Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2019 Sep 12.
Published in final edited form as: Dev Biol. 2014 Jul 19;393(2):236–244. doi: 10.1016/j.ydbio.2014.07.009

Identification of genes driving lineage divergence from single-cell gene expression data in C. elegans

Chen Xu 1, Zhengchang Su 1,*
PMCID: PMC6742513  NIHMSID: NIHMS843943  PMID: 25050933

Abstract

The nematode Caenorhabditis elegans (C. elegans) is an ideal model organism to study the cell fate specification mechanisms during embryogenesis. It is generally believed that cell fate specification in C. elegans is mainly mediated by lineage-based mechanisms, where the specification paths are driven forward by a succession of asymmetric cell divisions. However, little is known about how each binary decision is made by gene regulatory programs. In this study, we endeavor to obtain a global understanding of cell lineage/fate divergence processes during the early embryogenesis of C. elegans. We reanalyzed the EPIC data set, which traced the expression level of reporter genes at single-cell resolution on a nearly continuous time scale up to the 350-cell stage in C. elegans embryos. We examined the expression patterns for a total of 131 genes from 287 embryos with high quality image recordings, among which 86 genes have replicate embryos. Our results reveal that during early embryogenesis, divergence between sister lineages could be largely explained by a few genes. We predicted genes driving lineage divergence and explored their expression patterns in sister lineages. Moreover, we found that divisions leading to fate divergence are associated with a large number of genes being differentially expressed between sister lineages. Interestingly, we found that the developmental paths of lineages could be differentiated by a small set of genes. Therefore, our results support the notion that the cell fate patterns in C. elegans are achieved through stepwise binary decisions punctuated by cell divisions. Our predicted genes driving lineage divergence provide good starting points for future detailed characterization of their roles in the embryogenesis in this important model organism.

Keywords: C. elegans, Cell fate, Lineage divergence, Embryogenesis, Classification

Introduction

A central goal in developmental biology is to understand the molecular mechanisms of cell fate determination during embryogenesis. The nematode Caenorhabditis elegans displays an essentially invariant cell lineage during embryogenesis and gives rise to a constant number (558) of cell nuclei in the newly hatched lava (Sulston et al., 1983). This well-characterized developmental architecture has served as an excellent model system to study the relationship between cell lineage and fate. Cell fate specification generally depends on both lineage-based and spatiotemporal context-based mechanisms (reviewed in Labouesse and Mango (1999) and Maduro (2010)). In C. elegans, the spatiotemporal context for each embryonic cell is uniquely specified by a lineage history defined as a series of divisions leading to the cell. Over the last few decades, numerous efforts have been made in understanding how these two mechanisms are involved in driving the embryo patterning of C. elegans (Bowerman et al., 1997; Broitman-Maduro et al., 2006; Edgar et al., 2001; Good et al., 2004; Hunter and Kenyon, 1996; Liu et al., 2009; Maduro and Rothman, 2002; Robertson et al., 2004).

In principle, the functional state of a cell is determined by the expression of a specific combination of genes (Bertrand and Hobert, 2010). During the embryogenesis of C. elegans, the dynamic expression of regulatory genes results in a regulatory cascade, which drives the differentiation pathway from the zygote to a terminally differentiated cell through a succession of temporary functional states (Bertrand and Hobert, 2010; Maduro, 2010). Often the transitions of functional states are coupled to developmental decisions being made about each cell cycle and are closely related to asymmetric divisions directed by the binary cell fate decision systems (Baugh, 2003; Cowing and Kenyon, 1996; Edgar and McGhee, 1988; Kaletta et al., 1997). A general binary decision mechanism adopted by C. elegans is the Wnt/β-catenin (SYS-1) pathway (Mizumoto and Sawa, 2007; Phillips and Kimble, 2009). This pathway can cooperate with lineage-specifying factors to transit the functional states forward to the next layer (Bertrand and Hobert, 2010). For example, the GATA factors med-1/2 are first activated by the maternal factor SKN-1 in EMS to determine the fate of both the E and MS lineages (Bowerman et al., 1992). Because SYS-1 is asymmetrically enriched in the posterior nucleus compared to the anterior nucleus, the POP-1 level is low in E but high in MS. In the E cell, both end-1 and end-3 are transcribed under the combined control of SKN-1, the POP-1/SYS-1 complex, and MED-1/2, which ultimately leads to the intestinal differentiation. On the other hand, in the MS cell, a high level of POP-1 allows mesoderm development by repressing the endoderm promoting end-1/3 genes (Lin et al., 1998; Maduro et al., 2005).

To elucidate developmental processes in C. elegans, emerging studies started to characterize the expression profiles of individual progenitors and differentiated cells. For example, the transcriptomes of a few early-stage C. elegans blastomeres have been recently reported using single-cell RNA-seq techniques (Hashimshony et al., 2012). However, enormous technological development is still needed before the transcriptomes of every blastomere can be sequenced. Alternatively, Waterston and colleagues have developed a method to trace the cell lineages of embryos expressing fluorescent reporters, enabling the quantification of expression levels of a certain gene in individual cells on a nearly continuous time scale (Bao et al., 2006; Murray et al., 2008). Using this method, these authors have recently measured the expression levels of more than 100 genes, mostly transcription factors (TFs), at single-cell resolution during embryogenesis up to the 350-cell stage (Mace et al., 2013; Murray et al., 2012). This comprehensive dataset at high temporal and spatial resolution should provide a great opportunity to connect gene expression with lineage specification in a systematic manner. Although significant results have been drawn in the original study, in this paper we explored this dataset by a new computational method from a different perspective. Although the number of genes measured is far from complete to describe the entire embryogenesis, we found that a few of these genes are sufficient to explain the divergence of many sister lineages. Our analysis revealed many known and novel candidate genes driving lineage and fate divergences in early blastomere divisions. Intriguingly, using a small sample of these identified genes, we were able to accurately differentiate the developmental paths leading to later lineages starting from a certain embryo stage. Therefore, our study provides new insights into the regulatory mechanisms governing lineage and fate divergence, and facilitates elucidating the functions of these genes in C. elegans embryogenesis. For convenience of discussion in this paper, we refer to a pair of lineages with a common mother cell as ‘sister lineages’, and relate them to their mother cell by calling them ‘daughter lineages’. For example, ABal yields two sister blastomeres after its division: ABala and ABalp, which are the ancestor cells of the ABala and ABalp lineages, respectively. Therefore, the ABala and ABalp lineages are a pair of sister lineages and they are daughter lineages of ABal.

