An inducible system to study the regulatory functions of GSX2 in human lateral ganglionic eminence-like progenitors

Abstract

Animal models have demonstrated a critical role of the homeodomain transcription factor Genetic-Screened Homeobox 2 (Gsx2) in the developing basal ganglia. Moreover, recent clinical genetic studies have shown that GSX2 patient variants are associated with severe neurological symptoms and basal ganglia dysgenesis. Unfortunately, technical limitations with existing animal models, such as progenitor heterogeneity and limited temporal control, have impeded the investigation of direct regulatory targets. In this study, we engineered a Dox-inducible human embryonic stem cell (hESC) line to investigate the function of GSX2 in directed differentiation cultures that model developing lateral ganglionic eminence-like (LGE-like) progenitors. Transcriptomic, chromatin accessibility, and genomic binding studies revealed that GSX2: (1) binds both high- and low-accessibility chromatin using varying binding site preferences; (2) alters chromatin accessibility largely through indirect mechanisms; (3) functions primarily as a transcriptional repressor; and (4) regulates key conserved target genes that impact both neuronal progenitor maturation and regional specification. These results provide insight into the key regulatory roles and targets of GSX2, thereby establishing a new tractable experimental system to investigate basal ganglia development.

Summary Statement:

GSX2 represses key forebrain transcription factors related to regional specification and progenitor maturation.

Background:

Many transcription factors (TFs) play critical roles in the developing forebrain by regulating processes that include regional specification, progenitor maturation, and cellular migration (Acampora et al., 1995; Avilion et al., 2003; Corbin et al., 2000; Danesin et al., 2009; Hirata et al., 2006; Lagutin et al., 2003; Monuki, 2007; Pratt et al., 2000; Shinozaki et al., 2002; Sussel et al., 1999; Toresson et al., 2000; Ypsilanti et al., 2021). Embryonic brain structures such as the ganglionic eminences (GEs) that develop into the basal ganglia (e.g., striatum, globus pallidus, substantia nigra) utilize complex regulatory networks to specify cell types and control progenitor maturation (Carney et al., 2009; Catta-Preta et al., 2025; Flandin et al., 2011; Lindtner et al., 2019; Long et al., 2009; Sandberg et al., 2016; Su et al., 2022; Toresson et al., 2000; Wang et al., 2013; Xu et al., 2018; Zhao et al., 2022). Disruptions in these transcription networks can impair basal ganglia formation, with significant clinical implications (Bonneau et al., 2002; De Mori et al., 2019; Kato et al., 2004; Urel-Demir et al., 2024). One of the crucial TFs implicated in basal ganglia development is the Genetic-Screened Homeobox 2 (Gsx2) homeodomain factor (Corbin et al., 2000; Toresson and Campbell, 2001; Toresson et al., 2000; Waclaw et al., 2009; Wang et al., 2009; Yun et al., 2003; Yun et al., 2001).

During forebrain development, Gsx2 plays critical roles in establishing the pallial-subpallial (cortical-basal ganglia) boundary through repression of cortical factors (e.g., Pax6) and in specifying neurogenic precursors and their resultant neuronal subtypes in the lateral ganglionic eminence (LGE) (Corbin et al., 2000; Pei et al., 2011; Roychoudhury et al., 2020; Toresson and Campbell, 2001; Toresson et al., 2000; Tweedie et al., 2025; Waclaw et al., 2009; Yun et al., 2001). Accordingly, Gsx2 null mice show ventral expansion of cortical markers (e.g., Pax6 and Neurog2) and a significantly smaller LGE size, while recent human case studies have shown that homozygous GSX2 variants are associated with significant basal ganglia dysgenesis resulting in clinical phenotypes involving both movement and developmental disorders (De Mori et al., 2019; Urel-Demir et al., 2024). Thus, GSX2 has a conserved role in both forebrain patterning and promoting cell fates that contribute to the proper formation and function of the human basal ganglia and striatum.

At the molecular level, GSX2 functions as a homeodomain transcription factor that regulates target gene expression through direct DNA binding. Several lines of evidence suggest that GSX2, along with its paralog (GSX1) and homologues (Ind in Drosophila), function as sequence-specific transcriptional repressors. First, GSX factors encode two highly conserved domains: an N-terminal Engrailed homology-1 motif that binds Groucho/TLE co-repressor complexes that regulate chromatin accessibility (Brantjes et al., 2001; Chen et al., 1999), and a C-terminal homeodomain that binds AT-rich DNA sequences. Second, transcriptional reporter assays suggest that Ind and human GSX2 use multiple repression domains to inhibit transcription (Basu et al., 2020; Von Ohlen et al., 2007b). Third, ectopic expression studies in Xenopus found that fusing the Gsx2 homeodomain to the engrailed repression domain phenocopied wild-type Gsx2 over-expression, whereas ectopic expression of the VP16 activation domain fused to Gsx2 led to antimorphic phenotypes, implying that Gsx2 predominantly functions as a transcriptional repressor (Winterbottom et al., 2011). However, some studies have suggested that Gsx factors may also function as a transcriptional activator. Drosophila studies showed that Ind can directly auto-activate its own expression and induce ectopic reporter expression in heterologous protein experiments (Von Ohlen et al., 2007a; Von Ohlen and Moses, 2009). Additionally, recent mouse and Drosophila studies found that Gsx2 can mediate opposing transcriptional outcomes in a DNA binding site-dependent manner. This study revealed that Gsx2 binds DNA as both a monomer (TAATTA) and as a cooperative homodimer (TAAT – 7N – TAAT) (Salomone et al., 2021). Luciferase reporter assays using synthetic Gsx2 monomer or homodimer sites adjacent to UAS sequences that can be activated by Gal4-VP16 revealed context-dependent Gsx2 functions: repression of Gal4-VP16-mediated activation via monomer sites versus stimulation of Gal4-VP16-mediated activation via homodimer sites. However, the mechanisms underlying how Gsx2 stimulates Gal4-VP16-mediated activation via homodimer sites are unclear, as Gsx2 was not sufficient to activate gene expression in the absence of Gal4-VP16.

Consistent with Gsx2’s critical role as a transcription factor regulating LGE development, published murine Gsx2 null transcriptomic studies revealed significant dysregulation of numerous target genes, including many TFs (Salomone et al., 2021). This extensive dysregulation of TFs makes identifying direct versus indirect Gsx2 target genes challenging. For example, the existing literature has largely relied upon animal and cell models with limitations that confound investigating direct functional targets. First, the assayed tissues show significant cellular heterogeneity (Roychoudhury et al., 2020), representing cells that express a range of Gsx2 levels as well as more mature cells that no longer express Gsx2. Second, these models have limited temporal control of Gsx2 expression (Pei et al., 2011; Waclaw et al., 2009), complicating the investigation of the direct effects of Gsx2 before other TFs (e.g., ASCL1) become dysregulated. Thus, the existing literature underlines the importance of GSX2 in the developing basal ganglia, but leaves the unanswered question: What are the direct regulatory targets and functions of GSX2 in LGE progenitors? To address this question, we developed a doxycycline-inducible GSX2 human embryonic stem cell (hESC) system, together with a directed differentiation protocol, to study GSX2 function in an appropriate cellular context with near homogenous expression and precise temporal control. We show that GSX2 does not directly alter local chromatin accessibility but may do so indirectly through the transcriptional regulation of other TFs. Additionally, we found that GSX2 primarily acts as a transcriptional repressor of critical target genes that are conserved in both mouse and human and involved in both forebrain regional patterning and progenitor maturation.

Methods:

All commercially available reagents used throughout this study are provided in Supplemental Table 1.

Molecular Cloning:

The ^2xFLAGGSX2 construct was synthesized by GenScript (Piscataway, NJ) and the insert was PCR amplified using KOD Xtreme Hot Start Polymerase. To generate a AAVS1 plasmid with the ^2xFLAGGSX2 insert, we first used MluI and SalI digests to remove the EGFP insert from the donor AAVS1 TRE3G–EGFP plasmid (Qian et al., 2014). The ^2xFLAGGSX2 WT PCR fragment was then ligated into the AAVS1 TRE3G backbone using NEBuilder HiFi DNA Assembly kit. All new plasmids were confirmed by Sanger sequencing, and the final annotated sequences are provided in Supplemental Information.

Stem Cell Generation and Culturing:

Experiments using the human embryonic stem cell (hESC) line H9 (WA09) were performed under protocol EIP190101, approved by the Cincinnati Children’s Hospital Medical Center (CCHMC) Embryonic Stem Cell Research Oversight (ESCRO) Committee. The H9 line was obtained from the WiCell Research Institute under an appropriate license. All research was conducted in accordance with institutional guidelines, relevant regulations, and the principles outlined in the 2021 ISSCR Guidelines for Stem Cell Research and Clinical Translation.

Human embryonic stem cell (hESC) cultures were maintained on vitronectin-coated dishes in Essential 8 Flex Media as previously described (Tchieu et al., 2017). Cells for all experiments were used between passages 43–53 and cultures were passaged twice a week.

