Reverse engineering neuron-type-specific and type-orthogonal splicing-regulatory networks using diverse cellular transcriptomes

SUMMARY

Cell-type-specific alternative splicing (AS) enables differential gene isoform expression between diverse neuron types with distinct identities and functions. Current studies linking individual RNA-binding proteins (RBPs) to AS in a limited number of neuron types underscore the need for holistic modeling. Here, we use network reverse engineering to derive a map of the neuron-type-specific AS-regulatory landscape of 133 mouse neocortical cell types using pseudobulk transcriptomes derived from single-cell data. We infer the regulons of 350 RBPs and their cell-type-specific activities, among which we validate Elavl2 as a key RBP for medial ganglionic eminence (MGE)-specific splicing in GABAergic interneurons using an in vitro embryonic stem cell (ESC) differentiation system. We also identify a module of exons and candidate regulators specific to long- and short-projection neurons across multiple neuronal classes. This study provides a resource for elucidating splicing-regulatory programs that drive neuronal molecular diversity, including those that do not align with gene-expression-based classifications.

In brief

Moakley et al. use network reverse engineering to map comprehensive splicing-regulatory networks and predict their differential activity across mouse neocortical cell types by leveraging transcriptomic diversity derived from single-cell RNA-seq data. They validate predicted neuron-type-specific programs underlying both developmentally defined and morphology-associated cellular heterogeneity using in vitro models.

Graphical Abstract

INTRODUCTION

Normal brain function relies on a diverse repertoire of neurons that together enable the emergent properties of neural circuits. Neuron types vary in many physiological characteristics, including their spatial distribution, the axonal inputs and targets, the neurotransmitters they release, and their electrophysiological properties.¹ Some of these phenotypic differences are shaped by developmental-lineage-specific genetic programs. Alternatively, convergent or orthogonal mechanisms can dictate certain functional properties in specific neuronal subtypes across different lineages, suggesting a multifaceted array of factors determining neuronal identity and function. For example, glutamatergic excitatory neurons and GABAergic inhibitory interneurons are two principal neuronal classes with distinct morphological, functional, and connective properties, reflecting their developmental origins in the pallium and subpallium.^2–6 On the other hand, neuron types with long-projecting axons can be found in both glutamatergic and GABAergic neuronal classes, and these neurons have common biological needs associated with the upkeep of a long axon despite their difference in developmental lineages.^7,8

At the molecular level, neuronal identity and function are specified by the selective activity of gene-regulatory programs. Much progress has been made in elucidating the tissue- and cell-type-specific patterns of regulation that drive the RNA transcription levels required for gene function. In particular, technological advancements in single-cell RNA sequencing (scRNA-seq) have enabled unbiased concurrent examination of all brain cell types, which has led to the identification of hundreds of “transcriptional cell types” in the mouse and human brain^8–11 and the elucidation of the underlying transcriptional regulatory programs driven by DNA-binding transcription factors. However, the contributions of different steps of co- or post-transcriptional RNA processing to neuronal cell identity and function are less well studied.

Alternative splicing (AS) allows cells to customize gene products by selectively including or excluding exons from the mature RNA species, providing a major source of transcriptional diversity, especially in the brain.^12–15 In this process, RNA-binding proteins (RBPs) repress or activate the inclusion of target exons by binding to flanking cis-regulatory elements and interacting with the splicing machinery, thus altering the coding regions in the mRNA transcripts or fine-tuning mRNA stability, localization, and translation, as required by specific cellular contexts.^16,17 Although functions of most neuron-type-specific AS events remain unknown, proper AS in neurons has been shown to be critical for neural development and physiology by enabling modular roles for genes involved in specialized processes such as neuronal migration, axonal assembly, cell adhesion signaling, and synaptic transmission.^18–28 Dysregulation of AS in the nervous system has been linked to a range of neurological disorders,¹⁷ some of which are known to be caused by cell-type-specific deficiencies, including spinal muscular atrophy (SMA),²⁹ autism,²¹ and schizophrenia.³⁰ A full understanding of how the molecular identities of neuron types are established, and how they may be disrupted in disease, will require elucidation of the AS-regulatory landscape. Toward this goal, we and several other groups have recently used bulk or scRNA-seq data to catalog neuron-type-specific AS.^31–33 Efforts have also been made to reveal RBP regulators that dictate neuron-type-specific AS, but these studies have mostly focused on particular cell types and/or only a limited set of putative regulators.^{15,22,32–36} The field still lacks a comprehensive map of the splicing-regulatory networks mediating AS differences among diverse neuron types that constitute the mammalian brain.

Here, we present an end-to-end solution named master regulator analysis of alternative splicing (MR-AS) to reverse engineer splicing-regulatory networks using scRNA-seq data from diverse neuron types. MR-AS builds on the optimization of the ARACNe/VIPER framework we previously developed, which has been successfully employed to reverse engineer transcriptional regulatory networks in cancer and other cellular contexts using an information-theoretic approach.^37,38 Importantly, master regulators driving cellular states were successfully identified by estimating protein activity from the aggregated behaviors of the target genes (i.e., regulons).³⁹ Applying this pipeline to 133 adult mouse cortical cell types using AS and gene expression quantified in the same scRNA-seq data, we demonstrate that MR-AS can derive reliable splicing-regulatory networks and infer putative regulators that drive differential splicing between various groupings of neuronal types. We experimentally validated our computational predictions by focusing on the role of Elav-like 2 (Elavl2), a neuron-specific splicing factor, in directing differential splicing between medial ganglionic eminence (MGE)- and caudal ganglionic eminence (CGE)-lineage interneurons, the two primary subclasses of cortical inhibitory interneurons. We also identified a module of exons whose inclusion is specific to long- versus short-projecting neurons across a range of unrelated neuron types, as well as multiple RBPs, including Rbfox3, Khdrbs3, Qk, and Elavl3, as potential driving factors. We show that reprogramming neuronal identity to a long-projecting type is accompanied by consistent AS switches of these module exons. Together, these results suggest the validity of the inferred splicing-regulatory networks and neuron-type-specific RBP activities as a resource to elucidate the complex landscape of post-transcriptional gene regulation in the mammalian nervous system.

RESULTS

Reverse engineering of splicing-regulatory networks and estimation of RBP activity

To investigate the rich diversity of AS regulation across neuron types, we leveraged the scRNA-seq data from 19,858 mouse neocortical cells generated by the Allen Institute for Brain Science.⁸ This dataset used the SMART-seq protocol to achieve full-transcript read coverage, which can be used to quantify both gene expression and splicing simultaneously. After experimenting with different strategies for imputing missing data and inferring splicing-regulatory networks, we decided to infer the associations between exon inclusion and RBP expression levels among 133 neuronal cell types using pseudobulk samples pooled from single cells (Figures 1A, 1B and S1; see STAR Methods). The pseudobulk samples have an average of 254.8 million reads per cell type, which allowed precise quantification of both gene expression and splicing, as we established in a recent study.¹⁵ The algorithm for reconstruction of accurate cellular networks (ARACNe) was then used to identify regulatory relationships based on mutual information between regulators and their targets (Figure 1B; Table S1). The set of target exons regulated by an RBP was referred to as an RBP regulon. For each putative target exon, the direction and likelihood of regulation by the RBP (mode of regulation [MOR]) was determined by the direction and magnitude of RBP-exon correlation as described previously.³⁹ We then estimated cell-type-specific RBP activity using the virtual inference of protein activity by enriched regulon (VIPER) algorithm,³⁹ which utilizes a rank-based analysis of the relative inclusion levels of inferred regulon exons in each sample compared to a baseline (i.e., the average across cell types; Figure 1C; Table S2; see STAR Methods).

