There are more than a dozen approaches available for single-cell isolation and transcriptome amplification. Based on benchmark comparisons, Smart-seq2 generally has higher sensitivity than competing droplet-based approaches (Svensson et al., 2017; Ziegenhain et al., 2017). We reasoned that sensitive measurement is crucial in our study, given that single Toxoplasma gondii parasites are at least 50-fold smaller in volume than a typical mammalian cell, and thus the average parasite gene is likely expressed with much lower copy number per cell than a typical mammalian gene. For our initial studies, we used the common Type I lab strain of Toxoplasma, RH, grown in vitro in human foreskin fibroblasts (HFFs). Following such growth, individual tachyzoites were released by passage through a narrow-gauge needle and then purified by fluorescence activated cell sorting (FACS) into 384-well or 96-well plates. We then synthesized, amplified, and barcoded cDNA using Smart-seq2 and Illumina Nextera protocols. We reduced the reagent cost in 384-well plates effectively by ten-fold compared to the 96-well format. The sequenced reads were bioinformatically deconvolved and grouped into individual parasites for analysis using modified bcl2fastq and custom python scripts (Materials and methods). A schematic to illustrate our experimental workflow is shown in Figure 1.

To ensure that our workflow efficiently captures single Toxoplasma parasites, we mixed equal numbers of two transgenic lines of RH, one expressing GFP and the other expressing mCherry, and sorted individual parasites into a 384 well plate based on the presence of either green or red fluorescence. After Smart-seq2 amplification, we quantified the expression of GFP and mCherry mRNAs using quantitative polymerase chain reaction (qPCR). Across all 301 wells that we measured, we observed the presence of both GFP and mCherry mRNA in only one well, indicating that the rate of doublet events is below 1% (Figure 1—figure supplement 1a). To address the possibility that the reduced reagent volume in the 384-well format could potentially saturate the reaction chemistry and thus limit quantification range, we sorted varying numbers of RH and quantified with qPCR the mRNA of a gene encoding the abundantly expressed surface protein, SRS29B (SAG1) (Figure 1—figure supplement 1b). The detected amount of SAG1 mRNA present in wells containing single, eight, or fifty RH fell into the expected distributions based on the number of parasites sorted, without signs of saturation, indicating that the assay is capable of quantitative measurement at the single Toxoplasma level. We then proceeded to sort parasites into 384-well based on live/dead staining and sequence 729 RH (612 passed quality control) strain single Toxoplasma parasites from asynchronous populations grown under tachyzoite conditions. We also sorted and sequenced 287 RH parasites (237 passed quality control) into 96-well plate for comparison, which we will discuss further in another section. For Pru and ME49 strains, we collected parasites at several time points post alkaline treatment which induces differentiation from tachyzoites to bradyzoites to follow changes in their expression profiles during in vitro development (Materials and methods), yielding 2655 Pru (2198 passed quality control) and 1828 ME49 (1552 passed quality control) single parasites. RH reads were aligned to the GT1 strain genome, which is the most complete reference for Type I parasites, while Pru and ME49 were aligned to the ME49 strain Type II genome reference. Because many genes encoding Toxoplasma secretion factors and surface proteins are evolutionary products of gene duplication events (Reid, 2015), we expected high sequence similarity amongst a substantial portion of the parasite genes. Thus, we modified our gene counting pipeline to account for duplicated genes by distributing reads across all regions with equal alignment score that passed thresholds (Materials and methods). The reason why we adopted a correction scheme for multiply-mapped reads is because the analysis of co-occurrence, or lack thereof, of pathogenic factors (e.g. surface antigens) hinges on sensitive detection of their expression. We thus faced a choice of increasing the false positive in gene alignment by correcting for multiply mapped reads or increasing its false negative by counting only uniquely aligned reads. We chose to account for multiply mapped genes, which would otherwise be obscured, as we hypothesized that surface antigen expression may vary between individual parasites, which we further discuss in another section. A comparison of counting methods does not reveal significant differences in the observed counts (Figure 1—figure supplement 1c). Further analysis reveals that our modified pipeline recovered the detection of more parasite genes than default parameters (Figure 1—figure supplement 1d).

To ensure that poorly amplified or sequenced parasites did not confound our downstream analysis, we filtered samples based on several quality metrics including percent reads mapping to ERCC spike-in sequences, number of genes detected, and sequencing depth (Materials and methods; Figure 1—figure supplement 2a). On average, each sequenced parasite contains 30–50% reads that mapped to Toxoplasma genes encoding proteins (top panel in Figure 1—figure supplement 2b). Most of the unmapped reads are from Toxoplasma’s 28 s ribosomal RNA. The relatively high rate of rRNA contamination was also observed in single-parasite RNA sequencing of Plasmodium (Reid et al., 2018). We suspect this occurred due to non-specific priming as protozoan cells have low RNA input. We normalized for sequencing depth across cells by dividing each read count by the read sum of each corresponding cell and then multiplied by the median of read sum within each dataset to yield ‘count per median’ (CPM). After filtering ERCC spike-in and rRNA genes, we detected on average 862, 1290, and 970 genes per parasite with greater than or equal to two read counts (Materials and methods) in the RH, Pru, and ME49 datasets, respectively (bottom panel in Figure 1—figure supplement 2b). Characterization of our measurement sensitivity based on logistic regression modeling of ERCC spike-in standards (Materials and methods) (Lönnberg et al., 2017) reveals a 50% detection rate of 17, 17, and 26 molecules for RH, Pru, and ME49 datasets, respectively (top panels in Figure 1—figure supplement 2c). The sensitivity of our 384-well Smart-seq2 measurement is comparable to the previously reported range for the 96-well format (Ziegenhain et al., 2017). As expected from our qPCR titration experiment, scRNA-seq measurement of gene expression is quantitative at single parasite resolution based on ERCC standards. We determined that the linear dynamic range of our scRNA-seq measurement spans over three orders of magnitude (bottom panels in Figure 1—figure supplement 2c). Taken together, we demonstrate a scalable and cost-effective approach to measure the transcriptomic changes of individual parasites with high sensitivity and accuracy.

