1 complex in both high-calcium and low-calcium states. These structures reveal the DUOX1 complex is a symmetric 2:2 hetero-tetramer stabilized by extensive inter-subunit interactions. Substrate NADPH and cofactor FAD are sandwiched between transmembrane domain and the cytosolic dehydrogenase domain of DUOX. In the presence of calcium ions, intracellular EF-hand modules might enhance the catalytic activity of DUOX by stabilizing the dehydrogenase domain in a conformation that allows electron transfer.
|
IntroductionReactive oxygen species (ROS) are oxygen-containing chemical species that are highly reactive, such as hydrogen peroxide and superoxide anion1. They participate in many physiological processes and are implicated in several pathological conditions1. ROS can be generated by a class of dedicated enzymes called NADPH oxidase (NOX) in a highly regulated manner. These enzymes are multi-pass transmembrane proteins that catalyze the reduction of extracellular or luminal oxygen by intracellular NADPH to generate superoxide anion or hydrogen peroxide. NOX proteins are involved in many biological processes, including host defense, differentiation, development, cell growth and survival, cytoskeletal reorganization, and modification of the extracellular matrix2.Comprising the human NOX protein family are NOX1–5 and DUOX1–2 (ref. 2). NOX2 protein catalyzes the production of superoxide anion during phagocytosis in neutrophils and is essential for host defense3. DUOX1–2 proteins are highly expressed in thyroid tissue and they catalyze the production of hydrogen peroxide, which is important for the biosynthesis of thyroid hormones4. The function of DUOX protein requires physical interactions with an auxiliary protein called dual oxidase maturation factor (DUOXA)5. DUOXA promotes the maturation and proper plasma membrane localization of DUOX5. DUOX protein is encoded by two homologous genes in human, namely DUOX1 and DUOX2. Similarly, DUOXA protein is encoded by DUOXA1 and DUOXA2. Loss-of-function mutations of DUOX2 or DUOXA2 in human cause congenital hypothyroidism6. Because of the important role of DUOX in thyroid tissue, they are also named thyroid oxidase4.NOX family proteins share a common catalytic core, formed by a heme-coordinating transmembrane domain (TMD) and a cytosolic dehydrogenase (DH) domain7. The DH domain binds intracellular substrate NADPH and cofactor FAD, and shares sequence homology to the ferredoxin-NADP + reductase (FNR), which is composed of two subdomains8. In addition to the shared TMD-DH catalytic core of NOX, the functional DUOX protein has an additional large N-terminal extracellular peroxidase homology domain (PHD) and a long intracellular loop 0 containing two EF-hand domains, and it requires an auxiliary DUOXA protein for proper function. The activity of DUOX is regulated by intracellular calcium concentration4. Prior to our studies, the structures of NOX family members are only available in the form of isolated domains, including the DH domain (PDB ID: 5O0X)9 and TMD (PDB ID: 5O0T)9 of NOX5 from the algea Cylindrospermum stagnale (csNOX5) and a subdomain of human NOX2 DH domain (PDB ID: 3A1F). Despite the functional importance of DUOX and other NOX family members, their structures in the context of full-length functional protein complex are still unknown. Several open questions for DUOX remain elusive: How is the DH domain engaged with TMD to perform the catalytic redox reaction? How does DUOXA protein interact and co-assemble with DUOX? How is the activity of DUOX regulated by intracellular calcium? To answer these fundamental questions, we sought to characterize DUOX–DUOXA protein complex both structurally and functionally. Here, we present the cryo-EM structures of human DUOX1–DUOXA1 (hDUOX1–hDUOXA1) complex in both high-calcium and low-calcium states, providing insights into the structure and mechanism of calcium activation for DUOX.ResultsStructure determinationTo express the hDUOX1 protein, we constructed the matured hDUOX1 protein (20–1551) in frame with N-terminal GFP tag guided by a rat FSHβ signal peptide for efficient secretion10. The molecular weight of GFP-tagged hDUOX1 is 219 kDa. To monitor the formation of DUOX1–DUOXA1 complex, we fused the hDUOXA1 protein with a C-terminal MBP-mScarlet tag to increase its molecular weight to 106 kDa. Fluorescence size-exclusion chromatography (FSEC) showed the co-expression of hDUOXA1 effectively shifted the peak of hDUOX1 toward higher molecular weight, suggesting the formation of a stable hDUOX1–hDUOXA1 hetero-complex (Fig. S1a, b). The peak positions indicated hDUOX1 migrated as a monomer, while hDUOX1–hDUOXA1 complex migrated as a heterotetramer (Fig. S1b). Moreover, we found the co-expression of DUOX1 and DUOXA1 resulted in cell membranes that showed robust calcium-activated, NADPH-dependent hydrogen peroxide production detected by the Amplex Red assay11 (Fig. 1a–c). In the low-calcium condition, DUOX1–DUOXA1 complex showed low basal activity (Fig. 1b, c). Addition of calcium not only reduced the Km, but also increased the Kcat of DUOX1 complex, leading to the overall enhancement of enzymatic activity (Fig. 1b,c).Fig. 1Structure of human DUOX1–DUOXA1 complex in the high-calcium state.a Schematic of the DUOX enzymatic assay. In the presence of H2O2 (produced by DUOX), horseradish peroxidase (HRP) converts nonfluorescent Amplex Red to fluorescent resorufin, which is measurable and proportional to H2O2. b Calcium-dependent activation of hDUOX1–hDUOXA1 complex. Data are shown as means ± standard deviations, n = 3 biologically independent samples. Source data are provided as a Source data file. c Steady state enzyme activity of hDUOX1–hDUOXA1 complex as the function of NADPH concentration in the presence or absence of calcium. Data were fit to the Michaelis–Menten equation to obtain the Km and Kcat value. Data are shown as means ± standard deviations, n = 3 biologically independent samples. Source data are provided as a Source data file. d Side view of the cryo-EM map of hDUOX1–hDUOXA1 complex in the high-calcium state. The approximate boundaries of phospholipid bilayer are indicated as gray thick lines. One protomer of hDUOX1 and hDUOXA1 complex is colored as blue and green, the other one is colored as yellow and red, respectively. e A 90° rotated top view compared to d. f A 180° rotated bottom view compared to e. g Top view of the cross-section of the transmembrane layer at the position indicated as a dashed line in d. The large cavity in the transmembrane layer is indicated by dashed oval. For clarity, the cryo-EM map was low-pass filtered to 6 Å. h Topology of hDUOX1 and hDUOXA1 subunits. Transmembrane helices are shown as cylinders, unmodeled disordered regions are shown as dashed lines. The phospholipid bilayer is shown as gray layers. PHD peroxidase homology domain of hDUOX1, PHLD pleckstrin homology-like domain of hDUOX1, EF EF-hand calcium-binding module of hDUOX1, DH dehydrogenase domain of hDUOX1, CLD claudin-like domain of hDUOXA1. i Structure of one protomer of hDUOX1 and hDUOXA1 complex in cartoon representation. The colors of each individual domain are the same as in g. The approximate boundaries of phospholipid bilayer are indicated as gray thick lines. Sugar moieties, hemes, FAD, and NADPH are shown as black, yellow, pink, and green sticks, respectively.We solubilized and reconstituted hDUOX1–hDUOXA1 complex into peptidisc using NSPr12. Purified peptidisc sample showed a homogenous peak on SEC (Fig. S1c) and two major protein bands on SDS–PAGE, both of which could be trimmed upon PNGase F treatment (Fig. S1d), suggesting both of hDUOX1 and hDUOXA1 were modified by N-linked glycosylation. UV–vis spectrum showed the peptidisc sample has characteristic Soret band with peak at 415 nm (Fig. S1e), indicating proper Fe (III) heme incorporation. Moreover, the highly purified peptidisc sample recapitulated the calcium-activated NADPH-dependent hydrogen peroxide production observed on membrane (Fig. S1f), confirming that the calcium-dependent activation is a built-in mechanism of hDUOX1–hDUOXA1 protein complex. However, we found the maximum activity of purified peptidisc sample was lower than the activity measured using crude cell membrane (Fig. 1b and Fig. S1f), suggesting either membrane bilayer or endogenous lipids might play a role on DUOX activity. We prepared cryo-EM grids using the peptidisc sample, either in the presence of 2.5 mM ethylene glycol tetraacetic acid (EGTA; low calcium) or 0.5 mM free calcium (high calcium). Both samples contained 0.1 mM FAD as cofactor and 0.5 mM NADPH as substrate.Single particle cryo-EM analysis showed the purified protein was homogeneous and showed twofold symmetry (Figs. S2–S4). The overall resolution of cryo-EM maps in the low-calcium and high-calcium states reached 2.7 and 2.6 Å, respectively (Table S1). The