Ultra-deep profiling of alternatively spliced Drosophila Dscam isoforms by circularization-assisted multi-segment sequencing

Abstract

The Drosophila melanogaster gene Dscam (Down syndrome cell adhesion molecule) can generate thousands of different ectodomains via mutual exclusive splicing of three large exon clusters. The isoform diversity plays a profound role in both neuronal wiring and pathogen recognition. However, the isoform expression pattern at the global level remained unexplored. Here, we developed a novel method that allows for direct quantification of the alternatively spliced exon combinations from over hundreds of millions of Dscam transcripts in one sequencing run. With unprecedented sequencing depth, we detected a total of 18 496 isoforms, out of 19 008 theoretically possible combinations. Importantly, we demonstrated that alternative splicing between different clusters is independent. Moreover, the isoforms were expressed across a broad dynamic range, with significant bias in cell/tissue and developmental stage-specific patterns. Hitherto underappreciated, such bias can dramatically reduce the ability of neurons to display unique surface receptor codes. Therefore, the seemingly excessive diversity encoded in the Dscam locus might nevertheless be essential for a robust self and non-self discrimination in neurons.

Introduction

Alternative splicing of precursor messenger RNA (pre-mRNA) makes substantial contribution to the expansion of protein diversity (Nilsen and Graveley, 2010; Barbosa-Morais et al, 2012; Merkin et al, 2012). While most genes in metazoan genomes encode only a few isoforms of mRNA, some can produce a large number of splicing isoforms (Pan et al, 2008; Wang et al, 2008), such as CD44 (Screaton et al, 1992), neurexin (Ullrich et al, 1995), and protocadherins (Wu and Maniatis, 1999). The most extreme case is the Drosophila melanogaster homologue of Down syndrome cell adhesion molecule (Dscam) gene (Schmucker et al, 2000). The Dscam gene locus contains 115 exons, of which 95 are arranged into four clusters, that is, exons 4, 6, 9, and 17, consisting of 12, 48, 33, and 2 variable exons, respectively. The variable exons within each cluster are spliced in a mutually exclusive manner, thereby generating potentially up to 19 008 isoforms encoding different assortments of immunoglobulin domains with differential adhesive properties (exon 4, 6, and 9 clusters) as well as two different transmembrane domains (exon 17 cluster) (Schmucker et al, 2000). In addition, four different cytoplasmic domains could be generated by exon skipping (Yu et al, 2009). Importantly, a series of functional studies have demonstrated that a large isoform diversity is essential for its functions in both nervous and immune system (Zhan et al, 2004; Watson et al, 2005; Chen et al, 2006; Dong et al, 2006, 2012; Hattori et al, 2007, 2009; Hughes et al, 2007; Matthews et al, 2007; Soba et al, 2007; Watthanasurorot et al, 2011).

Specifically, it has been shown that dendrites that express identical Dscam isoforms on their surface repel each other (Hughes et al, 2007; Matthews et al, 2007; Soba et al, 2007). In wild-type conditions, neighbouring neurons with overlapping dendritic fields express different isoforms. This limits the Dscam–Dscam interactions to sister dendrite interactions supporting self-avoidance. If the diversity of Dscam isoforms is decreased such that neighbouring neurons also express identical isoforms, then heteroneuronal repulsion occurs leading to wiring defects (Hattori et al, 2009). This illustrates that there exists critical thresholds of isoform diversity as such and also suggests that there might be additional cellular control mechanisms that ensure that different neurons express non-overlapping sets of isoforms. While receptors closely related to Dscam exist in higher vertebrates, it is surprising that they do not show a high degree of alternative splicing. It has however been proposed, recently, that in vertebrate other diverse receptors and in particular the clustered protocadherin receptors provide the functional counterpart to the Dscam isoform diversity (Schmucker and Chen, 2009; Zipursky and Sanes, 2010). This hypothesis is supported by the recent finding that protocadherin-gamma receptors are important for self-avoidance in retinal cells in mice (Lefebvre et al, 2012) and that due to alternative splicing and tetramer formation of protocadherins tens of thousands of homophilic binding specificities can be generated (Schreiner and Weiner, 2010). Overall, this shows that the generation of receptor diversity by means of alternative splicing is of general importance for the process of neuronal wiring specificity, thereby emphasizing the importance of applying novel systematic and quantitative isoform expression analysis in order to dissect the underlying molecular mechanisms.

The alternative splicing of Dscam during development and in different tissues/cell types has been previously studied using customized microarrays and other PCR-based methods. These studies demonstrated both temporal and spatial regulation of splicing choices from exon 4, 6 and 9 clusters (Celotto and Graveley, 2001; Neves et al, 2004; Zhan et al, 2004; Watson et al, 2005). The observations suggested a ‘stochastic yet biased’ splicing model, in which Dscam isoform profiles arise from a series of stochastic splicing events (Neves et al, 2004). The inclusion probability of individual variable exons is determined by the interaction between various RNA elements and specific splicing factors expressed in different cell types (Park et al, 2004; Graveley, 2005; Kreahling and Graveley, 2005; Anastassiou et al, 2006; Olson et al, 2007; May et al, 2011; Yang et al, 2011; Wang et al, 2012).