Figure 1. Reconstruction of splicing-regulatory networks using transcriptomic data from adult mouse neocortical cell types.

(A–E) Schematic (A) and visualization (B–E) of the MR-AS splicing-regulation inference workflow. Adult mouse neocortical cell scRNA-seq data⁸ were pooled by cell type for the simultaneous quantification of RBP expression and cassette exon inclusion using OLego/Quantas. The ARACNe algorithm was applied to the cell-type-level RBP expression and exon-inclusion data to predict the regulatory targets (regulon) of each RBP based on their mutual information (B; see STAR Methods). The inferred regulons were applied to the exon-inclusion data from each cell type to estimate RBP activity using the VIPER algorithm, which incorporates regulon exon-inclusion signatures derived from quantile-transformed ranks of exon Z scores (C; see STAR Methods). The resulting network consisted of 174,274 edges to 8,336 exons, approximately following a power-law distribution with RBPs grouped into 50 bins based on regulon size (D). A representative subgraph of the inferred network highlighting high-confidence targets of several well-studied neuronal splicing factors is depicted in (E) with mode of regulation (MOR; estimated subsequently by VIPER) indicated.

(F) Scatterplot summarizing the first two principal components of RBP activity with cell types indicated by color (top) and the top contributing RBPs listed (bottom).

(G) Heatmap of glutamatergic and GABAergic neuron-type-specific exons with inferred positive and negative regulatory RBPs indicated on the right. Exon inclusion is represented as mean-centered percent spliced in (PSI) values.

(H) Similar to (G) but for CGE- and MGE-specific exons.

With this approach, ARACNe inferred a bipartite network consisting of 174,274 edges between 350 RBPs and 8,336 cassette exons, with an average of 498 target exons per RBP (Figures 1D and 1E). The regulon sizes followed approximately a power-law distribution, indicating the scale-free property of the inferred network (Figure 1D). Principal-component analysis (PCA) using RBP activities estimated by VIPER clearly separated glutamatergic excitatory neurons, GABAergic interneurons, and non-neuronal glial cells, and multiple known neuron-specific RBPs such as Khdrbs3 (Slm2), Rbfox, and Elavl proteins had near-maximal loadings on the first two principal components (Figure 1F). This is also consistent with the observation that the inferred regulators of neuron-type-specific exons between various clades of neuron types included a number of RBPs we have previously shown to play a role in neuron-specific AS regulation (Figures 1G and 1H).¹⁵

To further validate the inferred network, we first compared the predicted regulons to integrative splicing-regulatory network models of several well-studied neuronal RBP families, which were derived from analysis of multimodal datasets, including protein-RNA interactions and RBP-dependent splicing changes.^40–42 The target exon lists of ARACNe and the integrative models showed significant overlaps (Figures 2A–2D, left panels). The relatively moderate magnitude of overlap presumably reflects the intrinsic differences of each method, resulting in potential false-positive and false-negative predictions. For example, the integrative models assigned many ARACNe-inferred targets high but subthreshold probabilities of being targets due to the stringent cutoffs, while others were given low scores (Figure S2). Importantly, among common target exons identified by both methods, ARACNe/VIPER-predicted MOR typically agreed in the direction of regulation when compared to integrative modeling (Figures 2A–2D, right panels; Figure S2). This consistency was also seen between predicted regulons and exons showing altered splicing after RBP perturbation for several other RBPs without available integrative models^23,28,43 (Figures 2E and 2F), including Elavl2, Elavl3, and Elavl4 (Figure 2E). The concordance of the different Elavl family members is notable since they display distinct expression patterns among neuron types (Figure S3), suggesting the capability of ARACNe to delineate specific targets regulated by unique members of a homologous RBP family.

Figure 2. Validation of RBP regulons and activities inferred by ARACNe and VIPER algorithms.

(A–D) ARACNe-inferred regulons show significant overlap (Venn diagrams) with high-confidence lists of RBP family targets from integrative Bayesian models (IMs) based on analysis of multimodal splicing data.^44–46 Among the exons common to both lists, the MORs are also highly consistent with the integrative models (bar graphs). p values shown are derived from Fisher’s exact tests.

(E and F) Similar to (A)–(D) but additional regulons inferred by ARACNe were compared to RBP target lists identified from exon inclusion changes after RBP knockouts (KOs), as measured by RNA-seq or exon-junction microarray data.

(G) Bar plot of mean RBP activity difference estimates in RBP-depleted samples using the predicted regulons as compared to control samples. For depletions of multiple RBP family members, activity value sums of the multiple RBPs perturbed were used. The statistical significance of the directional changes was tested using a binomial test.

Some RBPs, including Rbfox and Mbnl, have been shown to modulate tissue- or cell-type splicing in a binding position-dependent manner (aka, RNA map), with those binding upstream of target exons tending to repress the exon and those binding downstream tending to activate it.^44–47 Therefore, we visualized the distribution of Rbfox and Mbnl binding sites as determined by bioinformatics motif predictions or RBP footprints mapped by crosslinking and immunoprecipitation (CLIP) tags in the upstream and downstream introns of the identified exons. Consistent with direct position-dependent AS regulation, we found that exons predicted to have RBP-dependent inclusion (positive targets) tend to have motif and CLIP tag peaks downstream of the alternative exon and those predicted to have RBP-dependent exclusion (negative targets) tend to have peaks upstream of or inside the exon (Figures S4A and S4B, left panels). Notably, these patterns are also evident when only considering novel targets that were not previously identified by integrative network modeling (Figures S4A and S4B, right panels), suggesting that ARACNe predictions have identified previously unknown AS-regulatory relationships.

We next evaluated the reliability of the estimated RBP activity, which aggregates the cell-type-specific splicing of hundreds of target exons in each regulon, an approach we previously demonstrated to be highly robust against false positives in the input regulons.³⁹ Estimated RBP activity across neuron types is correlated with RBP expression level, as one would expect. However, the extent of correlation varies, likely because mutual information used in ARACNe captures both linear and nonlinear regulatory relationships (Figure S5). As an independent and direct evaluation, RBP activity was estimated by VIPER using the inferred regulons and various independent RNA-seq datasets comparing RBP knockout (KO) or knockdown samples and their controls.^{19,20,42,48,49} This analysis confirmed that estimated activity was consistently lower in RBP depletion versus control samples (Figure 2G). In most cases we examined, the perturbed RBP was among those with the largest estimated activity reduction between the two compared conditions (Figure S6). These results suggest that the activity metric captures bona fide differences in RBP activity levels.