extracellular domains and TMD showed better local resolution than the cytosolic domains, suggesting the higher mobility of the cytosolic domains (Figs. S2g and S4g). To further improve the map quality of cytosolic domains, we exploited symmetry expansion13 and multibody refinement14 by dividing one protomer into the large body (the extracellular domain and TMD) and the small body (the cytosolic domains; Figs. S2c and S4c). The final resolutions of cytosolic domain reached 3.4 and 3.2 Å for the low-calcium and high-calcium states, respectively (Figs. S2–4 and Table S1). The high map quality and available homology structures allowed us to build the order regions of the complex, which encompassed 88% of DUOX1 and 79% of DUOXA1 (Figs. S5–8 and Table S1). In the following text, we will focus on the high-calcium state structure unless noted otherwise, because of its higher resolution.The architecture of hDUOX1–hDUOXA1 protein complexhDUOX1 subunits and hDUOXA1 subunits co-assemble into a 2:2 heterotetrameric protein complex with molecular weight ~457 kDa. The complex encompasses 140 Å × 105 Å × 160 Å 3D space and has an overall twofold rotational symmetry (Fig. 1d–f). Vertically, the complex can be divided into three layers: the extracellular layer, the transmembrane layer, and the cytosolic layer (Fig. 1d). In the extracellular layer, the two large N-terminal PHD domains of hDUOX1 pack against each other diagonally and are buttressed by the extracellular domain of DUOXA1 from beneath (Fig. 1d–f). The transmembrane layer is formed by 24 transmembrane helices and harbors the heme-binding sites that provide the electron transfer pathway across the membrane (Fig. 1g). At the center of the transmembrane layer, there is a large cavity without discernable protein densities. The interior surface of this cavity is highly hydrophobic (Fig. S3i) and there are several lipid molecules bound on this surface (Fig. S3i), suggesting this cavity is probably filled by phospholipids on the cell membrane. The cytosolic layer is comprised of the catalytic DH domain and regulatory domains for intracellular calcium sensing (Fig. 1f, i).Structure of the catalytic hDUOX1 subunithDUOX1 is the catalytic subunit of the complex (Fig. 2). On the extracellular side of hDUOX1 resides the large N-terminal PHD domain which shares sequence homology with several peroxidases, such as peroxidase A from Dictyostelium discoideum (DdPoxA, PDB ID: 6ERC)15 (Fig. S9a). Functional peroxidases utilize histidine-coordinated heme as the cofactor for catalysis. However, key residues for heme binding, such as the heme ligand histidine, are missing in the PHD of hDUOX1. Indeed, we did not observe any heme density in the structure of hDUOX1 PHD, suggesting PHD is probably not enzymatic functional in term of peroxidase activity. Close inspection of the map reveals two putative cation densities in PHD. One cation (cation binding site 1, CBS1) is coordinated by the side chains of D397 and T332, and the main chain carbonyl groups of V399, T332, and R395 (Fig. S9b). The second cation (CBS2) is coordinated by the side chain of D109, D174, S176, and T170, and the carbonyl groups of T170 and W172 (Fig. S9b). We observed strong densities in these two sites in both low-calcium and high-calcium conditions (Fig. S9b), suggesting the bound cations might be sodium ions which were present in large quantities in our protein sample or calcium ions that bind very tightly. Both CBS1 and CBS2 are evolutionary conserved in DUOX (Fig. S5) and DdPoxA15 (Fig. S9c), indicating their functional importance. Interestingly, we found both CBS1 mutant (D397A + T332A) and CBS2 mutant (D109A + D174A) of DUOX1 failed to co-assemble with DUOXA1 (Fig. S9d). Because CBS1 and CBS2 are away from the subunit interfaces in the DUOX1–DUOXA1 complex, we speculate these mutants probably affect the folding of PHD domain, suggesting the role of CBS1 and CBS2 in protein stability.Fig. 2Structure of hDUOX1 subunit.a Side view of hDUOX1 subunit in the high-calcium state, highlighting the key interfaces (boxed by dashed lines). Each domain is colored as in Fig. 1h. The surface of hDUOX1 is shown in transparency. b The binding site of outer heme in the TMD. Heme is shown as sticks and colored in yellow. Unrelated helices in TMD are omitted for clarity. The putative oxygen-reducing center is indicated by arrow. c The binding site of inner heme in the TMD. d The interface between PHD and TMD boxed in a. Disulfide bond between C118–C1165 is shown as golden sticks. e The interface between PHLD and TMD boxed in a, the hydrogen bonds are indicated with dashed lines. f The interface between PHLD and DH domain. g The interface between EF module and DH domain. h The FAD-binding site located at the interface between TMD and DH domain. Ligands and interacting residues are shown as sticks. i The NADPH-binding site located at the interface between TMD and DH domain.The PHD packs on top of the TMD of DUOX1 through multiple noncovalent interactions (Fig. 2a). Moreover, a disulfide bond between C118 on PHD and C1165 on loop C of TMD further staples the bottom of PHD onto the top of TMD (Fig. 2a, d). In the TMD, hDUOX1 has an extra bent M0 helix at the periphery of M3 and M4, together with the canonical six TM helices of NOX protein family. M1–M6 of hDUOX1 form two heme-binding sites within the TMD. H1144 on M3 and H1238 on M5 coordinate the outer heme (Fig. 2b). H1130 on M3 and H1225 on M5 coordinate the inner heme (Fig. 2c). These four histidines are absolutely conserved in NOX family proteins (Fig. S6). We observed a spherical density surrounded by the invariant R1087 on M2, H1148 on M3, and outer heme-coordinating residue H1144 (Fig. 2b and Fig. S3c). Previous studies showed mutations of the csNOX5 residues corresponding to R1087 and H1148 of hDUOX1 affected the reoxidation of dithionite-reduced TMD by oxygen and this site was proposed to be the oxygen substrate binding site, namely oxygen-reducing center (Fig. S10a, b)9. Our structure observations in hDUOX1 support the hypothesis.Preceding the M1 helix of DUOX1 TMD, an amphipathic preM1 helix floats on the inner leaflet of plasma membrane (Fig. 1h). This helix was previously observed in csNOX5 (ref. 9) and is probably a shared feature of NOX family proteins. Between M0 and preM1 is a long cytosolic fragment loop 0. Cryo-EM maps reveal that the N-terminal of loop 0 is a domain rich of β sheets (Figs. S3g and S10c). Structural search using DALI server16 identified the β sheets-rich domain is a crypto pleckstrin homology-like domain (PHLD) that shares little sequence homology, but high structural similarity to the PH domain proteins (Fig. S10c)17.Following the PHLD, two EF-hand type calcium-binding domains (EF1 and EF2) form a compact helical module that is connected to the PHLD through αC (Fig. S6). Residues predicted to be responsible for calcium binding in EF1 and EF2 are evolutionary conserved in DUOX family proteins (Fig. S6). Although we did not observe the strong densities for small calcium ions due to poor local resolution (Figs. S2–3), the structure of EF-hand module closely resembles the small subunit of calcium-dependent protein phosphatase calcineurin in the calcium-bound state (PDB ID: 4IL1)18 (Fig. S10d), suggesting both EF1 and EF2 are loaded with calcium in the high-calcium state. Based on the homology structure (4IL1), side chains of D828, D830, N832, and E839 and the main chain carbonyl group of Y834 chelate one calcium ion in EF1 (Fig. S10e) and side chains of D864, D866, N868, E875, and the main chain carbonyl group of L870 chelate another calcium ion in EF2 (Fig. S10f). It is reported that mutations of any of these calcium-binding sites abolished calcium activation19 and E879K mutation in hDUOX2 (E875 in hDUOX1) leads to congenital hypothyroidism20, emphasizing their importance in calcium activation.The C-terminal catalytic DH domain is connected to M6 of DUOX1 TMD via a short linker (Fig. 1h, i). The DH domain of hDUOX1 has a canonical DH fold and its structure is similar to csNOX5 (ref. 9; Fig.1i and Fig. S7). We observed strong densities for both FAD cofactor and NADPH substrate, and their binding sites were contributed from not only DH domain, but also TMD (Fig. 2h, i), as described later.Inter-domain interactions in the high-calcium stateIn the high-calcium state, individual domains of DUOX1 in the cytosolic layer are stabilized by multiple inter-domain interactions. The PHLD interacts with adjacent TMD and DH domains (Fig. 2e). The main chain carbonyl group of K653 on PHLD makes hydrogen bond with R1215 on loop D of TMD (Fig. 2e). Side chain of R674 of PHLD interacts with the main chain carbonyl group of E1348 and I1349 on α1 of the DH domain (Fig. 2f). The EF1–EF2 module in the high-calcium state shapes a crevice that embraces α4 and post α4 loop of the DH domain (Fig. 2g and