Phenotypic Convergence: Distinct Transcription Factors Regulate Common Terminal Features

Abstract

Transcription factors regulate the molecular, morphological, and physiological characteristics of neurons and generate their impressive cell-type diversity. To gain insight into the general principles that govern how transcription factors regulate cell-type diversity, we used large-scale single-cell RNA sequencing to characterize the extensive cellular diversity in the Drosophila optic lobes. We sequenced 55,000 single cells and assigned them to 52 clusters. We validated and annotated many clusters using RNA sequencing of FACS-sorted single-cell types and cluster-specific genes. To identify transcription factors responsible for inducing specific terminal differentiation features, we generated a "random forest" model, and we showed that the transcription factors Apterous and Traffic-jam are required in many but not all cholinergic and glutamatergic neurons, respectively. In fact, the same terminal characters often can be regulated by different transcription factors in different cell types, arguing for extensive phenotypic convergence. Our data provide a deep understanding of the developmental and functional specification of a complex brain structure.

Keywords: Drosophila optic lobe; cell-type evolution; gene regulation; modeling; neuronal development; neuronal diversity; neurotransmitters; scRNA-seq; single-cell sequencing; transcription factors.

Figure 1. Drop-seq experimental procedure, analysis, and clustering, see also Figure S1 and Table S1

(A) We dissected the optic lobes of the Drosophila central nervous system and dissociated them into single cells. The cell bodies can be seen by the DAPI staining in the cortex and rim of the three neuropils (visualized using an antibody against NCad). The single cells were then fed into the microfluidic device, alongside the beads (which were in lysis buffer) and the oil, in a setup that resulted in the generation of aqueous droplets in an oil background. Each droplet may be empty, carrying a bead and a single cell, or carrying one of the two. After lysis, transcript annealing, droplet breakage, cDNA preparation and PCR amplification, we sequenced the pooled single cell transcriptomes, analyzed the results using the Seurat package in R and clustered the single cells in 52 clusters. Scale bar, 20um. (B) We performed PCA to reduce the dimensions of the data for further analysis. Genes (rows) and cells (columns) are ordered according to their PCA scores and the 500 most extreme cells and 30 most extreme genes on both sides of the distribution are shown in the heatmap. The first PCs (as indicated here by PC1) were responsible for the separation of neurons from glia, as indicated by the positive contribution of glial genes (such as nrv2, Inx2, alrm, and ogre) in PC1 and the opposite one for genes enriched in neurons (VGlut and nicotinic acetylcholine receptors). Later PCs divide the neurons based on their neurotransmitter identity, as can be seen for PC6 (glutamatergic and GABAergic neurons are separated from the rest, mainly cholinergic ones) and PC11 (glutamatergic and GABAergic neurons are separated from each other). (C) The tSNE plot of all single cells included in our analysis shows the separation of different clusters. We used a k-nearest neighbor algorithm to call 61 clusters, which are shown in different colors on the tSNE plot. (D) Transcription factor-based hierarchical clustering of the Drop-seq cluster transcriptomes. Clusters are numbered from 0 to 58. The first split of the tree represents the separation of 7 glial clusters (red) from 45 neuronal ones (blue), as expected from the PCA. Numbers at the bottom of the tree indicate clusters. (E) The expression of 401 selected Drop-seq cluster markers (rows) is shown in all Drop-seq single cells (columns) (see Table S1). Clusters are separated by white lines and are arranged according to the tree in Figure 1D. Glial clusters are highlighted in red, while neuronal clusters are in blue. Interestingly, a single neuronal cluster that expresses elav but not repo (cluster 14 – red arrow) shares many common markers with glia (see Discussion).