H9 hESC cells were edited using TALE nucleases (TALENs; Addgene) as previously described (Qi et al., 2017; Tchieu et al., 2017). Briefly, 20µg of AAVS1 donor plasmid and 5µg each of TALEN pairs were nucleofected into 4 million cells using the Lonza Kit V with program B-016. Clones were selected using puromycin one week after targeting and then tested for homogeneous transgene expression. Two suitable colonies were tested for pluripotency using immunohistochemistry and flow cytometry (see below). Lines expressing all markers were verified by Sanger sequencing through the Cincinnati Children’s Hospital Medical Center (CCHMC) Genomics Sequencing Facility (RRID:SCR_022630). Karyotyping for the two sequenced lines showed no significant translocations or duplications (46, XX). Lines were negative for mycoplasma contamination.

Directed Differentiation Assays:

Cells were differentiated using a modified dual SMADi protocol (Qi et al., 2017; Tchieu et al., 2017). Plates were coated with Matrigel diluted in DMEM/F12 overnight at 4°C and then seeded with 400–450K cells/cm² in Essential 8 Flex (E8 Flex) + 5µM Y-27632. Cultures were assessed for confluent monolayers ~24 hours later and then washed twice with Dulbecco’s PBS (DPBS). For the first three days, cells were treated with Essential 6 (E6) Media supplemented with 500nM LDN193189, 10nM SB431542 and 5nM XAV939. The same media without XAV939 supplementation was used for days 3–5, after which cells were treated with E6 supplemented with 2.5µM IWP2 and 25ng/mL Activin A. Media was changed daily on all cultures, or earlier for shorter timepoints (i.e., 12-hour Dox induction). Induced cultures were treated with media supplemented with 100ng/mL of doxycycline at each timepoint. Doxycycline was sterile filtered after reconstituting in nuclease-free water and used with two weeks.

Immunohistochemistry:

Cells were washed once with DPBS and then fixed in 4% PFA in DPBS for 10 minutes at 4°C. DPBS + 0.5% Triton X-100 was used to permeabilize cells for 5 minutes at room temperature and then cells were maintained in DPBS + 0.2% Tween-20 (DPBS-T) until blocking. Samples were blocked in DPBS-T + 5% Normal Donkey Serum (NDS) for 1 hour at room temperature and then incubated in primary antibodies overnight at 4°C. Primary antibodies were prepared in DPBS-T + 5% NDS as follows: goat anti-POU5F1 (OCT4; 1/250); rabbit anti-Nanog (1/250); rat anti-SOX2 (1/250); rabbit anti-Gsx2 (1/500); mouse anti-NKX2.1 (1/200); rabbit anti-ASCL1 (1/500); mouse anti-FLAG (1/1,000); rabbit anti-Dlx2 (1/500); sheep anti-ARX (1/1,000); goat anti-SP8 (1/500); rabbit anti-Six3 (1/500); goat anti-ISL1 (1/500); rabbit anti-FoxP1 (1/1,000); and rabbit anti-Cleaved Caspase 3 (1/100). Cells were washed three times with DPBS-T for 5 minutes at room temperature and then were incubated in secondary antibodies for 45 minutes at room temperature. Secondary antibodies were prepared in DPBS-T + 5% NDS as follows: donkey anti-mouse Alexa Fluor 488 (1/1,000); donkey anti-goat Alexa Fluor 488 (1/1,000); donkey anti-rabbit Alexa Fluor 488 (1/1,000); donkey anti-rabbit Alexa Fluor 555 (1/1,000); donkey anti-sheep Alexa Fluor 568 (1/1,000); donkey anti-rat Alexa Fluor 594 (1/1,000); and/or donkey anti-rabbit Alexa Fluor 647 (1/1,000). Samples were washed three times with DPBS-T, with 1 µg/mL DAPI supplemented in the second wash. Following the final wash, cells were maintained in DPBS-T until imaging.

Cleaved Caspase 3:DAPI ratios were quantified using three images per treatment group and assessing three fields per image using Fiji (Schindelin et al., 2012). DAPI-positive nuclei and Cleaved Caspase 3-positive cells were manually counted, and then apoptotic cells (Cleaved Caspase3-positive) cells were normalized to total cells per field (DAPI+ nuclei).

Flow Cytometry:

All samples were washed once with DPBS, dissociated with Accutase in 37°C incubator for 5–30 minutes and then treated as follows for cell surface and intracellular markers:

Cell Surface:

Live cell pellets were resuspended in primary antibody solutions (Staining Buffer: DPBS, 0.09% Sodium Azide, 1% FBS): mouse anti-TRA-1-60 PE-conjugated (2 µL/reaction); mouse IgM Isotype Control PE-conjugate (2 µL/reaction); mouse anti-SSEA-4 Alexa Fluor 488-conjugated (2 µL/reaction); and mouse IgG3 Isotype Control Alexa Fluor 488-conjugated (2 µL/reaction). Cells were incubated for 30 minutes at 4°C and then washed twice with Staining Buffer. After the final wash, cells were resuspended in staining buffer and maintained at 4°C until acquisition.

Intracellular:

~500K live cells were spun down and resuspended in Human FcR Blocking Solution (1µg * 10⁶ cells) for 10 minutes at room temperature. Cell pellets were resuspended in LIVE/DEAD^™ Fixable Blue Dead Cell Stain Kit for 30 minutes at 4°C. Cells were washed once in Cell Staining Buffer and then fixed using the True-Nuclear Transcription Factor Kit for 1 hour at room temperature. After a single wash, cells were resuspended in primary antibody mixtures on ice for 30 minutes: rabbit anti-ASCL1 (0.1 µL * 10⁶ cells), mouse anti-FLAG (0.1 µL * 10⁶ cells), and rabbit anti-Gsx2 (0.05 µL * 10⁶ cells). After two washes, cells were resuspended in secondary antibody mixtures on ice for 20 minutes: donkey anti-rabbit Alexa Fluor 647 (1/10,000) and/or rat anti-mouse IgG1 PE (1/10,000). Cells were washed twice and then maintained in permeabilization solution at 4°C until acquisition.

All flow cytometric data were acquired using a BD LSRFortessa maintained by the Research Flow Cytometry Facility in the Division of Rheumatology at Cincinnati Children’s Hospital Medical Center. Data was processed using FlowJo^™ (Becton).

Chromatin Immunoprecipitation Sequencing (ChIP-Seq):

At harvest, wells were washed once with room temperature DMEM/F12 and then treated with Accutase for 30 minutes at 37°C. Each well was supplemented with DMEM/F12 + 10% Fetal Bovine Serum (FBS) and homogenized. Eight wells were combined for each biological replicate and counted using a manual hemocytometer. Thirty million cells were aliquoted and fixed at room temperature for 10 minutes using 1% Formaldehyde diluted in DPBS. Fixation was quenched using glycine to a final concentration of 125mM and cells were washed twice with ice-cold DPBS + 1% Bovine Serum Albumin (BSA). Cell pellets were snap frozen and stored at −80°C until sonication.

Fixed cells were lysed on ice for 10 minutes (10mM Tris-HCl, 10mM NaCl, 0.50% NP-40, 1x cOmplete Protease Inhibitor; ice-cold) to obtain nuclei, which were subsequently lysed for a further 10 minutes on ice (50mM Tris-HCl, 10mM EDTA, 0.32% SDS, 1x cOmplete Protease Inhibitor; ice-cold). Chromatin was sheared using a Covaris S220 Sonicator (200 Cycles, duty factor 10, 3.5 minutes). Each immunoprecipitation (IP) reaction used 50µL of Dynabead Protein G magnetic beads washed twice (DPBS + 0.02% Tween-20) and then incubated on a rotator with mouse M2 anti-FLAG antibody (1.5µg/reaction) for four hours at 4°C. Sheared chromatin was diluted (40mM Tris-HCl, 56mM NaCl, 7mM EDTA, 0.40% Triton X-100, 0.20% SDS; ice-cold) and 2% volume was taken as input. The antibody-conjugated beads were incubated with 15µg of diluted chromatin overnight at 4°C on a rotator and then washed for three minutes once each with: FA Lysis (50mM HEPES KOH, 150mM NaCl, 2mM EDTA, 1% Triton X-100, 0.10% sodium deoxycholate); NaCl Buffer (50mM HEPES KOH, 500mM NaCl, 2mM EDTA, 1% Triton X-100, 0.10% sodium deoxycholate); LiCl Buffer (100mM Tris-HCl, 500mM LiCl, 1% NP-40, 1% Sodium Deoxycholate); and Tris Buffer (10mM Tris-HCl).

Chromatin-protein complexes were eluted over three rounds (TE Buffer + 1% SDS) in a 65°C thermomixer at 800 rpm for 15 minutes/round. Chromatin was reverse-crosslinked at 65°C for eight hours in 200mM NaCl, treated with RNase A (9 µg/reaction) for 1 hour at 37°C and then Proteinase K (38 µg/reaction) for 2.5 hours at 37°C. Chromatin was extracted using Phasemaker tubes via a phenol/chloroform extraction and then precipitated overnight at −20°C (32 µg/mL Glycogen, 64% ethanol). The DNA pellet was washed with 80% ethanol and resuspended in TE Buffer. Libraries were prepared using 20ng of chromatin with the NEBNext Ultra II DNA Library Prep Kit for Illumina and Multiplex Oligos for Illumina kits. Agilent Tapestation results were used for QC and libraries were sent to Novogene (Sacramento, CA) for paired-end 150bp sequencing on a NovaSeq X instrument.