Fig. S10g). The interactions between the EF module and DH are mainly hydrophobic and involve F768, F772, F807, F819, F840, and F847 of the EF module, L1463, M1467, I1470, F1475, V1478, and F1484 of the DH domain (Fig. 2g). In addition, K814 of the EF module makes electrostatic interaction with E1281 on β2 of DH (Fig. 2g). The interactions between the EF module and the DH domain of hDUOX1 mimic the interactions between calcineurin subunit B and A in the calcium-bound state (PDB ID: 4IL1)18 (Fig. S10g, h).The linker between the EF module and preM1 helices binds in a groove on the surface of the DH domain (Fig. 1i). DH docks onto the bottom of TMD via polar interaction between R1270 on M6 and D1367 on β7, and between R1113 on loop B of TMD and N1550 of DH (Fig. 2h, i). It is reported that R1111Q mutation in hDUOX2 (R1113 in hDUOX1) was identified in congenital hypothyroidism patients20, highlighting the importance of this inter-domain interaction. Moreover, both the FAD cofactor and NADPH substrate bind at the interface between DH and TMD. R1214 and R1131 in TMD form electrostatic interaction with phosphate of FAD. D1128 makes hydrogen bonding with ribose of FAD (Fig. 2h). E1039 and N1040 in TMD make hydrogen bonding with adenosine ring of NADPH, and R1036 make cation-π interaction with both adenosine ring and electrostatic interaction with phosphate group of NADPH (Fig. 2i). Notably, R1495, R1424, and R1036 all participate in electrostatic interactions with the phosphate group of NADPH ribose, providing structural mechanism to distinguish NADPH from NADH (Fig. 2i). Through structural comparison, we found the NADPH-binding site in the csNOX5 structure was blocked by the artificially engineered C-terminal insertion, which was introduced into previous crystallization construct9 (Fig. S10i). Moreover, the adenosine group of FAD has a 180° flip compared with structure of the isolated DH of csNOX5 (Fig. S10j). This is probably because D1128 on TMD stabilizes the ribose of FAD in such a conformation to make the connecting adenosine group of FAD in close proximity with inner heme for electron transfer (Fig. 2h). Taken together, the binding of FAD and NADPH at the interface between DH and TMD of hDUOX1 would stabilize the docking of DH onto the bottom of TMD.The putative electron transfer pathwayThe measured edge-to-edge distances between NADPH and FAD, between FAD and inner heme, and between inner heme and outer heme are 8.2, 3.9, and 6.7 Å, respectively (Fig. 3a). Although, it is possible that there are additional protein residues on DUOX1 that rely electrons from NADPH to FAD, such as W378 between two hemes in csNOX5 (ref. 9), the distance between NADPH and FAD is larger than that in the canonical FNR protein, such as 3.2 Å in spinach FNR (sFNR, PDB ID: 1QFZ)21. Through structural comparison, we found the DH domain of DUOX1 shows a relaxed conformation, in which two subdomains are loosely packed, while both the DH of csNOX5 and sFNR show a tense conformation and their two subdomains are tightly packed against each other to bring FAD and NADPH into close proximity for electron transfer (Fig. 3b–e). Therefore, the electron transfer efficiency in the current structure of DUOX1 is not optimal. Because the DUOX1 complex on cell membrane exhibited higher activity (Fig. 1b and Fig. S1f), it is possible that lipids on cell membrane or the bilayer environment could somehow affect the structure of DUOX1 to enhance its electron transfer efficiency.Fig. 3Electron transfer pathway in hDUOX1 subunit in the high-calcium state.a The edge-to-edge distances between NADPH and FAD, FAD and inner heme, and two hemes are shown beside dashes. The ligands are shown as sticks, each domain of hDUOX1 are shown in surface, and colored the same as Fig. 1h. Only one hDUOX subunit is shown for clarity. The putative oxygen-reducing center is boxed by dashed lines. b–e The DH domain of hDUOX1 in a relaxed conformation (b), DH domain of csNOX5 (c), and sFNR (d) in a tense conformation. The ligands are shown as sticks, two subdomains (FAD-binding domain, FBD, and NADPH-binding domain, NBD) of DH are shown as cartoon with surface. Distances between Cα atoms of the Arg (Lys in sFNR) of FBD and the Cys of NBD (shown as spheres) are labeled. e Structural comparison of the hDUOX1 DH domain (cyan) and csNOX5 (purple). FBD is used for structural alignment. f–i The close-up view of the putative oxygen-reducing center. Four predicted tunnels for oxygen substrate entrance and product exit are shown as surface in yellow, green, magenta, and orange, respectively. Residues surrounding the tunnels are shown as sticks. j Calculated radii of tunnels shown in f–i. The putative oxygen-reducing center is used as the starting point for calculation.At the terminus of electron transfer chain near the extracellular side, the initial product of oxygen-reducing reaction is superoxide anion. We probed the possible pathways for oxygen entrance and for superoxide anion exit with CAVER22, using the oxygen-reducing center as the starting point. We located four possible tunnels: tunnel A is formed by M1, M2, M5, and M6 and is capped by loop E on top (Fig. 3f); tunnel B is surrounded by M2, loop A, loop C, and loop E (Fig. 3g); tunnel C is embraced by M3, M4, and loop C (Fig. 3h); and tunnel D is enclosed by M3, M4, loop C, and M0 (Fig. 3i). The bottleneck radii of these tunnels are ~1 Å (Fig. 3j), which may allow the permeation of small oxygen substrate under dynamic motion of DUOX1 protein. Further analysis showed tunnels B–D are all surrounded by hydrophobic residues (Fig. 3g–i), which are unfavorable for superoxide anion permeation. In contrast, tunnel A is gated by hydrophilic R1087 on M2, R1062 on M1, and R1248 and Q1245 on loop E (Fig. 3f). We speculate the highly positively charged constriction of tunnel A would strongly attract the negatively charged superoxide anions, and this might be essential for the dismutation reaction between two superoxide anions to generate uncharged hydrogen peroxide for diffusion. Therefore, manipulations that may alter the constrictions of tunnels A–D would affect superoxide anion intermediate leakage. Indeed, it is reported that mutations on DUOX1 loop A or on DUOXA1 N-terminus peptide (NTP) which interacts with and stabilizes loop A would change the ratio of superoxide anion and hydrogen peroxide produced, probably by affecting the leakage of superoxide anions through these tunnels23–25.Structure of hDUOXA1 and mechanism of complex assemblyDUOXA protein is an essential auxiliary subunit for DUOX enzyme5 (Fig. S11a) and it has an extracellular N-terminus that is important for hydrogen peroxide generation24,25. We observed the NTP of hDUOXA1 extends and packs onto the PHD–TMD junction of the distal hDUOX1 subunit (Fig. 4a–e). Side chains of F8, F10, and Y11 of NTP insert into the hydrophobic groove formed by loop C, loop A, and PHD of hDUOX1 (Fig. 4c). In addition, K15 of DUOXA1 NTP makes electrostatic interactions with D1077 of DUOX1 (Fig. 4d). This agrees with previous data showing DUOXA1 NTP interacts with DUOX1 loop A (ref. 23). hDUOXA1 has five transmembrane helices. Lower part of TM1 interacts with preM1 and M1 of hDUOX1 (Fig. S11a). The remaining four helices and associated extracellular loops share structural similarity with claudin superfamily members, such as claudin-9 (PDB ID: 6OV2)26 (Fig. S11). The extracellular loop between TM2 and TM3 folds into a compact claudin-like domain (CLD) composed of four β strands and two α helices (Figs. S8 and S11). CLD forms extensive interactions with both distal and proximal DUOX1 subunits (Fig. 4a, b), emphasizing its important role in the complex assembly. This agrees with previous studies showing that splicing variants at TM2–TM3 loop have distinct behavior in supporting the activity of DUOX1 (ref. 24). Moreover, we found an ordered N-linked glycosylation decoration on N109 of hDUOXA1 and its branched sugar moieties make extensive polar interactions with both DUOXA1 and DUOX1 subunits (Fig. 4a, b). The PHD of two DUOX1 subunits also interact with each other (Fig. 4a, b). Close to the dyad axis, R50 and R507 on one PHD make polar interactions with E41 and F313 on the opposite PHD (Fig. 4e). We further analyzed the effects of interface mutations on the tetramer assembly, and found mutations of R50E, R507E, and R507A all severely affect tetrameric peak formation on FSEC (Fig. 4f). These structural information and biochemical data revealed the detailed inter-subunit interactions that dictate the heterotetramer assembly.Fig. 4Mechanism of hDUOX1–hDUOXA1 tetramer assembly.a The side view of hDUOX1–hDUOXA1 protein complex shown in surface representation and colored the same as in Fig. 1d. b The open-book view of the inter-subunit interfaces. Residues of hDUOX1 subunits that interact with hDUOXA1 subunit are colored in green. Residues