Material and methods

Data processing

The expression data of a total of 141 genes in 345 embryos were downloaded from the Expression Patterns in C. elegans (EPIC) database (http://epic2.gs.washington.edu/Epic2/) (Mace et al., 2013; Murray et al., 2012) and the expression levels normalized by the ‘local normalization’ method (Murray et al., 2012) were used. To pursue a high confidence for the results, we excluded the cells with fewer than half-genes (70 out of 141) measured in the dataset (mostly the cells appear in embryos later than the 350-cell stage), which resulted in 735 conceptual embryonic cells. We used the median expression level over all time points measured for a cell to represent the expression level of the gene in the cell.

We defined that a cell expresses a gene if its expression level exceeds the background level (defined as 2000 units in the original paper). We excluded from the analysis the embryos having 5 or less than 5 cells expressing the tagged gene. For a gene with replicate embryos, we excluded embryos with inconsistent expressions (Supplemental Table S1). Specifically, we used a k-means clustering method to partition embryos with the same tagged gene in two clusters (k=2) based on the number of cells expressing the gene in an embryo. If the centroid of one cluster is smaller than 0.3 fold of the other, we excluded all embryos in the former cluster.

There is a lag time between the onset of fluorescence and the expression commitment of the corresponding gene, since it takes a few cell cycles for GFP or mCherry to mature before they emit fluorescence (Murray et al., 2012). This may affect the ability and sensitivity to identify differentially expressed genes. It has been shown that the histone::mCherry reporter protein is highly stable with a lifetime longer than a few cell cycles (Murray et al., 2012). Thus, its fluorescence persists even if the corresponding native gene is turned off. We took advantage of this fact to overcome the artifacts caused by the delay of fluorescence. Specifically, for each cell and gene, we used the median of the reporter levels in the cell and all its descendants to represent the expression level of the gene in the cell. The modified gene expression levels may better reflect a gene’s actual onset time and expression levels. In addition, the method may reduce the effect of inconsistent gene expression caused by technical variability. Another concern is that there are many expression values below the background level (2000) in the data set, which could cause false positives in selecting differentially expressed genes. Therefore, we reassigned them to 2000 to keep the consistency and continuity of the data (Supplemental Table S2).

Wilcoxon rank sum test and machine-learning classifiers

We used a combination of statistical and classification methods to identify sets of genes that best discriminate a pair of sister lineages. We excluded from this study the divisions leading to one or both sister lineages having fewer than 15 cells with expression data recorded, resulting in a total of 40 divisions up to the sixth round divisions starting from the zygote. The AB/P1 sister lineage pair is also excluded because the reporters of many ubiquitously expressed genes are not expressed in the P1 lineage. First, we performed a non-parametric test, Wilcoxon rank sum test (equivalent to Mann–Whitney U test), to select genes differentially expressed in two sister lineages at a division. To minimize the effect of variations between individual embryos and gene expression noise, we applied a strict selection criterion: a gene is selected for a division only if in all replicate embryos the gene has significantly different (p-value < 0.05) expression levels between the two sister lineages. The selected genes are referred to as informative genes in this paper.

Second, we employed machine-learning methods to further refine the selected informative genes that potentially drive lineage divergence. To this end, we computed the mean expression level for a gene in the same conceptual cell across replicate embryos, to represent the expression level of the gene in the conceptual cell. We employed a decision stump (Iba and Langley, 1992) to calculate the classification error rate for each informative gene in separating a pair of sister lineages. Next, we selected the genes with error rates smaller than 0.15 as important genes of a division. For each important gene, we assigned it to one of the two sister lineages in which it has a higher expression level than the boundary point (possible boundary location of two classes detected by the decision stump). To see how sister lineages could be distinguished by a group of genes collectively, we performed random forest (Breiman, 2001) using the TreeBagger command in MATLAB (R2013b). Before creating ensemble of bagged decision trees, we fixed the initial random seed. For each classification, a random forest was grown on 50 trees. The default setting of TreeBagger for classification was used. The minimal leaf size was set to 1 and the square root of total number of genes was selected for each split at random. We calculated the Out-of-Bag (OOB) error to get an unbiased estimate of the classification error.

Information content reduction

To quantify the fate constitution of a blastomere, we counted the number of its terminal descendants that fall into each fate category and calculated an information content (IC) for the blastomere defined as IC=−∑k=19(P(k)ln(P(k))), where k stands for each fate category and P(k) denotes the fraction of terminal cells within category k. The following nine fate categories were used for the analysis: glia, hypodermal, intestinal, body wall muscle, pharyngeal, arcade, rectal, seam, and neuron (except neurons in pharyngeal). Each terminal cell can only be assigned to at most one category; hence, the categories are disjoint. The degree of asymmetry of a division was quantified by the averaged amount of IC reduced in two daughter cells compared to the mother cell.

The relationship between important genes and the developmental history of lineages

Given a lineage tree derived from a root cell (e.g., AB), we want to compare the developmental paths for lineages that are generated after N (e.g. N=5) rounds of divisions, i.e., the sub-lineages starting from the (N+1)-th (e.g. 6th) level in the tree (e.g. ABalaaa, ABalaap, ABalapa, ABalapp, etc.). To achieve this goal, we first defined the developmental path from the root to a target lineage as a gene expression rule constituted by a series of binary decisions. At a division M along a path, we selected one gene gi with the minimum error rate in classifying the two sister lineages and identified the classification boundary (value Vi). We assigned an IF function Ei=IF(gi ≤ Vi) or Ei=IF(gi ≥ Vi) to each sister lineage according to its expression level of gi relative to Vi. The IF function returns TRUE when the specified condition is met. For a path L consisting of N divisions leading to a target lineage at the (N+1)-th level, we constructed a Boolean algebra RL by serially applying the logical operator AND to connect all the Ei functions: RL=(E1 ∨ E2 ∨ … ∨ EN), which returns TRUE only when all the Ei functions evaluate to TRUE. As a result, each of the 2N target lineages has a corresponding rule R. To reveal how different the 2N paths are from each other in generating descendant lineages, we scored each target lineage T by each R, and compared the scores. To score a target lineage T by RL, each cell in T is judged by RL and the final score of T on RL is equal to the number of cells that fail to follow RL. Hence, the score could be any integer between zero and the number of cells in T. In this way we converted the task of comparing paths to a well-known assignment problem. We solved it by adopting the Hungarian algorithm (MATLAB implementation), which finds the minimum weight matching of a bipartite graph and meanwhile optimizes the assignments by minimizing the sum of scores. Then from the assignment results, we could conclude how similar or diverse the paths are in defining target lineages.

