Main

Large-scale genetic studies have implicated hundreds of genes in autism spectrum disorder (ASD), with many identified through rare mutations that result in loss of function of one allele1,2,3. This suggests that gene dosage plays an important part in ASD11 and makes it critical to determine how appropriate expression levels of ASD genes are maintained. Characterization of gene regulatory networks (GRNs) could provide important insights into mechanisms mediating the proper transcriptional dosage of these ASD genes. In addition, ASD is significantly male-biased in prevalence, and although some ASD genes have been identified on the X chromosome12, X-linked mutations alone cannot fully explain this sex bias8,13. Moreover, it has been consistently demonstrated that female individuals require greater autosomal genetic risk to manifest ASD compared with male individuals6,7, but the genetic mechanisms of this female protective effect (FPE) are unknown8,9.

Here we use GRN reconstruction to identify key drivers of expression of ASD genes, including CHD8 and REST. These analyses also identify female-biased transcription factor ZFX, which is expressed from both the active and inactive X chromosome in female individuals10, as a key regulator of ASD genes, suggesting a direct transcriptional mechanism that could contribute to the FPE. This study provides a crucial framework for determining how variants in diverse genes can cause convergent molecular effects that ultimately result in ASD.

ASD genes exhibit broad transcriptomic convergence

Functional interrogation of ASD genes revealed strongly overlapping effects on downstream targets in human neural progenitor cells (NPCs). Knockdown (KD) studies were performed for 20 ASD genes, including 18 genes from the Simons Foundation Autism Research Initiative (SFARI) gene database14 and two long non-coding RNAs (lncRNAs) previously implicated in ASD (NR2F1-AS1 (ref. 15) and SOX2-OT16). RNA-degrading CRISPR enzyme CasRx17 and an array of three guide RNAs (gRNAs) per target (Supplementary Table 1) were used to induce individual KD of each target gene relative to non-targeting controls (NTCs) in NPC cultures from a control human male (XY) induced pluripotent stem (iPS) cell line (Fig. 1a,b and Extended Data Fig. 1). RNA sequencing (RNA-seq) analysis after 24 h of KD demonstrated that nearly all perturbations led to substantial numbers of significantly differentially expressed genes (DEGs; adjusted P (Padj) < 0.05), with a median of 920 DEGs per perturbation (25th percentile: 644, 75th percentile: 1,485) (Extended Data Fig. 2 and Supplementary Table 2). To ensure sufficient statistical power for downstream analyses of these DEGs, we focused on the 18 perturbations (18 of 20, 90%) that resulted in more than 20 DEGs, excluding the ARHGAP5 and ARID2 KD samples from further analysis. The remaining perturbations each led to a minimum of 287 DEGs, which enabled robust characterization of these samples.

Fig. 1: Depletion of ASD genes reveals widespread transcriptomic convergence.

a, Experimental design schematic showing CasRx-based KD in NPCs. b, Immunocytochemistry of iPS cell-derived human NPC cultures, representative of three independent analyses with similar results. c, Bar plot depicting enrichment of SFARI genes or intellectual-disability-only genes among the DEGs for each KD condition, excluding the targeted gene itself. Bars indicate the OR from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR. P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. d, Circos plot depicting transcriptomic convergence of HC-SFARI DEGs shared between different KD conditions (Extended Data Fig. 3a). e, Heatmap showing the relative expression of HC-SFARI genes that were DEGs in at least ten KD conditions (recurrent dysregulation) (Extended Data Fig. 3b,c). Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini–Hochberg correction for multiple hypothesis testing and are indicated by an asterisk. Top ASD risk genes from ref. 1 are shown in bold. f, Relative CytoTrack dye intensity (−log2) over time. Only KD conditions that exhibited statistically significant effects on NPC proliferation that were consistent over time are displayed. Plot depicts mean ± s.e.m. Statistical significance was determined using ordinary two-way ANOVA with Dunnett’s multiple comparisons test. At least three biologically independent samples per condition were analysed for each time point. See Supplementary Table 10 for full statistical results. Scale bar, 50 μm. ID, intellectual disability; NS, not significant.

Source data

Transcriptome-wide analyses showed that ASD gene KDs caused remarkably widespread disruption of other ASD genes. Most sets of DEGs exhibited significant enrichment for known ASD genes in the SFARI Gene database (15 of 18, 83.3%) (Fig. 1c and Supplementary Table 3). By contrast, there was no significant enrichment for genes implicated in intellectual disability but not ASD (intellectual-disability-only genes) (Fig. 1c and Supplementary Table 3) (Methods), despite their comparable expression in NPCs (intellectual-disability-only genes: 1,062 of 1,242, 85.5% expressed; SFARI genes: 797 of 1230, 64.8% expressed). This suggests that perturbation of ASD genes particularly affects other ASD genes, as opposed to neurodevelopmental genes more broadly.

We observed a striking pattern of transcriptomic convergence across perturbations, which we defined as statistically significant overlap of the same DEGs (Extended Data Fig. 3). This was particularly evident for high-confidence SFARI (HC-SFARI) genes, the SFARI genes in the highest evidentiary tier for association with ASD (Fig. 1d). Moreover, HC-SFARI genes tended to shift in a concordant direction as a result of different perturbations even when they were not significant DEGs within a particular KD sample (Extended Data Fig. 3c). Such genes included prominent ASD risk gene CHD8 (ref. 18), which was significantly downregulated by most (11 of 18, 61.1%) of the KDs. Notably, 27 HC-SFARI genes were DEGs in at least 10 of the KDs (27 of 183, 14.8%, P = 4.1 × 10−9), and all were consistently downregulated (Fig. 1e), indicating convergent dysregulation of many of the strongest risk genes for ASD.

High levels of CasRx have been reported to cause non-specific RNA degradation of a magnitude that is correlated with the expression level of the KD target, whereas use of lower CasRx levels avoids this problem19. In our NPC transcriptomic studies, CasRx was expressed from only a single safe-harbour locus (AAVS1), which limited its expression. We also performed several further experiments to confirm that non-specific RNA degradation did not affect our results. First, we found no relationship between expression levels of KD targets and numbers of resulting DEGs when considering either the whole transcriptome (P = 0.990, R2 = 8 × 10−6, linear regression) or SFARI genes specifically (P = 0.784, R2 = 0.0043, linear regression) (Extended Data Fig. 4a). We next generated NPCs expressing fluorophores mKO2, mScarlet and iRFP670 and targeted each individually with CasRx (Supplementary Table 4); flow cytometry showed no collateral effects on the fluorescence intensity of non-targeted fluorophores (Extended Data Fig. 4b and Supplementary Table 5). Finally, we confirmed strong transcriptomic convergence on shared downregulated DEGs by performing further KD studies of a subset of our target genes using short hairpin RNAs (shRNAs) in NPCs (Extended Data Figs. 4c,d and 5 and Supplementary Table 6). Re-analysis of data from a study that used Cas9 to knock out ASD genes in iPS cells and iPS cell-derived excitatory neurons (ExNs)20 similarly demonstrated strong transcriptomic convergence on shared downregulated DEGs both within each cell type and between cell types (Extended Data Fig. 5 and Supplementary Table 6). Recent studies published since our work was submitted have also noted convergence across diverse ASD models21,22. Together, these data confirm that disruption of ASD genes results in transcriptomic convergence across various cell types and experimental contexts.