Previous work posited a potential link between bradyzoite development and cell cycle, which poses a significant challenge to the bioinformatic analysis of either process (Radke et al., 2003). To characterize cell cycling changes without confounding contributions from developmental processes, we first analyzed an asynchronous population of Type I RH strain parasites grown under tachyzoite conditions; this extensively passaged lab strain is known to have little propensity to switch to bradyzoites under such conditions (Soête et al., 1994) (Materials and methods). After filtering out genes whose expression levels did not vary significantly between individual parasites, we projected the data with principal component analysis (PCA) (Materials and methods). Interestingly, the first two principal components (PCs) reveal a circular trajectory that coincides with relative DNA content, determined using a cell permeable DNA content stain (top panel in Figure 2a). Unsupervised neighborhood clustering identified five distinct clusters of parasites based on their transcriptional profiles (middle panel in Figure 2a) (Materials and methods). We computed RNA velocity to infer transcriptional dynamics (La Manno et al., 2018; Wolf et al., 2018) and the velocity vector field indicates a net ‘counter-clockwise’ flow of transcriptional changes (bottom panel in Figure 2a) (Materials and methods). We assigned cell cycle phase to the clusters based primarily on change in DNA content (Figure 2—figure supplement 1a) but also considering previous bulk transcriptomic characterization (Behnke et al., 2010; Figure 2—figure supplement 1b). Unsupervised clustering identified two distinct clusters in G₁ state, which we have designated as G₁a and G₁b. We found a list of differentially expressed genes between the two G₁ clusters. The G₁a cluster is highly enriched for the expression of metabolic genes such as phenylalanine hydroxylase (TGGT1_411100) and pyrroline-5-carboxylate reductase (TGGT1_236070), as well as invasion-related secreted factors such as MIC2 (TGGT1_201780), MIC3 (TGGT1_319560), and MIC11 (TGGT1_204530). On the other hand, G₁b cluster is enriched for the expression of 3-ketoacyl reductase (TGGT1_217740) and cytidine and deoxycytidylate deaminase (TGGT1_200430), as well as numerous uncharacterized proteins (Supplementary file 2). The relative abundance of G₁a, G₁b, S, M, and C states were determined to be 18%, 32%, 28%, 15%, and 7%, respectively. Without chemical synchronization, the correlation between the scRNA-seq data of asynchronous parasites and previously published bulk transcriptomic measurement suggests strong agreement in cluster assignment and cell cycle state identification (Figure 2—figure supplement 1b). This highlights a key advantage of scRNA-seq, as it enables identification of cell cycle status of a parasite without reliance on chemical induction, which may lead to unnatural cellular behavior.

To verify the cyclical nature of gene expression through the lytic cycle, we reconstructed a biological pseudotime of RH using Monocle 2 (Materials and methods). The results show a clear oscillatory expression pattern for the variably expressed genes along the pseudotime axis (Figure 2b). To further characterize cell cycle expression patterns, we clustered genes based on pseudotime interpolation and hierarchical clustering (Materials and methods). Some of the key organelles in tachyzoites are made at different times in the cell cycle (Behnke et al., 2010). To confirm and refine this finding, we calculated the mean expression values for each set of organelle-specific genes based on their annotation in ToxoDB (Gajria et al., 2008; Supplementary file 4). This showed the expected, strong oscillation with pseudotime (bottom panel in Figure 2—figure supplement 1c), which also strongly correlates with the oscillation of DNA and total mRNA content (top panels in Figure 2—figure supplement 1c). On the other hand, we also observed instances where a given gene’s expression was discordant to the dominant trend of its nominal organelle set (arrows in Figure 2—figure supplement 1d). For example, 63.5% of genes annotated as rhoptry (ROP) or rhoptry neck (RON) are assigned pseudotime cluster 3, while the remaining 36.5% rhoptry genes are assigned pseudotime clusters 1 or 2 (Figure 2—figure supplement 1e). Specifically, genes annotated as ROP33 and ROP34, based on their homology to genes encoding known rhoptry proteins, are assigned to cluster 2 instead of cluster 3 (left panel in Figure 2—figure supplement 1f). Recent reports have experimentally determined these two to be non-rhoptry-localizing proteins. This is consistent with our observation of discordance between their and known ROPs' expression profiles along the pseudotime (Beraki et al., 2019). Through analysis of pseudotime clustering, we also identified genes not annotated as ROPs within the ROP-dominated cluster 3, such as TGGT1_218270 and TGGT1_230350, that have recently been shown to encode bona fide rhoptry and rhoptry neck proteins, now designated as ROP48 and RON11, respectively (Camejo et al., 2014; Beck et al., 2013) (left panel in Figure 2—figure supplement 1f). As another example, IMC2a peaks in expression level in G₁, while the majority of inner-membrane complex (IMC) genes are expressed towards the M/C phase of the cell cycle (right panel in Figure 2—figure supplement 1f). A recent report has proposed reannotation of IMC2a as a dense granule (GRA) protein (GRA44) based on subcellular localization (Coffey et al., 2018), which is consistent with our unsupervised group assignment of IMC2a as falling in cluster one where GRA genes dominate. A list of 8590 RH genes with their corresponding pseudotime clustering assignment is provided (Supplementary file 5). We observe high discordance of pseudotime expression for several genes in each annotated organelle sets, suggesting that the current Toxoplasma annotation may need significant revision. Our scRNA-seq data provide an important resource to help identify mis-annotated genes and infer putative functions of uncharacterized proteins.