Results

In principle, a lineage specification process is driven by a regulatory cascade that diversifies daughter cells at each division along the lineage path and ultimately leads to a complete cell differentiation and fate determination. In this work, we attempt to understand how the cell fates are conferred by a series of blastomere divisions by identifying candidate genes that potentially drive lineage divergence. Toward this goal, we identified putative determinant genes at every cell division along a developmental time-line. Specifically, we looked for genes that could best explain the divergence between two sister lineages yielded at each of the selected 40 divisions.

An effective quality control method facilitates the analysis of single-cell gene expression data

All the analysis in this work are based on the EPIC data set (Mace et al., 2013; Murray et al., 2012), which were collected by tracing gene expression levels from 345C. elegans embryos, each was constructed to express a reporter for a specific gene (Supplemental Table S1). Among the 141 reporter genes, 42 were only recorded in a single embryo without replicates; the rest 99 genes were measured in more than one (2–16) embryo. Expression levels of replicates are highly concordant, but for some genes there still exists variability. The original paper examined the consistency of gene expression levels from reporters that were analyzed in multiple embryos, in regard to the number of strongly expressing cells (Murray et al., 2012). They found that most reporters (~80%) were highly concordant. For the remaining reporters, the variability was largely due to the replicate embryos constructed from different strains, where one strain was overall less bright than the other. Alternatively, the onset of the reporters was near the end of the embryo stage; thus the detection was less reliable. In addition, some genes in EPIC were measured in two reporter constructs (promoter fusion and protein fusion). For most genes, the expression patterns of the two different reporters are similar but not identical (Murray et al., 2012). Since expression levels from different embryos are inevitably variable owing to the stochastic nature of gene expression and technical variability, it is essential to control the data quality and minimize the technical variation before a meaningful conclusion can be drawn on the gene expressions.

First, we intended to identify and exclude embryos which are from strains or constructs that behave differently and are much less bright than other replicates. The measurements of reporter intensity from these embryos would not be reliable. In addition, since we used a rigid background level (2000 intensity units) for all embryos, replicates with different overall reporter intensity should not be considered together. Therefore, using the quality control procedure detailed in Materials and methods, we filtered out the embryos with very few cells expressing the gene (>2000 intensity units) or significantly less cells expressing the gene compared to replicate embryos (Supplemental Table S1). As a result, 58 embryos from 31 genes were filtered out; 287 embryos from 131 genes were included for further analysis. Among the 131 genes, 86 genes have two to nine replicate embryos; the rest 45 genes have only a single embryo included, mostly (35/45) because of a lack of replicate in the original data set. Although we did not exclude these 45 genes with a single recording from further analysis for completeness, we clearly labeled them throughout the following results to bring them to notice. Among the 58 embryos filtered out, 24% have no cell expressing; 50% have less than 10 cells with expression over the background level; the rest embryos have moderate number of cells expressing the gene but significantly less than other replicate embryos, mainly because of different constructs or strains used. It is worth noting that using an unsupervised clustering method could differentiate strains or constructs that act differently. For example, lin-26 was tagged in four embryos, each with 8, 15, 268 and 297 cells expressing the gene (level > 2000). The k-means clustering partitioned the four embryos into two clusters, [8, 15] and [268, 297]. Since the centroid of the former cluster (11) is significantly smaller (<0.3 fold) than that of the latter cluster (282), both embryos in the former cluster were filtered out. In fact, embryos in the former cluster were constructed from a different strain than the embryos in the latter cluster.

We next excluded genes with inconsistent expression patterns among replicate embryos and meanwhile selected informative genes that likely account for lineage divergence at a division. For each of the 40 divisions, we identified genes that are significantly differentially expressed by two sister lineages using the Wilcoxon rank sum test in each embryo; we only considered a genes as an informative gene if all replicate embryos reach the significance level (p-value < 0.05) at the division. As summarized in Supplemental Table S3, for most divisions, majority of the 131 genes do not satisfy this selection criterion and thus are excluded; the number of informative genes ranges from 15 (ABarppa/ABarppp) to 79 (EMS/P2) for a division. The sample size (number of cells in a lineage) positively correlates with the number of informative genes (ρ=0.60), which is expected since a larger sample size would make it easier to reach the significance level. This additional filtering procedure ensures that the informative genes selected for a division are the genes with high reproducibility and consistent differential expression patterns (p-value < 0.05) among all its replicates at the division.

Divergence of sister lineages is coupled with cell divisions

During C. elegans embryogenesis, developmental pathways are largely directed by asymmetric cell divisions, which give rise to two sister lineages with diverse fates (Baugh, 2003; Cowing and Kenyon, 1996; Edgar and McGhee, 1988; Kaletta et al., 1997). The divergence is executed and maintained by regulatory programs that are manifested as differential expression patterns between two sister lineages. To see whether the divergence of sister lineages could be explained by the selected informative genes, we adopted a supervised classification method, random forest, at each division. Here the classification objects are binary labeled cells according to which of the two sister lineages they belong to, and the features are the gene expression levels. As shown in Fig. 1A, the error rates of classifications for the 40 pairs all drop rapidly as the number of decision trees increases and are eventually stabilized near zero. By contrast, when the class labels on the cells are randomly permuted, the random forest with the same settings fails to classify the two lineages, as its performance is no better than random guessing. In the case of the ABarppa/ABarppp sister lineages, the classification on permuted data yields an error rate around 0.5 and could not be brought down when more trees are grown (Fig. 1B). To verify our criterion for selecting informative genes, we compared the classification results using all of the genes (131) to those using only the selected informative genes. As shown in Supplemental Table S4, although the informative genes selected for each division are fewer, they generally have at least the same classification power as all the 131 genes. For example, only 15 out of the 131 genes are selected as informative genes for the ABarppa/ABarppp sister lineages. However, the error rate drops even faster using the 15 informative genes than using all the genes (Fig. 1B). These results support the model that the transition of gene expression profiles is made about each cell cycle through asymmetric divisions in C. elegans. Moreover, our selection criteria can largely identify genes that likely contribute to the divergence of sister lineages.