of hDUOXA1 subunit that interact with hDUOX1 subunits are colored in yellow and blue. c The close-up view of the interactions between NTP of hDUOXA1 and hDUOX1 boxed in a. d The close-up view of additional interactions between NTP of hDUOXA1 and hDUOX1 boxed in c. e The top view of interactions between PHD of two opposing hDUOX1 subunits. f Representative FSEC traces of hDOUX1 R50E, R507E, and R507A mutants are compared to that of wild-type (WT) hDOUX1. The peak position of the hDOUX1 peak is denoted by the hollow circles. Asterisks denote the peak position of hDUOX1–hDUOXA1 protein complex.Conformational change of DUOX1 complex upon calcium activationThe consensus map in the low-calcium state showed the cytosolic layer had poor local resolution, which was improved by multibody refinement14 (Fig. S4). Further molecular flexibility analysis14 showed the cytosolic domains (small body) in the low-calcium state were sampling a broad range of orientations relative to the TMD, evidenced by the plateau-shaped distribution on the histogram of the major eigenvector (Fig. S4f). This is in great contrast to the normal distribution in the high-calcium state (Fig. S2f), suggesting the cytosolic layer in the low-calcium state is more flexible. We compared the structures in the low-calcium state and high-calcium state, and found structural changes in the extracellular layer and transmembrane layer are small (Fig. 5a). However, there are large conformational changes of the regulatory PHLD and EF-hand module in the cytosolic layer (Fig. 5a–c and Movie S1). In the absence of calcium, the EF module switches from an extended shape into a more contracted shape (Fig. 5d, e), which reconfigures the interface between the EF module and α4 of the DH domain, resulting in a loosely packed structure (Fig. 5f). In the low-calcium state, EF2 moves away from the DH domain. The Cα atom of A894 on αJ of EF2 has 40 Å displacement (Fig. 5b). PHLD rotates away from the TMD and DH domains, and αA of PHLD has 17.2° outward rotation (Fig. 5c). As a result, several inter-domain interactions observed in the high-calcium state were disrupted and therefore the docking of DH domain onto TMD is weakened by these structural changes, leading to a higher mobility of DH domain (Fig. S4g). We propose the increased mobility of DH domain negatively correlates with the electron transfer efficiency and thus the catalytic activity of DUOX. In addition, because TMD also contributes to FAD and NADPH binding, the increased mobility of the DH domain would result in the reduced affinity of NADPH as well. This is in agreement with the markedly reduced Kcat and moderately increased Km in the low-calcium state, as we observed (Fig. 1c).Fig. 5Conformational change of hDUOX1 complex during calcium activation.a Structural comparison of hDUOX–hDUOXA1 complex between the high-calcium state (colored) and the low-calcium states (gray). Protein is shown as cartoon. Regions with large conformational changes are boxed by dashed lines. b Close-up view of the conformational changes of EF-hand module. Cα atom of A894 on αJ helix is used as marker to measure the movement of EF2. c Close-up view of the conformational change of PHLD. The angle between αA helices in the high-calcium and low-calcium states was measured. d, e Conformational differences of EF-hand module between the high-calcium state and the low-calcium state. f Reconfiguration of the interface between EF-hand module and α4 helix of DH domain. Arrows denote movements from high-calcium state into the low-calcium state.During the preparation of this manuscript, another group reported the structures of mouse DUOX1–DUOXA1 complex27. Interestingly, they found mouse DUOX1 complex exists in both heterodimeric and heterotetrameric form, and they proposed the activity of DUOX1 complex is regulated by dimer–tetramer assembly27. This is in contrast to our observation that majority of hDUOX1 complex is in a homogenous tetrameric form (Fig. S1c). Whether this difference is due to different protein preparation procedure or different species (mouse vs human) remain elusive. Moreover, the intracellular PHLD and EF domains were not resolved in mouse DUOX1 complex structure because of insufficient map quality27. The overall structures of resolved parts between mouse and hDUOX1 complex are similar with root mean square deviation of 1.521 and 0.908 Å for DUOX1 and DUOXA1 subunit, respectively (Fig. S11c). However, detailed structural comparison revealed several differences especially in the atomic models of FAD (Fig. S11d) and NADPH (Fig. S11e), probably due to the poor local map quality of mouse DUOX1 complex (EMD-21964).DiscussionIn this study, we provided the structures of hDUOX1–hDUOXA1 as a peptidisc-stabilized heterotetrameric protein complex in both high-calcium and low-calcium states. The structure of hDUOX1 complex in the high-calcium state reveals multiple inter-domain interactions that orientate DH and TMD for electron transfer, and thus redox reaction. Removal of calcium ions results in the reconfiguration of cytosolic inter-domain interactions which in turn mobilizes the DH domain and lowers the electron transfer efficiency (Fig. 6). These structures provide mechanistic insights into the structure and mechanism of DUOX and other NOX enzymes.Fig. 6Activation mechanism of DUOX1 complex by calcium.Two DUOX1 and one DUOXA1 subunit are shown as cartoon, and colored the same as Fig. 1d. Calcium ions are presented as green spheres. Electron transfer pathways are indicted with gray arrows.MethodsCell cultureHEK293F suspension cells (Thermo Fisher Scientific) were cultured in Freestyle 293 medium (Thermo Fisher Scientific) supplemented with 1% FBS at 37 °C with 6% CO2 and 70% humidity. Sf9 insect cells (Thermo Fisher Scientific) were cultured in SIM SF (Sino Biological) at 27 °C. The cell lines were routinely checked to be negative for mycoplasma contamination.Protein expression and purificationWe constructed a modified BacMam vector28,29 with N-terminal GFP tag guided by rat FSHβ signal peptide10 and cloned hDUOX1 cDNA into this vector (please see Supplementary Table 2 for primer list). The hDUOX1 cDNA, we obtained from Prof. Han, has the same protein sequence as NP_059130.2 except the L1178F mutation, a SNP previously observed in AAI14939.1. The cDNAs of hDUOXA1 were cloned into a non-tagged BacMam vector or a modified BacMam vector with C-terminal MBP-mScarlet tag (please see Supplementary Table 2 for primer list)28,29. The hDUOX1 mutants were generated by Quick Change methods (please see Supplementary Table 2 for primer list). The two expression cassette were further merged into one bicistronic vector by the LINK sequence on the modified vector28,30. The baculoviruses were produced using the Bac-to-Bac system and amplified in Sf9 cells. For protein expression, HEK293F cells cultured in Freestyle 293 medium at density of 2.8 × 106 ml−1 were infected with 15% volume of P2 virus. A total of 10 mM sodium butyrate was added to the culture 12 h post infection and transferred to a 30 °C incubator for another 36 h before harvesting. Cells were collected by centrifugation at 3999 × g (JLA 8.1000, Beckman) for 10 min, and washed with 20 mM Tris (pH 8.0 at 4 °C), 150 mM NaCl, 2 mM EGTA, 2 μg ml−1 aprotinin, 2 μg ml−1 pepstatin, 2 μg ml−1 leupeptin, flash-frozen, and storage at −80 °C.For each batch of protein purification, cell pellet corresponding to 0.5 liter culture was thawed and extracted with 20 ml buffer A (20 mM Tris pH 8.0 at 4 °C, 150 mM NaCl, 5 μg ml−1 aprotinin, 5 μg ml−1 pepstatin, 5 μg ml−1 leupeptin, 20% (v/v) glycerol, and 2 mM EGTA) containing 1 mM phenylmethanesulfonyl fluoride and 1% (w/v) digitonin (Biosynth) at 4 °C for 50 min. A total of 1 mg ml−1 iodoacetamide (Sigma—I1149) was added during the detergent extraction procedure to reduce nonspecific cysteine crosslinking. The supernatant was ultracentrifuged at 135,300 × g (TLA100.3, Beckman) for 50 min. The solubilized proteins were loaded onto 5 ml Streptactin Beads 4FF (Smart-Lifesciences) column and washed with 20 ml buffer A + 0.1% digitonin. The column was washed with 100 ml buffer A + 0.1% digitonin plus 10 mM MgCl2 and 1 mM adenosine triphosphate (ATP) to remove contamination of heat shock proteins. Then the column was washed with 40 ml buffer A + 0.1% digitonin again to remove residual MgCl2 and ATP. The target protein was assembled into the peptidisc on the Streptactin Beads through washing with 4 ml 1 mg ml−1 NSPr in 20 mM Tris pH 8.0 (ref. 12). Then the column was washed with 100 ml buffer A to remove free NSPr. The assembled peptidiscs were eluted with 40 ml buffer A + 5 mM D-desthiobiotin (IBA). Eluted protein was concentrated using 100-kDa cutoff concentrator (Millipore) and further purified by Superose 6 increase (GE Healthcare) running in HBS (20 mM Hepes pH 7.5, 150 mM NaCl) + 0.5 mM