Toxoplasma has one of the most complicated developmental programs of any single-celled organism; however, it is unknown how synchronized the transition is between developmental states. To address this, we assessed the inherent heterogeneity within asexually developing Pru, a type II strain that is capable of forming tissue cysts with characteristics that resemble early ‘bradyzoites’ from in vivo source upon growth in in vitro alkaline conditions (Soete et al., 1993; Jones et al., 2017). We applied scRNA-seq to measure and analyze Pru parasites grown in HFFs as tachyzoites (‘uninduced’) and after inducing the switch to bradyzoites by growth in alkaline media for 3, 5, and 7 days. Projection of the first two PCs of uninduced Pru tachyzoites (Day 0) reveals the expected circular projection (Figure 3—figure supplement 1a), presumably reflecting cell cycle progression as seen for the RH tachyzoites, described above. To validate this, we developed a random forest classifier model based on our cell cycle assignment in RH (Materials and methods). Comparable to what we observed in RH, cell cycle prediction reveals that the uninduced population of Pru is composed of 28%, 41%, 21%, 7%, and 3% parasites in G₁a, G₁b, S, M, and C states, respectively. Consistent with previous observation (Jerome et al., 1998), our data show most induced Pru parasites (Day 3–7) are in the G₁ state with a predominance of G₁b (Figure 3—figure supplement 1b).

To identify transcriptomic changes associated with the tachyzoite-bradyzoite transition, we next projected data from both induced and uninduced Pru parasites onto two dimensions using UMAP, a nonlinear dimensionality reduction method (Materials and methods) (McInnes et al., 2018). Unsupervised clustering revealed six distinct clusters of parasites, which we label P1-6 (Figure 3a). Cluster formations partially correlate with treatment time points and cell cycle states (Figure 3b; Figure 3—figure supplement 1c), suggesting that the asexual differentiation program overlaps with cell cycle regulation in Toxoplasma, as proposed previously (Radke et al., 2003). We stratified the datasets by days post alkaline induction (dpi) and observed elevated expression of previously described ‘early’ bradyzoite marker genes in induced parasites, including SRS44 (CST1) and BAG1, with a concomitant reduction in expression of SRS29B (SAG1), a tachyzoite-specific surface marker gene (Figure 3—figure supplement 2). The abundance of SAG1⁺ parasites (72%) in the induced population suggests two possible interpretations: (1) the depletion of SAG1 mRNA is relatively slow and we are measuring SAG1 transcripts made when the parasites were still tachyzoites, or (2) the asexual transition induced by alkaline treatment is highly asynchronous. Interestingly, RNA velocity analysis suggests that P3 may be a fate decision point as the trajectory trifurcates into either P4 (cell cycle), P1, or P2 as evident by the net transcriptional flow (compare Figure 3a to right panel in Figure 3b).

To determine the gene modules specific to a given cluster, we conducted differential gene expression for each cluster (Figure 3c and Supplementary file 5). P1 cluster cells are enriched for expression of bradyzoite-specific genes while P2-5 are enriched for that of tachyzoite-specific or cell cycle-associated genes (Figure 3c–d). In our scRNA-seq data, we also observe a small portion of BAG1⁺ bradyzoites (7.1%) annotated as either S, M, or C states, indicating that they are replicating (Figure 3—figure supplement 1d). Our data supports the notion that bradyzoites can undergo cell cycle progression (Watts et al., 2015). We observe a family of transcription factors known as Apetala 2 (AP2) that are differentially expressed across different clusters, some of which are implicated in Toxoplasma development (De Silva et al., 2008; Radke et al., 2013; Walker et al., 2013; Figure 3e). In particular, we identify AP2Ib-1, AP2IX-1, AP2IX-6, and AP2VI-2 as over-expressed in P1, suggesting their potential roles in the regulation of developmental transition, while AP2IX-9, AP2X-8, AP2VIIa-6, AP2XI-1, AP2IX-3, AP2VIII-7, and AP2-domain protein TGME49_215895 (not yet assigned a formal AP2 number), are highly expressed in P6, hinting at their possible roles in defining this distinct cluster of parasites.

The most highly expressed genes in P6 include genes enriched in P2 as well as bradyzoite-specific genes found in P1 (Figure 3c). To identify genes that are specifically expressed in P6, we used Wilcoxon’s test (Figure 3—figure supplement 3a) (Materials and methods) between P6 and P2 or P1. Comparison of our data to previous bulk transcriptomic measurement in tachyzoites, tissue cyst, or isolates at the beginning or the end of sexual cycle showed no specific enrichment in known developmental stages (Figure 3—figure supplement 3b; Ramakrishnan et al., 2019). Instead, we show that based on their expression, P6 forms a distinct sub-population of parasites which suggests that alkaline induced Toxoplasma may be more heterogeneous than previously thought. Thus, scRNA-seq resolves a transcriptomic landscape of asexual development and suggests the existence of an otherwise hidden state.