RBP activity estimations reveal candidate drivers of neuron-type-specific AS programs

RBP activity estimations on the cell-type level allow the exploration of candidate RBPs mediating AS differences between various groups of cell types by performing differential activity tests (Table S3). As expected, known neuron-specific splicing factor RBPs tended to have higher activity in neurons versus glia (Rbfox1-3, Elavl2-4, Mbnl2, Khdrbs2/3, Nova1/2, Ptbp2, and Celf1-6), and the opposite was true for RBPs enriched in non-neuronal cells (Ptbp1, Srsf2/5, Hnrnpl, Qk; Figure 3A; Table S3). Importantly, differential activity tests between clades of neuron types identified known and novel putative neuron class- and subclass-specific splicing factors. For example, RBPs with the largest activity differences between the glutamatergic and GABAergic neuron classes included those we had previously identified based on their expression levels¹⁵ (Mbnl2, Celf1/2, and Khdrbs3 for glutamatergic; Elavl2 and Qk for GABAergic) as well as some previously unknown putative type-specific RBPs. These include Rbfox2/3 and Khdrbs2 for glutamatergic neurons and Elavl3 and Celf6 for GABAergic neurons (Figure 3B). Comparing the two GABAergic interneuron subclasses, we identified Elavl2 and Rbfox1 as candidates with MGE lineage-specific activity and Khdrbs2 and Celf1 as candidates with CGE-specific activity (Figure 3C).

Figure 3. Differential RBP activity analysis provides candidate regulators of neuron-type-specific splicing across different clades.

(A) Neurons and glia.

(B) Glutamatergic and GABAergic neurons.

(C) MGE- and CGE-lineage interneurons.

(D) Long- and short-projecting glutamatergic neurons. q values are derived from empirical Bayes-moderated t tests followed by multiple test correction. RBP activity was compared between two groups of cell types.

We previously identified a group of exons showing differential splicing between neurons with long- versus short-range projections.¹⁵ To identify potential regulators of this AS program, we performed differential RBP activity tests between long- versus short-projecting glutamatergic neurons. This analysis revealed distinct sets of RBPs with higher activity in long-projecting glutamatergic neurons (Mbnl2, Celf1/2, Khdrbs2/3, and Rbfox1-3) and short-projecting glutamatergic neurons (Elavl2/3, Qk, and Celf6) (Figure 3D). Interestingly, these RBPs also show differential activities between glutamatergic and GABAergic neurons, which differ in their range of projection.

Elavl2 modulates splicing of predicted targeted exons in MGE- but not CGE-lineage interneurons

Although transcription factors controlling MGE-versus CGE-specific gene expression have been identified,^9,10,50 RBPs regulating the AS differences between these lineages are still unknown. To demonstrate the utility of MR-AS to generate testable hypotheses on AS-regulatory programs in a specific cellular context, we decided to focus on Elavl2 (also known as HuB), which was predicted to have higher activity in MGE- than CGE-lineage interneurons (Figure 3C), consistent with its differential expression pattern quantified by scRNA-seq data from MGE and CGE neurons in vivo starting from early developmental stages^8,10,24 (Figure 4A).

Figure 4. Validation of Elavl2 as a key MGE-lineage-specific splicing factor.

(A) Mean Elavl2 expression RPKM (reads per kb per million) values in MGE- and CGE-lineage interneurons at various time points during cortical development.^8,10,24 The gene shows a consistent preferential expression in MGE-lineage neurons as early as embryonic day 12.5 (E12.5) and persists into adulthood.

(B) Construct designs for the MGE/CGE dual-reporter mouse ESC line. eGFP is positioned downstream of the MGE-specific Lhx6 promoter, and tdTomato is contained in an Ai9 reporter with Cre under the control of the CGE-specific marker 5ht3a. Ascl1/Mash1 overexpression is driven by the neural progenitor marker Nestin to promote interneuron differentiation.

(C) Experimental schema for testing the role of Elavl2 in interneuron-type-specific splicing regulation. Elavl2 was knocked out in the dual-reporter mouse ESC line using CRISPR-Cas9. WT and KO ESCs were differentiated into interneurons (ESC-INs) using an embryoid body-based protocol. On day 16 of differentiation, cells were isolated based on reporter fluorescence by FACS and RNA was isolated for RNA-seq.

(D) Centered exon inclusion (percent spliced in [PSI]) values of inferred Elavl2 targets ordered by predicted MOR. Inclusion differences of these exons in WT GFP+ (ESC-MGE cells) vs. tdTomato+ (ESC-CGE cells) samples and adult MGE-vs. CGE-lineage interneurons are shown at right.

(E) Scatterplots of mean exon inclusion differences (dPSI) in WT versus Elavl2 KO eGFP+ samples (x axis) compared to WT versus Elavl2 KO tdTomato+ samples (y axis). Positive and negative Elavl2 regulon exons predicted by MR-AS are indicated by red and blue dots, respectively.

(F and G) Overlap (F) and agreement in directionality (G) of DSEs in WT versus Elavl2 KO ESC-MGE cells and Elavl2 regulon exons predicted by MR-AS.

(H) Scatterplots of exon inclusion value differences in WT versus Elavl2 KO eGFP+ samples (x axis) compared to WT eGFP versus WT tdTomato+ samples (y axis).

(I) Agreement in directionality of DSEs in WT versus Elavl2 KO ESC-MGE cells and DSEs in WT ESC-MGE cells versus WT ESC-CGE cells.

(J and K) Genome browser views of Slit2 exon 31 (J) and Alcam exon 13 (K) inclusion in WT or KO ESC interneurons and adult MGE- or CGE-lineage neurons. Both exons are inferred targets of Elavl2. p values indicated in (F), (G), and (I) are calculated by Fisher’s exact test.

To confirm the predicted MGE-specific Elavl2 activity, we derived a mouse embryonic stem cell (ESC) line with two fluorescent reporter constructs to mark the MGE- and CGE-lineage cells when they are differentiated into interneurons (ESC-MGE and ESC-CGE) (Figures 4B and 4C). The transgenic line contained eGFP downstream of the Lhx6 promoter, which is specifically active in MGE-lineage cells, and tdTomato under the control of an Ai9 reporter allele driven by Cre downstream of the 5ht3a promoter, which has CGE-lineage-specific activity. To encourage neuronal differentiation, an Ascl1 (encoding Mash1) gain-of-function construct driven by Nestin was also incorporated (Figure 4B). We used CRISPR-Cas9 genomic engineering to generate a homozygous Elavl2 KO line from this parent line (Figures 4C and S7; see STAR Methods). Wild-type (WT) and Elavl2 KO lines were differentiated into ESC interneurons using an established embryoid-body-based differentiation protocol.^51,52 We then isolated the differentiated neurons using fluorescent activated cell sorting (FACS), collecting eGFP+ ESC-MGE and tdTomato+ ESC-CGE populations from both genotypes for RNA isolation, RT-qPCR, and RNA-seq (Figures 4C, S8A, and S8B). ESC-MGE and ESC-CGE WT and KO samples were enriched in previously known MGE- and CGE-lineage marker expression,^8,50 respectively (Figures S8C–S8E), indicating that the differentiated ESC interneurons had committed to the appropriate developmental lineages. ESC-MGE cells also showed higher expression of some MGE-specific RBPs we identified in adult, including Elavl2 (Figure S8F). Furthermore, although ESC interneurons are immature, we observed differential splicing between ESC-MGE and ESC-CGE cells, in directions consistent with those we observed in adult neocortex (Figures 4D and S9A; Table S4), indicating that the differential splicing-regulatory programs between MGE- and CGE-interneurons in vivo is recapitulated in the in vitro experimental system.