EGTA. Fraction 19 corresponding to DUOX1 + DUOXA1 peptidisc complex was concentrated to A280/415 = 4.4/1.4 with estimated concentration of 10.7 μM DUOX1 subunits (ε415 = 0.131 μM−1 cm−1).Enzymatic assayThe membrane fractions of DUOX1 for enzymatic assay were prepared as previously reported with minor modification31. Briefly, cells were washed with buffer A. After centrifuging at 3999 × g for 10 min at 4 °C, the cell pellets were broken using a needle for 12 times in 1 ml of 20 mM Tris pH 8.0 at 4 °C containing 0.1 mM dithiothreitol, 10 mM EGTA (pH 8.0), and the mixture of protease inhibitors. The pellet was removed by centrifuging at 3615 × g for 15 min, and the supernatant was collected and then centrifuged at 228,600 × g (TLA100.3, Beckman) for 1 h. The membrane pellet was resuspended in HBS containing 1 mM EGTA. Meanwhile, 10 μl membrane was solubilized by 100 μl TBS + 1% digitonin with the mixture of protease inhibitors for 1 h at 4 °C for FSEC. The protein concentrations in the membrane were estimated by comparing their GFP fluorescence signal to that of a purified GFP-tagged DUOX1 complex.The H2O2-generating activity of DUOX1 complex was determined using the amplex red assay32. The concentrations of H2O2 solution were determined by measuring UV–vis absorbance at 240 nm with spectrophotometer (Pultton) and calculated using molar extinction coefficient of 43.6 M−1 cm−1. The concentration of H2O2 solution was further validated by reacting with Amplex red to generate resorufin which has ε571 = 69,000 M−1 cm−1 (ref. 32). Then the H2O2 solution with known concentration was used to calibrate the resorufin fluorescence curve (excitation, 530 nm; emission, 590 nm) measured using a Microplate Reader (BioTek Synergy HT) at 37 °C.The H2O2-generating reaction of the membrane fraction containing DUOX1 complex was performed at 37 °C in 0.15 ml of HBS with 1 mM EGTA, 10 μM FAD, 100 μM NADPH, 50 μM amplex red, 0.067 mg ml−1 horseradish peroxidase, and 0.0576 mg ml−1 SOD. Ca2+ concentrations were determined using fluorescent indicators fura-2 or fluo3-FF. The Kcat and Km values of the membrane fraction containing DUOX1 complex were determined at 37 °C with different concentrations of NADPH in the presence or absence of 1.4 mM CaCl2. The H2O2-generating reaction of the purified DUOX1 complex in peptidisc was performed at 27 °C in 0.15 ml of HBS + 1 mM EGTA, 10 μM FAD, 100 μM NADPH, 50 μM amplex red, 0.067 mg ml−1 horseradish peroxidase, and 0.0576 mg ml−1 SOD in the presence or absence of 1.1 mM CaCl2. Progress of the reactions was monitored continuously by following the increase of the resorufin fluorescence, and the initial reaction rates were obtained by fitting the curve with linear equation. The activity of DUOX1 complex was determined by subtracting the background of the corresponding buffer without enzyme. The data were processed with Microsoft Excel-2013, SigmaPlot-12.0, and GraphPad Prism 6.Cryo-EM sample preparation and data acquisitionThe peptidisc sample was supplemented with 2.5 mM EGTA (low calcium) or 0.5 mM free calcium (high calcium) for cryo-EM analysis, respectively. Both samples contain 100 μM FAD as the cofactor and 500 μM NADPH as the substrate. To overcome the preferred orientation problem, 0.5 mM non-solubilizing detergent fluorinated octyl-maltoside was added to the sample before cyro-EM sample preparation. Aliquots of 1.5 μL protein sample were placed on graphene oxide-coated grids, as previously reported33. Grids were blotted for 3 s at 100% humidity and flash-frozen in liquid ethane cooled by liquid nitrogen using Vitrobot Mark I (FEI). Grids were then transferred to a Titan Krios (FEI) electron microscope that was equipped with a Gatan GIF Quantum energy filter and operated at 300 kV accelerating voltage. Image stacks were recorded on a Gatan K2 Summit direct detector in super-resolution counting mode using Serial EM at a nominal magnification of 130,000× (calibrated pixel size of 1.045 Å pixel−1), with a defocus ranging from −1.5 to −2.0 μm. Each stack of 32 frames was exposed for 7.12 s, with a total dose ~50 e− Å−2 and a dose rate of 8 e− pixel−1 s−1 on detector.Image processingThe image processing workflow is illustrated in Figs. S2 and S4. A total of 7076 super-resolution movie stacks of the high-calcium state sample and 2076 stacks of the low-calcium state sample were collected using Serial EM, and motion-corrected, dose weighted, and twofold binned to a pixel size of 1.045 Å using MotionCor2 (ref. 34). Contrast transfer function (CTF) parameters were estimated with Gctf35. Micrographs with ice or ethane contamination, and empty carbon were removed manually. Autopicking were performed using Gautomatch (kindly provided by Kai Zhang). All subsequent classification and reconstruction was performed in Relion 3.1 (ref. 36) unless otherwise stated. Reference-free 2D classification was performed to remove contaminants. Initial model was generated using cryoSPARC37. Particles were subjected to multi-reference 3D classification38,39 and random-phase 3D classification38,39. Phase-randomized models were generated from the model obtained from previous refinement using randomize software (from the lab of Nikolaus Grigorieff). Further CTF refinement was then performed with Relion 3.1 using C2 symmetry. The particles were then re-extracted, re-centered, and re-boxed from 256 pixels to 320 pixels for consensus refinement in Relion 3.1 (ref. 36) and cryoSPARC37. To improve the density of cytosolic layer, particles were symmetry expended13 for multibody refinement14. One soft mask (the large body) that covers the extracellular domain together with TMD of one protomer was generated from the consensus map, using UCSF Chimera and Relion 3.1 (ref. 36). The other soft mask (the small body) covers the cytosolic domains of the same protomer. 3D multibody refinements14 were performed using the two soft masks and the parameters determined from previous consensus refinement. The motions of the bodies were analyzed by relion_flex_analyse in Relion 3.1 (ref. 36). The two half-maps of each body generated by 3D multibody refinement were subjected to post-processing in Relion 3.1 (ref. 36). The masked and sharpened maps of each body were aligned to the consensus map using UCSF Chimera40, zoned to isolated nonoverlapping regions and summed using Relion 3.1 (ref. 36) to generate the composite maps for visualization and model building. All of the resolution estimations were based on a Fourier shell correlation of 0.143 cutoff after correction of the masking effect. B-factor used for map sharpening was automatically determined by the post-processing procedure in Relion 3.1 (ref. 36). The local resolution was estimated with Relion 3.1 (ref. 36).Model buildingThe composite maps derived from multibody refinement were used for model building. The structures of PHD, TMD, EF1–2, and DH domains of hDUOX1 were generated using phyre2 server41 based on PDB ID: 6ERC, 5O0T, 4IL1, and 5O0X, and manually docked into the cryo-EM maps using Chimera40. Initial models of PHLD were generated by Rosetta Web Server using ab initio mode42, manually selected according to the distances calculated by RaptorX Contact Prediction server43, and validated by the fitting between model and cryo-EM densities, especially the location of bulky aromatic residues. The partial model of hDUOXA1 were generated using EM builder44. The initial models were iteratively built using Coot45 and refined using Phenix in real space46. Figures were prepared using UCSF chimera40, Chimera X47, and Pymol.Reporting summaryFurther information on research design is available in the Nature Research Reporting Summary linked to this article.Supplementary informationSupplementary InformationDescription of Additional Supplementary FilesSupplementary Movie 1Reporting Summary
|
nature communications
|
[
"Article"
] |
[
"Enzyme mechanisms",
"Oxidoreductases",
"Cryoelectron microscopy"
] |