A major technical limitation in conventional profiling of Dscam isoforms lies in the fact that choices of the variable exons can only be investigated for each cluster separately. The frequencies of different transcript isoforms have then to be inferred based on various assumptions, for example, that alternative splicing occurs independently at different clusters. Two studies have suggested such an independent splicing mode (Neves et al, 2004; Chen et al, 2006). However, whether it indeed holds true awaits a more direct experimental examination, where the complete transcripts could ideally be quantitatively profiled. Recently, massive parallel shotgun cDNA sequencing (RNA-seq) has been used for high-throughput mRNA profiling (Cloonan et al, 2008; Lister et al, 2008; Mortazavi et al, 2008; Nagalakshmi et al, 2008; Wilhelm et al, 2008; Wang et al, 2009). But with limited read length, standard RNA-seq methods are unable to directly identify combinations of more than two Dscam variable exons. Moreover, given its enormous diversity, computational inference of Dscam isoform composition based on shotgun sequencing data would be impossible.

In this study, we developed CAMSeq (Circularization-Assisted Multi-Segment Sequencing), a novel method that enables to quantitatively profile the expression pattern of Dscam isoforms consisting of exons 4, 6, and 9. We analysed the splicing pattern of the three exon clusters at different developmental stages and in different cells/tissues. With unprecedented sequencing depth, out of 19 008 theoretically possible ones, we could detect 18 496 isoforms. They expressed across a broad dynamic range, and showed different splicing patterns at different stages as well as in different cells/tissues. Furthermore, we demonstrated that the alternative splicing between different exon clusters was largely independent. Finally, our data suggest a surprisingly strong bias in isoform expression. Taken together, our quantitative method and measurements enable now a thorough evaluation of how much protein diversity globally as well as per cell is essential to support a robust system distinguishing between self and non-self neurites.

Results

Development of CAMSeq, a novel method that enables the quantitative profiling of Dscam ectodomain isoforms

The genomic structure and splicing model of Dscam is shown in Figure 1A. In this study, we focused on the combinations of variable exons 4, 6, and 9, since alternative splicing in these three clusters could generate theoretically up to 19 008 different ectodomains, which contain the actual domains of the Dscam protein determining its recognition specificity. Since Illumina paired-end sequencing could only yield up to 150 nt sequences from both 5′ and 3′ ends of cDNA fragments shorter than 1 kb, it cannot be used to directly determine for each isoform the precise combination of the three variable exons. Therefore, we developed a new method termed as ‘CAMSeq’, the novelty of which consists of two major components: (1) circularization followed by another PCR reduces the size of cDNA fragments to be sequenced; (2) multi-segment sequencing yields multiple exon sequences from a same cDNA molecule. The scheme of CAMSeq is illustrated in Figure 1B. In brief, first, using RT–PCR with the barcode-indexed primers targeting constitutive exon 3 and exon 10, Dscam mRNA was reverse-transcribed and amplified. PCR products derived from different samples and labelled with different barcodes were pooled together. After circularization of the pooled 2 kb RT–PCR product and another round of PCR with the primers targeting constitutive exon 7 and exon 8, the amplification product of ∼1 kb in length was then sequenced. As shown in Figure 1B, we modified the standard Illumina sequencing procedure and obtained from every template DNA molecule four sequencing reads (quadruple-reads) derived from exons 4, 6, 9, and barcode, respectively (Materials and methods). Thereby, we could identify the exon usages simultaneously in the three clusters and unambiguously reveal the identity of the expressed isoforms.

Figure 1.

CAMSeq, a novel massive parallel sequencing-based method for quantitative profiling Dscam isoforms. (A) Alternative splicing of Drosophila Dscam gene. Constitutive exons are depicted in grey, whereas alternative exons from exon clusters 4, 6, 9, and 17 are depicted in orange, blue, green, and purple, respectively. The exon 4, 6, and 9 alternatives code for the variable extracellular Immunoglobin (Ig) domains. The two exon 17 alternatives code for the single transmembrane domain. (B) Outline of CAMSeq. In brief, first, using RT–PCR with the barcode-indexed primers targeting constitutive exon 3 and exon 10, Dscam mRNA was reverse-transcribed and amplified (barcodes are depicted in yellow and marked with ‘B’). After circularization of the ∼2 kb RT–PCR product and another round of PCR with the primers targeting constitutive exon 7 and exon 8, the amplification product of ∼1 kb in length was then sequenced. Here, using a modified sequencing procedure with four specific sequencing primers targeting constitutive exon 3, exon 5, exon 8, and exon 10, respectively (blue arrows), we obtained from every template DNA molecule four sequencing reads derived from exons 4, 6, 9, and barcode, respectively. (C) Analysis of sequencing data obtained from two controlled mixtures of eight in vitro synthesized Dscam mRNAs. The relative RNA abundance (X axis) was plotted against the normalized number of derived sequencing reads (RPM, reads per million total reads) (Y axis). (D) Comparison of sequencing data obtained from two replicate experiments on the same total RNA extracted from S2 cells showed a high correlation (R²=0.993). (E) Comparison of sequencing data obtained from CAMSeq to that from PacBio sequencing showed a high correlation (R²=0.978). (F) Estimation of chimeric rate. To estimate the rate of forming chimeras, we first count the number of chimeric reads derived from different samples with different barcodes joined together. Here, Y axis represent the frequency of such chimeric reads F_4,6,9,b,b′, whereas X axis is the product between the frequency of reads containing the same exons 4 and 6 as well as the same forward barcode, and that of reads containing the same exon 9 as well as the same reverse barcode, F_4,6,b × F_9,b′. Assuming a second-order reaction kinetics, the mean chimerical rate could be represented by the slope of regression line, i.e., 0.0097 (Materials and methods).