When we examined the impact of Elavl2 on neuron-type-specific splicing, we found that exons predicted to be activated by Elavl2 (with a positive MOR) show higher inclusion in ESC-derived or adult MGE cells, while exons predicted to be repressed by Elavl2 (with a negative MOR) show higher inclusion in ESC-derived or adult CGE cells (Figures 4D, S9A, and S9B). Importantly, consistent with a predicted higher activity, Elavl2 KO caused splicing changes in a much larger number of exons in ESC-MGE cells than in ESC-CGE cells (707 vs. 379 differentially spliced exons [DSEs], false discovery rate [FDR] <0.05, |ΔΨ|≥0.1; Figure 4E). Elavl2-dependent exons in ESC-MGE cells overlapped significantly with the predicted Elavl2 regulon or Elavl3/4 target exons previously identified in the mouse brain.⁴³ The direction of regulation also agrees very well, while this agreement is not seen in ESC-CGE cells (Figures 4F, 4G, S9C, and S9D). For example, upon Elavl2 depletion in ESC-MGE samples, exons predicted to be activated by Elavl2 mostly showed more skipping (88 out of 110 = 80%), while exons predicted to be repressed by Elavl2 showed more inclusion (55 out of 61 = 90.2%; p = 5.5e–20; Fisher’s exact test) (Figure 4G). Confirming our observation from the predicted Elavl2 regulon, exons showing reduced exon inclusion upon Elavl2 depletion in general have a higher inclusion in ESC-MGE cells, while exons showing increased exon inclusion on Elavl2 depletion in general have a higher inclusion in ESC-CGE cells (p = 3.2e–142, Fisher’s exact test) (Figures 4H and 4I); similar patterns were also observed when ESC-MGE Elavl2-dependent exons were compared with lineage-specific exons observable in adult neurons (Figures S9E and S9F). Together, these data validate that Elavl2 has a higher activity in the ESC-MGE cells than ESC-CGE cells and its depletion results in diminished differences between the two cell populations, supporting the hypothesis that Elavl2 plays a pivotal role in driving AS differences in MGE-versus CGE-lineage interneurons.

MGE- and CGE-lineage neurons differ in the migratory path they take during development and settle into different distributions among the cortical layers.⁵³ Exons predicted to be Elavl2 targets include those in genes that have roles in neuronal migration (Table S5). Among these, Slit2 exon 31, a predicted target activated by Elavl2, has much lower inclusion in KO than WT ESC-MGE samples; regulation of the exon by Elavl2 may thus underlie its higher inclusion in adult MGE-lineage interneurons (Figure 4J). Slit2 is part of the Robo-Slit pathway, which is critical for axon guidance and cell migration during development.^54–56 Another example, exon 13 of the Alcam gene, which encodes a cell adhesion molecule involved in migration, neurite extension, and axon guidance,^56,57 shows the opposite pattern of inclusion, consistent with its predicted negative regulation by Elavl2 in adult MGE-lineage interneurons (Figure 4K). Notably, these two targets also have clusters of U-rich elements as predicted Elavl2 binding sites in the upstream intron and downstream intron, respectively, consistent with the Elavl2 splicing-regulation RNA map.⁴³

Identification and validation of a splicing program differentially regulated in long- versus short-projecting neurons

As a second case study of MR-AS applications, we decided to test the impact of neuron-type-specific AS regulation on neuronal morphology, specifically the range of axonal projections, a property that can be regulated by mechanisms orthogonal to developmental cell lineages. In a recent study, we identified a subset of exons showing differential splicing between long-vs. short-range projecting neurons across multiple clades.¹⁵ However, their functional significance has not been validated, and the upstream regulators that drive differential AS are unknown. To address these questions, we first refined our analysis to identify a common module of 61 cassette exons that show differential inclusion between long- and short-range projecting neurons. The exons are differentially spliced between glutamatergic and GABAergic neurons (which mostly have long and short projections, respectively), long- and short-projecting glutamatergic neurons, and long- and short-projecting somatostatin (Sst)-positive interneuron types (Figures 5A and 5B; Table S6). The genes harboring these exons are enriched in Gene Ontology (GO) annotations related to the establishment or maintenance of axons including “cell projection organization” and cellular component terms related to axonal projections, including “presynapse,” “cell projection,” and “distal axon” (Figure 5C; Table S7).

Figure 5. Characterization and validation of an exon module specific to long- or short-projecting neurons.

(A) Top left: Venn diagram showing the overlap of exons differentially spliced between glutamatergic versus GABAergic neurons, long- versus short-projecting glutamatergic neurons, and long- versus short-projecting Sst interneurons. Bottom right: comparisons in the directionality of differentially spliced exons in the identified common exon between each of the neuron-type comparisons, including a comparison of ES-derived globus pallidus neurons (ESC-GPNs) and ES-derived interneurons (ESC-INs). p values are calculated by Fisher’s exact test.

(B) A heatmap showing the similar inclusion levels of the 61 exons identified in all three comparisons.

(C) Gene Ontology (GO) analysis of genes containing the 61 overlap exons shows an enrichment of genes related to axonogenesis and their maintenance.

(D) A schematic depicting Ptk2 exon 13, which is specifically included in short-axon neurons. The cartoon depicts the possible biological function of the exon, which contains a reverse calmodulin binding domain and may link the gene’s function to calcium signaling.

Most of the exons show consistent preferences between the comparisons for exon inclusion in long- or short-projection neurons (p = 3.6e–15, p = 3.8e–3, and p = 8.5e–3 in three pairwise comparisons by Fisher’s exact tests; Figure 5A, top three tables). Some reside in genes that suggest a long- or short-projecting neuron-type-specific tailoring of calcium-responsive gene function, such as those in calmodulin-dependent serine kinase II delta (Camk2d) and calmodulin-dependent serine protein kinase (Cask) (Table S6). A particularly interesting example is exon 13 of Ptk2 (also known as focal adhesion kinase [FAK]), which shows a strong preference for inclusion in short-projecting compared to long-projecting neuron types (Figure 5D). This alternative exon contains a reverse calmodulin (CaM) consensus sequence, a motif involved in binding CaM similarly to the forward sequence.^58,59 The differential inclusion of this exon raises the possibility that FAK signaling is modulated to allow activation by CaM in short-projecting neurons (Figure 5D). Among its many known roles in axon outgrowth, FAK has already been shown to modulate axonal branching in a calcium/CaM-dependent manner in hippocampal neurons.⁶⁰ It is possible that this mode of signaling is carried out by FAK isoforms containing exon 13, which are enriched in short-projecting neuron types.