ChIP-seq Analysis:

Quality-controlled FASTQ files for each dataset were cleaned using Trim Galore version 0.6.6 (Krueger, 2020) to remove remaining adapter sequences and trim low-quality reads. Reads were aligned to the hg38 genome using BowTie2 version 2.4.2 (Langmead and Salzberg, 2012), and duplicates were removed using Samtools version 1.18.0 (Danecek et al., 2021). Reads overlapping blacklisted regions (Amemiya et al., 2019) were removed using Bedtools version 2.30.0 (Quinlan and Hall, 2010) and peaks were called against 2% Input DNA control using MACS 3.0.2 (Zhang et al., 2008), with a p-value cutoff of 1 × 10⁻³. The consensus peak set was defined as peak presence in ≥2/3 biological replicates. Bigwig files were generated using deepTools version 3.5.6 (Ramirez et al., 2016).

Assay for Transposase-Accessible Chromatin Sequencing (ATAC-seq):

Cells were harvested as in the ChIP-seq experiment and 50,000 – 100,000 live cells were aliquoted. Samples were processed as previously described (Corces et al., 2017), with minor modifications. Lysed nuclei were incubated in 50–100µL of transposase mixture in a thermomixer at 800 rpm for 30 minutes at 37°C. Reactions were purified with Qiagen MinElute PCR Purification kit and then PCR-amplified for 6–11 cycles using the NEBNext Ultra II Q5 PCR Mastermix (72°C 5 min, 98°C 30 sec, (98°C 10 sec, 65°C 90 sec)*per cycle). The optimal number of cycles for each sample was determined by qPCR using SYBER Green to define the number of cycles required to achieve ¼ of maximal intensity. Libraries were purified using the Qiagen MinElute PCR Purification kit and an Agilent Tapestation was used for QC. Libraries were sequenced through the Genomics Sequencing Facility at CCHMC using a NovaSeq X for paired-end 100bp reads.

ATAC-seq Analysis:

Files for each dataset were preliminarily processed as in ChIP-seq analysis, using HMMRATAC for peak calling (Tarbell and Liu, 2019), and high-quality peaks were selected as having a score ≥10. The Day 10 and 12-hour datasets combined triplicate samples from two differentiations. Consensus peak sets for each group required presence in at least 2/3^rd of biological replicates and were then merged into a global peak set for differential analysis.

RNA-Sequencing:

At each timepoint, wells were washed twice with DPBS, mirVana miRNA Lysis Buffer was added to each monolayer and cells were incubated at 37°C for 5 minutes. Lysates were homogenized and transferred to cryovials for storage at −80°C until transfer to the Genomics, Epigenomic and Sequencing Core at University of Cincinnati.

Total RNA was extracted according to the mirVana miRNA protocol, RNA concentration was measured by nanodrop, and RNA integrity was determined using a Bioanalyzer (Agilent, Santa Clara, CA). Directional poly-A RNA-seq was performed as previously described (Qiu et al., 2023; Reigle et al., 2021), with minor modifications. Poly-A transcripts were enriched using the NEBNext Poly(A) mRNA Magnetic Isolation Module with 500ng of high-quality total RNA as input. Libraries were prepared using the NEBNext Ultra II Directional RNA Library Prep Kit with unique dual indices for 9 total PCR cycles. After QC and Qubit quantification, normalized libraries were sequenced using a NextSeq 2000 instrument for pair-end 61bp reads.

Modified hg38 FASTA and GTF files were generated by adding the inserted transgenic region as an additional chromosome. Quality-controlled FASTQ files were trimmed using Sickle version 1(Joshi and Fass, 2011) and aligned to the modified hg38 genome using STAR version 2.7.9 (Dobin et al., 2013).

Differential Expression/Accessibility Analyses:

Subread version 2.0.3 (Liao et al., 2014) was used to generate count matrices for differential analyses. All differential analyses used custom R code with DESeq2 (Love et al., 2014). The count matrix for the RNA-seq analysis used the modified hg38 GTF file and aligned to reverse-stranded exons, whereas the ATAC-seq count matrix used the global peak set. Analysis design used an interaction term of timepoint and Dox induction status to investigate the changes in Dox treatment over time. Pairwise comparisons and volcano plots used the Wald method (RNA-seq: |LFC|≥1 & FDR<0.05; ATAC-seq: |LFC|≥0.58 & FDR<0.01). Differentially expressed genes/accessible regions for Ward.D2 hierarchical clustering were defined using the Likelihood Ratio Test (LRT) comparing the design including the timepoint:Dox interaction term to a reduced formula (RNA-seq: FDR<0.05; ATAC-seq: FDR<0.01). The optimal number of clusters was empirically determined.

Gene Ontology (GO) Term Enrichment Analysis:

All GO analyses were conducted with the enrichR R package (Chen et al., 2013; Kuleshov et al., 2016) in R using the 2025 databases. An adjusted p-value < 0.05 was used to determine significance and top terms were provided in figures, unless otherwise specified.

Single Cell RNA-seq (scRNA-seq) Analysis:

A published reference atlas for first trimester human fetal brain scRNA-seq (Braun et al., 2023) was intersected with bulk RNA-seq datasets using Seurat v5.3.0 (Hao et al., 2024), adapting a provided vignette on Mapping and Annotating Query Datasets. Briefly, a subset of the reference atlas was processed using Sketch-based analysis (Hie et al., 2019), and then the provided atlas metadata was used to subset appropriate cell populations. We selected for neural cell types (CellClass = Radial glia, Neuronal IPC, Neuroblast, Neuron) and the earliest timepoint containing forebrain regional information (Age = 6.90000009536743 p.c.w = Carnegie Stage 20). The bulk RNA-seq count matrix was normalized using DESeq2, and then converted into a Seurat object, treating each bulk sample as a single cell. After standard Seurat normalization, the bulk samples were projected onto the reference atlas, followed by UMAP visualizations.

For the module score analyses, DEG lists for up- and down-regulated genes were taken from the timepoint ± Dox pairwise comparisons (discussed above). The resulting features were then plotted using Seurat’s FeaturePlot.

Motif Analysis:

HOMER version 4.9 was used for de novo motif analyses (Heinz et al., 2010), searching a 200bp region around peak centers. ChIP-seq analysis scanned for 6, 12 and 20bp motifs, while ATAC-seq analyses assessed for 10 and 20bp motifs. The motif scan analyses used the GSX2 monomer and homodimer PWMs generated by the de novo enrichment analysis on the consensus ChIP-seq peaks. The motif scan output files were processed using custom R code.

Murine Genomics Datasets:

Embryonic day 12.5 (E12.5) mouse LGE datasets for anti-^2xFLAGGsx2 CUT&RUN and bulk RNA-seq datasets of Gsx2^+/+ (wild-type) and Gsx2^EGFP/RA (germline KO) were obtained from a previous publication (Salomone et al., 2021). The published peak set for the ^2xFLAGGsx2 CUT&RUN was used and genes for the RNA-seq were defined as significantly different based on |LFC|≥0.58 and FDR<0.05 from the published differential analyses. E13.5 anti-Gsx2 ChIP-seq FASTQ from a previous publication (Catta-Preta et al., 2025) was processed similar to the human ChIP-seq datasets (discussed above), aligning to the mm10 genome and using a p-value cutoff of 5 × 10⁻² for MACS3 peak calling.

Genomics Integration:

Custom scripts were used to integrate the genomic binding (ChIP-seq), transcriptomic (RNA-seq) and chromatin accessibility (ATAC-seq) datasets. Murine and human RNA-seq DEG lists were intersected for commonality based on gene symbols. ChIP- and ATAC-seq datasets were intersected using Bedtools, segregating ChIP-seq peaks based on ≥1bp overlap with any ATAC-seq peak. ChIP-seq peak accessibility was defined through intersection with the 12-hour +Dox ATAC-seq group consensus peak set. Bigwigs and heatmaps of read enrichment around peak centers were generated using deepTools. Bigwigs were normalized using 1x reads per genome coverage (RPGC). ChIP-seq peaks were annotated to the genome using ChIPseeker package (Wang et al., 2022), either to the nearest transcription start site (TSS) or in a set window around each peak.

Data Visualization:

Genome visualizations were generated using Integrative Genomics Viewer (Robinson et al., 2011) with the provided hg38 reference genome. Hierarchical clustering heatmap was generated using the ComplexHeatmap package in R (Gu, 2022; Gu et al., 2016). Unless otherwise specified, plots and graphs were generated using ggplot2 (Wickham, 2016) with viridis color palettes (Garnier et al., 2024) in R.