To evaluate the accuracy of our method, we first generated a set of eight Dscam mRNAs with known concentrations by in vitro transcribing cloned Dscam cDNAs containing different combinations of exons 4, 6, and 9. We then prepared two reference samples by mixing these eight RNAs in different amounts spanning five orders of magnitude and applied CAMSeq on these two samples (Materials and methods). As shown in Figure 1C, a straight linear relationship spanning the full dynamic range was observed between the RNA amount and the number of sequencing reads derived from each RNA, demonstrating that our method provides an accurate estimation of the relative abundance of different isoforms. Furthermore, to examine the reproducibility of our method when applied to biological samples, we analysed twice the same RNA extracted from S2 cells, a cell line derived from Drosophila embryonic hemocytes. As shown in Figure 1D, the isoform profiles from the two replicates were highly correlated (R²=0.993). Finally, to assess a potential systematic bias caused by cDNA circularization, the second round of PCR as well as Illumina sequencing procedure, we also measured the isoform abundance in S2 cells by directly sequencing the 2 kb RT–PCR products using PacBio RS system (Materials and methods). A total of 63 109 PacBio reads could be used to reveal the identities of exons 4, 6, and 9 for 3725 Dscam isoforms and the isoform abundances estimated using the two approaches showed a high correlation (R²=0.978; Figure 1E).

During the preparation of the sequencing libraries, in the circularization step, in addition to self-circularization, chimeras could also form from intermolecular ligation events where two or more DNA molecules are joined together. Although occurring at a much lower frequency compared to self-circularization, these chimeras could nevertheless lead to an overestimation of the number of isoforms that were detected. To rule out the ‘chimera’ effect and obtain an accurate number of detected isoforms, we estimated the rate of forming chimeras by counting the number of apparent chimeras in which Dscam cDNAs derived from different samples and labelled with different barcodes joined together (Materials and methods). As shown in Figure 1F, the mean chimerical rate was ∼1%.

Detection of Dscam isoforms expressed at different developmental stages and in different cells/tissues

We used CAMSeq to analyse Dscam isoform expression at different developmental stages (embryo, first instar larvae (L1), second instar larvae (L2), and third instar larvae (L3), and pupae) and in adult brain. Here, Dscam cDNAs from each sample were amplified with the primers containing distinct barcode sequences at both 5′ and 3′ ends (Figure 1B; Materials and methods). The PCR products from different samples were then pooled in equal amounts and the mixture went through the remaining steps as described above. For each sample, we obtained between 5.71 and 15.22 million quadruple-reads that could be used to unambiguously identify the usage of exons 4, 6, and 9 as well as the barcode representing a specific sample (Supplementary Table S1). In all samples, we could detect the presence of all variable exons from exon 4, 6, and 9 clusters, except exon 6.11. During development, the exon usages in clusters 4 and 9 showed moderate to dramatic changes, whereas the differences in exon 6 clusters were relatively modest (Figure 2A).

Figure 2.

The relative expression of Dscam variable exons and Dscam isoforms during development and in S2 cells. (A) The relative expression of variable exons 4, 6, and 9 in different samples. (B) Cumulative distribution function of abundances (RPM) of Dscam isoforms in different samples.

After subtracting all potential chimerical reads, we detected with high confidence between 13 216 and 16 886 isoforms in each sample, and 18 496 isoforms in at least one sample (Materials and methods; Supplementary Table S1). The number was quite close to 18 612, the maximum number of potential isoforms if excluding the pseudo-exon 6.11, indicating that all the remaining Dscam isoforms were expressed.

In each sample, the relative abundance of different isoforms spanned at least four orders of magnitude (Figure 2B; Supplementary Table S2). The most abundant 10 and 100 isoforms derived 0.7–2.0%, and 5.3–12.2% of all reads from one sample (Supplementary Table S1). Importantly, our comparison between different samples showed that S2 cells express a significantly more restricted repertoire of Dscam isoforms in which only 7317 isoforms were detected with the most abundant 100 isoforms accounting for 25.6% of all reads (Supplementary Table S1). Such striking difference between S2 and all the other samples might be explained by the fact that S2 cells are a homogeneous cell population whereas other samples consist of different types of Dscam expressing cells with the splicing preferences towards different sets of variable exons. As a result, at a similar sequencing depth, we could detect much fewer isoforms in S2 cells. Interestingly, when we compared dynamic ranges of different exon usages in the three clusters between S2 and other samples, it turned out that exon 9 cluster expressed a relatively limited set of exons in S2 compared to other samples (Figure 2A). Given the observation that the splicing choice of exon 9 was most variable between different cell types (Figure 2A), this corroborates our hypothesis that the larger repertoire observed in the other samples was due to the much higher cell-type diversity.