Fig. 1.

Fig. 1.

Open in a new tab

The classification error rate changes as the number of decision trees increases in the random forests. (A) The Out-of-Bag (OOB) error rates as a function of the number of decision tree grown in the random forests for classifying the 40 pairs of sister lineages. (B) The random forest can successfully differentiate ABarppa/ABarppp sister lineages using either all the 131 genes or only the 15 informative genes selected by the Wilcoxon rank sum test. However, the classification fails when class labels on cells are permutated, as the OOB error rate fluctuates around 0.5 even when 50 trees are grown.

Divergence of sister lineages can be explained by a few important genes

To further narrow down the genes explaining the divergence of a pair of sister lineages, we evaluated each informative gene for its capability to separate a pair of sister lineages. We considered an informative gene as an important gene if it could discriminate a pair of sister lineages with an error rate less than 0.15. We assigned each important gene to one of the two sister lineages in which it has a higher expression level than the boundary point detected by the decision stump (Supplemental Table S5). About 40% of the informative genes are selected as important genes; the number of important genes identified for a division ranges from 0 (ABal/ABar and ABpl/ABpr) to 53 (MS/E). The inability to identify important genes for ABal/ABar and ABpl/ABpr is understandable as ABa and ABp are precursors of analogs, each producing a group of approximately equivalent cells that result in bilateral symmetry in the nervous system (Sulston, 1983). Remarkably, 108 (82.4%) of the 131 genes in our analysis are identified as important genes for at least one sister lineage pair. On average, a gene is identified as an important gene for about five divisions; about 10% of the genes are identified as important genes in more than 10 divisions. For example, tbx-11 is identified as an important gene for 17 pairs of sister lineages (Supplemental Table S5).

To illustrate how the important genes contribute to discriminating a sister lineage pair, we conducted principal component analysis (PCA) on the expression levels of the important genes (Supplementary Fig. S1). Shown in Fig. 2A is the case for the ABala/ABalp sister lineages, in which 11 of the 43 informative genes are identified as important genes. Intriguingly, the cells from the ABala/ABalp pair are well separated into two clusters by the first and second principle components (PCs) of the 11 important genes (Fig. 2A). In contrast, when the rest 32 genes with error rates over 0.15 are used for PCA, the resulting boundary between the ABala and ABalp lineages is much less clear (Fig. 2B). In addition, as indicated earlier, although the sample size correlates positively with the number of informative genes selected by the Wilcoxon rank sum test (ρ=0.60), it does not show such correlation with the number of important genes selected by decision stump (ρ= −0.21).These results suggest that decision stump is able to effectively identify genes making the biggest contribution to separating sister lineages, thereby accurately refining the selection of genes that potentially drive the lineage divergence.

Fig. 2.

Fig. 2.

Open in a new tab

Few important genes can explain the divergence between sister lineages. (A) Projections of cells in the ABala/ABalp lineages on the first (x-axis) and second (y-axis) principle components of the 11 important genes with error rates below 0.15. Cells are colored according to their lineage origin (ABala: red, ABalp: blue). (B) PCA projections using the rest 32 informative genes with error rates over 0.15.

Importantly, as shown in Supplemental Table S5, many of the important genes identified for discriminating sister lineages are in excellent agreement with their known functions. We illustrate this using a few well-studied examples. First, we have identified med-2 and pal-1 to be important genes in discriminating the EMS/P2 sister lineage pair. Based on the expression level relative to the boundary point, we assigned med-2 to the EMS lineage and pal-1 to the P2 lineage. Consistent with these predictions, it has been shown that in the P2 lineage, the C and D fate specifications require the CAUDAL-like TF PAL-1 for body wall muscle and hypodermal development (Edgar et al., 2001; Hunter and Kenyon, 1996). On the other hand, the body wall muscle cells derived from the MS cell in the EMS lineage do not depend on the activation of pal-1 but med-1/2, which are necessary and sufficient to program mesendoderm development (Maduro et al., 2001). Moreover, in med depleted embryos, EMS descendants adopt the C fate. Hence, our results are consistent with these early findings that the differentiation program of P2 is driven by pal-1 whereas that of EMS is driven by med-2. However, since med-1 is not measured in the original study, it is not included in our analysis here. Second, we identified the med-1/2 target genes end-1/3 and tbx-35 as important genes in discriminating the MS/E sister lineage pair, and assigned tbx-35 to the MS lineage and end-1/3 to the E lineage. In consistent with these predictions, it has been shown that end-1/3 are the earliest expressed genes in the E lineage and the END-1/3 regulation defines the separation of the E lineage from its sister lineage MS by contributing to the intestinal fate commitment (Zhu et al., 1997). On the contrary, tbx-35, a T-box TF, is repressed in the E lineage but activated in the MS lineage to specify the MS-derived pharynx and body wall muscle fates (Broitman-Maduro et al., 2006). Third, it has been reported that TBX-35 acts through regulators PHA-4 and HLH-1 to specify pharynx and muscle fates, respectively (Gaudet and Mango, 2002; Krause et al., 1990; Lei et al., 2009). PHA-4 first appears close to the time point at which a pharyngeal clonal forms and then present in all pharyngeal cells derived from the ABalp, ABara, MSaa and MSpa lineages (Horner et al., 1998; Kalb et al., 1998). HLH-1 is a potent myogenic factor whose expression is detected in the C, D and MS blastomeres and descendants (Fukushige et al., 2006; Krause et al., 1990). Our results are in good agreement with these early findings, as we assigned pha-4 to the ABalpa, ABara, MSaa and MSpa lineages which are all precursors of the pharynx tissue. Besides, hlh-1 was assigned to MSap and Cpp lineages which are both muscle clones. Fourth, in addition to hlh-1, we also identified another previously described body wall myogenic factor gene hnd-1 (Fukushige and Krause, 2005) as an important gene in the Cpp, MSap and MSpp lineages, which are developmental clones for body wall muscles. Fifth, previous studies have characterized nhr-25 as a hypodermal tissue marker (Asahina et al., 2000; Moore et al., 2013) and we assigned it to the Cpa lineage, a clone fated to be hypoderm. Finally, we found tbx-37/38 as important genes in segregating the ABa lineage from the ABp lineage, which is in agreement with their known roles in mesodermal induction for specifying the pharynx fate in the ABa lineage (Good et al., 2004). In conclusion, our results indicate that the important genes picked by our algorithm are likely to play a role in the lineage specification process; their divergent expression levels in two sister lineages might activate distinct transcription programs that drive further fate specification processes.