Convergence nominates candidate ASD risk genes

The marked transcriptomic convergence we observed in our KD studies suggested that candidate ASD risk genes could be nominated by the following features: (1) consistent downregulation following ASD gene KDs; and (2) dosage sensitivity, as indicated by high genic constraint scores (Supplementary Table 7 and Methods). This approach yielded 397 genes (Supplementary Table 8), including 113 SFARI genes (odds ratio (OR) = 5.87, Padj = 1.85 × 10−39) and 49 HC-SFARI genes (OR = 11.17, Padj = 1.11 × 10−29), demonstrating that these criteria effectively prioritize known ASD genes. Notably, this set of genes showed no enrichment for intellectual-disability-only genes (OR = 1.01, Padj = 0.509). All of the remaining 284 non-SFARI genes exhibited significant co-expression with SFARI genes (all Padj < 0.05), which has previously been shown to help prioritize novel ASD risk genes4, further supporting their relevance to ASD. Moreover, at least 40 of these genes showed some association with ASD in at least 1 major sequencing study (Supplementary Table 9). For example, mutations in AKT3, although classically associated with brain overgrowth disorders such as megalencephaly, have recently been associated with ASD; other genes including KIF11, KIF2A and SOS1 have also been associated with ASD in multiple studies (see references in Supplementary Table 9). Therefore, the remaining genes merit consideration as candidate ASD risk genes.

ASD genes regulate NPC proliferation

Many ASD gene perturbations also caused similar phenotypic effects in NPCs. Proliferation analysis using heritable dye (CytoTrack) showed that 9 of 18 (50%) of the KDs led to a statistically significant decrease in proliferation (Fig. 1f, Extended Data Fig. 6a, Supplementary Table 10 and Supplementary Fig. 1), including several genes that have been implicated in microcephaly (SETD1B23, SETD2 (ref. 24), SETD5 (ref. 25), KMT2E26 and WAC27). By contrast, only 1 of 18 (5.6%) of the KDs led to a significant increase in NPC proliferation rate. This suggests that reduced NPC proliferation is a common phenotype that results from disruption of several different ASD genes, consistent with previous findings28. Deciphering the specific relationships between these molecular phenotypes and abnormal head size in ASD is complex, as reports implicate increased rates of both microcephaly and macrocephaly across individuals with autism29,30. Indeed, two KD targets that we found to decrease proliferation in NPCs, SETD2 (ref. 24) and KMT2E26, have been reported to be associated with either microcephaly or macrocephaly depending on the specific variants studied in these genes. Thus, although disruption of NPC proliferation is a convergent phenotype observed across many of our perturbations, how this relates to downstream changes in head size remains to be determined.

Disruption of ASD genes alters organoid development

Inducible KD in cerebral organoids also revealed shared neurodevelopmental phenotypes. For these organoid studies, we created the FLEx-based inducible CRISPR knockdown (FLICK) construct. FLICK enables control over the timing of KD through doxycycline-inducible genetic recombination that results in constitutive expression of CasRx and an EGFP reporter (Fig. 2a), providing robust KD (Extended Data Fig. 7a). The FLICK constructs also contain identifying sequences that can be captured during single-cell RNA-seq (scRNA-seq) and used to determine the KD target in a given cell; this enables analysis of mosaic cerebral organoids.

Fig. 2: Disruption of ASD genes alters cerebral organoid development.

a, Design of FLICK constructs for doxycycline-inducible CasRx KD. FLICK constructs also express gRNAs (not shown). b, FLICK organoid after 2 months of development, representative of 3 independent analyses with similar results. Immunohistochemistry was performed to visualize expression of the following markers: EGFP (FLICK construct), DAPI (nuclei), PAX6 (NPCs), TBR2 (intermediate progenitor cells) and CTIP2 (ExNs). c, Expression (expr.) of canonical brain marker genes used to identify cell types by scRNA-seq. d, UMAP showing organoid cell clustering from integrated long-term organoid scRNA-seq data collected at 1 month and 2 months. e, Relative abundance of cells for each KD target from mosaic organoids at 1 month or 2 months. Results were normalized to the starting proportion for each sample at day 0 and are displayed relative to NTCs from the same time point. The plot depicts mean ± s.e.m. The FLICK constructs were induced by treatment with doxycycline on days 14–15. Statistical significance was assessed using repeated measures one-way ANOVA with Dunnett’s multiple comparisons test. n = 3 biologically independent samples per condition. f, Cell type composition per KD target in 1-month organoids. Dotted lines indicate the NTC proportions of main cell types (radial glia, intermediate progenitors and neurons). Samples for which the ExN/RG proportion was significantly different from the NTC are indicated with an asterisk (Padj < 0.05). Statistical analysis was performed using ordinary one-way ANOVA with Dunnett’s multiple comparisons test, considering all long-term samples from 1 month (shown here) and 2 months (see Extended Data Fig. 7f). Scale bar, 1 mm. Avg., average; Dox., doxycycline; ExN-Early, early ExN; ExN-Imm, immature ExN; ExN-Mat; mature ExN; IPC, intermediate progenitor cell; IPC-Div, dividing IPC; IN, inhibitory neuron; RG, radial glia; RG-Div, dividing RG.

For long-term studies, we targeted ASD genes whose KDs were not found to impair NPC proliferation in our CytoTrack experiments (Supplementary Table 10). FLICK constructs targeting these genes or expressing NTC gRNAs were each separately transposed into iPS cells. We then used fluorescence-activated cell sorting (FACS) to isolate cells that had integrated the FLICK constructs on the basis of their expression of fluorescent reporter TagBFP. Following FACS, we pooled an equal number of FLICK+ iPS cells for each KD target and for the NTCs, and cultured this mixed iPS cell population for 4 days to allow the cells to recover before generating mosaic organoids (day 0). Recombination of the FLICK construct, which leads to constitutive expression of CasRx and an EGFP reporter, was induced in a subset of the cells (approximately 10%) by treatment with doxycycline from days 14–15, when the organoids predominantly consisted of PAX6+ NPCs (Extended Data Fig. 7b). These organoids were subsequently grown until either 1 or 2 months of development (Fig. 2b) and analysed by scRNA-seq; this analysis identified all expected cell types (Fig. 2c,d and Extended Data Fig. 7c,d) including radial glia, intermediate progenitor cells, inhibitory neurons and ExNs. Further scRNA-seq analysis confirmed that the FLICK construct could produce efficient KD (Extended Data Fig. 7e).