Independent splicing choice between the different exon clusters

To address whether the splicing choices at different exon clusters are independent or not, we first estimated the relative abundance of all variable exons, and then assuming an independent splicing model, calculated the expected relative frequencies of different isoforms by simply multiplying the frequencies of their respective variable exons 4, 6, and 9. Comparison between the observed and expected isoform frequencies in different samples showed mixed results (Figure 3A; Supplementary Figure S1A). Whereas a straight linear relationship was observed in S2 cells, demonstrating unambiguously the independent splicing choice among different exon clusters, other samples showed only weak to modest correlations between the observed and expected frequencies (Figure 3A; Supplementary Figure S1A).

Figure 3.

Independent splicing choice between the different variable exon clusters. (A) Observed isoform frequencies were depicted in X axis. Expected frequencies were calculated by multiplying the frequencies of their respective variable exons 4, 6, and 9, and depicted in Y axis. To determine whether the splicing between the three clusters was independently controlled, the two frequencies were compared. (B–D) In a similar way, we determined whether the splicing choices were independent between exons 4 and 6 (B), exons 6 and 9 (C), and exons 4 and 9 (D), respectively. (E) In adult brain sample, the variable exon 4s and 9s were clustered based on their expression patterns, the exon 9s could be clearly divided into two groups, one containing only five exons and the other consisting of the remaining 27. (F) Given the differential usages of variable exon 4 within the two groups, the whole brain data were in silico decomposed into two sets with different usages of exons 4 and 9, the yellow and blue groups. (G) The splicing choice within each group was largely independent between exons 4 and 9. X axis depicted the observed isoform frequencies from the whole brain data set, whereas in Y axis, the expected isoform frequencies were the sum of the expected frequencies of the two groups.

Given that the splicing choice of exons 4 and 9, especially the latter, was quite variable between different cell types, we hypothesized that the different observation between S2 and other samples was due to the fact that other samples consisted of different cell types expressing distinct sets of exon 4s and 9s. To corroborate this hypothesis, we further analysed splicing choices between exons 4 and 6, exons 6 and 9, as well as exons 4 and 9, separately. Indeed, whereas the splicing appeared to be independent between exons 4 and 6, as well as between exons 6 and 9 in all samples (Figure 3B and C; Supplementary Figure S1B and C), the splicing between exons 4 and 9 showed different patterns between S2 and other samples (Figure 3D; Supplementary Figure S1D). Notably, we could cluster the variable exon 4s and 9s based on their expression patterns in adult brain and other samples. As shown in Figure 3E and Supplementary Figure S2A, exon 9s could be clearly divided into two groups, one containing only five exons and the other consisting of the remaining 27. Given the differential usages of variable exon 4s within the two groups, we could in silico decompose the whole adult brain data into two sets with different usages of exon 4s and 9s, and the splicing choices within each data set being largely independent between the two clusters (Figure 3F and G). In a similar way, other samples could also be decomposed into two or three groups expressing distinct sets of exon 4s and 9s, and all with independent splicing choices among different exon clusters (Supplementary Figures S1E, F and S2B).

Discussion

We developed CAMSeq, a new massive parallel sequencing-based approach for quantitatively profiling Dscam isoform expression. All previous global analyses of the alternative splicing of Dscam using microarrays measured the relative abundance of variable exons from different clusters separately. In contrast, our new method allows to identify the expression of Dscam isoforms directly by determining for each isoform the precise combination of exons 4, 6, and 9. Furthermore, our sequencing approach provided an accurate quantitative measurement, demonstrated by several control experiments. Finally, the sequencing depth achieved in this study enabled us to detect almost all the possible isoforms except those containing pseudo-exon 6.11. This is consistent with the previous findings (Celotto and Graveley, 2001; Neves et al, 2004) and the observation that the amino-acid sequence of exon 6.11 lacks critical residues essential for proper Immunogloblin (Ig) domain folding (Dietmar Schmucker, personal communications). Notably, we could detect a very minor fraction of isoforms skipping either of exons 4, 6, or 9, consistent with previous observations (Kreahling and Graveley, 2005). Taken together, with the unprecedented sequencing depth, we achieved an ultra-high sensitivity of detecting lowly expressed isoforms without detection of any false positive sequences.

We demonstrated that the alternative splicing between different exon clusters is independent in a uniform cell population (S2 cells). In a previous study, using a genetic approach, Chen et al (2006) generated two fly lines in which different parts of exon 4 clusters were deleted. Subsequent expression analysis of the splicing pattern in exon 6 and 9 clusters revealed no significant difference between the larval central nervous system (CNS) of the control and that of the two mutant strains, implicating that splicing of exons 6 and 9 is independent from that of exon 4. In another study, using S2 cells, Neves et al analysed the relative abundance of variable exon 4s and 6s in the isoforms containing two different exon 9s. They did not find specific exon 4 and 6 alternatives associated with either of the two exon 9s and therefore suggested that splicing choices of the three clusters were independent (Neves et al, 2004). Our quantitative data are consistent with these findings, and, for the first time, provided direct and comprehensive experimental evidence for the independent splicing regulation of the three exon clusters in a distinct cell type.