Effective lineage classifications are associated with fate divergences

In essence, gene expression has a high correlation with cell fate, as terminally differentiated cells must express determinant/signature genes to fulfill certain functions. It would be interesting to evaluate the relationships between gene expression and fate divergences. We began by determining the degree of asymmetry at each division, since an asymmetric division that produces two daughter lineages with different developmental potentials is a general mechanism for fate specification. We computed an information content (IC) for each embryonic cell based on the tissue type constitution of its terminal descendants (Supplemental Table S6). Given that the IC could capture the differentiation potential of a cell, the level of IC reduction in the daughter cells relative to their mother cell is a measurement of the degree of asymmetry of the division. For example, Cp generates eight hypodermal cells and 16 body wall muscle cells, having an IC of 0.64 nats. Its daughter cells, Cpa which generates only hypodermal cells (eight), and Cpp which generates only body wall muscle cells (16), both have an IC of 0 nats. This significant reduction of IC from 0.64 to 0 is consistent with the fact that the division is highly asymmetric. In addition, the 0 IC of the daughter cells reflects the fact that both of them become clonal producing precursor cells of single fate. In contrast, the E cell (produces 20 intestinal cells) undergoes a symmetric division, as its daughter cells Ea and Ep carry no difference in their progeny constitution (each produces 10 intestinal cells). In agreement with the symmetric division, there is no IC reduction in both of the daughter cells (E, Ea and Ep all have an IC 0). Therefore, the IC reduction can quantify the degree of asymmetry for a division.

We further found that the number of important genes of a sister lineage pair is positively correlated (ρ=0.49) with the mean IC reduction in its daughter cells (Fig. 3). The fact that many genes are identified to be important genes at some divisions is associated with functional asymmetry of the sister lineages. For example, among all of the 40 classifications, MS/E has the largest number of important genes (53). In fact, MS and E sister lineages are derived from a highly asymmetric division and the mean IC reduction (0.71) is the highest among all the 40 divisions. By contrast, the fact that few genes are identified to be important genes at some divisions is associated with functional symmetry of sister lineages. The divisions with the lowest number of important genes identified (0–3) are the ones that yield functionally symmetric lineages, such as MSa/MSp, Ca/Cp, ABpl/ABpr, and ABal/ABar. In the 13 cases where the numbers of important genes are below 10, the divisions have relatively low mean IC reductions (0.12 in average), while in the remaining 27 cases with equal to or more than 10 important genes, the divisions have relatively high mean IC reductions (0.20 in average). These results indicate that the important genes identified at asymmetric divisions might play roles in fate specifications. However, as shown in Fig. 3, there are also some relatively symmetric divisions having many important genes. In these cases, the important genes might act as upstream regulators functioning before the asymmetric divisions take place and the fate divergence can be observed.

Fig. 3.

Fig. 3.

Open in a new tab

Number of important genes is positively correlated with the degree of asymmetry of a division. The x-axis is the number of important genes selected at a division. The y-axis is the mean IC reduction of the division.

Lineage identities can be uniquely defined by a few important genes

It is generally believed that lineage specifications in C. elegans are largely achieved by a series of stereotyped cell divisions from the zygote through to terminally differentiated cells. We hypothesize that the specification/differentiation pathways are uniquely defined for each lineage, in which a cascade of binary decisions made at ancestral nodes/cells collectively define and restrict the characters of descendant cells. If this is the case, we wonder what genes are required to uniquely distinguish the developmental paths leading to descendant lineages. We addressed this question using the AB lineages, as we have learned the important genes at all the divisions up to the fifth round starting from AB (e.g., the divisions of AB, ABa, ABal, ABala and ABalaa). We compared the developmental paths of the 32 lineages derived from the fifth round of AB divisions (i.e., the lineages of ABalaaa, ABalaap, ABalapa, etc., and we refer to them as the AB32 lineages hereafter). Specifically, for each developmental path from AB to an AB32 lineage, e.g., the ABalaaa lineage, we defined a corresponding gene expression rule consisting of binary decisions of one important gene (with the lowest classification error rate) at each of the five divisions along the path. For example, the path from AB to ABalaaa is collectively defined by the expression patterns of the five important genes at their respective divisions, i.e., divisions of AB, ABa, ABal, ABala and ABalaa. Because some genes were repeatedly identified as important genes at different divisions, we only used a total of 14 important genes, all measured and analyzed in multiple embryos, to define all the 32 rules (Supplemental Table S7). We scored each AB32 lineage by each rule to see how well the 32 lineages could be differentiated from one another by the scores using the Hungarian assignment algorithm (see Materials and methods). Remarkably, the lineages are all correctly assigned to their corresponding paths. This finding demonstrates that the combined expression patterns of a small set of genes are sufficient to distinguish developmental paths starting from a common ancestor cell.

Discussion

In this study, we reanalyzed the EPIC data set containing expression levels for 141 genes measured in individual cells in C. elegans embryos up to the ~350-cell stage. Although the relatively small number of genes and short period of monitoring time are clearly not sufficient to explain all developmental details occurring during embryogenesis, this large-scale dataset provides a substantial resource for analyzing the relationships between gene expression and lineage specification in early embryogenesis. However, careful data processing must be taken in order to draw meaningful conclusions due to the unavoidable technical artifacts in the data set. Specifically, since there is a time lag of about 30 min for a detectable fluorescence after the expression commitment (Murray et al., 2012), a modification of the expression levels is necessary to overcome the delay. Considering the fact that the fluorescent proteins are highly stable after being produced, and that newly produced cells do not grow in size during embryogenesis, the concentration of a reporter in a cell is largely the same as in its descendant cells if divisions are symmetric. On the other hand, if divisions are not necessarily symmetric, the concentration of the reporter in the cell is roughly the median of the reporter levels in the descendant cells. Thus we used the median level of the reporters in the cell and all its descendants to infer the concentration of the corresponding gene in the cell. In principal, using the median could also reduce the effect of technical variability in the measurements of fluorescence levels. Indeed, we found that this modification largely enhanced the sensitivity to identify differentially expressed genes by the Wilcoxon rank sum test compared to using the direct measurements (data not shown), suggesting that the modification could largely correct the delay of fluorescence without introducing considerable new artifacts.