Results:

Generation of human cell lines to study inducible GSX2 expression in LGE-like progenitors

Previous genomic studies focused on Gsx2’s role in basal ganglia gene regulation using germ line knock-out mouse models have significant cellular heterogeneity and a lack of temporal control of Gsx2 action. To address these limitations and study human GSX2 function, we modified the H9 (WA-09) human embryonic stem cell (hESC) line by inserting a doxycycline- (Dox-) inducible cassette driving ^2xFLAGGSX2 expression at the AAVS1 safe-harbor locus using TALENs (Fig 1A). Two independent transgenic lines were generated and sequence verified by Sanger Sequencing (Supp Fig 1A). Flow cytometry of edited lines confirmed homozygous expression of the cell surface markers TRA-1-60 and SSEA-4 (Supp Fig 1B) and immunofluorescence analysis further showed high expression of the intracellular pluripotency markers NANOG, POU5F1 (OCT4) and SOX2 (Supp Fig 1C). To demonstrate the responsiveness and homogeneity of our Dox-inducible system, we used flow cytometry in undifferentiated hESC cultures to show homogeneous transgene expression within 4 hours of Dox induction (Supp Fig 1D).

Figure 1: Generation of a hESC line carrying a Dox-inducible GSX2 expression cassette capable of robust transgene expression within LGE-like progenitor cells.

(A) TRE3G Dox-inducible cassette driving ^2xFLAGGSX2 transgene expression was inserted at the AAVS1 safe-harbor locus on chromosome 19 of H9 hESCs. (B) Modified dual SMAD inhibition directed differentiation protocol (Qi et al., 2017; Tchieu et al., 2017). Supplementation with TGF-β agonist (Activin A) and WNT antagonist (IWP2, PORCN inhibitor) starting at Day 6 induces endogenous GSX2 expression between Days 10 and 12. (C-D’’’) TRE3G – ^2xFLAGGSX2 WT hESCs differentiated as in Fig 1B show heterogeneous expression of endogenous GSX2 at Day 12. Absence of significant NKX2.1-expressing cells suggest cultures are not being patterned to a more ventral MGE-like identity. (E) Comparing Dox induction paradigms of pulsatile (12hr +Dox) and continuous (48hr +Dox) GSX2 induction versus non-induced controls (Days 10 and 12). (F-I’’’) FLAG expression from the ^2xFLAGGSX2 WT transgene expression is only detectable upon Dox supplementation (100 ng/mL), which induces robust transgene expression detectable for at least 36hrs after induction removal (H’’). Continuous Dox treatment induces near homogeneous ^2xFLAGGSX2 expression (I’’). In non-induced cultures, ASCL1 expression is minimally detectable at Day 10 but is expressed in scattered cells by Day 12. A 12-hour pulse of transgene induction shows moderate effects on ASCL1 expression at Day 12, while continuous induction (i.e., 48-hours) shows complete absence of detectable ASCL1 protein. Scale bars represent 100µm. Abbreviations: HA = homology arm; TRE3G = Tet-On 3G promoter; CAG = CMV-chicken β-actin composite promoter; E6 = Essential 6 Media; Dox = Doxycycline

To study GSX2 function in ventral forebrain neural progenitors, we adapted an existing dual SMAD inhibition (dSMADi) directed differentiation protocol that is commonly used to generate dorsal telencephalic fates (Qi et al., 2017; Tchieu et al., 2017). By supplementing Essential 6 media with Activin A (TGF-β agonist) and IWP2 (PORCN Inhibitor, WNT antagonist) starting at Day 6 (Fig 1B), we found that endogenous GSX2 (LGE marker) was minimally detectable at Day 10 (Fig 1C–C’’’) but showed heterogeneous expression by Day 12 (Fig 1D–D’’’). In contrast, we did not observe significant expression of the medial ganglionic eminence (MGE) marker NKX2.1 at either timepoint (Fig 1C’’, 1D’’). To further characterize this directed differentiation system, we stained against additional GE markers known to be downstream of Gsx2 in murine LGE development (Fig 1F’’’–1G’’’; Supp Fig 2). The bHLH factor ASCL1 (Fig 1F’’’, Fig 1G’’’), homeodomain factors ARX (Supp Fig 2A) and DLX2 (Supp Fig 2A), and the LIM-homeodomain factor ISL1 (Supp Fig 2B), all show increasing, heterogeneous expression between Days 10 and 12. The zinc finger factor SP8 and homeodomain factor SIX3 show relatively weak ubiquitous expression throughout the cultures (Supp Fig 2B–C). The absence of the late striatal marker FOXP1 (Supp Fig 2C), suggests the assayed timepoints represent early/intermediate progenitor populations. Together, these data indicate that this directed differentiation protocol patterns cells to a LGE-like progenitor state between Days 10 & 12.

Based on these findings, we selected Day 10 to examine the impact of GSX2 induction within LGE-like neural progenitors, as these cells have not begun to significantly express LGE markers but are primed to do so within the next 2 days. To establish GSX2 induction conditions in these LGE-like progenitors, we treated Day 10 cultures with doxycycline (Dox, 100 ng/mL) for variable lengths of time (Fig 1E) and performed immunostaining for the FLAG epitope tag at Day 12. These studies revealed the following: (1) as expected, no significant transgene expression was detected in non-induced cultures (Fig 1F’’, Fig 1G’’); (2) transgenic GSX2 protein persisted for at least 36 hours after a 12-hour induction (Fig 1H”); and (3) robust, near homogeneous transgene expression was observed under continuous Dox treatment (Fig 1I”).

To define the impact of GSX2 expression on these progenitor cells, we performed immunostains for a cortical-enriched (PAX6) and basal ganglia-enriched marker (ASCL1). First, we found that GSX2 induction resulted in the down-regulation of the dorsal telencephalic cortical marker PAX6 in Day 5 cultures (Supp Fig 3A). This finding is consistent with murine studies showing that Gsx2 establishes the early pallial-subpallial boundary, at least in part, by repressing Pax6 (Corbin et al., 2000; Toresson et al., 2000; Yun et al., 2001). Second, we found that in non-induced LGE-like cultures, ASCL1 expression was nearly undetectable at Day 10, but by Day 12, a timepoint in which heterogeneous endogenous GSX2 expression was observed, scattered ASCL1-positive cells were found throughout the culture (Fig 1E’’’, Fig 1F’’’). We further found a mild increase in the proportion of ASCL1-positive cells after 12 hours of Dox-induction relative to the time-matched non-induced control (Fig 1F’’’, Fig 1H’’’), as confirmed by flow cytometry against ASCL1 (Supp Fig 3B). In contrast, ASCL1 expression was minimally detectable after 48 hours of continuous GSX2 induction (Fig 1I’’’). Staining for the apoptotic marker Cleaved Caspase 3 (Supp Fig 3C–D) showed no significant difference in cell survival between Dox-induced and control cultures at Day 12, suggesting the absence of ASCL1 expression in the continuously-induced cultures is not due to cell death. These results suggest that moderate GSX2 concentrations mildly increase ASCL1 expression and thus promote LGE-like differentiation, whereas high levels of continuous GSX2 protein inhibits ASCL1 expression and LGE progenitor maturation, consistent with GSX2’s proposed role in progenitor maintenance (Pei et al., 2011). Taken together, these results demonstrate that this inducible hESC system provides a robust, temporally-controlled platform to study GSX2 regulatory functions in a human LGE-like context.

GSX2 expression induces transient, indirect changes in chromatin accessibility

Using the LGE-like differentiation platform, we performed a series of genome-wide studies across multiple timepoints to investigate the regulatory targets and roles of GSX2 (Fig 2A). First, we performed triplicate ChIP-seq using an antibody against the FLAG epitope tag on cultures induced with Dox for 12 hours starting at Day 10. Analysis of the anti-FLAG datasets shows strong, reproducible enrichment against 2% Input controls around key regulatory targets, such as the PAX6 locus (Fig 2B). Genomic annotation of a consensus peak set, requiring overlapping peaks in at least two of the three biological replicates (N = 7,162), showed that most peaks are located within promoter/intronic regions near neurogenesis-associated genes (Fig 2C–D). Importantly, as in published murine embryonic day 12.5 (E12.5) anti-^2xFLAGGsx2 CUT&RUN and E13.5 pan-basal ganglia anti-Gsx2 ChIP-seq datasets (Catta-Preta et al., 2025; Salomone et al., 2021), human ChIP-seq HOMER de novo motif analysis significantly enriched for both the GSX2 monomer and homodimer motifs as the top two motifs (Fig 2E). Together, these data support the quality and reproducibility of the ChIP-seq data, and its suitability for intersection with additional datasets.

Figure 2: ^2xFLAGGSX2 ChIP-seq datasets show enrichment for monomer and homodimer sites near neurogenic genes.