However, in more complex samples, we observed some potential splicing dependence, especially between exons 4 and 9. We attributed such observation to the cellular heterogeneity of these samples. While the splicing is independent within a distinct cell type, different types of cells with differential usages of exon 4s and 9s, combined together, could give the misleading impression of dependence, as demonstrated by our in silico data decomposition (Figure 3; Supplementary Figures S1 and S2).

Dscam diversity is essential for neurite self-avoidance and plays a profound role in wiring the fruit fly brain. Using an elegant genetic approach, the Zipursky laboratory demonstrated that thousands of isoforms are essential to provide neurons with a robust mechanism to distinguish between self and non-self during self-avoidance (Hattori et al, 2009). Moreover, they used mathematical modelling to support the hypothesis that the full molecular diversity encoded by the Dscam gene locus is almost five times larger than what may be considered as necessary. However, in such a model, all the potential isoforms were randomly sampled with equal probability. Apparently, such assumption of uniform isoform expression is an oversimplification. Starting with a realistic in vivo data set, we performed a similar modelling study using our actual quantitative data sets. First, we estimated the number of different isoforms that could be obtained by randomly sampling certain numbers of Dscam mRNA copies. As expected, this number is dependent on the biased choice of the variable exons. Due to the biased exon usage, the number of distinct isoforms that could be present in a certain number of neurons is much lower than that under the assumption that all isoforms expressed with equal probability (Figure 4A). For example, Drosophila mushroom body (MB) comprises some 2500 neurons. If each individual MB neuron expresses 20 Dscam mRNA copies (Zhan et al, 2004) and all possible isoforms express with equal probability, then about 17 300 different isoforms would be present in one MB. In contrast, if the isoforms express based on the pattern we measured from the adult brain, then only 12 300 different isoforms would be present in one MB (Figure 4A; Materials and methods). With such reduced repertoire, obviously the number of neurons with unique Dscam identity also decreases (Figure 4B; Materials and methods). For instance, if up to 20% of Dscam isoforms were allowed to share between two neurons, under the assumption of uniform isoform expression, 68 500 neurons could be distinguished from each other. But with the more realistic size of adult brain Dscam repertoire evaluated by our quantitative CAMSeq analysis, only 3200 neurons could be uniquely labelled (Figure 4B; see Supplementary Figure S3B for the conditions in which up to 0 or 10% of isoforms were allowed to share between two neurons). The same labelling capacity could also be coded by about 5500 uniformly expressed isoforms. To facilitate the comparison of labelling capacities between cell types with different splicing biases, we suggest to define the effective size of a certain Dscam repertoire as the number of uniformly expressed isoforms that could label the same number of neurons with unique identity (Materials and methods).

Figure 4.

Monte Carlo simulation of Dscam repertoire and the number of neurons that could be labelled with unique Dscam identity. (A) The number of different isoforms (Y axis) could be obtained by randomly sampling different numbers of Dscam mRNA molecules (X axis) based on the distribution of Dscam isoform abundances in adult brain, S2 cells or a hypothetical uniform distribution (Materials and methods). (B) The number of neurons that obtain unique identities at >95% likelihood (Y axis) when each neuron expresses different numbers of Dscam mRNA molecules (X axis), if allowing 20% of isoforms shared between any pair of neurons, calculated based on the distribution of Dscam isoform abundances in adult brain, S2 cells or a hypothetical uniform distribution (Materials and methods). See Supplementary Figure S3B for the condition in which up to 0 or 10% of isoforms are allowed to share between any pair of neurons.

Obviously, the fruit fly nervous system consists of many different cell types expressing different Dscam splicing repertoires. As suggested by our decomposition analysis of the brain data set (the yellow group in Figure 3F), it is very likely that in some distinct types of neurons, the Dscam isoform repertoire might be similarly small as observed in S2 cells. Due to the usage of a rather limited set of variable exons, the number of different Dscam isoforms present in a certain number of cells would be quite small and only dozens of cells could be labelled with unique Dscam identities when any pair of neurons were allowed to share 20% of their expressed Dscam isoforms (Figure 4B). Indeed, based on these calculations we would suggest that the effective size of Dscam repertoire of S2 cells is only around 800. On the other hand, different types of neurons would manifest the preferences towards different sets of exons, thereby lowering the probability to share too many of the same isoforms. The low effective size of isoform repertoire is then counteracted by a cell type-specific splicing bias. Therefore, in spite of a smaller effective size of isoform repertoire within a distinct type of neurons, the interconnecting neurons, consisting of different cell types, can still easily discriminate self from non-self.

In general terms, we would like to speculate that in any complex nervous system two ‘identity-labelling’ strategies could be used for the proper wiring in a large group of interconnecting neurons. That is, they can either be a homogeneous cell population with low bias in exon usage and thus expressing randomly from a relatively large surface receptor repertoire, or consist of different cell groups with each distinctly controlling the expression of a limited but selective set of receptor isoforms. Notably, with the second strategy, surface receptor isoforms could be used not only to distinguish self and non-self (‘individual identity’), but also potentially to differentiate between different groups of cells (‘group identity’). In this scenario, the neurites from different types of neurons are allowed to connect with a higher probability than those from the same cell type.