Long-term organoid KDs revealed shared phenotypes across several perturbations. Several KDs, particularly those targeting BAZ2B, CLASP1 and EHMT1, were depleted from the organoids compared with NTCs (Fig. 2e), even though these KDs did not impair proliferation in NPC cultures (Extended Data Fig. 6a and Supplementary Table 10). This discrepancy between our NPC cultures and organoids suggests that long-term organoids, which support progenitor proliferation as well as differentiation, can be used to identify neurodevelopmental phenotypes that are not apparent in short-term NPC cultures. Notably, microcephaly has been associated with pathogenic variants in these genes31,32,33, consistent with the organoid phenotypes. Furthermore, these KDs also altered cell type proportions in organoids, producing relatively greater numbers of ExNs and fewer radial glia (Padj = 0.021 for BAZ2B, 0.018 for CLASP1 and 0.012 for EHMT1; ordinary one-way analysis of variance (ANOVA) with Dunnett’s multiple comparisons test) (Fig. 2f and Extended Data Fig. 7f). Together, these results indicate that disruption of these genes may result in increased ExN differentiation at the expense of self-renewal of radial glial progenitors. Most of the other KDs also resulted in a trend towards an increased proportion of ExNs by the 2-month time point, suggesting that this is a convergent neurodevelopmental phenotype, in line with previous observations34,35.

Depletion of ASD genes also resulted in prominent transcriptomic convergence in cerebral organoids (Fig. 3a and Extended Data Fig. 7g). Radial glia with different KDs exhibited highly overlapping DEGs, further supporting our findings from NPC cultures. Indeed, the DEGs identified in organoid radial glial progenitors strongly overlapped with the DEGs identified from KD of the same target in NPC cultures (Extended Data Fig. 7h), suggesting that these effects on gene expression are not specific to a particular experimental context. Transcriptomic convergence was also observed in other cell types, including intermediate progenitor cells and ExNs (Fig. 3a and Extended Data Fig. 7g), demonstrating that the convergence of ASD genes on shared downstream effects is not unique to neural progenitors. This was consistent with findings of studies that identified shared ASD DEGs in postmortem brain tissue from individuals with distinct causes of ASD36,37,38, as well as with widespread observations of convergent ExN phenotypes such as disordered synaptic structure or activity upon disruption of distinct ASD genes20, suggesting that convergence does not reflect merely a shared alteration of cell type or fate of progenitor cells. Furthermore, we found that detection of overlapping DEGs using scRNA-seq was heavily dependent on the number of cells sequenced for each perturbation, as demonstrated by downsampling analysis (Fig. 3b). This sample-size dependence may partially explain the lack of robust transcriptomic convergence observed at the individual gene level in some previous studies with smaller sample sizes34,39. Our use of an isogenic background is also likely to have contributed to our detection of transcriptomic convergence, as different genetic backgrounds have been found to significantly affect neurodevelopmental phenotypes resulting from the disruption of ASD genes in organoids34. Finally, although some convergently disrupted genes were cell-type specific, many genes were consistently disrupted across several cell types, including SFARI DEGs (Fig. 3c). This suggests that although cell-type-specific features affect the transcriptomic consequences of ASD gene perturbations, certain effects could be broadly shared by diverse cell types.

Fig. 3: Transcriptomic convergence upon ASD gene perturbations in cerebral organoids.

a, Heatmap of cross-enrichment of DEGs for radial glia and ExNs showing convergence in long-term organoids. b, Pie charts denoting proportions of pairwise sets of DEGs that showed significant cross-enrichment (transcriptomic convergence). Downsampling of the number of cells per condition to the number indicated above each pie chart decreased the transcriptomic convergence observed. c, Network visualization of SFARI DEGs in long-term organoids shared among different KDs conditions and cell types. The thickness of edges represents the number of shared DEGs. Nodes were plotted on the basis of clustering with the Fruchterman–Reingold algorithm (igraph). Nodes that are closer together had more shared DEGs on average. d, Proportions of cells with each KD target from scRNA-seq of mosaic organoids with acute KD. e, UMAP showing clustering of cells from acute KD mosaic organoids from scRNA-seq. f, Heatmap of cross-enrichment between different KD samples showing transcriptomic convergence in acute KD organoids.

Convergence upon acute perturbations in organoids

Acute disruption of ASD genes in 2-month organoids revealed even greater convergence on shared DEGs. A further set of mosaic FLICK organoids was grown for 2 months before induction of KD with doxycycline treatment from days 61–62; this was followed by scRNA-seq 2 days later, to determine the acute effects of disruption of these genes in organoids that had developed normally. This approach resulted in similar numbers of cells per KD (Fig. 3d and Extended Data Fig. 7i), enabling deeper analysis of the KDs that had been depleted from the organoids with long-term KD (Fig. 2e). As expected, these organoids contained a mixture of cell types (Fig. 3e and Extended Data Fig. 7j), with ExNs by far the most abundant cell type at this developmental time point (Extended Data Fig. 7k). Many KDs resulted in ExN DEGs that were enriched for the DEGs identified from NPC cultures (Extended Data Fig. 7l), again suggesting shared changes in gene expression across cell types upon KD of ASD genes. Furthermore, acute KD of the target genes that had been depleted from organoids over the course of long-term KD (Fig. 2e) allowed us to obtain much greater cell numbers for these perturbations, which showed that these KDs also resulted in strong transcriptomic convergence (Fig. 3f).

Regulators such as REST implicated in convergence

Transcriptional regulators that could mediate transcriptomic convergence in ASD were identified through MAGIC analysis (Methods), which uses chromatin immunoprecipitation followed by sequencing (ChIP–seq) data from ENCODE to identify transcription factors and cofactors whose binding targets are enriched within a gene set of interest. This identified transcriptional regulators whose target genes were recurrently differentially expressed across many of the ASD gene KDs in NPCs (Extended Data Fig. 8a and Supplementary Table 11), including HC-SFARI genes such as CHD2 and SIN3A. Of the recurrent MAGIC regulators not previously implicated in ASD, several were both highly constrained (Supplementary Table 7) and co-expressed with known ASD genes (Extended Data Fig. 8a), features that have been demonstrated to be predictive of novel ASD risk genes4.

MAGIC analysis also identified consistent drivers of convergence in organoids (Extended Data Fig. 8a and Supplementary Table 11). Of the top recurrent transcriptional regulators identified in NPCs, half (13 of 26, 50%) were also identified in organoid radial glial progenitors (Extended Data Fig. 8b). MAGIC also identified recurrent ExN regulators shared across developmental stages and across long-term versus acute KD timescales, including many SFARI genes. Nine of these genes were also identified as top transcriptional regulators in neural progenitors, indicating that they may be key drivers of convergence in various cellular contexts.