(A) Schematic of the longitudinal differentiation protocol used to investigate GSX2 genomic binding and impact on chromatin accessibility and gene expression. (B) ^2xFLAGGSX2 ChIP-seq peaks primarily localize near gene bodies (i.e. promoters, introns, etc.). (C) Replicate samples show reproducible enrichment at many loci. Example shown is the genomic region that spans the PAX6 locus, which displays strong enrichment from background (2% Input). (D) The nearest TSS’s of consensus GSX2 peaks enrich for neurogenic gene GO terms. (E) HOMER de novo motif enrichment of monomer (TAATKA) and homodimer (TAAT – 7N – TAAT) DNA-binding motifs is consistent with published Gsx2 genomic binding datasets (Catta-Preta et al., 2025; Salomone et al., 2021). Abbreviations: TSS = Transcription Start Site; UTR = Untranslated Region; GO = Gene Ontology; PWM = Position Weight Matrix; % of Bg = Percent of Background

To evaluate if GSX2 genomic binding is associated with open versus closed chromatin, we performed bulk ATAC-seq experiments after a 12-hour pulse of transgene induction (Fig 2A). Of note, we refer to ATAC-seq peaks as ‘high-accessibility’ (i.e., open) regions and those regions not overlapping a peak as ‘low-accessibility’ (i.e., closed). Intriguingly, >80% of consensus GSX2 ChIP-seq peaks were in low-accessibility regions at the 12-hour timepoint (Fig 3A). To confirm this finding, we performed comparative read enrichment analysis of the ChIP-seq and ATAC-seq data centered on each consensus GSX2 ChIP-seq peak, with the 1,404 highly accessible sites at the top and the 5,738 low accessibility regions at the bottom (Fig 3B). This analysis revealed that the overall GSX2 ChIP-seq read enrichment was highly similar between the high versus low accessible sites (Fig 3B). Interestingly, de novo motif analysis on the GSX2-bound high- and low-accessibility regions revealed that while the GSX2 monomer site was enriched in both regions, the GSX2 homodimer motif was selectively enriched in only low-accessibility peaks (Fig 3C). Further analysis of the motifs enriched in the low- versus high-accessibility peaks did not identify any additional obvious non-homeodomain motifs (Fig 3C). These findings suggest that GSX2 binds genomic regions of varying chromatin accessibility, with distinct monomer versus homodimer DNA-binding motifs potentially mediating this differential targeting.

Figure 3: ^2xFLAGGSX2 binding does not directly alter local chromatin accessibility.

(A-B) At the 12hr +Dox timepoint, GSX2 primarily binds regions with low-accessibility as defined by overlap with ATAC-seq peaks and heatmaps of ATAC-seq read pile-up centered on ChIP-seq peaks. (C) ChIP-seq peaks in low- vs. high-accessibility regions show differing enrichment for DNA-binding motifs in HOMER de novo motif analysis. Only peaks in low-accessibility regions show enrichment for homodimer motifs (-TAAT-7N-TAAT-), whereas both subsets enrich for monomer motifs (-TAATTA-). (D-E) Volcano plots of differentially accessible peaks at 12- and 24-hours ± Dox (|LFC| ≥ 0.58, FDR < 0.01) shows that GSX2 expression initially reduces chromatin accessibility, followed by a transient increase. Note, 48-hour ± Dox analysis showed no significantly differentially accessible peaks. GSX2 binding in high-accessibility regions is not associated with significant changes in chromatin accessibility (~92% unchanged). Abbreviations: PWM = Position Weight Matrix; HD = homeodomain; HMG = high-mobility group; RFX = regulatory factor X; ZF = zinc finger; RMM = RNA recognition motif

Next, we utilized the ATAC-seq data to determine if GSX2 significantly alters chromatin accessibility. Direct pairwise comparisons of the ATAC-seq peaks from the 12-hour ± Dox samples revealed that GSX2 expression results in a general reduction in chromatin accessibility, with ~3,700 regions showing decreased accessibility, whereas only ~200 regions showed increased accessibility (Fig 3D). To determine if the regions with altered chromatin accessibility were associated with direct binding of GSX2, we performed two additional analyses. Direct comparison of the ATAC sequence read depth from the 12-hour ± Dox treated samples revealed that GSX2 binding to highly accessible regions caused a moderate, but statistically significant decrease (p<1E-16, Wilcoxon signed rank test) in overall accessibility when normalized to the median read depth across all 1,404 high-accessibility bound regions. However, when analyzing each independent GSX2-bound peak for significant changes in accessibility, we found that less than 7% of GSX2-bound regions showed a significant decrease in accessibility, whereas nearly 92% of the GSX2-bound regions showed no significant change in accessibility (Fig 3D, bar graph at right). In fact, only 94 of the 3,693 regions (less than 3%) that showed decreased accessibility were directly bound by GSX2, indicating that most of the chromatin accessibility changes following GSX2 expression were caused by indirect mechanism(s), potentially involving secondary TFs. To identify potential secondary TFs responsible for the accessibility changes, we performed de novo motif analysis on the differentially accessible peaks from the 12- and 24-hour ± Dox samples and found motif enrichment for several protein families known to be important in forebrain development (e.g., HD, ZF, HMG; Supp Fig 4A). In sum, these results indicate that while GSX2 expression is associated with widespread changes in chromatin accessibility, only a small fraction of these changes occur at directly bound loci. The motif enrichment for other forebrain development-associated transcription factors in differentially accessible regions suggests that GSX2 expression may exert its impact on chromatin through secondary or indirect mechanisms.

Lastly, we performed longitudinal ATAC-seq analysis following Dox removal to assess if the chromatin accessibility changes induced by a 12-hour pulse of GSX2 expression were sustained and to assess if GSX2 binding to high- versus low-accessibility regions altered their accessibility at subsequent steps of maturation. Within 12 hours of Dox withdrawal (i.e., the 24-hour timepoint), the pattern of differentially accessible regions inverts and starts to normalize to non-induced controls (Fig 3E). Notably, there were no differentially accessible regions in the 48-hour ± Dox comparison. Moreover, comparative analysis across time revealed that the GSX2-bound high- and low-accessibility regions at the 12-hour timepoint behaved similarly after Dox removal in the 24-hour and 48-hour ± Dox-treated samples (Supp Fig 4B). Taken together, these results suggest that GSX2-induced chromatin accessibility changes are transient and insufficient to stably alter the chromatin landscape.

GSX2 represses neuronal differentiation genes and non-LGE patterning programs

To determine how GSX2 expression in human LGE-like progenitors alters gene expression, we performed triplicate bulk RNA-seq experiments at the same timepoints as the ATAC-seq samples (Fig 2A). First, we focused on comparative analysis at the 12-hour timepoint to define differentially expressed genes (DEGs) that are activated and repressed immediately following GSX2 induction. This analysis revealed that GSX2 expression causes significant changes in over 1,400 DEGs, with approximately half being up- (708) versus down-regulated (780) (Fig 4A, top). However, the down-regulated genes had much stronger statistical significance than the activated genes (Fig 4A, top). Similar pairwise analysis at the 24-hour timepoint shows significant residual transcriptional dysregulation (>1,000 DEGs) that largely normalizes by the 48-hour timepoint (<200 DEGs; Fig 4A, middle & bottom). These results further support the transient effect of the 12-hour GSX2 pulse in this in vitro differentiation system. Gene Ontology (GO) term enrichment analysis on the DEGs at each ±Dox timepoint reveals that down-regulated genes were consistently associated with neurogenic processes, while up-regulated genes had weaker enrichment for more non-specific terms (Fig 4B). For example, while normalized read counts for GE and cortical progenitor genes revealed that the non-induced cultures have sustained expression of numerous GE progenitor genes (e.g., ARX, ASCL1, DLX1/2, SP8; Supp Fig 5A) and increasing expression of more mature striatal markers (e.g., DLX5/6, FOXP1; Supp Fig 5B), GSX2 induction (Supp Fig 5) caused a sharp and consistent repression of many of these genes at the 12-hour timepoint, followed by distinct recovery patterns after Dox removal. These analyses are consistent with the ATAC-seq data, suggesting that GSX2 initially causes both chromatin accessibility and gene expression changes that alter neurogenic processes, but these changes are largely transient following the removal of Dox-mediated GSX2 induction. Hence, while these studies demonstrate that a 12-hour pulse of GSX2 is insufficient to fully reprogram these neural progenitor cells, the initial widespread mis-regulation of numerous neuronal genes provides an opportunity to better understand how GSX2 directly regulates target genes.

Figure 4: GSX2 expression in human LGE-like progenitors represses neurogenic genes.