Genetic studies have been instrumental in understanding why the enormous Dscam molecular diversity is required in neuronal wiring. In these studies, connectivity phenotypes in different nervous systems were assessed in the strains with different sets of variable exons deleted (Chen et al, 2006; Hattori et al, 2009). Often, the effects of different deletions on the Dscam repertoire were implicitly assumed to be solely dependent on the number of deleted exons. Such assumption would hold true if all the variable exons express with equal probability. However, due to splicing bias, the effect will also depend on identities of the deleted exons. Importantly, we observed that the effect will be unequal in different cell types with distinct splicing patterns (Supplementary Table S3). Counter-intuitively, there might be some extreme scenarios in which the effective Dscam repertoire could even increase when the exons with predominant splicing bias are removed (Supplementary Table S3). In addition, the number of neurons that could be labelled with unique Dscam identity will become sensitive to the Dscam expression level when the repertoire gets sufficiently small and there is an optimal range of Dscam mRNA copies per cell that maximize the total number of labelled neurons (see S2 sample in Figure 4B). Normal wiring pattern could break down if the expression of Dscam fluctuates out of such a range. Therefore, due to all these complications, the results in the genetic studies need to be complemented by quantitative expression data in order to better interpret the influence of molecular diversity on neuronal wiring specificity.

Taken together, due to the biased usage of variable exons, which could be surprisingly strong in some distinct cell types, the accessible Dscam isoform repertoire is more restricted than previously appreciated. Moreover, the splicing of Dscam is determined by the interaction between its various RNA elements and specific splicing factors. Therefore, it seems clear that, dependent on the expression levels of the splicing factors and other interacting RNAs, the abundances of Dscam-accessible splicing complexes could fluctuate and thus lead to uncertainties in Dscam splicing outputs. To accommodate these limitations, during evolution, which, as Francois Jacob put, is a tinkering process, Drosophila Dscam gene locus might have adapted to this limitation by way of expanding exon number to encode an extremely high isoform diversity (Jacob, 1977). Such diversity, although seemingly beyond the necessity, is nevertheless essential to assure the neurons with a robust discrimination system to distinguish between self and non-self.

Materials and methods

RNA from fruit fly samples

Fruit flies from D. melanogaster J5 strain were raised on standard fruit fly medium at room temperature or at 25°C. Fruit flies from embryonic, first larval, second larval, and third larval stages were collected according to the time period after egg laying (embryos 13–18 h, first stage larvae 24–36 h, second stage larvae 60–72 h, and third stage larvae 96–108 h). Fruit fly pupae were collected 0–48 h after puparium formation. Adult brains were dissected from 1- to 3-day-old female after eclosion. S2 cells were maintained in Schneider’s medium with 10% fetal bovine serum and 100 ng/μl of penicillin/streptomycin at room temperature. Total RNAs from fruit fly samples and S2 cells were isolated using TriZOL reagent according to manufacturer’s instruction (Life Technologies).

Dscam reference RNA samples

Reverse transcription (RT) was performed on 5 μg of embryonic fruit fly total RNA with a specific primer annealed to the constitutive exon 19 (5′-TGTCCTGGTGGAAGCATAG-3′) using SuperScript III system with a reaction volume of 20 μl (Life Technologies). PCR was followed using 2 μl of RT product as template in 25 μl of GoTaq PCR system (Promega). The PCR primers were targeted at constitutive exons 3 and 11 (DsRef-1-F: 5′-GAGGTCCATGCCCAGGTGTACG-3′; DsRef-1-R: 5′-GTCGACATGCAGAGTGCCCTC-3′). PCR was run as following, 2 min at 95°C, followed by 30 cycles of 30 s at 95°C, and 2.5 min at 72°C, and a final elongation of 10 min at 72°C. PCR product was purified using Agencourt AMPure XP system (Beckman Coulter) and then cloned into pGEM-T Easy Vector, transformed into JM109 competent cells and plated onto LB/ampicillin/IPTG/X-gal plates according to manufacturer’s instruction (Promega). Plasmids from positive colonies were purified using GeneJET plasmid DNA purification kits (Thermo Scientific) and sequences of inserted Dscam isoform cDNAs were confirmed using Sanger sequencing method. Plasmids from eight colonies containing different combinations of exons 4, 6, and 9 were selected. Using these eight plasmids as templates, another PCR was performed in 25 μl of Advantage 2 PCR system (Clontech) using the forward and reverse primers targeted at constitutive exons 3 and 11, with T7 promoter sequence attached at the end of forward primer (Dsref-2-T7-F: 5′-GGATCCTAATACGACTCACTATAGGGATCCATTATCTCCCGGGACGTCCATGT-3′; DsRef-2-R: 5′-GTCGACATGCAGAGTGCCCTC-3′). After purification and measurement of concentrations and fragment sizes using Qubit system (Life Technologies) and Agilent 2100 Bioanalyzer (Agilent), the eight PCR products were used as templates for in vitro transcriptions with mMESSAGE mMACHINE T7 kit (Life Technologies). The resulting RNA samples were purified using Agencourt RNAClean system (Beckman Coulter) and quantified by Qubit system. The eight RNAs were then mixed together in different amounts.