To validate the functional impact of the AS module on neuronal projection, we utilized comparison of ESC interneurons and ESC-derived globus pallidus neurons (ESC-GPNs), modeling an intriguing pair of cell lineages that share a common developmental origin in the MGE but diverge into short-projecting cortical interneurons and long-projecting globus pallidus midbrain neurons, respectively. We recently demonstrated that, while overexpressing the transcription factors Nkx2.1 and Dlx2 in ES cells promotes interneuron differentiation, adding the overexpression of St18 is sufficient to drive them to acquire a globus pallidus-like state, including long and elaborated axons.⁵² We leveraged this system to investigate whether differential AS regulation may covary with neuronal projection length by examining the state of the putative long- or short-projection-specific AS module using RNA-seq data from the two cell populations. Importantly, we found that these exons showed consistent splicing switches in reprogrammed ESC-GPNs as compared to ESC-derived interneurons, as one would expect from the splicing differences between the other types of long- and short-range projecting neurons. The majority of exons with increased inclusion in ESC-GPN are more included in glutamatergic neurons as compared to GABAergic neurons, as well as in long-range projecting glutamatergic and GABAergic neurons, than their counterparts with short-range projection (p < 0.05 in all comparisons by Fisher’s exact tests; Figures 5A and 5B). This result provides an independent line of evidence that the identified splicing module is likely important for tailoring transcriptomes to the needs of establishing or maintaining long- or short-projecting neurons across diverse neuronal classes.

We next examined possible RBP splicing regulators of the long- or short-projection-specific AS module by comparing the estimated RBP activity between each of the long- versus short-projecting neuron types. Intriguingly, PCA of estimated neuron-type-specific RBP activity placed near-projecting glutamatergic and long-projecting Sst neurons in an intermediate state between those of the majority of the glutamatergic and the majority of the GABAergic neuron types along PC1, indicating large differences in RBP activity along this axis (Figures 6A and S10). RBP activity differences between each of the long- versus short-projecting neuron comparisons revealed a number of RBPs with consistent preference for a projection type, including Rbfox3, Khdrbs2, and Khdrbs3 for long-projecting neuron-specific activity and Qk, Srsf1, Elavl2, and Elavl3 for short-projecting neuron-specific activity (Figures 6B–6D and S10). The discriminative activity of these RBPs is also consistent with the direction of differential inclusion levels of their target exons common to the long- or short-projection-specific module. For example, exons activated by Rbfox3 and Khdrbs3 in general showed higher inclusion in long-projecting neurons, while exons activated by Qk and Elavl2 showed higher inclusion in short-projecting neurons; the opposite patterns were seen for exons repressed by these RBPs (Figures 6E and 6F). These data suggest that neuron-type-specific activity of these RBPs plays a role in driving the differential splicing programs. This case study again demonstrates the power of the inferred splicing-regulatory networks and RBP activity estimations to discover functionally relevant exon modules and the underlying regulatory mechanisms.

Figure 6. Inferred networks identify candidate drivers of projection-length-associated splicing.

(A) PCA plot of the first two principal components of RBP activity with neuron projection types indicated by color.

(B–D) Comparisons of differential RBP activity in long- versus short-axon neuron types correlate across different neuronal classes (Pearson).

(E and F) Concordance of candidate RBP regulons with the module of projection-length-associated exons identified in Figure 5. Each scatterplot shows the predicted MOR by the RBP on the x axis and change in exon inclusion between projection types on the y axis. Positive and negative target exons overlapping with the projection-length-associated module and that are differentially spliced between the groups are colored and counted in red or blue, respectively, and all other exons in the regulon are shown with reduced opacity. p values are from Fisher’s exact test.

DISCUSSION

Recent efforts on transcriptome profiling using deep sequencing have revealed an enormous diversity of AS among brain cell types,^{13,15,24,32,34,61,62} but technical and analytical challenges have limited the pace of progress elucidating the regulatory programs driving these patterns. Previously, splicing-regulatory networks are typically inferred by focusing on the regulons of individual RBPs, which can be interrogated using multifaced assays, such as mapping RBP-binding sites by CLIP or bioinformatic predictions of motif sites, and identifying exons with altered splicing upon perturbation of RBPs. These datasets can then be integrated (e.g., using Bayesian networks) to define RBP regulons with confidence.^40–42 Together with cell-type-specific RBP expression and exon splicing patterns, these regulons allowed us to evaluate glutamatergic or GABAergic neuron-specific activity for a select set of well-studied RBPs, such as Rbfox1-3, Elavl2-4, Mbnl2, Khdrbs2/3, and Qk.¹⁵ Despite the high accuracy of this approach, it involves substantial efforts to produce the required datasets for network inference, which is challenging to scale. Furthermore, experiments are also frequently performed in bulk tissues so that cell-type specificity of the regulons is difficult to delineate.

This study presents an alternative strategy to infer splicing-regulatory networks by correlating exon inclusion and RBP expression across a large panel (>100) of cell types with rich regulatory dynamics and then estimate neuron-type-specific RBP activity levels based on aggregated splicing patterns of the regulon for each RBP. A major advantage of this strategy is that it requires only full-transcript-coverage RNA-seq data, which are nowadays routinely obtained with advancements in deep-sequencing technologies. The effectiveness of this approach for reverse engineering of transcriptional regulatory networks and master-regulator analysis has been well established.^37–39,63 With the availability of similar scRNA-seq datasets that can simultaneously quantify RBP expression and exon splicing across more than 100 neuronal types, here we demonstrate that the same strategy, as implemented in MR-AS, can also be used to systematically reverse engineer splicing-regulatory networks and identify driving factors of neuron-type-specific splicing-regulatory programs. While full-transcript coverage is less common than 3′ end sequencing for single-cell approaches, it is becoming more routine due to new technologies and reduced costs, which would broaden the applicability of this approach. MR-AS can also be used on bulk RNA-seq datasets with sufficient samples and transcriptional diversity.

We present a comprehensive network composed of 8,336 cassette exons and 350 RBPs connected by 174,274 edges inferred from neocortical cell types in the adult mouse brain, with an average of 498 target exons per RBP, together with neuron-type-specific activities for each RBP. Multiple lines of evidence support the validity of the inferred splicing-regulatory network and cell-type-specific RBP activities. The inferred RBP targets overlap significantly with high-confidence target lists defined by integrative modeling using independent, multimodal datasets. They also exhibit enriched binding sites shown by CLIP and motif data in positions consistent with the direction of splicing regulation. Importantly, this is also true for targets that were not previously identified by the integrative modeling approach (Figure S4), suggesting discovery of bona fide targets. Overlapping RBP target exons predicted by both ARACNe and integrative modeling are highly concordant in the inferred direction of regulation (Figures 2A–2F). Confirming the accurate detection of differential RBP activity, estimated values are consistently lower in RBP-depleted samples than their WT counterparts using eight independent datasets (Figure 2G). Together, these data suggest that our analyses provide a reliable resource to generate testable hypotheses of neuron-type-specific splicing-regulatory programs.