oxygen species (ROS) are oxygen reactive hydrogen peroxide superoxide participate in processes pathological ROS generated by NADPH oxidase (NOX) multi-pass transmembrane proteins catalyze reduction oxygen superoxide anion hydrogen peroxide NOX proteins in biological processes host defense differentiation development cell growth survival cytoskeletal reorganization modification extracellular human NOX protein family NOX1–5 DUOX1–2 NOX2 protein catalyzes superoxide anion essential for host DUOX1–2 proteins expressed in thyroid tissue catalyze production hydrogen peroxide important for thyroid DUOX protein interactions with dual oxidase maturation factor (DUOXA promotes maturation plasma membrane localization DUOX encoded by genes DUOX1 DUOX2. encoded by DUOXA1 DUOXA2. mutations of DUOX2 DUOXA2 cause congenital hypothyroidism6 named thyroid oxidase4 family proteins share catalytic core heme-coordinating transmembrane domain cytosolic dehydrogenaseDH domain binds NADPH FAD shares sequence homology ferredoxin-NADP + reductase two TMD-DH catalytic core functional DUOX protein large N-terminal extracellular peroxidase homology domain) long intracellular loop 0 two EF-hand domains requires auxiliary DUOXA protein for function activity DUOX regulated by intracellular calcium structures NOX family members isolated domains DH domain 5O0X TMD 5O0T)9 NOX5 from Cylindrospermum stagnale subdomain human NOX2 DH domain ID 3A1F). functional importance DUOX NOX family members structures full-length functional protein complex unknown open questions DUOX DH domain with TMD DUOXA protein with DUOX activity regulated by intracellular calcium? DUOX–DUOXA protein complex structurally functionally cryo-EM structures human DUOX1–DUOXA1 complex in high-calcium low-calcium states structure calcium activation DUOX constructed matured protein (20–1551) N-terminal GFP tag guided rat FSHβ signal peptide for molecular weight GFP-tagged hDUOX1 219 kDa.DUOX1–DUOXA1 complex fused hDUOXA1 protein C-terminal MBP-mScarlet tag molecular weight 106 kDa chromatography co-expression hDUOXA1 shifted peak hDUOX1 higher molecular weight stable hDUOX1–hDUOXA1 hetero-complex (Fig. S1a peak positions hDUOX1 migrated monomer heterotetramer co-expression DUOX1 DUOXA1 cell membranes robust calcium-activated-dependent hydrogen peroxide production (Fig. low-calcium DUOX1–DUOXA1 complex low basal activity Addition calcium reduced Km increased Kcat DUOX1 enzymatic activity 1Structure DUOX1–DUOXA1 complex high-calcium state Schematic DUOX enzymatic assay H2O2 horseradish peroxidase converts nonfluorescent Amplex Red to fluorescent resorufin H2O2. Calcium-dependent activation hDUOX1–hDUOXA1 complex Data means ± standard deviations n = 3 samples Source data Steady state enzyme activity NADPH concentration presence absence calcium Data Michaelis–Menten equation Km Kcat valueData shown as means ± standard deviations = 3 independent samples Source data file Side view cryo-EM map hDUOX1–hDUOXA1 complex high-calcium state boundaries phospholipid bilayer gray lines One protomer hDUOX1 colored blue green other yellow red 90° rotated top view 180° rotated bottom view cross-section transmembrane layer dashed line large cavity dashed cryo-EM map low-pass filtered to 6 Å Topology of hDUOX1 hDUOXA1 subunits Transmembrane helices cylinders disordered regions dashed lines phospholipid bilayer gray layers PHD peroxidase PHLD pleckstrin EF-binding DH dehydrogenase CLD Structure of protomer hDUOX1 cartoon colors same boundaries phospholipid bilayer gray lines Sugar moieties hemes FAD NADPH as black yellow pink green sticks solubilized reconstituted hDUOX1–hDUOXA1 complex into peptidisc using NSPr12 Purified peptidisc sample homogenous peak on SEC (Figtwo protein bands SDS–PAGE trimmed PNGase F treatment (Fig. hDUOX1 hDUOXA1 modified N-linked glycosylation UV–vis spectrum peptidisc sample Soret band peak 415 nm (Fig. Fe (III) heme incorporation purified peptidisc sample recapitulated calcium-activated-dependent hydrogen peroxide production. calcium-dependent activation hDUOX1–hDUOXA1 protein complex maximum activity purified peptidisc lower crude cell membrane. 1b membrane bilayer endogenous lipids DUOX activity prepared cryo-EM grids peptidisc sample 2.5 mM ethylene glycol tetraacetic acid 0.5 mM free calcium samples 0.1 mM FAD cofactor 0.5 mM NADPH substrate particle cryo-EM analysis purified protein homogeneous twofold symmetry (Figs. S2–S4) resolution cryo-EM maps low-calcium high-calcium states 2.7 2.6 Å S1) extracellular domains TMD better resolution cytosolic domains higher mobility (Figs. S2g quality exploited symmetry multibody protomer large smallS2c S4c). final resolutions cytosolic domain reached 3.4 3.2 Å low high-calcium states (Figs. S2–4 Table S1) high map quality homology structures order regions complex 88% DUOX1 79% DUOXA1 (Figs. S5–8 Table S1) focus high-calcium state structure higher resolution hDUOX1–hDUOXA1 protein co-assemble 2:2 heterotetrameric protein complex molecular weight ~457 kDa complex encompasses 140 Å × 105 Å 160 Å 3D space twofold rotational symmetry (Fig. 1d–f). divided three layers extracellular transmembrane layer cytosolic layer extracellular layer two N-terminal PHD domains hDUOX1 pack buttressed DUOXA1 transmembrane layer 24 transmembrane helices heme-binding sites electron transfer pathway (Fig. center large cavity protein densities interior surface hydrophobic (Fig. S3i lipid molecules bound filled phospholipids cytosolic layer catalytic DH domain regulatory domains intracellular calcium sensingcatalytic hDUOX1 catalytic subunit complex (Fig. 2) extracellular side large N-terminal PHD domain shares sequence homology peroxidases peroxidase A Dictyostelium discoideum (Fig. S9a). Functional peroxidases utilize histidine-coordinated heme catalysis key residues heme binding ligand histidine missing PHD hDUOX1 heme density hDUOX1 PHD not enzymatic functional peroxidase activity reveals two cation densities PHD One cation CBS1) coordinated side chains D397 T332 chain carbonyl groups V399 T332 R395 (Fig. second cation (CBS2) coordinated side chain D109 D174 S176 T170 carbonyl groups T170 W172 observed strong densities sites low-calcium high-calcium conditions cations sodium ions calcium ions CBS1 CBS2 evolutionary conserved in DUOX (Fig. S5) DdPoxA15 functional importance CBS1 mutant CBS2 mutant (D109A D174A) DUOX1 failed co-assemble with DUOXA1CBS1 CBS2 subunit interfaces DUOX1–DUOXA1 complex mutants affect folding PHD domain role protein stability.Fig. 2Structure hDUOX1 subunit Side view high-calcium state key interfaces domain colored Fig. 1h surface transparency site outer heme TMD sticks colored yellow Unrelated helices omitted oxygen-reducing center arrow site inner heme interface between PHD TMD Disulfide bond between C118–C1165 golden sticks interface PHLD TMD hydrogen bonds dashed lines interface PHLD DH domain EF module DH domain FAD-binding site interface TMD DH domain Ligands interacting residues sticks NADPH-binding site interface TMD DH domain PHD packs top TMD DUOX1 noncovalent interactions (Fig. disulfide bond between C118 PHD C1165 loop C staples bottom PHD top TMD (Fig. hDUOX1 extra bent M0 helix periphery M3 M4 six TM helices NOX protein family M1–M6 form two heme-binding sites TMD H1144 on M3 H1238 on M5 coordinate outer hemeH1130 M3 H1225 M5 coordinate inner heme. histidines conserved NOX proteins. S6) observed spherical density R1087 M2 H1148 M3 heme residue H1144 (Fig. 2b mutations csNOX5 residues R1087 H1148 reoxidation dithionite-reduced TMD oxygen substrate site oxygen-reducing center. S10a structure observations hDUOX1 support hypothesis amphipathic preM1 helix inner leaflet plasma membrane. observed csNOX5 shared feature NOX proteins Between M0 preM1 cytosolic fragment loop 0 Cryo-EM maps N-terminal loop domain rich β sheets (Figs. S3g Structural search DALI identified β sheets-rich domain pleckstrin homology-like domain high structural similarity PH domain proteins. S10c EF-hand calcium-binding domains (EF1 EF2) compact helical module connected PHLD αC (Fig. S6) Residues calcium binding EF1 EF2 conserved DUOX proteins S6) strong densities small calcium ions poor local resolutionstructure EF-hand module resembles calcium-dependent protein phosphatase calcineurin state ID 4IL1)18 EF1 EF2 loaded calcium high-calcium state homology structure side chains D828 D830 N832 E839 carbonyl group Y834 chelate one calcium ion EF1 chains D864 D866 N868 E875 carbonyl L870 chelate another EF2 mutations calcium-binding sites abolished calcium E879K mutation hDUOX2 (E875 hDUOX1) leads congenital hypothyroidism20 activation C-terminal catalytic DH domain connected to M6 DUOX1 TMD short linker (Fig. 1h DH domain hDUOX1 canonical DH fold structure similar csNOX5 strong densities FAD cofactor NADPH substrate sites contributed from DH domain TMD 2h-domain interactions high-calcium domains DUOX1 stabilized inter-domain interactions PHLD interacts with TMD DH domains main chain carbonyl group K653 PHLD bond with R1215 loop D TMDR674 PHLD interacts E1348 I1349 α1 DH domain (Fig. EF1–EF2 module high-calcium α4 α4 loop DH domain (Fig. 2g interactions EF module DH hydrophobic involve F768 F772 F807 F819 F840 F847 L1463 M1467 I1470 F1475 V1478 F1484 DH K814 EF E1281 β2 DH interactions EF module DH domain hDUOX1 mimic calcineurin subunit B A calcium-bound state (Fig. S10g linker EF module preM1 helices binds groove surface DH domain (Fig. DH docks bottom TMD interaction R1270 M6 D1367 β7 R1113 B TMD N1550 DH (Fig. 2h R1111Q mutation hDUOX2 (R1113 hDUOX1) identified congenital hypothyroidism inter-domain interaction FAD cofactor NADPH substrate bind interface DH TMD R1214 R1131 TMD electrostatic interaction FAD D1128 hydrogen ribose FADE1039 N1040 TMD hydrogen bonding adenosine ring NADPH R1036 cation-π interaction adenosine electrostatic interaction phosphate group NADPH (Fig. R1495 R1424 R1036 participate electrostatic interactions phosphate group NADPH ribose distinguish NADPH from NADH NADPH-binding site csNOX5 blocked by engineered C-terminal previous crystallization (Fig. adenosine group FAD 180° flip DH csNOX5 D1128 TMD stabilizes ribose FAD adenosine group inner heme electron transfer binding FAD NADPH interface DH TMD hDUOX1 stabilize docking DH bottom TMD putative electron transfer edge-to-edge distances between NADPH FAD FAD inner heme outer heme 8.2 3.9 6.7 Å (Fig. 3a). additional protein residues on DUOX1 rely electrons NADPH to FAD W378 hemes csNOX5 distance between NADPH FAD larger than canonical FNR protein 3.2 Å spinach FNRcomparison DH domain DUOX1 relaxed conformation two subdomains loosely packed DH csNOX5 sFNR tense conformation tightly packed FAD NADPH electron transfer (Fig. 3b–e). electron transfer efficiency not optimal DUOX1 complex cell membrane higher activity (Fig. 1b lipids bilayer environment affect structure electron transfer efficiency. 