CAMSeq

RT was performed on either 5 μg of total RNA from fly sample or 10 pg of the mixture of Dscam reference RNA samples with a primer annealed to the constitutive exon 11 (5′-GTCGCTCTTCTTTAGATCCTTGTAC-3′) using SuperScript III system with a reaction volume of 20 μl. The first round PCR was followed using 2 μl of RT product as template in 25 μl of Advantage 2 PCR system. The PCR primers were targeted at constitutive exons 3 and 10 with indexed barcode sequences attached at 5′ ends (CAMSeq-1-F: 5′-AGNNNNACCATTATCTCCCGGGACGTCCATGTGC-3′; CAMSeq-1-R: 5′-GTNNNNACCTTATCGGTGGGCTCGAGGATCCA-3′; NNNN represents barcode sequences). PCR was run as following, 2 min at 95°C, followed by 22 cycles of 30 s at 95°C, and 2.5 min at 72°C, and a final elongation of 10 min at 72°C. The products of first round PCR were purified and eluted into 10 μl of water using Agencourt AMPure XP system. After the measurement of concentrations and fragment size on Qubit system and Agilent 2100 Bioanalyzer, the purified first round PCR products obtained from different samples were then mixed together in equal amounts. The mixture was run on agarose gel, and DNA fragments with sizes between 1500 and 2500, bp were excised, purified and eluted into 20 μl of water using Qiaquick gel extraction kit (Qiagen). The product was then end-repaired using NEBNext End Repair Module (NEB), purified using Agencourt AMPure XP system. After measuring the concentration with Qubit system, 60 ng of the end-repaired product was used for circularization reaction following manufacturer’s instruction (Illumina). The circularization product was purified using Agencourt AMPure XP system and quantified using Qubit system. Using 1 ng of purified circularization product as template, the second round PCR was then performed in 100 μl of Phusion PCR system (Thermo Scientific). The PCR primers were targeted at constitutive exons 7 and 8 with Illumina adapters attached to the 5′ ends (CAMSeq-2-F: 5′-AATGATACGGCGACCACCGAGATCTACACTGGATACTCTGCTCGAGGATCTCTGGAAGTGC-3′; CAMSeq-2-R: 5′-CAAGCAGAAGACGGCATACGAGATCGGTCCAGCTTGTTTACGGGTTGTTCCTTCGATGA-3′). PCR was run as following, 1 min at 98°C, followed by 15 cycles of 30 s at 98°C, and 1.5 min at 72°C, and a final elongation of 5 min at 72°C. The product of second round PCR was purified and eluted into 10 μl of water using Agencourt AMPure XP system. After the measurement of concentration and fragment size by Qubit system and by Agilent 2100 Bioanalyzer, the purified product was sequenced using Illumina GAIIX following manufacturer’s instruction with the following modifications. On one flowcell, we performed a total of four sequencing using four specific sequencing primers targeting constitutive exon 10, exon 8, exon 5, and exon 3, respectively (CAMSeq-barcode-seq-primer: 5′-CCTCCCAGATGGATCCTCGAGCCCACCGATAAG-3′; CAMSeq-ex9-seq-primer: 5′-GATACTCTGCTCGAGGATCTCTGGAAGTGCAAGTCA-3′; CAMSeq-ex6-seq-primer: 5′-CGATTAAGTGCCACAAAAGGACGATTGGTCATCA-3′; CAMSeq-ex4-seq-primer: 5′-CCATTATCTCCCGGGACGTCCATGTGCGAG-3′). After each sequencing, the sequencing primer and the synthesized strand were washed away. By running the four sequencing for 25, 36, 36, and 36 cycles, respectively, we obtained for each DNA template molecule four sequencing reads derived from the barcode, variable exon 9, exon 6, and exon 4, respectively.

PacBio sequencing of Dscam isoforms

The 2 kb first round RT–PCR product obtained from S2 cells, as described in the previous section, was directly sequenced using PacBio RS system according to manufacturer’s instruction (Pacific Biosciences).

Processing of CAMSeq data

Each Illumina sequencing read was split into four segments derived from barcode (1st–25th nt), exon 9 (26th–61st nt), exon 6 (62nd–97th nt), and exon 4 (98th–123rd nt), respectively. The three segment sequences corresponding to exons 4, 6, and 9 were aligned to reference Dscam exon sequences ( http://www.ncbi.nlm.nih.gov/nucleotide/AF260530?tool=FlyBase) using bowtie2 (parameters: --very-sensitive-local -5 3). Only the reads with all the three segments that could be uniquely mapped to the respective exons were retained. The barcode segment was used to extract the two barcode sequences derived from the 5′ end of either forward or reverse primer in the firsts round PCR. The two barcode sequences were then compared with those used in the experiments. The reads containing the two barcodes with at most one mismatch from the used barcodes were retained. Those with the two barcodes derived from a same sample were used to calculate the isoform frequency, whereas those with the two barcodes derived from different samples were used to estimate the rate of forming chimeras.

Processing of PacBio sequencing data

Circular consensus reads obtained from PacBio sequencing were aligned to Dscam exons using BLAT (parameters: -tileSize=8 -stepSize=5 -oneOff=1 -minScore=20 -minIdentity=70). We retained the sequences if and only if the identity of exon 4, exon 6, and exon 9 could all be unambiguously revealed.