This method allowed us to identify known and novel candidate regulators of neuron-type-specific splicing for well-defined neuronal subclasses based on their estimated RBP activity differences. To validate our predictions, we focused on Elavl2 showing MGE-lineage interneuron-specific activity. The Elav-like genes encode a highly conserved RBP family that have emerged as critical regulators of neural development and function. Three of the four family members, Elavl2, Elavl3, and Elavl4 (also known as HuB, HuC, and HuD), are selectively expressed in neurons throughout neuronal development, and they function as important splicing factors for proper neuronal differentiation and synaptic plasticity.^64–69 However, specific functions of individual factors on neuronal identity have not been elucidated. Elavl2, a risk gene for schizophrenia,⁷⁰ is expressed on a different timeline and in different cell types compared to Elavl3 and Elavl4. During neurogenesis, Elavl2 is expressed in neurons immediately after their birth in the ventricular zone, whereas Elavl3 and Elavl4 are expressed at later time points.^64,71 In the hippocampus, Elavl3 and Elavl4 show broad expression among neuron types, while Elavl2 is specifically expressed in CA3 pyramidal neurons and hilar interneurons.⁷² These observations imply potential functional differences among Elavl family members in different neuron types, supporting our hypothesis on the role of Elavl2 in MGE-lineage interneurons. Indeed, when we abrogated the expression of Elavl2 in an ES-derived model of MGE- and CGE-lineage interneurons, predicted Elavl2 targets showed specific splicing changes in ESC-MGE neurons as compared to ESC-CGE neurons such that differential splicing between the two cell populations became diminished, confirming our hypothesis. GO term enrichment analysis of genes containing these Elavl2-regulated exons suggests that the RBP may have a role in establishing distinct cell projection or migratory properties of MGE- and CGE-lineage interneurons in accordance with the needs of each lineage (Table S5). These results validate the utility of MR-AS in identifying candidate RBPs driving cell-type-specific AS-delineating canonical neuronal types.

Epigenetic and transcriptional regulation have been repeatedly demonstrated to play a dominant role in specifying developmental cell lineages.^{9,50,53,73–77} Intriguingly, while much of splicing varies by tissue in concordance with gene expression at the transcriptional level,^78–80 it has been a long-standing observation that tissue- or cell-type-specific genes are frequently regulated at the transcriptional and splicing levels in distinct patterns,^19,80,81 raising the possibility that AS provides an orthogonal regulatory mechanism to specify specific neuronal properties shared across developmental lineages.¹⁵ AS is uniquely situated to modulate genetic programs in response to various signaling inputs since post-transcriptional processing occurs at a shorter timescale compared to transcriptional regulation. For example, multiple neuronal RBPs have been shown to modulate AS in response to synaptic activity.^82–85 The neuronal projection-specific AS module we identified across several disparate neuronal clades is another possible example of orthogonal AS regulation to meet biological needs of particular cell types or states shared across different developmental lineages. We note that the identification of the long- or short-projection neuron-specific AS program required a wide diversity of cortical neurons to be sampled in both glutamatergic and GABAergic lineages for systematic comparison, and its generality would have been missed in an analysis using a more limited number of cell types. As groups of long- and short-projection neuron types are found separately within the larger classes of glutamatergic neurons and Sst-positive GABAergic interneurons, traditional lineage marker-based isolation of the groups to examine their AS differences would be a tedious and resource-intensive process. The MR-AS framework offers unique flexibility to compare groups of neuron types that differ in certain functional properties, regardless of their developmental lineages, and generate testable hypothesis of the underlying molecular mechanisms.

In summary, the differential AS programs we inferred between canonical neuronal subclasses and within multiple unrelated classes illustrate the utility of an unbiased, comprehensive mapping of cell-type-specific splicing-regulatory networks. As ever-larger and higher-resolution scRNA-seq datasets become available, we expect such analyses will likely reveal AS-regulatory programs both parallel and orthogonal to the transcriptional cell type hierarchy, a key aspect to understanding neuronal function in physiological and pathological conditions.

Limitations of the study

ARACNe-inferred networks may contain false positives, as indicated by their relatively moderate degree of overlap with previously defined integrative networks. While this may be an expense, the method pays for its wider applicability, and ARACNe likely also identifies targets missed by the previous approach, such as those that are regulated only in specific cell types that represent minor populations in bulk tissues, only by specific members of a splicing factor family, or by both direct and indirect regulatory mechanisms. RBP estimation by VIPER mitigates the effect of false positives on its predictions by aggregating the splicing patterns of all target exons in the regulon as a “multiplexed reporter assay.” Our previous study demonstrated that VIPER can tolerate a substantial degree of noise in the input network without compromising the inferred transcription factor activity estimates as long as the true targets are implemented using the correct regulatory directions.³⁹ This feature demonstrates the power of the method in master regulator analysis, now extended to splicing factors in MR-AS, in addition to studies of individual targets.

RESOURCE AVAILABILITY

Lead contact

Further information and request for cell lines may be directed to, and will be fulfilled by, the lead contact, Chaolin Zhang (cz2294@columbia.edu).

Materials availability

The dual-reporter ESC line generated in this study can be provided upon request.

Data and code availability

• Illumina sequencing data from the Elavl2 WT and KO ESC-MGE and ESC-CGE cells have been deposited to NCBI Short Read Archive (SRA). Their accession number and those of publicly available datasets analyzed in this paper are listed in the key resources table.

• The MR-AS software package is available as an all-in-one pipeline at Github: https://github.com/chaolinzhanglab/mras (https://doi.org/10.1002/advs.202414493). The analysis described in this paper applied the pipeline’s standard settings as described in the STAR Methods.

• Any additional information required to reanalyze the data produced in this paper is available from the lead contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
fetal bovine serum	Hyclone	Cat#SH30071.03E
ESGRO recombinant leukemia inhibitory factor	Sigma Millipore	Cat#ESG1107
XAV-939	Tocris	Cat#3748
SAG	Tocris	Cat#6390
Alt-R S.p. HiFi Cas9 Nuclease V3 protein	Integrated DNA Technologies	Cat#1081058
Critical commercial assays
Mouse Embryonic Stem Cell Nucleofector Kit	Lonza	Cat#VPH-1001
RNAeasy Micro Kit	Qiagen	Cat#74004
Deposited data
scRNA-seq from adult mouse visual cortex (V1) motor cortex (M1)	Tasic et al.8	SRA: GSE115746
RNA-seq from WT and Elavl2-KO ES-derived interneurons	This paper	SRA: PRJNA1237387
RNA-seq from ESC-interneurons and ESC-GPNs	Nunnelly et al.52	SRA: SRP329886
Experimental models: Cell lines
MGE/CGE dual-reporter ESCs (WT)	This paper	N/A
MGE/CGE dual-reporter ESCs (Elavl2-KO)	This paper	N/A
Experimental models: Organisms/strains
5HT3aR-BACCRE/+; Ai14/Ai14 mice	Jackson	RRID: MMRRC_036680-UCD; RRID: IMSR_JAX:007909
Lhx6-eGFP mice	MMRRC	RRID: MMRRC_000246-MU
Oligonucleotides
Elavl2 knockout synthetic guide RNAs: UGGAAACACAACUGUCUAAU UGCUCCUCACCAGUUGACUC GAGGAAGGUAGUUGACUAUU	Synthego	Gene Knockout Kit v2 - mouse - Elavl2 - 1.5 nmol
Full list of primer oligos in Table S8	Integrated DNA Technologies	custom order
Recombinant DNA
Nestin-Ascl1-ires-tTA22; TRE-Dlx2 (NAIT) construct	Modified in this paper from Au et al.51
Software and algorithms
OLego (v1.1.9)	Wu et al.86	https://github.com/chaolinzhanglab/olego
Quantas (v1.1.1)	Yan et al.87	https://github.com/chaolinzhanglab/quantas
forked ARACNe-AP	Margolin et al.37	https://github.com/chaolinzhanglab/ARACNe-AP
VIPER (v1.26.0)	Alvarez et al.39	https://bioconductor.org/packages/release/bioc/html/viper.html
MR-AS	This paper	https://github.com/chaolinzhanglab/mras
topGO (v2.48.0)	Alexa et al.88	http://bioconductor.org/packages/release/bioc/html/topGO.html
Other
Nucleofector II device (program A-24)	Amaxa/Lonza	N/A