(A) Volcano plot of 12-hour samples with and without Dox induction shows a similar number of up- (708) and down-regulated (780) genes. However, the DEGs most significantly dysregulated trend strongly towards down-regulation. 24-hour ± Dox comparison shows significant but reduced dysregulation. By 48-hour comparison, majority of transcriptional dysregulation has resolved. Labelled genes based on biological relevance and meeting significance threshold at each timepoint. 19,805 passed filters against low transcript numbers. (B) Gene Ontology (GO) Biological Process term enrichment analysis shows that down-regulated genes across all three timepoints are associated with neurogenesis-related terms, while up-regulated DEG terms are non-specific. (C) Intersection of ChIP-seq and RNA-seq data at the 12-hour timepoint reveals enrichment for GSX2 binding to downregulated DEGs. N = number of genes bound by GSX2 in each significance criteria (i.e. LFC ≤ 1, FDR < 0.05). Percentages indicate the fraction of DEGs bound by GSX2 (e.g. 359/780 = 46.03%). Abbreviations: FDR = False Discovery Rate (Adjusted p-value); log2FC = log2 Fold Change; GO = Gene Ontology; DEG = Differentially Expressed Gene

To determine which of the dysregulated genes are direct GSX2 targets, we intersected our ChIP-seq and RNA-seq datasets at the 12-hour timepoint. Annotating the consensus GSX2 ChIP-seq peaks to the nearest transcription start site (TSS) shows that 46% of the downregulated genes (359/780) at the 12-hour timepoint were directly bound by GSX2, whereas less than 15% of the activated (68/708) and unchanged genes (2,537/18,317) were associated with a nearby GSX2 peak (Fig 4C). Moreover, the down-regulated genes included notable transcription factors (e.g., PAX6, FEZF1, and NEUROG1) that have been shown to specify cortical progenitors. Many of these factors are within families identified in the ATAC-seq de novo motif analysis, further supporting the conclusion that chromatin dysregulation in hESC-derived, LGE-like progenitors is likely due to indirect effects, such as the mis-regulation of secondary transcription factors after GSX2 expression. These transcriptomic studies are consistent with prior Xenopus and mouse work that GSX2 patterns the forebrain by repressing key neural patterning and differentiation genes (Corbin et al., 2000; Toresson and Campbell, 2001; Toresson et al., 2000; Winterbottom et al., 2011; Yun et al., 2001).

To study transcriptomic changes over time, we performed hierarchical clustering on the DEGs from the ±Dox comparisons across all timepoints. This analysis identified 11 distinct clusters of genes (Fig 5A). Since the majority of the DEGs directly bound by GSX2 were initially down-regulated (Fig 4C), we focused on the clusters of down-regulated genes that segregated into four distinct groups based upon their recovery patterns (RNA_8 – RNA_11; Fig 5A). Genes in cluster 8 were characterized by rapid recovery/over-expression by the 24-hour timepoint, cluster 9 genes were continually repressed throughout, and genes in clusters 10/11 showed delayed recovery and mild over-expression only at the 48-hour timepoint (Fig 5B). Additionally, we found that the genes in clusters RNA_8–9 enrich for neurogenesis-related GO terms (Supp Fig 6), and the intersection of the bound DEGs with the hierarchical clustering results showed the highest enrichment for genes in clusters RNA_8 and 9 (Fig 5C).

Figure 5: GSX2 target genes show differing evolution of expression patterns over time.

(A) Hierarchical clustering of DEGs showing the evolution of gene expression patterns over time. Note, the genes are clustered into groups with the early repressed clusters (RNA_8–11) showing multiple distinct patterns of recovery. However, the majority of genes approach baseline expression compared to non-induced controls within 36 hours. DEGs defined by Likelihood Ratio Test (LRT; FDR<0.05; N=9,631) as dependent on the interaction of Dox induction + timepoint. (B) Log2FoldChange (LFC) values of ±Dox pairwise comparisons shows specific recovery trends after Dox removal: (1) RNA_8 shows rapid recovery and mild-moderate over-expression within 12 hours; (2) RNA_ 9 is continually repressed; (3) RNA_10 displays delayed inhibition at 24-hours and mild up-regulation at 48-hours; and (4) RNA_11 shows strong initial inhibition that gradually recovers to mild up-regulation by 48-hours. (C) GSX2-bound DEGs are enriched in clusters RNA_8, 9 and 11 from hierarchical clustering. Numbers below bars indicate total number of DEGs in each cluster.

To assess potential differences in GSX2 motif contributions to the cluster regulatory patterns, we first performed motif scans for GSX2 monomer and homodimer motifs on the hierarchical cluster-associated ChIP-seq peaks and found similar proportions of GSX2 DNA-binding motifs in each of the early repressed clusters (Supp Fig 7A). Second, we focused on the peaks associated with repressed genes found in RNA clusters 8 and 9, which were both similarly repressed at the 12-hour timepoint but had distinct recovery patterns at the 24- and 48-hour timepoints. This analysis again revealed enrichment for expected monomer and homodimer motifs, with moderately increased homodimer enrichment in the RNA_9-associated peaks (Supp Fig 7B). However, we did not identify any obvious additional TF motifs that may help explain the distinct recovery profiles (Supp Fig 7B). Taken together, these results suggest that GSX2 primarily serves as a transcriptional repressor of neurogenic target genes, and the distinct recovery patterns of the DEGs following Dox removal suggest that the target genes found within each cluster may share additional cis-regulatory mechanisms, whereas those genes in different clusters were more likely to have distinct regulatory mechanisms.

GSX2 induction promotes striatal radial glia identity by repressing alternative developmental programs and modulating progenitor progression

To assess how our directed differentiation protocol relates to in utero development and to better understand the impact of GSX2 expression on human neural gene expression, we compared our bulk RNA-seq samples to a recently published first trimester human brain single cell RNA-seq (scRNA-seq) atlas (Braun et al., 2023). Using samples from Carnegie Stage 20 (~6.9 weeks post-conception), we created UMAP projections that separated cells based on maturation (Neural Cell Type, Fig 6A) and regional identity (Forebrain Region, Fig 6C). We then projected our triplicate ±Dox bulk RNA-seq samples from each timepoint onto the reference atlas (Fig 6B). This analysis revealed that regardless of Dox-induction status, all samples were most similar to radial glia cells (RGCs), consistent with our culture conditions producing neural progenitor cells (Fig 6A). Moreover, the forebrain regionality of the reference atlas showed significant overlap with striatal and cortical progenitors (Fig 6C, Supp Fig 8). Interestingly, while each of the non-induced samples were predicted to be most similar to cortical progenitors, all the Dox-treated GSX2-induced samples were most similar to striatal radial glia (Fig 6B–C). These data are consistent with GSX2 promoting a striatal (i.e., LGE-like) progenitor cell type.

Figure 6: GSX2 expression promotes striatal progenitor identity and development.

(A, C) UMAP projections of scRNA-seq samples from human fetal brain atlas for Hansen Stage 20 (~6.9 weeks post-conception) based on neural cell type maturity (A) and regional identities (C). (B) Overlay of the bulk RNA-seq projections from our directed differentiation assays shows significant overlap with radial glial cells, consistent with a neural progenitor status regardless of timepoint or Dox condition. Regional identity of our bulk samples project to a region consistent with cortical and striatal identities with the non-induced cells more closely matching cortical fates. However, the induction of GSX2 expression via Dox promotes striatal identity from the default cortical program. (D) Module scores of up- and down-regulated DEGs from ±Dox comparisons show that GSX2 expression causes transient repression of non-striatal lineages and cyclical regulation of progenitor maturation, with initial down- (24-hr down-regulated DEGs) and subsequent up-regulation (48-hr up-regulated DEGs) of more mature markers.

To investigate how GSX2 expression favors a striatal RGC gene signature, we calculated module scores for the DEG lists from our ±Dox timepoint comparisons (Fig 6D–E). At our 12-hour timepoint, the most significant down-regulated DEGs correspond with both radial glial populations of alternative forebrain regions and more mature cellular populations (Fig 6E, left). At the 24- and 48-hour timepoints, the down-regulated DEGs correlate with more mature cell populations (Fig 6E, middle/right). Taken together, these findings indicate that GSX2 promotes striatal identity by repressing alternative forebrain RGC fates and maintaining progenitors by repressing pro-differentiation genes. Interestingly, while the up-regulated DEGs at both the 12- and 24-hour timepoints have little correlation with any specific cell population, the 48-hour up-regulated DEGs most closely correlate with maturing striatal progenitors/neuron populations (Fig 6D). This analysis further supports a role for GSX2 in human LGE progenitor specification and maturation.

GSX2 Binds Conserved Target Genes Critical for Forebrain Regional Specification and Progenitor Maturation

To identify the conserved regulatory functions of GSX2 across species, we integrated the human datasets generated in this study with published murine genomics datasets. To define putative conserved regulatory targets, we compared bulk RNA-seq from mouse E12.5 Gsx2 wild-type and Gsx2 germline KO LGEs to our human datasets (Salomone et al., 2021). Note, the opposing directionality of these two experimental systems, as the human data identifies DEGs that are regulated by induction of GSX2, whereas the mouse data identifies DEGs that are caused by loss of Gsx2 function. Comparing the DEG lists for each organism, we identified 342 genes that were mis-regulated in both species, with the majority of these DEGs (~75%) being repressed in the 12-hour ± Dox comparison in human cells (Fig 7A). Moreover, there is a trend towards opposing regulation, especially for those genes that are repressed by GSX2 induction in human cells, as 76% of the homologues of these genes (196 of 258 genes) were found to be up-regulated by loss of Gsx2 in mice (Fig 7A). Additional analysis of cross-species conserved DEGs revealed that the majority are present in the rapid-recovering RNA_8 and continually repressed RNA_9 hierarchical clusters (Fig 7B).

Figure 7: GSX2 binds near key conserved regulatory targets in the developing forebrain.