Of particular interest, the RE1 silencing transcription factor (REST) was significant according to MAGIC analysis in 17 of 18 (94.4%) NPC perturbations (Extended Data Fig. 8a) and consistently emerged as a top MAGIC regulator across all tested conditions (Extended Data Fig. 8b). Although direct mutations in REST have not previously been associated with ASD, REST has critical neurodevelopmental roles and functions as a master negative regulator of neurogenesis40,41. Dysregulation of REST has also been reported following disruption of ASD genes including CHD8 (ref. 42) and Rett syndrome gene MECP2 (refs. 43,44). Re-analysis of human and mouse REST ChIP–seq binding sites (Methods) identified strong enrichment of SFARI genes (Padj = 3.6 × 10−67 and 7.7 × 10−50, respectively), including HC-SFARI genes (Padj = 2.2 × 10−13 and 3.7 × 10−8, respectively), among REST targets (Extended Data Fig. 8c), supporting a key role for REST in regulation of ASD-relevant gene expression.

GRN analysis reveals highly interconnected ASD genes

De novo GRN reconstruction from our NPC RNA-seq data using ARACNe enabled association of further transcription factors and chromatin modifiers (together referred to as ‘regulators’) with their downstream targets (regulons) (Methods). We validated this approach by separately constructing ‘held-out’ GRNs that excluded data from samples in which the KD target itself was a potential regulator, thereby preventing circularity in regulon inference. Regulons predicted from these held-out GRNs strongly overlapped with the DEGs observed following KD of the corresponding target genes in NPCs (Padj < 3.98 × 10−116 for each, combined P = 4.08 × 10−937) (Fig. 4a and Extended Data Fig. 9a). Thus, ARACNe effectively captures bona fide regulatory relationships.

Fig. 4: GRN analysis identifies CHD8 as a top driver of convergence in ASD.

a, Volcano plot depicting differential expression upon KD of KMT2E. Genes predicted by GRN analysis to be positively regulated by KMT2E are highlighted, with statistically significantly decreased DEGs shown in blue, increased DEGs shown in red (none) and targets with no significant change shown in yellow. Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini–Hochberg correction for multiple hypothesis testing. Statistically significant enrichment of second-degree (2°) regulon genes among the downregulated DEGs is indicated by the Padj value (one-sided Fisher’s exact tests with Benjamini–Hochberg correction for multiple hypothesis testing). b, Depiction of a subset of the NPC GRN constructed using ARACNe. All regulators are shown along with first-degree (1°) target genes that are SFARI genes. c, Gene set enrichment analysis (weighted Kolmogorov–Smirnov running-sum statistic, permutation test, with Benjamini–Hochberg correction for multiple hypothesis testing) demonstrating enrichment of genes with high PageRank scores within the gene sets shown. d, Bar graph depicting mean normalized (norm.) degree, a measure of interconnectedness, across regulators from different gene sets. e, Depiction of modules of top regulators identified through iterative walktrap clustering. Hashed lines indicate that the lines are not to scale. Genes of particular interest are highlighted with large arrows. Red indicates genes directly targeted for KD in NPCs. f, Bar plot depicting enrichment of CHD8 target genes within several gene sets. Data from CHD8 ChIP–seq in NPCs from ref. 45. Bars indicate the OR from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR. P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. Rel., relative.

Source data

The full NPC GRN revealed a high degree of interconnection between ASD genes and their upstream regulators (Fig. 4b). Indeed, ASD genes were strongly enriched among central genes of the NPC GRN (Padj = 9.81 × 10−6 for SFARI genes, 1.99 × 10−4 for HC-SFARI genes, centrality assessed using the PageRank algorithm, Methods) (Fig. 4c). By contrast, intellectual-disability-only genes did not exhibit enrichment among the central genes of the GRN (Padj = 0.19), consistent with the lack of enrichment that we observed among the DEGs resulting from ASD gene KDs (Fig. 1c). In addition, central GRN genes were enriched for pathways associated with the G2M checkpoint (Padj = 6.0 × 10−4) and mitotic spindle (Padj = 3.6 × 10−5), consistent with the NPC proliferation phenotypes we observed (Fig. 1f and Supplementary Table 10).

CHD8 implicated askey driver of convergence in ASD

We next identified GRN regulators that acted as key drivers of ASD-associated transcriptomic convergence. This analysis revealed 70 top drivers of convergence whose activity was consistently altered by ASD gene KD (Methods) and whose targets were both enriched for SFARI genes and central to the GRN (Supplementary Table 12). Nearly half of these top drivers of convergence were SFARI genes (31 of 70, OR = 6.35, Padj = 3.9 × 10−11), and nearly one-third were HC-SFARI genes (22 of 70, OR = 10.61, Padj = 6.7 × 10−12). By contrast, only 7 of these top drivers of convergence were intellectual-disability-only genes (7 of 70, OR = 2.06, Padj = 0.078). Moreover, the top drivers of convergence were consistently downregulated across nearly all ASD gene KDs (Extended Data Fig. 9b). They were also highly interconnected in the GRN (P < 1 × 10−6) (Fig. 4d), as well as in the STRING database of protein–protein interactions (P = 2.0 × 10−9) (Methods). Taken together, these findings suggest that these genes both act as central regulators of many ASD genes and cross-regulate each other, driving transcriptomic convergence in ASD.

Iterative walktrap clustering identified 4 gene modules among the 70 top drivers of convergence (Fig. 4e), one of which (the blue module) was especially strongly interconnected. This module exhibited very high cross-interaction through STRING analysis (P < 1 × 10−16) and included REST (Extended Data Fig. 8) and several notable ASD genes including CHD8, one of the strongest known risk genes for ASD18. Notably, the inferred CHD8 regulon from the GRN was strongly enriched (first-degree regulon Padj = 4.6 × 10−7; second-degree regulon Padj = 3.9 × 10−20) for direct CHD8 target genes previously identified through ChIP–seq from NPCs45 (Fig. 4f), further supporting the validity of ARACNe-predicted regulatory relationships. Moreover, consistent with previous reports45,46, CHD8 ChIP–seq targets were also highly enriched for NPC-expressed SFARI genes (Padj = 2.6 × 10−8) and HC-SFARI genes (Padj = 8.5 × 10−8). The CHD8-containing module also included three transcriptional regulators (SETD2, SETD5 and KMT2E) that we had directly perturbed, and DEGs from these KDs exhibited significant enrichment for CHD8 ChIP–seq targets (Padj = 1.2 × 10−5, 7.8 × 10−8 and 2.6 × 10−4, respectively) (Fig. 4f), further demonstrating convergent regulation of shared targets by critical ASD genes. This analysis highlights CHD8, a high-confidence ASD risk gene, as a key driver of transcriptomic convergence in ASD.

ZFX is a female-biased regulator of ASD genes

Previous studies have found that genes downregulated in ASD tend to be more highly expressed in the neurotypical female versus male brain47,48, raising the possibility that this baseline female-biased expression may contribute to the FPE8,47,48. We identified 82 genes that were consistently downregulated by ASD gene KD and exhibited female-biased expression in the brain (Methods); 28 of these were highly constrained (Extended Data Fig. 10a), including 12 highly constrained SFARI genes (OR = 2.79, P = 8.65 × 10−3) (Fig. 5a). Re-analysis of single-nucleus RNA-seq (snRNA-seq) data (Methods) from the frontal cortex of neurotypical male controls and autistic male participants—including individuals with 15q duplication syndrome and genetically idiopathic ASD—revealed that many of these highly constrained genes were significantly downregulated in ExNs from autistic male individuals, further supporting their relevance to ASD (Fig. 5a and Extended Data Fig. 10a).