STAR★METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Mice used in this study to generate dual-reporter ESCs were maintained at Columbia University Medical Center under standard conditions. The MGE and CGE dual-reporter ESC line was derived from super ovulated 5HT3aR-BACCRE/+; Ai14/Ai14 (MMRRC:036680-UCD; JAX 007909) females mated with Lhx6-eGFP (MMRRC 000246-MU) males and the line is female in sex. All experiments were carried out in accordance with Columbia University IACUC approved protocols.

Elavl2 knockout dual-reporter ESCs were derived from the parent dual-reporter line as described in the method details. Mouse ESCs were cultured at 37°C in 5% CO₂ for at least two passages on 0.1% gelatin and fed every 1-2 days with ESC medium consisting of DMEM (EMD Millipore SLM220B) supplemented with ESC-screened FBS (Hyclone SH30071.03E), 1x Modified Eagle Medium Non-Essential Amino Acids (Gibco), 1x Sodium Pyruvate (Gibco), 1x Glutamax (Gibco), 1x Penicillin/Streptomycin (Gibco), 10 μM beta-mercaptoethanol (Fisher Scientific), and 10⁴ U/mL ESGRO LIF (Sigma Millipore). ESCs were differentiated into interneurons as described previously.^54,69 The cells were not authenticated, and mycoplasma testing confirmed that the cells were free from contamination.

METHOD DETAILS

RNA-seq data preprocessing

Quantification of cassette exon inclusion and RBP expression using adult cortex scRNA-seq data for 133 cell clusters (cell types)(ref. 8, NCBI/GEO accession: GSE115746) was obtained from our previous study.¹⁵ Briefly, scRNA-seq reads⁸ were mapped to the mm10 genome using OLego.⁸⁶ Cells with low read count (those with <1000 exons with sufficient coverage for quantification) were excluded, leaving 19,858 of 21,154 core cells for our analysis. Aligned single-cell reads were pooled before quantification per cluster as defined by the original study authors,⁸ resulting in a median of 254.8 million reads per cluster (cell type). Gene expression and exon inclusion values were quantified using the Quantas pipeline (https://zhanglab.c2b2.columbia.edu/index.php/Quantas) as previously described⁸⁶ using a database of 16,034 annotated cassette exons.^40,42,87 Exon inclusion (percent spliced in, or PSI) was quantified using exon junction reads by requiring a junction read coverage ≥20 and estimated standard deviation <0.1 to reduce uncertainty. The reliability of the PSI estimates was evaluated extensively in the previous study.¹⁵ To be included in network inference using ARACNe, we require exons to be quantified in ≥25 clusters, and for the remaining 12,903 exons after this filtering, missing values of exon inclusion were imputed by the knn method in the impute package (k = 10) in R. For RBPs, we used 376 genes from RBPdb (http://rbpdb.ccbr.utoronto.ca/), supplemented by literature search that identified ~20 additional RBPs not included in RBPdb, such as Srrm4/nSR100.

scRNA-seq data of E12.5 and E13.5 postmitotic interneuron progenitors and from E17 cortical interneurons were obtained from previous studies^10,24 (SRA accessions: SRP131682 and SRP166440), and analyzed similarly using the Quantas pipeline. RNA-seq data of ESC-GPNs obtained from a previous study⁵² (SRA accession: SRP329886) and ESC-interneurons we generated for this study (SRA accession: PRJNA1237387) were also analyzed similarly using the Quantas pipeline.

ARACNe and regulon inference

Gene expression data have been used to reconstruct transcriptional regulatory networks in various biological contexts. ARACNe (algorithm for the reconstruction of accurate cellular networks)^37,38 and its companion algorithm VIPER (virtual inference of protein activity by enriched regulon analysis)³⁹ uses mutual information estimated across diverse cellular conditions to associate regulators with their targets, whose collective behavior in a sample is then used to estimate the regulator’s activity. This approach has the advantage of aggregating large numbers of targets to estimate a regulator protein’s activity rather than relying on single gene expression values of the regulators, which can be an unreliable approximation of activity. While it has been successfully used to infer the downstream target networks of transcription factors and signaling molecules and estimate cell state-specific regulator activity, this method has not been used to model AS regulation, which has to be optimized due to different noise structures in the scRNA-seq data at the splicing level.

We used ARACNe-AP^37,38 to infer splicing-regulatory networks at the cell type level based on the mutual information of RBP expression and target exon inclusion with several modifications. First, we did not apply data processing inequality (DPI) to eliminate potential indirect regulations. This is because estimation of mutual information is noisier for scRNA-seq data compared to bulk RNA-seq. Importantly, estimation of gene expression and exon splicing is subject to different levels of noise, which complicates DPI assessment. Second, we examined a set of well-studied RBPs to calibrate the number of targets predicted by ARACNe using varying mutual information thresholds. The number tends to vary dramatically for different RBPs. For results presented in the paper, we set mutual information p-value = 1e-8, while limiting the maximum number of targets per RBP to be 1000 for each bootstrapped network. We then consolidated 100 bootstrapped networks to keep only reproducible edges by Bonferroni-corrected Poisson p-value = 0.05. After removing RBPs with less than 25 inferred edges, this analysis resulted in a total of 174,274 edges in the inferred regulatory network between 350 RBPs and 8,336 cassette exons.

After the regulatory network was derived, we used the aracne2regulon function implemented in the R package “viper” to infer the mode of regulation (MOR, i.e., the probability of a target being activated or repressed by an RBP). MOR is calculated based on the Spearman rank correlation between the expression of an RBP and the inclusion of its target exons. A 3-component Gaussian mixture (representing activation, with positive correlation, repression, with negative correlation, and uncertain direction, with correlation around zero) was then used to infer MOR. The mutual information is also recorded to represent the likelihood of regulation.

In addition to cluster-level analysis, we also inferred the regulatory network by correlating RBP expression and exon inclusion at the single-cell level. We noticed that the regulatory network and MOR inferred at the cluster level provided the most accurate results, as judged by comparison with targets derived by integrative modeling or perturbation experiments.^40–42 Therefore, for the results presented in this study, we used the network inferred at the cell type level.