(A) Overlap of DEG list from murine E12.5 Gsx2 KO LGE bulk RNA-seq versus generated bulk RNA-seq datasets from human LGE-like progenitors shows significant number of conserved DEGs between model systems. Majority of conserved DEGs are down-regulated in 12-hour ±Dox comparison (~75%), with a majority of those showing opposite dysregulation in the Gsx2 loss-of-function mouse model. Percentages given in terms of total conserved DEGs from ±Dox comparison at each timepoint. (B) Majority of conserved DEGs (~60%) correspond to clusters 8–9 from hierarchical clustering, suggesting these clusters contain key target genes for GSX2 regulation. Bottom number indicates number of conserved DEGs in each cluster and percentages are the proportion of conserved DEGs compared to the total DEGs in each cluster. (C) Comparing transcriptional regulation of conserved, dysregulated TFs within ±100kb of GSX2-bound loci in loss-of-function mouse and gain-of-function human datasets. TFs critical for regional specification are generally repressed in both organisms, while progenitor maturation TFs often show conflicting regulation between mouse and human. Numbers indicate total number of conserved genes by binding classification, including TFs in table. (D) Genes dysregulated and bound in both mouse and human (122 genes from “Both” in subfigure C) strongly enrich for neurogenesis-related GO terms.

To investigate which of the 342 DEGs were associated with conserved GSX2 genomic binding, we compared our human ^2xFLAGGSX2 ChIP-seq data to the published E12.5 ^2xFLAGGsx2 CUT&RUN and E13.5 Gsx2 ChIP-seq datasets (Catta-Preta et al., 2025; Salomone et al., 2021). We scanned a ±100kb window around each of the murine and human peaks, and found 272 of the conserved DEGs were associated with at least one genomic binding peak in one or more of the datasets, with 122 genes associated with peaks in both organisms (Fig 7C). Intriguingly, many of these conserved regulatory targets encode transcription factors known to contribute to forebrain development, including PAX6, ASCL1, GSX1/2, EMX1/2, NEUROG1/2, and SOX2 (Fig 7C). Overall, the 122 conserved DEGs show strong enrichment for neurogenic-related GO terms (Fig 7D), and most of the genes (104/122) were down-regulated in the 12-hour ± Dox comparison, consistent with GSX2 predominantly repressing these targets. To better understand the regulatory features of these GSX2-bound regions, we first separated all the peaks associated with these 122 genes for comparative studies and de novo motif analysis. Interestingly, the additional filters of conserved dysregulation and binding in mouse datasets enrich for DEGs associated with >1 human ChIP-seq peak, suggesting GSX2 uses multiple loci to regulate critical target genes (Supp Fig 9B). Comparative analysis of ATAC-seq overlap shows a similar proportion of high-accessibility peaks (~25%) as the overall peak set (~20%; Supp Fig 9C). Next, we performed de novo motif analysis of all peaks, the highly accessible peaks, and the low accessible peaks and found enrichment for monomer motifs, with weak enrichment for homodimer motifs only in the low-accessibility peaks (Supp Fig 9D). These results further support that GSX2 primarily functions as a transcriptional repressor and suggests that it uses multiple binding loci to target key genes.

Discussion:

New model to study GSX2 regulatory functions in LGE-like progenitors

Determining how GSX2 regulates gene expression to control ventral forebrain fates is critical to better understand its impact on basal ganglia development. In this study, we developed a human stem cell system to investigate GSX2 function using genome editing, directed differentiation cultures, and genomic assays. First, we genetically modified an existing human ESC line by inserting an inducible epitope-tagged GSX2 expression construct that permits temporal induction of transgene expression. Second, we modified a dual SMADi directed differentiation protocol to generate LGE-like progenitors within 12 days in culture. These findings are consistent with previous studies that TGF-β agonism can specify striatal neurons (Arber et al., 2015; Miura et al., 2020). However, this protocol differs from other protocols that include sonic hedgehog (SHH) agonism (Amimoto et al., 2021; Chambers et al., 2009; Delli Carri et al., 2013; Ma et al., 2012). While prior studies have shown that SHH signaling plays a key role in specifying NKX2–1-positive ventral progenitors of the MGE, SHH’s role in specifying GSX2-positive LGE progenitors is less clear. Hence, further studies are required to determine whether combining TGF-β and SHH agonism would improve differentiation efficiencies and/or specify distinct neuronal subtypes. Third, we used Dox to temporally induce GSX2 expression, and genomic assays to map GSX2 binding to the human genome and assess its impact on chromatin accessibility and gene expression. These studies suggest that human GSX2 binds genomic regions through both monomer and homodimer sites and functions predominantly as a transcriptional repressor of neural gene expression. Moreover, by intersecting our human genomic data with previously published mouse datasets, we identified a highly conserved core of known regulators of forebrain development that are early, direct targets of GSX2 regulation (Fig 8). In total, this work provides new insights into the gene regulatory and genomic functions of GSX2, as well as into how it may regulate striatal development.

Figure 8: GSX2 serves as a major regulator of LGE development.

Proposed model of early GSX2 direct regulatory targets in the developing forebrain highlights two key functions: (1) Repression of non-LGE developmental programs to establish LGE progenitors (dashed interactions); (2) Maintenance of progenitor state through modulation of downstream neurogenic factors (solid interactions). All highlighted genes met the following criteria: (1) Dysregulation in both mouse and human RNA-seq datasets; (2) At least one GSX2 genomic binding peak within ±100kb, or nearest TSS if no gene within window. Abbreviations: VZ = ventricular zone; SVZ = sub-ventricular zone; MZ = mantle zone; MGE = medial ganglionic eminence; LGE = lateral ganglionic eminence

GSX2 uses differing DNA motifs based on chromatin accessibility

Our genomic binding studies showed that GSX2 binds both monomer and homodimer motifs, consistent with our recent work using both in vitro and mouse models (Cain et al., 2023; Salomone et al., 2021; Webb et al., 2024). Intriguingly, we found that ~80% of the GSX2-bound genomic sites had relatively low accessibility, and that GSX2 binding to these regions did not dramatically alter their accessibility over the next 36 hours after induction removal. Similarly, GSX2 binding to loci with relatively high accessibility was not associated with significant changes in their accessibility over time. These findings have two implications: First, while GSX2 is capable of binding relatively inaccessible genomic regions, it does not appear to function as a bona fide pioneer factor by subsequently increasing their accessibility (Cirillo et al., 2002; Gassler et al., 2022; Soufi et al., 2015; Zaret, 2020). Second, while GSX2 functions as a potent transcriptional repressor, GSX2 genomic binding to most accessible regions does not dramatically reduce their chromatin accessibility via recruiting Groucho/TLE co-repressor complexes (Von Ohlen and Moses, 2009; Von Ohlen et al., 2007b). However, it should be noted that GSX2 expression was only induced for 12-hours in this study, and it is possible that longer duration GSX2 expression and subsequent genomic binding may be needed to significantly increase or decrease the chromatin accessibility of GSX2 bound regions. Nevertheless, in this study, we did find that a 12-hour pulse of GSX2 was sufficient to transiently decrease the accessibility of ~3,700 genomic regions via indirect mechanisms. Our analyses suggest that these changes are likely due to altered transcriptional regulation of other TFs, such as ASCL1, NEUROD1, and SOX2, which have each previously been shown to exhibit some form of pioneer activity (Chronis et al., 2017; Domcke et al., 2015; Park et al., 2017; Pataskar et al., 2016; Raposo et al., 2015; Soufi et al., 2015; Takahashi et al., 2007; Wapinski et al., 2017).

A fundamental question that arises from the GSX2 genomic binding data is what mechanisms drive GSX2 binding to high- versus low-accessible regions. De novo motif analysis on GSX2-bound peaks in low- versus high-accessibility chromatin found that while monomer sites were enriched in both contexts, homodimer sites were predominately enriched in relatively low-accessibility regions. Intriguingly, prior in vitro selection assays using purified GSX2 protein and nucleosome-bound DNA (NCAP-SELEX) found that GSX2 (and GSX1) enriched for homodimer sites (Zhu et al., 2018). This data suggests that cooperative GSX2 binding to homodimer sites may help drive binding to nucleosomal DNA. However, it should be noted that purified GSX2 can cooperatively bind non-nucleosome wrapped DNA in vitro (Salomone et al., 2021; Webb et al., 2024) and that high-throughput in vitro selection assays (HT-SELEX) on naked DNA enriched for both monomer and homodimer sites (Cain et al., 2023; Jolma et al., 2013). These studies demonstrate that GSX2 can bind accessible DNA via both monomer and homodimer sites. Intriguingly, a recent study identified a point mutation in the GSX2 homeodomain, p.I233E, that selectively disrupts homodimerization, without affecting monomer binding (Webb et al., 2024). Hence, future studies using human cells engineered to express the dimer-deficient variant are needed to understand the significance of GSX2 binding to low-accessibility chromatin using homodimer sites.