Fig. 5: X-linked transcription factor ZFX is a female-biased regulator of ASD genes.

a, Heatmaps showing relative expression of SFARI genes that are candidate contributors to the FPE in ASD. Left, expression in NPCs across KD conditions relative to NTCs, with statistically significant DEGs (indicated by an asterisk) identified using DESeq2 (Wald test) with Benjamini–Hochberg correction for multiple hypothesis testing. Right, expression in ExNs from male participants with ASD relative to male idiopathic controls, re-analysed from data from ref. 64, with statistically significant DEGs (indicated by an asterisk) identified using MAST (likelihood ratio test) with Bonferroni correction for multiple hypothesis testing. Lack of detection is indicated by X. See also Extended Data Fig. 10a. b, Bar plot depicting enrichment of ASD-relevant gene sets within the ZFX regulon, as identified by GRN analysis. Bars indicate the OR from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR. P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. c, Depiction of regulatory relationships between ZFX and the top drivers of convergence in ASD (previously identified in Fig. 4e), as identified by GRN analysis. Statistically significant enrichment of the top drivers of convergence within the ZFX regulon is indicated by the Padj value (one-sided Fisher’s exact tests with Benjamini–Hochberg correction for multiple hypothesis testing). d, Bar plot depicting enrichment of ASD-relevant gene sets within the following ZFX target sets: (1) targets in C4-2B cells, identified in ref. 53; (2) common targets identified in ref. 65; and (3) targets in MCF-7 cells, identified in ref. 53. Bars indicate the OR from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR. P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. e, Schematic depicting a model of the contribution of ZFX to the FPE in ASD. expr., expression.

Source data

ZFX is of particular interest because it is expressed from both the active and inactive X chromosome in female individuals10, making it one of the most highly and consistently female-biased genes across diverse tissues49, including the developing cerebral cortex (Extended Data Fig. 10b,c). Moreover, a recent report has implicated pathogenic ZFX variants in ASD50, and other coding variants in the ZFX gene have been identified in ASD probands3,51,52. ZFX encodes a potent transcriptional activator of genome-wide targets53, and we identified many regulatory relationships between ZFX and ASD-relevant gene sets in our GRN (Fig. 5b), including the top drivers of transcriptomic convergence in ASD that we had previously identified (OR = 9.8, Padj = 2.2 × 10−6). We also found that ZFX was a downstream target of many of these top drivers, with several reciprocal regulatory relationships (Fig. 5c). The strong enrichment of ZFX targets for ASD-relevant genes was further confirmed by re-analysis of ZFX ChIP–seq data (Fig. 5d). As ZFX is recurrently downregulated upon disruption of ASD genes, its female-biased expression, in combination with its role as a regulator of ASD-relevant genes, could result in cascading and broad sex-differential impacts on ASD gene expression, potentially contributing to the FPE in ASD (Fig. 5e).

Discussion

Here we have shown widespread transcriptomic convergence following depletion of ASD genes and identified key chromatin modifiers and transcription factors, including CHD8 and REST, as critical drivers of this convergence. These drivers were highly interconnected with each other, often through mutual regulatory relationships; these findings demonstrate how disruption of a single regulator can lead to changes that propagate throughout a GRN. We also identified the transcription factor ZFX, which is expressed from both the active and inactive X chromosome in female individuals, as a key transcriptional regulator of ASD genes; the female-biased expression of ZFX could thus buffer the effects of ASD gene disruptions in female individuals, contributing to the FPE in ASD. Taken together, these findings provide a model for how diverse genes can converge on core transcriptomic changes and generate shared ASD phenotypes, as well as how sex-differential activity of these genes can lead to male-biased risk for ASD.

Our detection of widespread transcriptomic convergence probably benefited from several experimental design choices that were intended to minimize variability. For instance, to avoid confounding effects of different genomic backgrounds, we perturbed the ASD genes in an isogenic background. In addition, we performed these perturbations by introducing KD constructs in a polyclonal manner instead of creating different cell lines for each ASD target gene, as the process of clonal selection can introduce line-to-line variability54. For our cerebral organoid studies, we created the FLICK KD construct so that we could generate mosaic organoids containing all KD conditions as well as the NTCs, minimizing any effects from organoid-to-organoid variability. By reducing variability across our experimental designs, we probably enhanced our sensitivity to detect convergence.

We also optimized our RNA-seq approaches for the different experimental contexts. Initially, we used highly homogeneous NPC cultures, enabling us to perform bulk RNA-seq on this single cell type. This allowed deep transcriptomic profiling, avoiding the problems of ‘dropout’ that affect scRNA-seq55. However, for samples that contained a heterogeneous mix of different cell types, such as cerebral organoids, bulk RNA-seq results could reflect differences in relative cell type abundance as well as changes in gene expression. Thus, we used scRNA-seq to analyse our organoids and sought to compensate for the decreased sensitivity of this approach by sequencing a large number of cells for each KD condition. As our downsampling analysis demonstrated (Fig. 3b), these large cell numbers were critical for detection of transcriptomic convergence through scRNA-seq. Furthermore, for our NPC experiments, we performed RNA-seq 24 h after introducing the KD constructs. We selected this short timeframe to focus our analysis on the initial effects of each KD, rather than long-term consequences that could be more strongly influenced by compensatory responses and stochastic effects. This short timeframe may have been particularly well suited for detection of transcriptomic convergence, which would have further benefited our GRN reconstruction.

Building an understanding of the complex relationships in GRNs is critical for identification of key ‘rate-limiting’ genes that affect the GRN as a whole and lead to particular biological outcomes. Previous studies of Hirschsprung disease have identified an analogous GRN centred on rate-limiting gene RET, which can be affected by direct mutations in RET itself or disruptions to upstream regulators of RET56. Here we similarly identified critical genes, including CHD8 and REST, the effects of which can propagate throughout the GRN and drive transcriptomic convergence in ASD. Our findings are consistent with those of previous studies identifying CHD8 as an important regulator of ASD gene expression45,46,57. Moreover, haploinsufficiency of CHD8 has been found to result in dysregulation of REST in the mouse brain42, consistent with the highly interconnected relationships we found in our GRN. In addition, disruption of Rett syndrome gene MECP2 affects REST43,44, and REST is dysregulated in ASD brain tissue relative to neurotypical controls42, further supporting its relevance to ASD. Given that REST is a master negative regulator of neuronal genes40,41 and that CHD8 affects proliferation and neurogenesis58, it is plausible that the convergent dysregulation of CHD8/REST targets could directly underlie the altered proliferation we found in NPCs (Fig. 1f), as well as the increased relative abundance of ExNs that we observed following some of the ASD gene KDs in organoids (Fig. 2f). Several previous studies have also identified imbalances in neural proliferation versus differentiation following ASD-relevant perturbations28,58,59; thus, future work is needed to further characterize the molecular mechanisms linking our observed convergent changes in gene expression to these shared neurodevelopmental phenotypes.