Estimation of RBP activity

We estimated RBP protein activity across 133 neuron types using regulons (i.e., regulatory networks and MOR) inferred at the cell type level. The analysis was conducted with the R package “viper” using the “scale” method, which first standardizes exon inclusion levels to z-scores across cell types, which effectively evaluates differential splicing between each sample and the average across all cell types used as a baseline. Then, RBP activity was estimated by analytic rank-based enrichment analysis (aREA), as described previously for gene expression regulation estimation.³⁹ Briefly, the analysis assesses the enrichment of regulon exon inclusion by incorporating two enrichment scores. Exon inclusion signatures are obtained for each sample by using quantile-transformed exon ranks based on their inclusion z-scores. The first enrichment score is calculated as the sum of absolute values of the regulon exon inclusion signature (a “one-tailed” approach), and the second as the sum of the inclusion signature for regulon exons inferred as positive targets and sign-inverted inclusion signatures for exons inferred as negative targets (a “two-tailed” approach). These two scores are then integrated, weighting the contribution of the one- and two-tailed enrichment scores based on the absolute value of each exons’ MOR and weighting the exons’ contribution based on its interaction confidence (mutual information) with the RBP. For each RBP regulon, this sum is weighted by the total magnitude of regulon interaction confidences, returning the normalized enrichment score used as the RBP activity estimate. The analysis was performed without pleiotropy correction, which was designed to punish enrichment of targets that are co-regulated by other regulators, since this feature led to an overly conservative set of inferred RBP targets contributing to the enrichment scores.

Gene ontology analysis

Genes containing exons of interest were tested for gene ontology (GO) term enrichment using the R package “topGO”. GO terms with a Bonferroni-corrected Fisher’s test p-value of ≤0.05 were shown in the figures.

MGE and CGE dual reporter ESC line construction

To derive the MGE and CGE dual-reporter ESC line, super ovulated 5HT3aR-BACCRE/+; Ai14/Ai14 (MMRRC:036680-UCD; JAX 007909) females were mated with Lhx6-eGFP (MMRRC 000246-MU) males, and late-stage blastocysts were harvested and cultured for outgrowth following established protocols.⁸⁹ Genomic DNA isolated from generated lines were genotyped for Cre, Ai14, or eGFP according to genotyping protocols for corresponding mouse lines. Lines with all alleles were further propagated and tested for faithful recapitulation of reporter expression following differentiation. Nestin-Ascl1-ires-tTA22; TRE-Dlx2 (NAIT) constructs were cloned by modification of constructs previously published⁵¹ using standard methods and introduced into the dual reporter ESC background by co-nucleofection with puromycin selection cassette and antibiotic selection, clonal isolation, and genotyping.

CRISPR/Cas9 genome engineering

Elavl2 knockout synthetic guide RNAs (sgRNA; Synthego multi-guide CRISPR Gene Knockout Kit v2) and Alt-R S.p. HiFi Cas9 Nuclease V3 protein (IDT, 1081058) were introduced into low-passage dual-reporter ESCs using the Mouse Embyronic Stem Cell Nucleofector Kit (Lonza VPH-1001). ESCs were treated with 0.025% trypsin/1% EDTA (Gibco) to form a single cell suspension in the appropriate kit reagents, electroporated with sgRNAs and Cas9 protein (Lonza Nucleofector, program A-24), and replated into 24 well plates. After 6 days in culture, they were sorted into 96 well plates at clonal density. Individual clones were then expanded for genotyping and screening for Elavl2 knockout by Sanger sequencing. Two clones containing the homozygous, frameshifting Elavl2 knockout were identified and expanded over the course of 4–5 passages before downstream experiments.

Neuronal differentiation of ESCs

ESCs were differentiated into ESC-interneurons as described previously.^54,69 Briefly, ESCs were dissociated using 0.025% trypsin/1% EDTA (Gibco) and suspended at a density of 50,000 cells/well in non-TC-treated 24 well plates in 1mL of differentiation media consisting of Glasgow’s Modified Eagle Medium (Gibco), 1x Penicillin/Streptomycin (Gibco), 1x Modified Eagle Medium Non-Essential Amino Acids (Gibco), 1x Sodium Pyruvate (Gibco), and 1x Glutamax (Gibco). The Wnt inhibitor XAV-939 (Tocris, 1.5uM) was added to this initial suspension to promote a telencephalic neural fate. On day 4, medium was replaced by the same mixture with both XAV-939 and sonic hedgehog agonist (SAG; Tocris, 0.1uM) to ventralize the cells, taking care not to disturb the newly formed EBs. On day 6 and every 3 days until day 15, medium was replaced with the mixture with SAG but not XAV-939.

ESC-interneuron FACS sorting, RNA collection and sequencing

On day 16 of ESC-interneuron differentiation, WT and Elavl2 knockout EBs were collected into Eppendorf tubes where the medium was replaced with papain solution containing 20 U/mL papain and 67 U/mL DNase (Worthington). Tubes were incubated for 45–60 min at 37°C under constant agitation. Papain solution was replaced with diluted ovomucoid inhibitor mixture with 67 U/mL DNase and gently triturated 10 times using M937 syringe needle tips. The supernatant cell suspension was then removed from residual EB clumps and layered over non-diluted ovomucoid inhibitor mixture and centrifuged at 100g for 6 min. Cell pellets were then resuspended in FACS sort buffer consisting of DPBS with no calcium or magnesium supplemented with 20mM HEPES (Thermo Scientific), 5mM EDTA (Thermo Scientific), 20 U/mL DNase I (Worthington), and DAPI (Sigma-Aldrich). The suspension was passed through a 70 μM cell strainer, and the viability of the single cell suspensions were confirmed to be at least 95% viable using a Countess II cell counter (Thermo Scientific). These single cell suspensions were then FACS sorted into eGFP+ and tdTomato+ populations (Figure S6) in differentiation medium supplemented with 5% FBS, using standard gating to remove debris, doublets, and nonviable cells. The collected cells were kept on ice and then resuspended in Trizol for RNA isolation. Samples were processed using either the standard Trizol (Thermo Scientific) protocol or using an RNeasy Micro Kit (Qiagen) using the aqueous RNA phase from samples centrifuged in Trizol. Isolated RNA was evaluated using a Bioanalyzer to confirm that RNA integrity numbers (RIN) were above 8. High-quality samples were submitted for RNA sequencing at the Columbia Genome Center, using TruSeq chemistry (Illumina) with poly-dT selection. Samples were multiplexed in each lane of a NovaSeq 6000 (Illumina) and 40 million 2 × 100 bp paired-end reads were generated per sample using RTA (Illumina) for base calling. RNA data were processed and analyzed as described above.

QUANTIFICATION AND STATISTICAL ANALYSIS

Statistical methods and software for gene expression and splicing quantification and downstream analyses are described in their respective results or method details sections. Details of the statistical tests conducted are indicated in the figure legend for each panel.

Supplementary Material

Supplemental information can be found online at https://doi.org/10.1016/j.celrep.2025.115898.

Highlights.

• MR-AS enables the inference of differential AS regulation across diverse cell types

• Inferred AS-regulatory map captures known and novel relationships

• Elavl2 regulates MGE-lineage interneuron-specific AS

• An AS program is associated with axon projection length across diverse lineages