GSX2 regulates striatal development through conserved target genes in humans

Gsx2 has previously been shown to be a critical regulator of striatal development in animal models (Corbin et al., 2000; Toresson and Campbell, 2001; Toresson et al., 2000; Waclaw et al., 2009; Wang et al., 2013; Yun et al., 2003; Yun et al., 2001), while case reports suggest a similar, conserved role in human neurodevelopment (De Mori et al., 2019; Urel-Demir et al., 2024). A role in human LGE progenitor specification is consistent with the data presented here and the single cell projection analysis that transgenic expression of GSX2 promotes a striatal radial glia identity from the default cortical progenitor state of the uninduced cultures. By incorporating module scores for the ±Dox comparisons, we found that genes related to regional specification/alternative forebrain regions show strong initial repression in both progenitor and more mature cell populations. However, only genes related to the more mature cell populations show longitudinal repression as transgenic GSX2 levels decay. Meanwhile, genes related to striatal progenitor maturation are similarly initially down-regulated but then are up-regulated at our latest timepoint (48 hours). This human model is consistent with mouse studies, where LGE progenitor maturation occurs as cells migrate through the ventricular zone with concomitant Gsx2 down-regulation (Corbin et al., 2000; Roychoudhury et al., 2020; Toresson et al., 2000), whereas continuous Gsx2 expression was found to disrupt progenitor progression (Pei et al., 2011).

At the gene regulatory level, animal studies have found that Gsx2 and its homologues primarily function as transcriptional repressors (Salomone et al., 2021; Von Ohlen et al., 2007b; Winterbottom et al., 2010; Winterbottom et al., 2011). Similarly, our human directed differentiation data provides strong evidence that GSX2 largely functions as a transcriptional repressor of key neural genes. Moreover, by integrating our human cell data with prior mouse studies, we identified conserved regulation of and binding near numerous TFs important for forebrain development. For example, consistent with Gsx2 being known to regulate early dorsal-ventral patterning of the forebrain, the mouse and human genomic binding and RNA expression analyses reveal that GSX2 represses both dorsal factors that promote cortical development, such as PAX6, DBX1, EMX1/2, NEUROG1/2, and FEZF2, as well as key ventral factors associated with the MGE, such as NKX2–1 and GSX1 (Fig 8). In addition, we found that GSX2 also predominantly repressed factors associated with basal ganglia progenitor maturation, such as ASCL1, SP8, OLIG2, and VAX1 (Fig 8). Interestingly, the finding that DLX1/2 do not show early down-regulation, but instead are up-regulated at later timepoints, suggests a more complex regulatory interplay and is consistent with prior mouse studies on the interaction of Gsx2 and Dlx1/2 (Wang et al., 2013). This core of common genes forms a set of high-confidence targets consistent with GSX2 functioning to both specify and maintain these cells as LGE progenitors that ultimately give rise to cells that populate the basal ganglia.

While most of the conserved GSX2-bound repressed genes in our gain-of-function human data showed opposing regulation in a knock-out mouse model, a subset of genes related to progenitor maturation were similarly down-regulated in both mouse and human (e.g., ASCL1). Two mechanisms could explain these findings. First, since the mouse loss-of-function studies prevented GSX2 protein expression at any developmental timepoint, the LGE progenitors are mis-specified as cortical progenitors, resulting in the indirect loss of gene expression needed for striatal development. Second, it is possible that dose-dependent GSX2 regulation results in the differential expression of progenitor maturation genes. For example, high levels of GSX2 could repress target genes, such as ASCL1, to promote progenitor maintenance and prevent precocious cell maturation, whereas lower levels of GSX2 could subsequently activate these genes to promote striatal fates. This latter model is consistent with the observed ASCL1 modulation under varying Dox treatment conditions and prior studies suggesting that GSX2 not only represses gene expression but also has transcriptional activation/stimulation functions (Salomone et al., 2021; Von Ohlen et al., 2007a; Von Ohlen and Moses, 2009). In contrast, other GSX2 target genes may be continually repressed at lower levels of GSX2 expression. While our studies do not demonstrate that GSX2 directly activates target gene expression, we did find that the initially repressed genes can be subdivided into two groups: genes that recover rapidly (RNA_8) and those continually repressed (RNA_9). Thus, the genes that recover quickly may be activated by GSX2 as its protein levels decrease, while the continually repressed genes are inhibited at similar lower GSX2 concentrations. However, cis-regulatory mechanisms are typically complex, and hence the distinct evolutions of gene expression are likely to be influenced through both direct GSX2 effects and indirect effects caused by the mis-expression of other TFs. For example, Catta-Preta et al. found Gsx2 ChIP-seq co-enrichment with varying combinations of eleven other TFs expressed in the mouse forebrain (e.g., Arx, Ascl1, Dlx1/2). Here, we attempted to better understand the cis-regulatory logic underlying the rapid-recovering versus the continually-repressed genes using motif scans for the GSX2 monomer and homodimer motifs and de novo motif analyses on each group. Unfortunately, we found similar patterns of GSX2 monomer and homodimer sites within peaks associated with rapid-recovering and continually-repressed genes and did not identify other significant differences in TF motif enrichment between groups. Hence, future genomic studies using ChIP-seq and/or CUT&RUN methods are needed to assess for both co-binding of additional GE-expressed TFs and for histone chromatin marks consistent with active versus repressed chromatin (e.g., H3K27ac, H3K27me3, etc.) to better understand the cis-regulatory logic underlying the distinct transcriptional dynamics of these GSX2 target genes.

Conclusions:

In sum, our Dox-inducible system described here enables investigation of GSX2 regulatory functions in human LGE-like progenitors. Our studies demonstrate that GSX2 primarily functions as a direct transcriptional repressor of key neuronal targets within these human neural progenitor cells. Moreover, this system is ideally suited for future studies focused on better understanding how GSX2 patient variants associated with brain malformations and clinical symptoms impact their ability to bind the genome and regulate target genes. One such variant, p.Q251R (De Mori et al., 2019; Tweedie et al., 2025), has been shown to alter DNA binding specificity in vitro and cause LGE and hindbrain defects in mouse models.(Tweedie et al., 2025). GSX2 is a member of the Q50 homeodomain subfamily and this variant also shows impaired ability to bind lower-affinity, Q50-preferred monomer sites (Tweedie et al., 2025), which have previously been shown to play critical roles in the regulatory output of homeodomain factors (Crocker et al., 2015). A second variant, p.W249C (Urel-Demir et al., 2024), has unknown molecular significance. However, conservation analysis shows that the tryptophan at position 48 of the homeodomain is amongst the most conserved residues across homeodomain proteins (Kock et al., 2024), suggesting the variant is likely to impact GSX2 DNA binding. Altogether, our platform will allow comparative studies between different GSX2 alleles and thereby provide new insights into the effects of disease-associated variants on GSX2 molecular functions within human forebrain progenitor cells.

List of Abbreviations:

Gsx2

Genetic-Screened Homeobox 2

ASCL1

Achaete-Scute Family bHLH Transcription Factor 1

NKX2–1

NK2 Homeobox 1

GSX1

Genetic=Screened Homeobox 1

Ind

Intermediate Neuroblast Defective

PAX6

Paired Box 6

ARX

Aristaless Related Homeobox

DLX2

Distal-Less Homeobox 2

ISL1

ISL LIM Homeobox 1

SP8

Sp8 Transcription Factor

SIX3

SIX Homeobox 3

FOXP1

Forkhead Box P1

Gal4

galactose-responsive transcription factor GAL4

VP16

Herpes Simplex Virus 1 Viral Protein 16

TLE

TLE Family Member 1

hESC

human embryonic stem cell

AAVS1

adeno-associated virus integration site 1

LGE

lateral ganglionic eminence

MGE

medial ganglionic eminence

VZ

ventricular zone

SVZ

sub-ventricular zone

MZ

mantle zone

TF

transcription factor

Dox

doxycycline

TRA-1-60

Podocalyxin

SSEA-4

Stage Specific Embryo Antigen 4

dSMADi

dual SMAD inhibition

HD

homeodomain

ZF

zinc finger

HMG

high mobility group

bHLH

basic Helix-Loop-Helix

RFX

regulatory factor X

TEF

transcriptional enhancer factor

HTH

helix-turn-helix

RRM

RNA recognition motif

HD-ZIP

homeodomain-leucine zipper

Dof

DNA-binding with one finger

GRF

gene regulatory factor

CUT&RUN

Cut Under Targets & Release Under Nuclease

ChIP-seq

Chromatin Immunoprecipitation Sequencing

HOMER

Hypergeometric Optimization of Motif EnRichment

ATAC-seq

Assay for Transposase-Accessible Chromatin Sequencing

DEG

differentially expressed gene

TSS

transcription start site

PCA

principal component analysis

UMAP

uniform manifold approximation and projection

KO

knockout

WT

wild-type

GO

gene ontology

SELEX

Systematic Evolution of Ligands by Exponential Enrichment

NCAP-SELEX

Nucleosome Consecutive Affinity-Purification SELEX

HT-SELEX

high throughput SELEX

Availability of data and materials:

All datasets generated in this study were uploaded to the NCBI Gene Expression Omnibus (GEO) database as GSE306348 (RNA-seq), GSE306345 (ChIP-seq) and GSE306350 (ATAC-seq). Published mouse datasets are available from the NCBI GEO as GSE162590 (bulk RNA-seq); GSE162589 (E12.5 LGE anti-^2xFLAGGsx2 CUT&RUN) and GSE222183 (E13.5 pan-basal ganglia anti-Gsx2 ChIP-seq). The human single cell RNA-seq atlas is available from the European Genome-Phenome Archive (EGPA) as EGAD00001006049.