To determine the reproducibility of the phenomena we observed in the differentiating Pru strain parasites, we repeated the analysis with another widely used Type II strain, ME49, examining 1828 single ME49 parasites exposed to alkaline conditions to induce switching to bradyzoites. Data from the two experiments were computationally aligned using Scanorama to remove technical batch effects while retaining sample-specific differences (Hie et al., 2019). Unsupervised clustering revealed five distinct clusters in ME49 which share significant overlap in expression patterns with Pru (Figure 3—figure supplement 4a). Matrix correlation of batch-corrected expression across the two strains demonstrate analogous mapping for most, but not all cluster identities (Figure 3—figure supplement 4b). To simplify the visualization and comparison across the two datasets, we next applied Partition-Based Graph Abstraction (PAGA) to present clusters of cells as nodes with connectivity based on similarity of the transcriptional profiles between clusters (Wolf et al., 2019). A side-by-side comparison of expression of tachyzoite, bradyzoite, and sexual stage specific genes reveals some key similarities and dissimilarities (Figure 3—figure supplement 4c). Clusters P1 and M1 are both enriched for the expression of bradyzoite marker genes, while clusters M4-5 and P4-5 are both predicted to be S/M/C phases of the cell cycle. Curiously, P6-specific genes (green panels in Figure 3—figure supplement 4c) are not enriched in any cluster in ME49, suggesting that P6 state is not conserved across the two type II strains. Such differences may not be surprising, however, as Pru and ME49 have entirely distinct passage histories.

A unique advantage of scRNA-seq over bulk RNA-seq is its ability to measure cell-to-cell variation that is independent of known biological processes. The Toxoplasma gondii genome encodes a family of over 120 SAG1-related sequence (SRS) proteins that fall into distinct subfamilies; most or all of these are presumed to be surface antigens based on their sequence similarities, including the presence of a predicted GPI-addition signal (Manger et al., 1998b). Whether SRSs constitute an antigenic repertoire that contribute to evasion of host adaptive immunity response is unclear; however, existing data on developmentally regulated expression of SRSs including SAG1 (SRS29B) and SRS16B lend support to that hypothesis (Kim and Boothroyd, 2005; Kim et al., 2007). In the experiments using 384-well plates, we noted that most SRS genes were detected in only a few sporadic cells (Figure 4—figure supplement 1a); however, biological variation was difficult to distinguish from measurement dropout (failure to capture a mRNA molecule) given the limited sensitivity of 384-well assay and the relatively low abundance of SRS transcripts. To increase our sensitivity, therefore, we performed scRNA-seq of extracellular Toxoplasma in 96-well plates which achieved ~40% sensitivity of detection for single molecules of ERCC spike-ins, compared to 14% sensitivity that we obtained in the 384-well experiments, at roughly equivalent sequencing depth (Figure 4a). To determine the extent of transcriptional variation in SRS genes independent of cell cycle or asexual development, we isolated RH parasites in G1 state using DNA content stain and FACS (Figure 4—figure supplement 1b). We noticed that the distributions of DNA content in single cells varied across the two experiments (compare Figure 4—figure supplement 1b to Figure 3—figure supplement 1a), suggesting that the distribution may depend on external factors such as tissue culture confluence, parasite load, or the amount of time passed since infection. We also measured on average 984 genes with read counts equal to or greater than two in 96-well plates, up from 862 genes in 384-well plates (Figure 4—figure supplement 1c). Analysis of the resulting 96-well plates data show that the vast majority of SRS genes are detected in only a small fraction of the population (Figure 4b), similar to what we observed in 384-well experiments. To control for measurement dropout as a cause of such variation, we first determine the non-zero median for those genes, that is, the median expression level in cells where any transcript for that gene is detected. We then assess how often we fail to detect an ERCC spike-ins that had a similar non-zero median expression level. If the failure to detect a SRS was due to measurement dropout, then the fraction of cells without detection for its transcript should be about the same as for the ERCC spike-in with similar non-zero median expression. If, on the other hand, the failure to detect the SRS transcript is due to underlying biological variation between cells, then we will find a lower frequency of cells with detectable transcript for that gene than the ERCC spike-in with similar expression. The results show that SRSs are indeed detected at a substantially lower rate when compared to ERCC spike-ins (Figure 4c), indicating that the variation of SRS detection cannot be explained by measurement noise. Compared to most other genes, SRSs are expressed in a significantly smaller fraction of the population and at a relatively lower abundance (Figure 4—figure supplement 1d-e). To determine whether variation of SRS expression is due to cell cycle or asexual development, we developed a bootstrapping approach to quantify the dependence of expression on a topological network that represents either process (Materials and methods). We show that apart from a few known SRSs, most SRSs do not co-vary with cell cycle or asexual development, which suggests that the sporadic nature of SRS expression may hint at a different biological role for their variation, beyond cell cycle and asexual development, as discussed further below (Figure 4d). These results reveal new insights into the antigenic repertoire of Toxoplasma, which appears far more heterogeneous, on a parasite-to-parasite basis, than previously known.