Our results also provide insight into the male bias of ASD. Downregulation of ZFX by KD of many different ASD genes in NPCs and lower ZFX levels in ExNs from individuals with ASD (Fig. 5a) suggest that decreased ZFX expression may be a recurrent transcriptional feature of ASD. The higher baseline level of ZFX in the female brain due to its expression from both X chromosomes could thus buffer against this ASD-associated downregulation. Consistent with this, a recent study of sex-biased gene expression across cell types in the cortex found that promoters of autosomal genes exhibiting female-biased expression were enriched for ZFX transcription factor motifs, whereas promoters of male-biased genes were not60. Male individuals have a Y-linked homologue of ZFX called ZFY; however, this encodes a weaker transcriptional activator than ZFX53, and the GRN regulon of ZFY showed no enrichment for SFARI genes (OR = 1.1, Padj = 0.36), suggesting divergent roles with respect to ASD. Further functional characterization of ZFX and other genes expressed from the inactive X chromosome may provide critical insights into how female-biased expression affects ASD risk13. Moreover, it will be valuable to directly compare the transcriptional effects of ASD gene perturbations between male cell lines, as used in this study, and female cell lines. Indeed, our use of a male rather than female founder cell line may have contributed to the striking changes in gene expression we observed, as female cells may be able to better tolerate the same ASD gene KDs owing to higher baseline expression of ZFX and/or other factors that contribute to the FPE.

As many genes that have been implicated in ASD show rare heterozygous loss-of-function mutations1, we prioritized ASD genes with a high probability of loss-of-function intolerance in our study. We also focused on genes that exhibited a ‘head-to-head’ configuration with an adjacent lncRNA; several lncRNAs in such a configuration have been found to fine-tune the expression of their neighbours61, meaning that this configuration could be used to putatively identify particularly dosage-sensitive genes. We further included two lncRNAs that had been implicated in ASD through heterozygous loss-of-function mutations15,16. Our Cas13-based KDs on average produced a 48.2% decrease in expression of the target genes (Supplementary Table 2); thus, this experimental approach seems to be suitable for study of how heterozygous loss of function of haploinsufficient genes can contribute to ASD.

Although our perturbations were focused on genes associated with ASD through rare single-gene mutations, it is likely that other classes of ASD-associated variants disrupt the same convergent network. Supporting this, studies of postmortem brain tissue have demonstrated that idiopathic ASD cases share core transcriptomic changes38 and have overlapping DNA methylation signatures62 with cases of ASD caused by rare mutations of large effect size, such as chromosome 15q11–13 duplication. Moreover, reports suggest that common alleles of smaller effect size that have been implicated in ASD can also drive changes that converge with those caused by rare ASD variants63. This raises the possibility that distinct causes of ASD, including rare single-gene mutations, large copy number variants and common genetic variation, could result in similar phenotypes by converging on disruption of shared ASD-relevant GRNs. Further characterization of these GRNs has broad potential to clarify our understanding of the genetic architecture of ASD across a wide variety of genetic variant classes and to identify potential therapeutic targets that could benefit individuals with diverse causes of ASD. More broadly, although this work was focused on ASD, the approaches used here are widely applicable to different cell types and conditions. Thus, this study provides a framework for identification of critical transcriptional regulators across diverse contexts, from development through disease.

Methods

KD with CasRx

CasRx17, specifically NLS-RfxCas13d-NLS (Addgene, 109049), was used in combination with target-specific gRNAs for KD of targeted transcripts. A deep-learning-based tool66 was used to design gRNA spacer sequences targeting the genes of interest. These spacer sequences were filtered to exclude homopolymers (more than four As, more than three Cs, Gs or Ts). Spacer sequences with potential off-target interactions were identified using BLAST67 and excluded. For each target gene, three separate spacer sequences were selected and combined into a tandem pre-gRNA array, with each spacer sequence preceded by the CasRx direct repeat (DR36) sequence66: CAAGTAAACCCCTACCAACTGGTCGGGGTTTGAAAC. Tandem pre-gRNA arrays were processed by CasRx into individual gRNAs for KD17. See Supplementary Table 1 for full sequences.

Molecular cloning

To create constructs for the initial NPC KD experiments, we used the lentiviral pSicoR plasmid68 (Addgene, 11579) as the backbone. The mouse U6 promoter was replaced with the human U6 promoter, which was used to drive expression of the CasRx gRNA arrays. These constructs were used in combination with a CasRx knock-in cell line (described below).

To obtain enough NPCs for CasRx KD validation by auto-western blot (described below), an all-in-one construct was created that contained CasRx and EGFP expressed from the CAG promoter element and gRNA arrays expressed from the human U6 promoter. The backbone also contained inverted terminal repeats, which allowed the construct to be integrated into the genome by piggyBac transposase69. These constructs were based on the LiOn design70 (Addgene, 154019), such that expression of CasRx and EGFP was dependent on recombination of the construct by piggyBac transposase, preventing expression from the episomal plasmid. This all-in-one construct was also used for the fluorophore KD assays and NPC proliferation assays. For expression of fluorophores in NPCs, iOn-based constructs were created with mKO2, mScarlet or iRFP670 driven by the CAG promoter element.

For shRNA-based KD, shRNAs were designed using the Horizon siDesign Center tool (https://horizondiscovery.com/en/ordering-and-calculation-tools/sidesign-center). Predicted shRNAs starting with G were selected, and those with potential off-target interactions were identified using BLAST67 and excluded. The selected shRNAs (Supplementary Table 13) were cloned into the pSicoR plasmid68 (Addgene, 11579).

For organoid experiments, we created the FLICK construct by engineering several features into a backbone incorporating the FLEx design71. In particular, we included a modified Cre recombinase that included an intron, which prevented recombination of the plasmid during cloning72. We also included CasRx17 as well as all necessary components of the Tet-On 3G system73 to enable doxycycline-inducible recombination. Furthermore, we included identifying barcode sequences a short distance upstream of the polyA sites for the fluorescent reporters, which enabled us to determine which KD construct was incorporated in each cell through scRNA-seq. Finally, we incorporated inverted terminal repeats, allowing the construct to be integrated into the genome by piggyBac transposase69. The sequences of all constructs were verified by whole-plasmid sequencing (Plasmidsaurus).