Moreover, studies of differential gene expression typically select causal genes according to fold change or test statistic rank. However, they both have limitations for this study. First, in the fold-change method, a gene is considered to be significantly differentially expressed if the ratio of its mean expression level in one sample to that in the other sample exceeds an arbitrary cutoff value (Cui and Churchill, 2003). One drawback using fold-change is the lack of associated statistic indicating the confidence level of the results. Moreover, when genes are expressed at low levels, fold-change suffers from high noise ratio which can result in high false discovery rates (FDRs). It may also fail for highly expressed genes, where small changes may be real but are rejected (Cui and Churchill, 2003; Tusher et al., 2001). Second, since there are only few replicate embryos for most genes in the dataset, to fully utilize the invaluable information in each embryo, we chose not to simply take averages of replicates followed by a ranking of genes by p-value. Instead, we applied statistical hypothesis testing in each embryo, and selected a gene as an informative gene only if all replicates reach the significance level. This reduces the effect of variations, thereby ensuring genes selected have a high consistency of expression patterns among replicate embryos. However, we wanted to only consider genes that could have sister lineages well separated with a clear boundary. Therefore, we further narrowed down the genes that are likely to drive the divergence of sister lineages according to their classification error rates. We resorted to the feature selection method, decision stump, to identify the important genes for each sister lineage. The task of feature selection is a widely addressed problem, where one has class-labeled data and wants to figure out which features best discriminate among the classes. Here the classes are lineages and the features are gene expression levels. The task is to select sets of genes that can best discriminate sister lineages. After applying an error rate cutoff in gene selection, the number of important genes is no longer correlated with the sample size (the number of cells in lineages). Besides, PCA projections on important genes result in a larger distance between two sister lineages compared to PCA on other informative genes, suggesting that important genes contribute significantly to the divergence of sister lineages. In summary, our results demonstrate that the combination of statistical hypothesis testing and decision stump could lead to better results than what would be resulted by using either one alone.

To evaluate the selected genes for discriminating a pair of sister lineages, we performed classification analyses on the sister lineage pairs using random forest, which is a strong classifier working by growing many decision trees and choosing the classification as the mode of outputs of individual trees (Breiman, 2001). Our successful classification of the sister lineages at each division strongly supports the model that cell lineage specification in C. elegans occurs sequentially through a cascade of binary decisions with each division diverging the daughter cells further, eventually leading to complete differentiations of terminal cells. Moreover, we found that some genes were repeatedly identified as important genes in various divisions. It raises an interesting question of whether their asymmetric expressions could be downstream effects of the transcriptional regulation by the POP-1/TCF, which is part of a general anterior/posterior coordinating system that acts in an iterative manner to differentiate sister cells (Lin et al., 1998). Based on the classification boundary point set by the decision stump, we assigned each important gene to the sister lineage where it shows a higher expression. This allows a further investigation of the correlation between gene expression and lineage fates. It is worth noting that some important genes are clonally expressed in one sister lineage (bold typed in Supplemental Table S5). It would be interesting to reveal their roles in lineage/fate specification by either gain-of-function or loss-of-function experiments.

In C. elegans, multiple lines of evidence support the strong effect of lineages on fate specification (reviewed in Maduro, 2010). Thus, in our study, we sought to associate differentially expressed genes with fate divergent processes. It has been observed that organ/tissue identity genes are active at as early as 50- or 80-cell stage well before any overt cell differentiation and from then on their activities are maintained through adulthood (Labouesse and Mango, 1999; Maduro, 2010). In this regard, although the EPIC is limited to the 350-cell stage and does not include the final round of divisions when complex cell migrations and tissue/organ formations take place, the dataset can be a useful resource to reveal possible genes related to fate specifications. To identify novel genes related to fate divergence, we determined whether or not a division is likely to be asymmetric based on IC reductions in the daughter cells. Indeed, we found a positive correlation between the number of important genes identified and the degree of asymmetry of a division. The result is expectable based on the knowledge of the relationship between gene expression patterns and cell fates that, if majority of the progenies in sister lineages are fated to the same tissue, the gene expression in the two lineages should be similar. The result also endorses our method for identifying important genes, and indicates that the EPIC data set contains many genes related to fate specification processes. Furthermore, we found that many of the important genes selected at divisions with relatively high IC reductions ( > 0.3) are known for their roles in initiating the fate specification processes in the corresponding lineages. We suppose those of unknown functions at such divisions to be novel lineage-specifying regulators and thus warrant further experimental investigations. However, the method is not valid for tissues without a clear lineage relation. One example is the nervous system where the majority of neurons are derived non-clonally from multiple lineages (Hobert, 2005). It is also worth noting that the IC is calculated based on an arbitrary and incomplete categorization of cell types, and the major tissue categories could be further subdivided. For example, pharynx is constituted by multiple types of cells such as neurons, epithelial and muscle cells. Clearly, different categorizations of the cells can lead insights into different aspects of the lineage and fate specification programs.

We modeled the developmental paths as a binary decision tree, with each level of divisions separating the lineages further. In this way, the entire tree can be built into a complete scoring system, where all the internal decisions contribute to the scoring rule for the lineages at the target level. For example, the expression pattern of the ABalaaa lineage (6th level) must follow all the binary decisions made at the division of AB (1st level), ABa (2nd level), ABal (3rd level), ABala (4th level) and ABalaa (5th level). We found that a small number of genes (14) are sufficient to uniquely define all the developmental paths derived from a common ancestral cell. This result is in excellent agreement with the notion that lineage histories play crucial roles in cell fate specifications. In other words, the cell fates are specified by stepwise instructions directed by the binary decisions occurring at each division leading to the cell. We expect that with more single-cell gene expression data available, more biological insights into cell fate specifications can be revealed by similar analyses.