3Electron transfer pathway hDUOX1 subunit high-calcium state edge-to-edge distances between NADPH FAD FAD inner heme two hemes shown ligands sticks domain hDUOX1 surface colored same Fig. 1h one hDUOX subunit clarity putative oxygen-reducing center boxed dashed lines DH domain hDUOX1 relaxed conformation DH csNOX5 sFNR tense conformation ligands sticks two subdomains (FAD NADPH-binding cartoon Distances between Cα atoms Arg FBD Cys NBD labeled Structural comparison hDUOX1 DH domain csNOX5 FBD for structural alignment close-up view putative oxygen-reducing center Four tunnels for oxygen substrate entrance product exit yellow green magenta orangeResidues tunnels as sticks Calculated radii tunnels f–i oxygen-reducing center starting point terminus electron transfer chain extracellular initial product oxygen-reducing reaction superoxide anion probed pathways oxygen entrance superoxide anion exit with CAVER22 center starting point located four tunnels A formed by M1 M2 M5 M6 capped by loop E B surrounded by M2 A C E C embraced by M3 M4 C D enclosed by M3 M4 C M0 bottleneck radii ~1 Å permeation small oxygen substrate DUOX1 protein tunnels B–D surrounded by hydrophobic residues unfavorable for superoxide anion permeation tunnel A gated by hydrophilic R1087 R1062 R1248 Q1245 loop E positively charged constriction tunnel A negatively charged superoxide anions dismutation reaction uncharged hydrogen peroxide constrictions tunnels A–D affect superoxide anion leakagemutations DUOX1 loop A N-terminus peptide) ratio superoxide hydrogen peroxide leakage hDUOXA1 complex essential auxiliary subunit DUOX enzyme5. S11a extracellular N-terminus important hydrogen peroxide NTP extends packs PHD–TMD junction distal hDUOX1 subunit (Fig. 4a–e). Side chains F8 F10 Y11 NTP insert into hydrophobic groove loop C A PHD hDUOX1 K15 DUOXA1 NTP interactions with D1077 DUOX1 (Fig. data NTP interacts DUOX1 loop A hDUOXA1 five transmembrane helices TM1 interacts with preM1 M1 hDUOX1. remaining four helices extracellular loops structural similarity with claudin superfamily members extracellular loop between TM2 TM3 into compact claudin-like domain (CLD) four β strands two α helices (Figs. S8 S11) CLD interactions distal proximal DUOX1 subunits. 4a role complex assemblyagrees studies splicing variants TM2–TM3 loop activity DUOX1 ordered N-linked glycosylation decoration N109 hDUOXA1 branched sugar moieties polar interactions with DUOXA1 DUOX1 subunits (Fig. 4a PHD two DUOX1 subunits interact (Fig R50 R507 PHD polar interactions with E41 F313 opposite PHD (Fig. 4e). analyzed effects interface mutations tetramer assembly mutations R50E R507E R507A affect tetrameric peak formation FSEC (Fig. 4f). structural biochemical data inter-subunit interactions heterotetramer assembly.Fig. 4Mechanism hDUOX1–hDUOXA1 tetramer assembly side view hDUOX1–hDUOXA1 protein complex Fig 1d open-book view inter-subunit interfaces Residues hDUOX1 subunits hDUOXA1 colored green yellow blue close-up view interactions NTP hDUOXA1 hDUOX1 interactions top view interactions between PHD two hDUOX1 subunitsFSEC traces hDOUX1 R50E R507E R507A mutants compared to wild-type hDOUX1 peak position hDOUX1 denoted hollow circles Asterisks denote peak position hDUOX1–hDUOXA1 protein complex.Conformational change DUOX1 complex upon calcium low-calcium state cytosolic layer poor local resolution improved by multibody refinement14 (Fig S4) molecular flexibility cytosolic domains broad orientations TMD plateau-shaped distribution histogram major eigenvector contrast normal distribution high-calcium state cytosolic layer more flexible compared structures structural changes extracellular transmembrane layer small (Fig. large conformational changes regulatory PHLD EF-hand module in cytosolic layer calcium EF module switches extended contracted reconfigures interface α4 DH domain loosely packed structure low-calcium state EF2 moves away from DH domain Cα atom A894 on αJ EF2 has 40 Å displacement PHLD rotates away from TMD DH domains αA 17.2° outward rotationinter-domain interactions high-calcium state disrupted docking DH domain TMD weakened higher mobility DH (Fig. S4g). increased mobility DH correlates with electron transfer efficiency catalytic activity DUOX TMD contributes to FAD NADPH binding increased mobility reduced affinity NADPH with reduced Kcat increased Km low-calcium state (Fig. 1c).Fig. 5Conformational change hDUOX1 complex during calcium activation Structural comparison hDUOX–hDUOXA1 complex high-calcium low-calcium states Protein cartoon large conformational changes dashed lines Close-up conformational changes EF-hand module Cα atom A894 on αJ helix movement EF2. conformational change PHLD angle between αA helices high low-calcium states measured Conformational differences EF-hand module high-calcium low-calcium state Reconfiguration interface between EF-hand module α4 helix DH domain Arrows movements group reported structures mouse DUOX1–DUOXA1 complex27 heterodimeric heterotetrameric form activity regulated by dimer–tetramer assembly27contrast majority hDUOX1 complex homogenous tetrameric form (Fig. difference protein preparation procedure species elusive intracellular PHLD EF domains not resolved mouse DUOX1 complex insufficient map structures parts mouse hDUOX1 complex similar mean deviation 1.521 0.908 Å DUOX1 DUOXA1 subunit (Fig. comparison differences atomic models FAD NADPH due poor map quality DUOX1 structures hDUOX1–hDUOXA1 peptidisc-stabilized heterotetrameric protein complex high-calcium low-calcium states structure high-calcium state reveals inter-domain interactions DH TMD electron transfer redox reaction Removal calcium ions cytosolic-domain interactions mobilizes DH domain lowers electron transfer efficiency (Fig. 6) structures insights structure mechanism DUOX NOX enzymes. 6Activation mechanism DUOX1 complex by calcium DUOX1 DUOXA1 subunit Calcium ions green spheres Electron transfer pathways gray arrows cultureHEK293F suspension cells cultured Freestyle 293 medium 1% FBS 37 °C 6% CO2 70% humiditySf9 insect cells Fisher Scientific cultured SIM SF 27 °C cell lines checked negative mycoplasma contamination expression constructed modified BacMam N-terminal GFP tag rat FSHβ cloned hDUOX1 cDNA Table 2 Prof. Han same protein sequence_059130.2 L1178F mutation AAI14939.1. cDNAs hDUOXA1 cloned non BacMam C-terminal MBP-mScarlet tag hDUOX1 mutants generated Quick Change methods merged bicistronic vector LINK sequence baculoviruses produced Bac-to-Bac system amplified Sf9 cells HEK293F cells cultured Freestyle 293 medium × 106 ml−1 infected 15% P2 virus 10 mM sodium butyrate added culture 12 h post infection transferred 30 °C incubator 36 h before harvesting Cells collected centrifugation 3999 × g 10 min washed 20 mM Tris (pH 8.0 4 150 mM NaCl 2 mM EGTA 2 aprotinin pepstatin leupeptin flash-frozen storage −80 °Cprotein purification cell pellet 0.5 liter culture extracted 20 ml buffer pH 8.0 4 °C 150 mM NaCl 5 μg aprotinin pepstatin leupeptin 20% glycerol 2 mM EGTA 1 mM fluoride 1% digitonin 4 °C 50 min 1 mg iodoacetamide added cysteine crosslinking supernatant ultracentrifuged 135,300 × g 50 min solubilized proteins loaded 5 ml Streptactin Beads column washed 20 ml buffer 0.1% digitonin 100 ml buffer 0.1% digitonin 10 mM MgCl2 1 mM adenosine triphosphate 40 ml buffer 0.1% digitonin MgCl2 protein peptidisc 4 ml 1 mg NSPr 20 mM Tris pH 8.0 washed 100 ml buffer NSPr peptidiscs eluted 40 ml buffer A 5 mM D-desthiobiotin Eluted protein concentrated 100-kDa concentrator purified Superose 6 HBS (20 pH 7.5 150 mM NaCl 0.5 mM EGTAFraction 19 DUOX1 A280/415 4.4/1.4 10.7 μM DUOX1 subunits (ε415 0.131 μM−1 cm−1) membrane DUOX1 cells washed buffer A centrifuging 3999 10 min cell pellets broken 12 times 1 ml 20 mM Tris pH 8.0 0.1 mM dithiothreitol 10 mM