Figure 2. Comparison of Drop-seq cluster transcriptomes and FACS-sorted cell type transcriptomes shows striking similarities between certain clusters and cell types and is used to annotate the Drop-seq clusters, see also Figure S2

(A) The expression levels of 401 selected Drop-seq cluster markers (rows) is shown for “simulated single cells” (columns) representative of FACS-sorted cell types (green – cell type name indicated on top) and for single cells of the respective Drop-seq clusters (red – cluster number indicated on top) (see Table S1). Each FACS-sorted cell type corresponds clearly to one Drop-seq cluster. (B) Histograms showing the Pearson correlation of the transcriptome of each FACS-sorted cell type with the transcriptome of the more correlated clusters. Most of the cell types map to one cluster. Error bars represent standard error of the mean of the triplicates’ Pearson correlation with the more related clusters. (C) R74G01-Gal4>UAS-myrGFP is expressed in two cell types: Tm1, whose projections in the lobula are indicated by the arrowhead, and T1, whose projections in the lamina are marked by the arrow. NCad is used to visualize the neuropils. Scale bar, 20um. (D) The heatmap shows the expression levels of the 401 selected markers (rows) in “simulated single cells” (columns) that represent R74G01-Gal4 and in single cells of the respective Drop-seq clusters, 12 and 23. Since R74G01-Gal4 is expressed in two different cell types, T1 and Tm1, its transcriptome matches two Drop-seq clusters. (E) Histogram showing the Pearson correlation of the T1 and Tm1 mixed population transcriptome with the more correlated Drop-seq clusters. It maps to two clusters, 12 and 23.

Figure 3. Annotation of glial (red) and neuronal (green) clusters, see also Figure S3

(A) Annotation of all glial clusters using glial markers. Repo is expressed in all glial clusters. AdamsTS-A is expressed in perineurial and subperineurial glia, while CG4797 and gemini are only expressed in perineurial glia. Gs2 is expressed in astrocyte-like glia and neuropile glia (see also Figure 3C). Drpr is mainly expressed in phagocytic ensheathing glia. Wrapper is only expressed in cortex glia (see also Figure 3D) and hoe1 mainly in chiasm glia (see also Figure S3C) (B) Annotation of neuronal clusters using three different techniques: 1) based on their correspondence to the FACS-sorted cell type transcriptomes (see also Figure 2), 2) based on known markers (Mi1 expresses bsh and Pm1, Pm2, and Pm3 express Lim3, Pm1 and Pm2 express svp, Pm1 expresses tsh, and Pm3 expresses Vsx1), 3) based on newly identified markers (kn is expressed in cluster 15 and corresponds to TmY14, and CG42458 is mainly expressed in cluster 12 and corresponds to TmY8). (C) A swapped MIMIC line expressing Gal4 in the pattern of Gs2 was used to drive MCFO (Nern et al., 2015). Single cell clones were generated in the adult brain and are shown in red and green. Gs2 is expressed in neuropile glia (arrow) and astrocyte-like glia (arrowhead). (D) A wrapper-Gal4 line was used to drive MCFO (Nern et al., 2015). Single cell clones were generated in the adult brain and are shown in green. Wrapper is expressed in cortex glia. (E) A swapped MIMIC line expressing Gal4 in the pattern of kn was used to drive MCFO (Nern et al., 2015). Single cell clones were generated in the adult brain and are shown in red. Kn is expressed in TmY14. (F) A swapped MIMIC line expressing Gal4 in the pattern of CG42458 was used to drive MCFO (Nern et al., 2015). Single cell clones were generated in the adult brain and are shown in green. CG42458 is expressed in TmY8 (arrowhead). NCad labels the neuropils in C–F. Scale bar, 20um.

Figure 4. Transcription factor and neurotransmitter expression in the Drop-seq clusters, see also Figure S4