In examining our data, we were struck by a lone cell in the 384-well dataset of RH that had no detectable level of SRS29B (SAG1), which was otherwise ubiquitously and abundantly expressed in all tachyzoites (Figure 5a). We wondered if this cell was damaged or unhealthy but its gene count and read depth were similar to its cohort (Figure 5b). Further analysis of this SAG1^- cell’s transcriptome revealed that it lacks expression of the most abundant tachyzoite-specific and bradyzoite-specific genes; instead, its gene expression most closely resembles the sexual stages from cat intestinal isolates, including abundant expression of a cat-stage SRS, SRS22C (Figure 5c). Within the set of genes uniquely expressed in the SAG1^- cell are AP2IX-1 and AP2III-4 which suggest a possible role for one or both of these transcription factors in regulating the transcript differences observed in this outlier. To test this hypothesis, we transiently expressed AP2IX-1 under the control of a strong promoter in RH parasites (Figure 5d). Consistent with our hypothesis, immunofluorescence assay (IFA) revealed a reduction of SAG1 surface protein expression within ~18–20 hr after transfection (Figure 5e–f), while quantitative RT-PCR showed significantly higher mRNA expression for several of the genes that were upregulated in the SAG1^- cell: namely, 207965, 222305, 205210, and SRS22C, the latter being over 1000-fold higher in the transfected population than control (Figure 5g). Note that the SAG1 transcript levels in the transfected population were not substantially lower, as expected because there remains a large number of untransfected cells in the population which still express high levels of this gene. As SRS22C is predicted to be a surface antigen like the rest of SRS (Gajria et al., 2008), this suggests that AP2IX-1 induction can control the switching of surface antigens. We also attempted to express AP2III-4 that is upregulated in the SAG1^- cell but were unable to obtain an epitope-tagged version of the gene (which is over nine kbp, including the promoter). Overall, these results demonstrate the ability to infer transcriptional regulation from a single parasite cell that has an unusual co-expression pattern, revealing AP2IX-1 as a novel transcriptional factor that can alter antigen expression in Toxoplasma.

Plasmodium and Toxoplasma are both unusual for their prevalence in humans and their complex developmental transition, despite the fact that humans are non-definitive hosts for both pathogens. Toxoplasma has been used as an experimental model for other apicomplexans including Plasmodium, yet the replication modes amongst apicomplexans can be very different. For example, asexual replication of Toxoplasma involves endodyogeny, which is similar to canonical binary fission, in order to replicate and divide. Asexual division of Plasmodium, on the other hand, involves schizogony, a process that entails multiple rounds of DNA replication and nuclear division followed by a mass cytokinesis (Aly et al., 2009; Cowman et al., 2016). Despite their fundamental differences in cell cycle progression, we wondered if our Toxoplasma dataset can be combined with the Malaria Atlas (Howick et al., 2019) to provide insights into apicomplexan biology. Toward this end, we first identified 1830 one-to-one orthologous genes between Plasmodium berghei and Toxoplasma gondii Pru (Figure 6—figure supplement 1a). As expected, the vast majority of orthologous genes are not surface adhesion factors or effector molecules in Toxoplasma (Figure 6—figure supplement 1b). After removing non-orthologous genes, we combined the two species’ datasets to produce an integrated UMAP projection with Scanorama (Hie et al., 2018). In this integrated projection space, mutually similar clusters of cells in the two datasets are brought close together, while organism-specific cell types are not (Figure 6a). Transitional dynamics of the parasitic development are preserved in the integrated projection, as separation of the original cluster assignment indicates (Figure 6b). We highlight comparative similarity across the two organisms by calculating the fraction of cells that share the same topological neighborhood in the integrated network of Plasmodium and Toxoplasma (Figure 6c), revealing striking similarity of expression pattern between the two organisms. We discover the concerted transcriptional expression of several conserved orthologous gene sets in the life cycles of these two parasites (Figure 6d). For example, the pre-erythrocytic stages of Plasmodium (sporozoite and merozoites) most closely match the ‘G1 a’ stage of Toxoplasma. Both parasites at these stages are not actively replicating DNA and express high levels of ribosomal and mitochondrial genes. Meanwhile, the ring stage of Plasmodium most closely matches the ‘G1 b’ phase of Toxoplasma because they both express high levels of ribosomal genes with minimal expression of IMC-related and microtubule-related genes. Exo-erythrocytic form (EEF), microgametes (male), and trophozoite stages of Plasmodium berghei most closely match the ‘S’ phase of Toxoplasma cell cycle. This is because they express DNA-replication factors and centrosome components, which are required for condensation and segregation of chromosomes. Intriguingly, the schizont stage of Plasmodium most closely resembles the ‘M’ and ‘C’ phases due to the expression of microtubule, centrosome, and IMC genes. A more detailed illustration of these gene sets expression is shown in Figure 6—figure supplement 1c. Our results indicate that while the cellular morphology and replication strategies may differ drastically between the two parasites, the cellular state, as defined by the transcriptomic profiles, can bear striking analogy and resemblance in the course of asexual replication.

Toward enabling the larger scientific community to take advantage of the scRNA-seq data we collected on the three Toxoplasma strains, we have constructed an interactive single-cell atlas for Toxoplasma gondii (http://st-atlas.org) using Bokeh and Javascript. Our atlas resource allows users to visualize the expression pattern of individual cells by providing a gene ID of interest and using the built-in graphical interface toolset. A factor plot displays expression levels within the parasite population based on cell cycle status, cluster identities, days post induction, or other categories. A sub-panel on the bottom left shows additional information for each input gene including translated product, mean, and standard deviation of the expression. Users can selectively highlight a subset of parasites by using a simple click-and-drag interface for further analysis. Example use cases are provided in Figure 6e–f. We have also made the raw data available through the atlas website. We hope this will help make our data and results readily accessible for others interested in exploring Toxoplasma parasitology.

All Toxoplasma gondii strains were maintained by serial passage in human foreskin fibroblasts (HFFs) cultured at 37 C in 5% CO₂ in complete Dulbeco’s Modified Eagle Medium (cDMEM) supplemented with 10% heat-inactivated fetal bovine serum (FBS), 2 mM L-glutamine, 100 U/ml penicillin, and 100 ug/ml streptomycin. T. gondii strains used in this study were RH, Pru-GFP, and ME49-GFP-luc.