EGTA protease inhibitors pellet removed 3615 g 15 min supernatant collected centrifuged 228,600 × g 1 h membrane pellet resuspended HBS 1 mM EGTA 10 μl membrane solubilized 100 μl TBS 1% digitonin protease inhibitors 1 h 4 °C protein concentrations estimated GFP fluorescence purified GFP-tagged DUOX1 complex H2O2-generating activity DUOX1 determined amplex red concentrations H2O2 UV–vis absorbance 240 nm molar extinction coefficient 43.6 M−1 cm−1 validated Amplex red resorufin ε571 69,000 M−1 cm−1H2O2 solution resorufin fluorescence curve 530 emission 590 Microplate Reader 37 °C H2O2-generating reaction DUOX1 complex 37 °C 0.15 ml HBS 1 mM EGTA 10 μM FAD 100 μM NADPH 50 μM amplex red 0.067 horseradish peroxidase 0.0576 mg SOD Ca2+ concentrations determined Kcat Km values determined 37 °C concentrations NADPH 1.4 mM CaCl2. H2O2-generating reaction purified DUOX1 27 °C 0.15 ml HBS 1 mM EGTA 10 μM FAD 100 μM NADPH 50 μM amplex red 0.067 horseradish peroxidase 0.0576 mg SOD 1.1 mM CaCl2. monitored resorufin fluorescence initial reaction rates curve activity DUOX1 complex subtracting background buffer enzyme data processed Excel-2013 SigmaPlot-12.0 GraphPad Prism 6.Cryo-EM sample supplemented 2.5 mM EGTA 0.5 mM free calcium samples 100 μM FAD 500 μM NADPH substrateorientation 0.5 mM fluorinated octyl-maltoside added before cyro-EM 1.5 μL protein sample on oxide-coated grids blotted 3 s 100% humidity flash-frozen ethane cooled nitrogen Vitrobot transferred to Titan Krios electron microscope Gatan GIF Quantum energy filter 300 kV voltage Image stacks recorded Gatan K2 Summit detector super-resolution Serial EM magnification 130,000× pixel size 1.045 Å defocus −1.5 to −2.0 μm stack 32 frames exposed 7.12 s total dose ~50 e− Å−2 rate 8 e− pixel−1 s−1 Figs. S2 S4 7076 super-resolution stacks 2076 low-calcium collected Serial EM motion-corrected dose weighted twofold binned pixel size 1.045 Å MotionCor2 Contrast transfer parameters estimated Gctf35 Micrographs ice ethane contamination empty carbon removed Autopicking Gautomatch classification reconstruction Relion 3.1 36) Reference-free 2D classification contaminants Initial model generated cryoSPARC37 Particles subjected multi-reference 3D random-phase 3DPhase-randomized models generated from previous refinement CTF refinement performed Relion 3.1 symmetry particles re-extracted-centered re-boxed from 256 to 320 pixels for consensus refinement Relion 3.1 cryoSPARC37 density cytosolic layer particles symmetry for multibody soft mask large extracellular domain TMD protomer generated from consensus map UCSF Chimera Relion 3.1 other covers cytosolic domains protomer 3D multibody performed using two soft masks parameters previous consensus refinement motions analyzed by relion_flex_analyse Relion 3.1 two half-maps post-processing Relion 3.1 masked sharpened maps aligned to consensus map UCSF zoned regions summed Relion 3.1 composite maps for visualization model building resolution estimations on Fourier correlation 0.143 cutoff after correction masking effect B-factor map sharpening determined post-processing Relion 3.1 local resolution estimated with Relion 3.1 composite maps multibody refinement used for model buildingstructures PHD TMD EF1–2 DH domains hDUOX1 generated PDB ID 6ERC 5O0T 4IL1 5O0X docked cryo-EM maps Chimera40 models PHLD generated Rosetta Web Server ab initio selected distances RaptorX Contact Prediction validated fitting model cryo-EM densities bulky aromatic residues partial model hDUOXA1 generated EM builder44 initial models built Coot45 refined Phenix Figures prepared UCSF chimera40 Chimera X47 Pymol Nature Research Reporting Summary.Supplementary Additional
| 50.3
| 0.740802
|
10.1038/s41467-020-17590-x
|
PMC7387547
| "In this study the authors show that monotonous basaltic volcanoes can host a range of melts in thei(...TRUNCATED)
| "Many volcanoes erupt compositionally homogeneous magmas over timescales ranging from decades to mil(...TRUNCATED)
| "IntroductionVolcanoes are underlain by complex and dynamic magmatic systems that often span tens of(...TRUNCATED)
|
nature communications
|
[
"Article"
] |
[
"Geochemistry",
"Geology",
"Petrology",
"Volcanology"
] | "by complex magmatic systems kilometres crust from Moho to near-surface1,2 melts ascend undergo proc(...TRUNCATED)
| 49.9
| 0.838315
|
10.1038/s41467-020-17400-4
|
PMC7371698
| "Assessments of future virtual water trading are still lacking. Here the authors estimated the globa(...TRUNCATED)
| "Water stressed regions rely heavily on the import of water-intensive goods to offset insufficient f(...TRUNCATED)
| "IntroductionVirtual water trade (VWT) is the amount of water, either green (soil moisture) or blue (...TRUNCATED)
|
nature communications
|
[
"Article"
] |
[
"Sustainability",
"Hydrology"
] | "IntroductionVirtual water trade (VWT) is water green or blue consumed in production agricultural go(...TRUNCATED)
| 49.5
| 0.424733
|
10.1038/s41467-021-21363-5
|
PMC7910301
| "Here, advanced scanning transmission electron microscopy techniques are used to image the atomic st(...TRUNCATED)
| "The atomic structure at the interface between two-dimensional (2D) and three-dimensional (3D) mater(...TRUNCATED)
| "IntroductionFollowing the success of moiré engineering in modulating (opto-)electronic properties (...TRUNCATED)
|
nature communications
|
[
"Article"
] | ["Surfaces, interfaces and thin films","Electronic properties and materials","Two-dimensional materi(...TRUNCATED)
| "success moiré engineering properties/hexagonal boron nitride heterostructures1,2 twisted bilayer s(...TRUNCATED)
| 49
| 0.584603
|
10.1038/s41467-020-17515-8
|
PMC7393140
| "Here, the authors apply live-cell and in situ fluorescence imaging at the single-molecule level to (...TRUNCATED)
| "Spatial organization of biological processes allows for variability in molecular outcomes and coord(...TRUNCATED)
| "IntroductionOrganization is a fundamental part of life. For complex organisms, the spatial developm(...TRUNCATED)
|
nature communications
|
[
"Article"
] |
[
"Bacteriophages",
"Cellular noise",
"Single-cell imaging"
] | "IntroductionOrganization fundamental life complex organisms spatial development controlled for func(...TRUNCATED)
| 48.8
| 0.917512
|
10.1038/s41467-020-14616-2
|
PMC7005897
| "Humans are normally not aware that their eyes are always in motion, even when attempting to maintai(...TRUNCATED)
| "High visual acuity is essential for many tasks, from recognizing distant friends to driving a car. (...TRUNCATED)
| "IntroductionHumans critically rely on high visual acuity. Although fine spatial resolution is restr(...TRUNCATED)
|
nature communications
|
[
"Article"
] |
[
"Neural encoding",
"Oculomotor system",
"Sensorimotor processing",
"Visual system"
] | "rely on high visual acuity fine spatial resolution restricted to foveola tiny region retina less th(...TRUNCATED)
| 46.6
| 1.105571
|
10.1038/s41467-020-17401-3
|
PMC7374574
| "Orexin signaling is provided by diffusely distributed fibers and involved in different brain circui(...TRUNCATED)
| "The relationship between orexin/hypocretin and rapid eye movement (REM) sleep remains elusive. Here(...TRUNCATED)
| "IntroductionStable vigilance states, which fundamentally serve vital brain functions, depend on the(...TRUNCATED)
|
nature communications
|
[
"Article"
] |
[
"Hypocretin",
"Hypocretin",
"Orexin",
"Orexin",
"REM sleep"
] | "vigilance states serve brain functions depend on neurochemical signals1 hypothalamic neuropeptide o(...TRUNCATED)
| 47.3
| 1.381432
|
10.1038/s41467-020-19306-7
|
PMC7595090
| "A conclusive account on how the brain translates audiovisual evidence into a rapid decision is stil(...TRUNCATED)
| "Despite recent progress in understanding multisensory decision-making, a conclusive mechanistic acc(...TRUNCATED)
| "IntroductionIn everyday life, we often encounter situations that demand rapid decisions based on am(...TRUNCATED)
|
nature communications
|
[
"Article"
] |
[
"Neuroscience",
"Psychology"
] | "life situations rapid decisions ambiguous sensory information Consolidating evidence requires proce(...TRUNCATED)
| 47
| 1.331961
|
10.1038/s41467-021-20907-z
|
PMC7846738
| "It is known that invasive lung adenocarcinomas evolve from pre-cancerous dysplastic lesions. In thi(...TRUNCATED)
| "The evolution of DNA methylome and methylation intra-tumor heterogeneity (ITH) during early carcino(...TRUNCATED)
| "IntroductionLung cancer remains the leading cause of cancer-related death worldwide, yet it is cura(...TRUNCATED)
|
nature communications
|
[
"Article"
] |
[
"Cancer genomics",
"Non-small-cell lung cancer",
"Tumour heterogeneity",
"Epigenomics"
] | "cancer leading cause death curable if treated early cancers preceded by precancers Treating precanc(...TRUNCATED)
| 49.1
| 0.889898
|
End of preview. Expand
in Data Studio
README.md exists but content is empty.
- Downloads last month
- 1