Cell culture

All human cell culture work was performed under protocols approved by Boston Children’s Hospital IRB (S07-02-0087). The control male human iPS cell line 280 was previously generated and fully characterized74. Human iPS cells and iPS cell-derived NPCs were grown on tissue culture plates precoated with Matrigel (Corning, 354277) or Geltrex (Thermo Fisher, A1413302). The cells were passaged as needed using Accutase (Millipore Sigma, SCR005) or ReleSR (STEMCELL Technologies, 100-0483). Human iPS cells were cultured in mTeSR Plus medium (STEMCELL Technologies, 100-0276). Human NPCs75 were generated from iPS cells using a STEMdiff SMADi Neural Induction Kit (STEMCELL Technologies, 08582) by following the monolayer culture protocol for at least 3 weeks. NPCs were continuously grown in neural induction medium plus SMADi and were not switched to neural progenitor medium. The medium was supplemented with 10 μM Y-27632 (STEMCELL Technologies, 72304) for approximately 1 day following passaging of iPS cells or NPCs.

Cerebral organoids were generated from 280-iPS cells using a STEMdiff Cerebral Organoid Kit (STEMCELL Technologies, 08570), following some recommendations from Giandomenico et al.76 with further modifications. In particular, the organoids were embedded in Matrigel (Corning, 354277) at day 7 and grown in expansion medium until day 14, at which point they were removed from Matrigel and switched to maturation medium on an orbital shaker. Full details are available at protocols.io (https://doi.org/10.17504/protocols.io.x54v95p5ql3e/v1).

None of the cell lines was authenticated. All cell lines routinely tested negative for mycoplasma contamination.

Derivation of CasRx knock-in iPS cells

CasRx17 was knocked into the AAVS1 safe-harbour locus of 280-iPS cells using targeting constructs that have been previously described73. Individual iPS cell colonies were picked and expanded. Southern blot analysis was performed by Celplor to confirm correct targeting of the AAVS1 locus. Karyotype analysis was performed using KaryoStat (Thermo Fisher); this revealed a 2.4-Mb gain on chromosome 1, described in microarray nomenclature as arr[GRCh37] 1q32.1(204108163_206539638)x3. This region did not contain any of the target genes.

Lentivirus

Lentiviruses were generated using standard protocols that have previously been described77. Lentiviruses were applied at a multiplicity of infection of 4.0, as this value yields more than 98% transduced cells. RNA was extracted 24 h later.

Library preparation for RNA-seq

Bulk RNA was extracted using a Direct-zol RNA Microprep Kit (Zymo Research, R2062), and 3′ RNA-seq libraries were prepared following a modified version of the TagSeq protocol78, incorporating recommendations and pooling strategies from the CheapSeq protocol79. Briefly, 200 ng of total RNA from each sample was combined with an indexed oligo(dT) primer and buffer (NEB, M0466L) as described for CheapSeq. RNA was fragmented in a preheated thermocycler at 94 °C for 2.5 min without the lid, after which the samples were moved immediately to ice. Each sample was then mixed with a template-switching oligo that contained an index and a unique molecular identifier (UMI), as well as other reagents required for reverse transcription, as described for CheapSeq. The reverse transcription reaction was performed at 42 °C for 90 min, without heat inactivation. Remaining single-stranded RNA and oligos were then digested by addition of 0.5 μl of buffer EB (Qiagen, 19086), 0.5 μl of thermolabile ExoI (NEB, M0568S) and 0.25 μl of RNase A (Thermo Fisher, EN0531) and incubation of the samples at 37 °C for 30 min, followed by heat inactivation at 85 °C for 5 min and a hold step at 4 °C.

Up to 8 samples with different index sequences were then pooled, and cDNA was purified by performing a 1.4× left-sided solid-phase reversible immobilization (SPRI) selection (Beckman Coulter, B23317). Q5 UltraII Mastermix (NEB, M0544L) was then used, with partial read 1 and partial read 2 primers, to amplify the cDNA for 7 cycles on a C1000 Touch Thermal Cycler (Bio-Rad, 1851197), following the CheapSeq PCR parameters. Amplified cDNA was purified through a 0.7× left-sided SPRI selection (Beckman Coulter, B23317). Final library construction was performed using Q5 UltraII Mastermix (NEB, M0544L) with standard TruSeq indexing primers (NEB, E6440S), with amplification for eight cycles following the CheapSeq PCR parameters. Libraries were purified by 0.4× to 0.65× double-sided SPRI selection (Beckman Coulter, B23317). Purified libraries were assessed using a TapeStation HSD1000 (Agilent, 5067-5584 and 5067-5585). For oligo sequences, see Supplementary Table 14.

The CheapSeq PCR parameters were as follows: 98 °C for 30 s; X cycles (see above) of 98 °C for 10 s, 65 °C for 75 s; 65 °C for 5 min; and 4 °C hold.

DEG analysis of bulk RNA-seq data from CasRx KD NPCs

Three distinct NPC samples were analysed per condition. Individual KD conditions were compared with three NTC conditions, which were considered together as a set of nine control samples. Gene expression was quantified using alevin-fry with UMI deduplication (salmon-v1.9). A salmon index and transcript-to-gene map were built using GENCODE v.45. Counts from distinct transcripts that mapped to the same gene were summed. Only genes with a baseMean expression greater than 10 in NTC samples were considered for further analyses. Differential expression analysis was performed using DESeq2 (v.1.42.0), with the model ~Batch + sgRNA.

For KD confirmation, P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing for the set of 20 directly targeted genes. The raw P value for SRCAP was derived from droplet digital PCR (ddPCR), as SRCAP was filtered from the RNA-seq analysis owing to low expression. See Supplementary Table 2 for details. For transcriptome-wide analyses, P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing for the set of 11,168 analysed genes (baseMean > 10 in NTCs).

Droplet digital PCR

To confirm KD of SRCAP, the RNA samples that had been used for RNA-seq (above) were assessed by ddPCR. RNA was first converted to cDNA using SuperScript IV VILO Master Mix (Thermo Fisher Scientific, 11756050). Next, we performed ddPCR by combining cDNA at an RNA-equivalent concentration of 0.25 ng μl−1 (5 ng per 20-μl reaction), primers at a final concentration of 0.1 μM and QX200 ddPCR EvaGreen Supermix (Bio-Rad, 1864033). The ddPCR reactions were performed using a QX200 AutoDG Droplet Digital PCR System (Bio-Rad). Expression levels were normalized using the expression of housekeeping gene HPRT1. The results were analysed using the model Expr ~ Batch + type, where type was either NTC or the targeted gene (SRCAP). The resulting raw P value was adjusted with those for the other KD targets (from RNA-seq) using Benjamini–Hochberg correction for multiple hypothesis testing for the set of 20 directly targeted genes. See Supplementary Table 2 for details.

Primer design

Primers (Supplementary Table 15) were designed with Primer-BLAST80, which assesses potential off-target interactions using BLAST67. Oligos were ordered through Azenta or IDT. Primer specificity was confirmed by Sanger sequencing of PCR products (Azenta).

Auto-western blot

For analysis of protein expression levels following CasRx KD, LiOn-based KD constructs (‘Molecular cloning’) were nucleofected into 280-NPCs with a piggyBac-transposase-expressing plasmid (see ‘Nucleofection’ for details). Four days later, EGFP+ cells expressing the KD constructs were collected by FACS (BD FACSymphony S6). The cells were pelleted and flash-frozen in liquid nitrogen. Protein extraction and auto-western blot were performed by RayBiotech using antibodies against vinculin (R&D systems, MAB6896, clone 728526; diluted 1:25) and WAC (Abcam, AB109486; diluted 1:50).