Estimation and correction of the chimeric effect

To estimate the rate of forming chimeras, we first identify the reads derived from the intermolecule ligation between two different molecules in the circularization step (see Processing of CAMSeq data). In these reads, the sequences of exon 4, exon 6, and one barcode b (forward barcode) originate from one molecule, while the sequences of exon 9 and the other barcode b′ (reverse barcode) are from a second molecule. Assuming a second-order reaction kinetics, the rate of forming chimeras r is given by

where F_4,6,9,b,b′ is the frequency of the chimeric product containing a distinct set of exons 4, 6, 9 as well as forward and reverse barcodes, F_4,6,b is the frequency of reads containing the same exon 4, exon 6 and forward barcode b, F_9,b′ is the frequency of reads containing the same exon 9 and reverse barcode b′. We calculated values r=r_4,6,9,b,b′ for all exon/barcode combinations with adequate expected numbers of reads (i.e., T × F_4,6,b F_9,b′?100, where T is the total number of mappable reads). Assuming that the chimera rate is independent from actual exon/barcode combination, we treated the calculated r_4,6,9,b,b′ values as a set of independent variables. The slope r_avg of a linear regression line with intercept 0 through the points x=F_4,6,b × F_9,b′ and y=F_4,6,9,b,b′, b≠b′ was used as an average chimeric rate (Figure 1F).

For b=b′ and a given chimera rate r, we could calculate the expected number of chimeric reads by

We then corrected the observed number of reads per isoform and barcode by subtracting the (rounded up) number of chimeras given by Formula (2) for r=r_avg.

In order to estimate the total number of expressed isoforms with high confidence, we applied Formula (2) to compute the number of potential chimeric reads using a highly conservative estimate of chimeric rate, that is, an upper α-quantile from the distribution of all r_4,6,9,b,b′ values, where α=1/n, and n=19 008, the number of theoretically possible isoforms. We then counted for each data set, the number of different isoforms for which the number of observed reads was higher than that of chimerical reads estimated in this conservative way. For each isoform, the probability to be a false positive is at most α, thus the expected number of false positives per data set is at most n × α=1.

Computation of the effective Dscam isoform repertoire

Given the relative frequencies f_i for all n combinations of exon 4, exon 6, and exon 9, we can compute the probability P₁₁ for two identical isoforms independently sampled from the same f_i distribution as

This probability gets minimal if all isoforms express with the same probability (uniform distribution); in this case, P₁₁=1/n. If the splicing on the other hand is biased towards certain isoforms, then it is more likely that two independently sampled isoforms are identical, so in this case P₁₁ is greater than 1/n and the ability of the cell to create distinctive Dscam identities is decreased. We define the effective size n_eff of the Dscam repertoire to be the number of uniformly expressed isoforms needed to get P₁₁:

The probability for a single Dscam transcript to have the same isoform identity as one or more of k independently expressed transcripts is

The probability that more than h out of k Dscam transcripts independently expressed in two cells share the same isoform identity is given by a binomial distribution:

If we assume for example that two distinct cells are allowed to share up to 20% of their Dscam transcripts, then we would set h=0.2 × k. P_kk can then be interpreted as the probability for two distinct cells getting the same Dscam identity by chance, see Supplementary Figure S3A.

For a set of m cells, the probability that each cell gets a unique identity is

where m × (m−1)/2 is the number of all possible pairwise combinations of the m cells. If we set for example Q=0.95, given P_kk, then m could be computed, see Supplementary Figure S3B.

Clustering of exons 4 and 9 based on the expression patterns

We created heatmaps using the heatmap.2 function from the R package gplots to visualize the expression pattern of exon 4 alternatives (row) and exon 9 alternatives (column). The numbers of sequencing reads were first normalized column-wise for each exon 4 alternative, and then scaled row-wise by using the parameter scale=‘row’. The rows and the columns were hierarchical clustered by complete-linkage clustering using distance metric d=(1−R)/2 where R is the Pearson correlation coefficient.

Decomposition of Dscam isoform distribution data sets

If exon 4, exon 6, and exon 9 are selected independently during splicing in a homogenous cell population, then the expected frequency f_4,6,9 of each isoform is given by

where f₄, f₆, and f₉ are the (marginal) frequencies for the exon 4, exon 6, and exon 9, respectively. If on the other hand the cell population consists of two cell types A and B with distinct splicing bias, Equation (8) may not hold. Instead we assume

where μ is the ratio in which the Dscam transcripts from A and B are mixed together. Assuming the f₆ is very similar between different cell types, Equation (9) could be simplified to

where f_4,9 is the expected frequency for a combination of exon 4 and exon 9.

For different fixed values μ we tried to find distributions f₄^A, f₉^A, f₄^B, f₉^B fitting to Equation (10) with minimum total log squared error. Starting with f₄^A=f₄^B=f₄ and f₉^A=f₉^B=f₉ we optimized the distributions in up to 500 rounds, where in each round we optimized for each exon 4 and exon 9 separately in random order, i.e., we adjusted the frequency of each variable exon such that the objective function was minimized.