Differentiation to bradyzoite was induced by growth under low-serum, alkaline conditions in ambient (low) CO₂ as previously described (Weiss et al., 1995). Briefly, confluent monolayers of HFFs were infected with tachyzoites at a multiplicity of infection (MOI) of 0.025 in RPMI 1640 medium (Invitrogen) lacking sodium bicarbonate and with 1% FBS, 10 mg/ml HEPES, 100 U/ml penicillin, and 100 g/ml streptomycin at pH 8.2. The infected HFFs were cultured at 37°C without supplemented CO₂.

HFF monolayers infected with parasites overnight were scraped, and the detached host cells were lysed by passing them through a 25-gauge needle three times or a 27-gauge needle six times. The released parasites were spun down at 800 rpm for 5 min to pellet out host cell debris, and the supernatant was spun down at 1500 rpm for 5 min to pellet the parasites. The parasites were then resuspended in 500 µL of FACS buffer (1x phosphate-buffered saline, PBS, supplemented with 2% FBS, 50 ug/ml DNAse I, and 5 mM MgCl₂*6H₂O), passed through both a 5 µm filter and a filter cap into FACS tubes, and stored on wet ice until it was time to sort. In samples stained for DNA content, the parasites were resuspended in 500 µL of FACS buffer plus 1.5 µL of Vybrant DyeCycle Violet (from ThermoFisher, catalog number V35003) and incubated at 37 C and 5% CO₂ for 30 min.

The parasites were also stained with either propidium iodide (PI), Sytox Green, or the live/dead fixable blue dead cell stain kit (catalog number L34962) prior to sorting in order to distinguish live cells from dead cells. To stain with PI, 10 µL of 0.5 mg/ml PI was added to every 500 µL of parasite suspension in FACS buffer, and the parasites were incubated covered on ice for at least 15 min. To stain with Sytox Green, 1 drop of Sytox Green per ml was added to the parasite suspension in FACS buffer, and the parasites were incubated at room temperature for at least 15 min. To stain with the live/dead fixable blue dead cell stain kit, 1.5 µL of the kit’s viability dye was added to every 500 µL of parasites along with the secondary antibody, and parasites were washed and resuspended in FACS buffer as usual.

Eight mL of lysis buffer was prepared by mixing together: 5.888 mL of water, 160 µL recombinant RNase inhibitor (Takara Clonetech), 1.6 mL of 10 mM dNTP (ThermoFisher), 160 µL of 100 uM oligo-dT (iDT; see attached Supplementary file 1 for oligos), 1:600,000 diluted ERCC spike-in (ThermoFisher), and 32 µL of 10% Triton X-100. All reagents are declared RNase free. Lysis plates were prepared by dispensing 0.4 µL of lysis buffer into each well of a 384 well hard-shell low profile PCR plate (Bio-rad) using liquid handler Mantis (Formulatrix). Single parasites were sorted using the Stanford FACS Facility’s SONY SH800s sorter or BD Influx Special Order sorter into the 384-well plates loaded with lysis buffer. Single color and colorless controls were used for compensation and adjustment of channel voltages. The data were collected with FACSDiva software and analyzed with FlowJo software. RH parasites were index sorted with fluorescence signal of cell permeable DNA stain, DyeCycle Violet.

Smart-seq2 protocol was carried out as previously described (Picelli et al., 2014) using liquid handlers Mantis and Mosquito (TTP Labtech) with slight modifications. 384-well Smart-seq2 was performed with a 2 µL final reaction volume, while 96-well format was carried out in 25 µL final reaction volume as recommended by the original protocol. For 384-well Smart-seq2, we performed 19 rounds of cDNA pre-amplification after reverse transcription with oligo-dT primers. Each well is then diluted with 1 to 4 v:v in RNAse free elution buffer (QIAgen) to a total volume of 8 µL. For 96-well Smart-seq2, we performed 30 rounds of cDNA pre-amplification after reverse transcription. PCR is performed with ‘IS_PCR primers’. Each well is then purified with Ampure XP beads at 0.8X volume ratio and resuspended in 20 µL RNAse free elution buffer. We measured the size distribution and concentration of each well in 96-well plate using Fragment Analyzer High-sensitivity NGS kit (Agilent). We normalized the concentration of cDNA of each well to a concentration of 0.4 ng/µL. Finally, for both 384-well and 96-well Smart-seq2 measurements, we conducted library preparation with in-house Tn5 tagmentation using custom cell barcode and submitted for 2 × 150 bp paired-end sequencing on NovaSeq 6000 at the Chan Zuckerberg Biohub Genomics core. All primer sequences are provided as a supplementary file.

The pGRA-AP2I × 1-V5 plasmid for AP2I × 1 transient expression was created using Gibson assembly (NEB) from the pGRA-V5 (Panas et al., 2019). RH parasites were transfected with pGRA-AP2I × 1-V5 using the Amaxa 4D Nucleofector (Lonza). Tachyzoites were mechanically released in PBS, pelleted, and resuspended in 20 µL P3 Primary Cell Nucleofector Solution (Lonza) with 7 or 15 µg DNA for transfection. After transfection, parasites were allowed to infect HFFs in DMEM. After 18–20 hr of infection, parasites were prepared for Immunofluorescence Assay (IFA) or qRT-PCR. To compute statistical independence between the transfected and control samples, we applied Student’s t-test by assuming unequal variance. One sigma (*) indicates one standard deviation in mean difference assuming null hypothesis.