Conclusions

In the classic lineal control model of C. elegans embryo development, most blastomere and terminal identities stem from consecutive binary diversifications (Kaletta et al., 1997). It would be highly valuable to dissect the architecture of regulatory cascades and reveal genes that play essential roles in driving the divergence of two lineages generated at each cell division. The major challenge to the goal is how to develop an effective approach to analyze highly noisy single-cell gene expression data with no or few replicates. Using a combination of careful data processing, non-parametric statistical test and classification methods, we were able to identify potential genes that distinguish sister lineages generated in the early embryogenesis in C. elegans. Intriguingly, we found that only a small set of genes is sufficient to discriminate a pair of sister lineages. With the availability of single-cell expression data for more genes and cells in later embryogenesis, more biological insights into cell lineage/fate specification can be revealed by similar analysis. Such decoding of regulatory architecture during embryogenesis can eventually lead to a comprehensive understanding of the lineage/fate specification processes in embryogenesis.

Supplementary Material

Figure S!
Table S1
Table S2
Table S3
Table S4
Table S5
Table S6
Table S7

Acknowledgments

This work is supported by the UNC Charlotte (Faculty Research Grants 1-11227) and the National Science Foundation (EF0849615 and CCF1048261). We would like to thank members of the Su lab for discussions. We are thankful to Dr. Robert Waterston and Dr. John Murray for assistance in downloading their datasets. We highly appreciate the valuable suggestions from the anonymous reviewers.

Footnotes

Appendix A. Supporting information

Supplementary data associated with this article can be found in the online version at http://dx.doi.org/10.1016/j.ydbio.2014.07.009.