(A) Heatmap showing the expression of transcription factors (rows) in all Drop-seq clusters (columns) – green in different intensities indicates expression in different levels, black corresponds to no expression. The transcription factors that are expressed in the adult optic lobe neurons and glia can be separated into two categories: 72 ubiquitous/pan-neuronal transcription factors are found at similar levels in most cell types, while 598 cell type-specific transcription factors are expressed in significantly higher levels in only one or few cell types. (A’) Heatmap showing the expression of transcription factors (rows) in all Drop-seq clusters (columns). Transcription factors are organized in modules (color-coded on the right), which were defined by weighted-gene co-expression network analysis (WGCNA). We observe that each module of transcription factors is mainly expressed in a single cluster, indicating that similar cell types have different compositions of transcription factors. (B) Heatmap showing the expression of neurotransmitter related genes (rows) in all Drop-seq clusters (columns) – green indicates expression, black corresponds to no expression. ChAT is expressed in cholinergic neurons, Gad1 in GABAergic, VGlut in glutamatergic, Eaat1 is an excitatory aminoacid transporter that is used to uptake glutamate, Vmat marks aminergic neurons and ple (tyrosine hydroxylase) is expressed in dopaminergic neurons. Most of the neurons in the optic lobe are cholinergic. (C–C’) Antibody staining against ChAT and VGlut showing the presence of cholinergic and glutamatergic neurons in the Drosophila optic lobe. (C’’) A Vmat-Gal4 line was used to drive MCFO (Nern et al., 2015). Single cell clones were generated in the adult brain and are shown in green. Two aminergic neuronal cell types can be seen with their cell bodies in the medulla rim (arrowhead) and the lobula cortex (arrow). Scale bar, 20um.

Figure 5. A ‘random forest’ model identifies transcription factors that regulate terminal genes involved in neurotransmitter expression, see also Figure S5

We trained a ‘random forest’ model using 39 clusters as a training set and 13 clusters as a test set. (A) The generated model can faithfully predict the expression of all genes in the test clusters given the transcription factor expression. The accuracy of the prediction was 98% for the neuronal clusters and 77% for the glial clusters (mainly due to the fewer clusters that led to incomplete training of the model). (B) The Pearson correlation between the predicted and the actual transcriptome ranged from 84.9% to 97.5% in the neuronal clusters and was 70.2% in the glial cluster. (C) Using the ‘random forest’ model, we identified the transcription factors that are mainly responsible for the generation of each of the four neurotransmitter identities: cholinergic (apterous), glutamatergic (traffic-jam), GABAergic (Lim3), and monoaminergic (CG33695). (D) The expression of ChAT was predicted to rely on the expression of ap in a subset of the clusters. Knock-down of ap in the adult optic lobe led to the downregulation of ChAT in specific medulla layers, M6 and M10, as well as in the lobula. (E) Effect of tj knock-down in the expression of VGlut. The expression of VGlut in synaptic boutons in medulla layers M1 and M6 is reduced upon downregulation of tj. (F) The expression of ChAT, Gad1, and VGlut is predicted to be regulated by Ap, Lim3, and Tj, respectively. Quantitative-PCR for the mRNA of ChAT, Gad1, and VGlut shows that the genes encoding these transcription factors are significantly downregulated upon activation of RNAi against their respective predicted regulators. Data are represented as mean ± SEM. Scale bar, 20um.

Figure 6. Cell type-specific transcription factors regulate terminal genes in different cells, see also Table S2

(A) We used the ‘random forest’ model to identify the top transcription factors that better correlated with the expression of effector genes. ChAT expression correlated best with the expression of three transcription factors, ap, chn, and CG16779. (A’) Similarly, the expression pattern of VGlut may be generated by the combination of four different transcription factors (traffic-jam, fd59A, CG32105, and CG4328). (A’’) The candidates for generating Gad1 expression pattern are Lim3 and ey. (A’’’) One transcription factor, CG33695, was needed to explain the expression of Vmat. (B) Two extant cell types that express an effector gene (indicated by the bold line) may either share a common ancestor that was expressing this gene or they may have independently evolved the capacity to express it. In the latter case, the expression of this gene may rely on the same or different transcription factor. (B’) As a consequence of the evolutionary history of the effector gene, its expression in different cell types of the optic lobe may either rely on the same transcription factor (single activator) or different transcription factors may regulate its expression in different cell types (neuronal type-specific activator). (C) Out of the 9738 genes that are expressed in the adult Drosophila optic lobe (excluding transcription factors), 3085 genes support a single activator model, while 6653 genes are better explained by a cell type-specific activator model. (C’) The genes that supported the single activator model are expressed in few clusters (2.2 clusters on average), while the 6653 genes explained by the cell type-specific activator model covered a larger range of clusters (22.2 on average).