Monolayers of infected cell on glass coverslips were fixed with cold methanol for 12 min. Samples were washed with PBS and blocked using 3% bovine serum albumin (BSA) in PBS for at least 30 min. SAG1 was detected with rabbit anti-SAG1 polyclonal antibody and V5 was detected with mouse anti-V5 tag monoclonal antibody (Invitrogen). Primary antibodies were detected with goat polyclonal Alexa Fluor-conjugated secondary antibodies (Invitrogen). Primary and secondary antibodies were both diluted in 3% BSA in PBS. Coverslips were incubated with primary antibodies for 30 min, washed, and incubated with secondary antibodies for 30 min. Vectashield with DAPI stain (Vector Laboratories) was used to mount the coverslips on slides. Fluorescence was detected using wide-field epifluorescence microscopy and images were analyzed using ImageJ. All images shown for any given condition/staining in any given comparison/dataset were obtained using identical parameters.

To quantify the purity of single parasite sort and to ensure the cDNA synthesis reaction was not saturated, GFP, mCherry, or SAG1 mRNA expression were measured using commercial qPCR mastermix, SsoAdvanced Universal SYBR Green mastermix (Bio-rad). Briefly, 0.1 µL of diluted cDNA was added in a total of 2.1 µL reaction volume per well on a 384 well plate with qPCR mastermix and 200 nM PCR primers. The reaction was incubated on a Bio-rad qPCR thermal cycler with the following programs: 5 min of 95°C, 45 cycles of 95°C for 5 s and 56°C for 1 min, and imaging. To quantify the transcriptional effects of AP2IX-1 transient expression in RH parasites, infected cell monolayers were first lysed with Trizol (Invitrogen) and RNA was extracted using standard molecular biology technique. cDNA from each sample was generated with Smart-seq2 protocol using 20 µL total reaction volume and roughly 200 ng RNA input followed by 0.8X AMPure XP beads (Beckman Coulter) purification. For qPCR, 1 ng of cDNA was added in a total of 2.1 µL reaction volume per well on a 384 well plate with qPCR mastermix and 200 nM PCR primers as described above. The reaction was incubated on a Bio-rad qPCR thermal cycler with the following programs: 5 min of 95°C, 60 cycles of 95°C for 5 s and 61°C for 30 s, and imaging. Each gene was measured four times with samples collected from at least two separate wells. The transfection and qRT-PCR experiments were performed twice in separate experiments. Fold change of gene expression was calculated as shown previously (Livak and Schmittgen, 2001) using ACT1 expression as an internal control for samples. All primer sequences are provided in Supplementary file 1.

BCL output files from sequencing were converted into gzip compressed FastQs via a modified bcl2fastq demultiplexer which is designed to handle the higher throughput per sequencing run. To generate genome references with spike-in sequences, we concatenated T. gondii ME49 NCBI genome assembly version 11/1/2013 or T. gondii GT1 T NCBI Genome assembly version 7/19/2013 genome references with ERCC sequences (obtained from ThermoFisher website). The raw fastq files from sequencing are aligned to the concatenated genomes with STAR aligner (version 2.6.0c) using the following settings: '--readFilesCommand zcat --outFilterType BySJout --outFilterMutlimapNmax 20 --alignSJoverhangMin 8 --alignSJDBoverhangMin 1 --outFilterMismatchNmax 999 --outFilterMismatchNoverLmax 0.04 --alignIntronMin 20 --alignIntronMax 1000000 --alignMatesGapMax 1000000 --outSAMstrandField intronMotif --outSAMtype BAM Unsorted --outSAMattributes NH HI AS NM MD --outFilterMatchNminOverLread 0.4 --outFilterScoreMinOverLread 0.4 --clip3pAdapterSeq CTGTCTCTTATACACATCT --outReadsUnmapped Fastx'. Transcripts were counted with a custom htseq-count script (version 0.10.0, https://github.com/simon-anders/htseq) using ME49 or RH GFF3 annotations (version 36 on ToxoDB) concatenated with ERCC annotation. Instead of discarding reads that mapped to multiple locations, we modified htseq-count to add transcript counts divided by the number of genomic locations with equal alignment score, thus rescuing measurement of duplicated genes in the Toxoplasma genome. Parallel jobs of STAR alignment and htseq-count were requested automatically by Bag of Stars (https://github.com/iosonofabio/bag_of_stars) and computed on Stanford high-performance computing cluster Sherlock 2.0. Estimation of reads containing exonic and intronic regions is computed with Velocyto estimation on the BAM output files and requested automatically by Bag of Velocyto (https://github.com/xuesoso/bag_of_velocyto) on Sherlock 2.0. Gene count matrix is obtained by summing up transcripts into genes using a custom python script. Scanpy velocyto package is then used to estimate transcriptional velocity on a given reduced dimension. Parameters used for generating the results are supplied as supplementary python scripts. Sample code to generate the analysis figures are provided in supplementary jupyter notebooks.

To filter out cells with poor amplification or sequencing reaction and doublet cells, we discarded cells based on gene counts (>0 reads), total reads sum, percent reads mapped to Toxoplasma genome, percent ERCC reads, and percent ribosomal RNA reads. We reported ‘% mapped’ based on the meta-alignment output from STAR aligner. We checked for some of the unmapped reads on BLASTn and found the majority of them to map to Toxoplasma 28S ribosomal RNA. Next, we filtered ‘ribosomal RNA’ genes from the gene count matrix. Gene count matrices are normalized as counts per median (CPM):(1)Xnorm=x∑(x)⋅medium(∑(x))where X is the gene count matrix, sum(X) is the read sum for each cell, and median(sum(X)) is the median of read sums. Normalized data are added with a pseudocount of 1 and log transformed (e.g. log₂(X_norm+1)). To determine the detection limit (e.g. 50% detection rate), we modeled the detection probability of ERCC standards with a logistic regression as a function of spike-in amount (Svensson et al., 2017).