References

  1. Asahina M, Ishihara T, Jindra M, Kohara Y, Katsura I, Hirose S, 2000. The conserved nuclear receptor Ftz-F1 is required for embryogenesis, moulting and reproduction in Caenorhabditis elegans. Genes Cells 5, 711–723. [DOI] [PubMed] [Google Scholar]
  2. Bao Z, Murray JI, Boyle T, Ooi SL, Sandel MJ, Waterston RH, 2006. Automated cell lineage tracing in Caenorhabditis elegans. Proc. Natl. Acad. Sci. USA 103, 2707–2712. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Baugh LR, 2003. Composition and dynamics of the Caenorhabditis elegans early embryonic transcriptome. Development 130, 889–900. [DOI] [PubMed] [Google Scholar]
  4. Bertrand V, Hobert O, 2010. Lineage programming: navigating through transient regulatory states via binary decisions. Curr. Opin. Genet. Dev 20, 362–368. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Bowerman B, Eaton BA, Priess JR, 1992. skn-1, a maternally expressed gene required to specify the fate of ventral blastomeres in the early C. elegans embryo. Cell 68, 1061–1075. [DOI] [PubMed] [Google Scholar]
  6. Bowerman B, Ingram MK, Hunter CP, 1997. The maternal par genes and the segregation of cell fate specification activities in early Caenorhabditis elegans embryos. Development 124, 3815–3826. [DOI] [PubMed] [Google Scholar]
  7. Breiman L, 2001. Random forests. Mach. Learn 45, 5–32. [Google Scholar]
  8. Broitman-Maduro G, Lin KT-H, Hung WWK, Maduro MF, 2006. Specification of the C. elegans MS blastomere by the T-box factor TBX-35. Development 133, 3097–3106. [DOI] [PubMed] [Google Scholar]
  9. Cowing D, Kenyon C, 1996. Correct Hox gene expression established independently of position in Caenorhabditis elegans. Nature 382, 353–356. [DOI] [PubMed] [Google Scholar]
  10. Cui X, Churchill GA, 2003. Statistical tests for differential expression in cDNA microarray experiments. Genome Biol. 4, 210. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Edgar LG, Carr S, Wang H, Wood WB, 2001. Zygotic expression of the caudal homolog pal-1 is required for posterior patterning in Caenorhabditis elegans embryogenesis. Dev. Biol 229, 71–88. [DOI] [PubMed] [Google Scholar]
  12. Edgar LG, McGhee JD, 1988. DNA synthesis and the control of embryonic gene expression in C. elegans. Cell 53, 589–599. [DOI] [PubMed] [Google Scholar]
  13. Fukushige T, Brodigan TM, Schriefer LA, Waterston RH, Krause M, 2006. Defining the transcriptional redundancy of early bodywall muscle development in C. elegans: evidence for a unified theory of animal muscle development. Genes Dev. 20, 3395–3406. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Fukushige T, Krause M, 2005. The myogenic potency of HLH-1 reveals widespread developmental plasticity in early C. elegans embryos. Development 132, 1795–1805. [DOI] [PubMed] [Google Scholar]
  15. Gaudet J, Mango SE, 2002. Regulation of organogenesis by the Caenorhabditis elegans FoxA protein PHA-4. Science 295, 821–825. [DOI] [PubMed] [Google Scholar]
  16. Good K, Ciosk R, Nance J, Neves A, Hill RJ, Priess JR, 2004. The T-box transcription factors TBX-37 and TBX-38 link GLP-1/Notch signaling to mesoderm induction in C. elegans embryos. Development 131, 1967–1978. [DOI] [PubMed] [Google Scholar]
  17. Hashimshony T, Wagner F, Sher N, Yanai I, 2012. CEL-Seq: single-cell RNA-Seq by multiplexed linear amplification. Cell Rep. 2, 666–673. [DOI] [PubMed] [Google Scholar]
  18. Hobert O, Specification of the nervous system (August 8, 2005), WormBook, ed. The C. elegans Research Community, WormBook, 10.1895/wormbook.1.12.1, http://www.wormbook.org. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Horner MA, Quintin S, Domeier ME, Kimble J, Labouesse M, Mango SE, 1998. pha-4, an HNF-3 homolog, specifies pharyngeal organ identity in Caenorhabditis elegans. Genes Dev. 12, 1947–1952. [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Hunter CP, Kenyon C, 1996. Spatial and temporal controls target pal-1 blastomere-specification activity to a single blastomere lineage in C. elegans embryos. Cell 87, 217–226. [DOI] [PubMed] [Google Scholar]
  21. Kalb J, Lau K, Goszczynski B, Fukushige T, Moons D, Okkema P, McGhee J, 1998. pha-4 is Ce-fkh-1, a fork head/HNF-3alpha,beta,gamma homolog that functions in organogenesis of the C. elegans pharynx. Development 125, 2171–2180. [DOI] [PubMed] [Google Scholar]
  22. Iba W, Langley P, 1992. Induction of one-level decision trees, In: Proceedings of the Ninth International Conference on Machine Learning, Aberdeen, Scotland, San Francisco, CA: Morgan Kaufmann, 233–240. (1–3 July). [Google Scholar]
  23. Kaletta T, Schnabel H, Schnabel R, 1997. Binary specification of the embryonic lineage in Caenorhabditis elegans. Nature 390, 294–298. [DOI] [PubMed] [Google Scholar]
  24. Krause M, Fire A, Harrison SW, Priess J, Weintraub H, 1990. CeMyoD accumulation defines the body wall muscle cell fate during C. elegans embryogenesis. Cell 63, 907–919. [DOI] [PubMed] [Google Scholar]
  25. Labouesse M, Mango SE, 1999. Patterning the C. elegans embryo: moving beyond the cell lineage. Trends Genet. 15, 307–313. [DOI] [PubMed] [Google Scholar]
  26. Lei H, Liu J, Fukushige T, Fire A, Krause M, 2009. Caudal-like PAL-1 directly activates the bodywall muscle module regulator hlh-1 in C. elegans to initiate the embryonic muscle gene regulatory network. Development 136, 1241–1249. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Lin R, Hill RJ, Priess JR, 1998. POP-1 and anterior–posterior fate decisions in C. elegans embryos. Cell 92, 229–239. [DOI] [PubMed] [Google Scholar]
  28. Liu X, Long F, Peng H, Aerni SJ, Jiang M, Sánchez-Blanco A, Murray JI, Preston E, Mericle B, Batzoglou S, Myers EW, Kim SK, 2009. Analysis of cell fate from single-cell gene expression profiles in C. elegans. Cell 139, 623–633. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Mace DL, Weisdepp P, Gevirtzman L, Boyle T, Waterston RH, 2013. A high-fidelity cell lineage tracing method for obtaining systematic spatiotemporal gene expression patterns in Caenorhabditis elegans. G3 3, 851–863. [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Maduro MF, 2010. Cell fate specification in the C. elegans embryo. Dev. Dyn 239, 1315–1329. [DOI] [PubMed] [Google Scholar]
  31. Maduro MF, Kasmir JJ, Zhu J, Rothman JH, 2005. The Wnt effector POP-1 and the PAL-1/Caudal homeoprotein collaborate with SKN-1 to activate C. elegans endoderm development. Dev. Biol 285, 510–523. [DOI] [PubMed] [Google Scholar]
  32. Maduro MF, Meneghini MD, Bowerman B, Broitman-Maduro G, Rothman JH, 2001. Restriction of mesendoderm to a single blastomere by the combined action of SKN-1 and a GSK-3β homolog is mediated by MED-1 and −2 in C. elegans. Mol. Cell 7, 475–485. [DOI] [PubMed] [Google Scholar]
  33. Maduro MF, Rothman JH, 2002. Making worm guts: the gene regulatory network of the Caenorhabditis elegans endoderm. Dev. Biol 246, 68–85. [DOI] [PubMed] [Google Scholar]
  34. Mizumoto K, Sawa H, 2007. Two betas or not two betas: regulation of asymmetric division by beta-catenin. Trends Cell Biol. 17, 465–473. [DOI] [PubMed] [Google Scholar]
  35. Moore JL, Du Z, Bao Z, 2013. Systematic quantification of developmental phenotypes at single-cell resolution during embryogenesis. Development 140, 3266–3274. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Murray JI, Bao Z, Boyle TJ, Boeck ME, Mericle BL, Nicholas TJ, Zhao Z, Sandel MJ, Waterston RH, 2008. Automated analysis of embryonic gene expression with cellular resolution in C. elegans. Nat. Methods 5, 703–709. [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Murray JI, Boyle TJ, Preston E, Vafeados D, Mericle B, Weisdepp P, Zhao Z, Bao Z, Boeck M, Waterston RH, 2012. Multidimensional regulation of gene expression in the C. elegans embryo. Genome Res. 22, 1282–1294. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Phillips BT, Kimble J, 2009. A new look at TCF and beta-catenin through the lens of a divergent C. elegans Wnt pathway. Dev. Cell 17, 27–34. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Robertson SM, Shetty P, Lin R, 2004. Identification of lineage-specific zygotic transcripts in early Caenorhabditis elegans embryos. Dev. Biol 276, 493–507. [DOI] [PubMed] [Google Scholar]
  40. Sulston JE, 1983. Neuronal cell lineages in the nematode Caenorhabditis elegans. Cold Spring Harb. Symp. Quant. Biol 48, 443–452. [DOI] [PubMed] [Google Scholar]
  41. Sulston JE, Schierenberg E, White JG, Thomson JN, 1983. The embryonic cell lineage of the nematode Caenorhabditis elegans. Dev. Biol 100, 64–119. [DOI] [PubMed] [Google Scholar]
  42. Tusher VG, Tibshirani R, Chu G, 2001. Significance analysis of microarrays applied to the ionizing radiation response. Proc. Natl. Acad. Sci. USA 98, 5116–5121. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Zhu J, Hill RJ, Heid PJ, Fukuyama M, Sugimoto A, Priess JR, Rothman JH, 1997. end-1 encodes an apparent GATA factor that specifies the endoderm precursor in Caenorhabditis elegans embryos. Genes Dev. 11, 2883–2896. [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Figure S!
NIHMS843943-supplement-Figure_S!.pdf (375.1KB, pdf)
Table S1
NIHMS843943-supplement-Table_S1.xlsx (199KB, xlsx)
Table S2
NIHMS843943-supplement-Table_S2.xlsx (1.3MB, xlsx)
Table S3
NIHMS843943-supplement-Table_S3.xlsx (23.7KB, xlsx)
Table S4
NIHMS843943-supplement-Table_S4.xlsx (10.2KB, xlsx)
Table S5
NIHMS843943-supplement-Table_S5.xlsx (40.9KB, xlsx)
Table S6
NIHMS843943-supplement-Table_S6.xlsx (14.9KB, xlsx)
Table S7
NIHMS843943-supplement-Table_S7.xlsx (8.8KB, xlsx)

RESOURCES