We calculated an estimate of absolute molecular abundance for all genes by fitting a linear regression to ERCC spike-ins:(2)log2(y)=m⋅log2(Xnorm+1)+bwhere X_{norm, ERCC>0.5} is the observed CPM value for ERCC spike-ins above the detection limit, Y is the amount of ERCC spike-in, m is the regression coefficient, and b is the intercept. To reduce the influence of measurement noise, we fit the model only to ERCC spike-ins with mean expression above the detection limit.

To determine the transcriptional variation associated with cell cycle, we applied Self-Assembling Manifolds (SAM) (Tarashansky et al., 2019) to filter for highly dispersed gene sets (>0.35 SAM weights) in asynchronous RH population. Principal components analysis (PCA) is then applied to the filtered and normalized RH data, and the nearest neighbor graph (K = 50) is computed using ‘correlation’ as a similarity metric. We identified the putative ‘G1’ clusters with 1N based on DNA content stain. Parasites in ‘G1’ cluster are further sub-clustered with Louvain Clustering, in which we identified ‘G₁a’ and ‘G₁b’ clusters with distinct transcriptional profiles. Pearson correlation between single-cell and bulk transcriptomic data is computed between bulk assignment (Behnke et al., 2010) and the scRNA-seq cluster assignment through which each cluster is uniquely assigned with a cell cycle state. To quantify genes that are differentially expressed across cell cycle clusters, we applied Kruskal-Wallis test. Genes are considered differentially expressed if their p-values are less than 0.05 and they are at least 2-fold over-expressed in a cluster compared to the average expression level of other clusters. We computed differential expression across all cell cycle clusters as well as between the ‘G₁a’ and ‘G₁b’ clusters; the results are uploaded as Supplementary files 2 and 3, respectively. To enable cell cycle assignment transfer from RH to Pru and ME49 data, we implemented a random forest classification model trained on RH data. Briefly, this is done by training a model with 1000 estimators on L2-normalized RH expression data containing only cell cycle associated genes in a 60–40 split scheme. Then the model is applied to predict cell cycle labels of L2-normalized Pru or ME49 data containing the homologous cell cycle associated genes. The testing accuracy was over 95%.

Pseudotime analysis is conducted with Monocle two package in R on preprocessed dataset with highly dispersive genes as described previously. A cell in ‘G₁a’ is designated as the root cell, and all other cells are placed after this cell in order of their inferred pseudotime. To cluster genes based on their pseudotime expression pattern, high frequency patterns are removed through a double spline smoothing operation. The interpolated expression matrix is then normalized by maximum expression along pseudotime such that the maximum value of gene expression along pseudotime is bound by 1. We then applied agglomerative clustering on this interpolated and normalized expression matrix using ‘correlation affinity’ as similarity metric and ‘average linkage’ method to predict three distinct clusters of genes.

To quantify the dependence of expression variation on a two-dimensional projection, we developed a novel approach based on k-nearest neighbor (KNN) averaging. First, a KNN graph is computed by locating nearest neighborhood in a projection using euclidean distance. We then generated a null expression matrix by shuffling the gene expression matrix along each cell column, such that its correlation with respect to the coordinate on projection is completely lost. Next, we compute an updated gene expression value by taking the average of expression values across the KNN. This is equivalent to:(3)XKNN=Mk⋅Xnormwhere X_KNN is the updated KNN averaged expression, M is the nearest-neighbor graph with k being the number of nearest neighbor, and X is the log-transformed CPM of observed or null expression matrices. We chose a k of 5 for all our analysis as varying k did not have a large effect on the results (data not shown). In our experiments, we have shown that the first two principal components (PCs) of PCA on RH correspond to the projection of cell cycle progression, and a two-dimensional UMAP projection of Pru corresponds to asexual development and cell cycle progression. We thus computed X_KNN for both the original, observed expression matrix and the shuffled, null matrix on either projection to reflect dependence on cell cycle progression and/or asexual development. X_KNN is further normalized to have identical sum as the original expression values. A Kolmogorov-Smirnoff two sample test is then computed between the normalized X_KNN of the observed matrix and that of the shuffled matrix based on 100 random permutations. The projection-dependence score for each gene is then computed as:(4)Sg=−log(p¯g)where S_g is the projection-dependence score for gene g and p¯g is the average p-values of 100 tests. We present S_g normalized by the maximum score within each respective data set.

To integrate scRNA-seq data of Plasmodium berghei from Malaria Atlas (Howick et al., 2019) with our Toxoplasma Pru dataset (measured both induced and induced population in 384-well), we first identified one-to-one orthologous genes obtained from PlasmoDB (https://plasmodb.org/) and ToxoDB (https://toxodb.org/toxo/). Next, we filtered each dataset with the ortholog genes. Using scanpy library, we filtered for the intersect of genes that are within the top 800 most dispersed genes in each dataset, resulting in a list of 403 genes. Finally, we used Scanorama (Hie et al., 2019) with default parameter settings to integrate the two datasets. We calculated co-clustering similarity as follows. We first computed a Leiden (Traag et al., 2019) clustering on the integrated graph and returned a cluster co-occurrence matrix to the original cluster assignment in Plasmodium berghei (‘ShortenedLifeStage4’) or Toxoplasma Pru (‘cell_cylcle’). Then, the dot product between the two matrices was calculated and normalized such that it has a maximum of one.