Gene sets

  1. (1)

    SFARI genes: all genes included in the SFARI gene database14 (3 April 2025 release).

  2. (2)

    HC-SFARI genes: SFARI genes with a score of 1.

  3. (3)

    Intellectual disability genes: genes from the SysNDD database81 (accessed 23 August 2025) that were: (i) associated with an intellectual disability term (HP:0001249, HP:0001256, HP:0002342, HP:0010864, HP:0002187); (ii) in the ‘definitive’ category; (iii) not SFARI genes.

Enrichment analyses

Enrichment analyses were performed using one-sided Fisher’s exact tests (fisher.test, base R v.4.3.1), and P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing (p.adjust, method = “BH”, base R v.4.3.1). In bar plots, main bars indicate the OR from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR (no upper bound is defined for a one-sided test).

KD of fluorophores

Fluorophore-expressing constructs (‘Molecular cloning’) were nucleofected into 280-iPS cells with a piggyBac-transposase-expressing plasmid (see ‘Nucleofection’ for details). iPS cells were expanded, and cells expressing all three fluorophores were isolated by FACS using a Cytek Aurora CS spectral cell sorter. The sorted iPS cells were further expanded and induced into NPCs (‘Cell culture’). LiOn-based CasRx KD constructs were then nucleofected into these NPCs with a piggyBac-transposase-expressing plasmid. Four days later, cells expressing all four fluorophores (with EGFP indicating expression of the KD constructs) were analysed for fluorescence intensity of the reporters using a Cytek Aurora CS spectral cell sorter. Ordinary two-way ANOVA with Bonferroni’s multiple comparisons test was used to test for significant differences in the geometric mean fluorescence intensity (log2) across samples and reporters.

KD by shRNA

For each target, pSicoR-based constructs with three different shRNAs were pooled and then nucleofected into 280-NPCs (see ‘Nucleofection’ for details, with modifications as described here). As these constructs are not transposable, the plasmid expressing piggyBac transposase69 was omitted, and the amount of pSicoR-based plasmid was increased to a total of 15 μg. The day after nucleofection, EGFP+ cells expressing the constructs were isolated by FACS (BD FACSymphony S6) and replated to allow recovery overnight. The following day, RNA was collected using Zymo DNA/RNA Shield. Finally, 3′ RNA-seq was performed by Plasmidsaurus (DGE RNA-Seq service).

DEG analysis of bulk RNA-seq data from shRNA KD NPCs

We performed differential expression analysis using the same analytical pipeline as for our CasRx KD experiments: DESeq2 (v.1.42.0) was used with the model ~targetGene to calculate differential expression relative to NTC shRNAs (designed to target non-expressed transcripts luciferase and lacZ). For transcriptome-wide analyses, P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing for the set of 14,601 genes with baseMean > 10 in NTC cells. DEGs were defined as genes that were differentially expressed with adjusted P < 0.05.

Analysis of gene expression data from ref. 20

Gene expression counts from ref. 20 were downloaded from GEO (GSE107878), and differential expression analysis was performed with the same analytical pipeline as for our CasRx KD experiments: DESeq2 (v.1.42.0) was used with the model expression ~ targeted gene. DEGs were identified relative to control cells for both neurons and iPS cells. Only genes with mean expression greater than 10 in control samples were considered to be potential DEGs, and P values were adjusted accordingly using Benjamini–Hochberg correction. DEGs were defined as genes that were differentially expressed with adjusted P < 0.05.

Cross-enrichment analysis

We quantified the overlap between DEGs identified from: (1) our CasRx KD experiments in NPCs, (2) our shRNA KD experiments in NPCs, (3) knockout iPS cells from ref. 20 and (4) knockout ExNs from ref. 20. For each pairwise comparison of DEGs, enrichment analyses were performed using one-sided Fisher’s exact tests with a universe of genes that were expressed by both sample types. P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing.

Analysis of co-expression with SFARI genes

To determine whether a gene of interest was significantly co-expressed with SFARI genes in our NPC RNA-seq data, we calculated the median Pearson correlation between the expression of the gene of interest and that of each SFARI gene. To assess whether the degree of correlation was significant, we created an empirical null distribution by calculating the median correlation between the gene of interest and 1,000 randomly selected gene sets of the same size as the SFARI gene set, using the universe of NPC-expressed genes (baseMean > 10 in NTCs). A one-sided empirical P value was then calculated as the proportion of null values greater than or equal to the observed median correlation. Regulators with P < 0.05 were considered to be significantly correlated with SFARI genes.

Identification of consistently downregulated DEGs

Consistently downregulated DEGs were defined as genes that exhibited: (1) statistically significant downregulation by at least three of the KDs, and (2) at least 80% concordance in the direction of significant differential expression across all KDs.

Constraint analysis

LOEUF scores (oe_lof_upper) were obtained from gnomad v.2 (ref. 82) for evaluation of genic constraint. For each gene, we calculated a per-gene LOEUF score by calculating the median of the LOEUF values of each gene’s associated transcripts. Although LOEUF is a continuous metric, we used LOEUF < 0.35 as recommended to identify highly constrained genes.

Further support for candidate ASD risk genes

Candidate ASD risk genes (Supplementary Table 8) that were identified in any of the following major sequencing studies are included in Supplementary Table 9: refs. 2,52,83,84,85,86.

Further references provide additional support for some of these candidate ASD risk genes: refs. 87,88,89,90.

Nucleofection

Nucleofections were performed using a Neon Transfection System (Thermo Fisher Scientific). For NPCs, 2 million cells were nucleofected with 10 μg of transposable plasmid and 2 μg of a plasmid expressing piggyBac transposase69 using 100-μl tips. For iPS cells, 1 million cells were nucleofected with 5 μg of transposable plasmid and 1 μg of a plasmid expressing piggyBac transposase69 using 100-μl tips. For small-scale nucleofections, 10-μl tips were used, and cell numbers and plasmid amounts were reduced proportionally. Nucleofection of NPCs was performed using the following settings: voltage = 1,600 V, width = 10, pulses = 3. Nucleofection of iPS cells was performed with voltage = 950 V, width = 30, pulses = 2.

NPC proliferation assays

LiOn-based KD constructs (‘Molecular cloning’) were nucleofected into 280-NPCs with a piggyBac-transposase-expressing plasmid (see ‘Nucleofection’ for details). Four days later, the cells were treated for 20 min with dye from a CYTOTRACK Red 628/643 Cell Proliferation Assay Kit (Bio-Rad, 1351205) diluted 1:500 in medium. The cells were then rinsed with phosphate-buffered saline (PBS; Thermo Fisher Scientific, 10010031) and passaged, with one-fifth of each sample plated into a well of each of four separate plates to be used for the subsequent four time points. The remaining cells were analysed by flow cytometry (BD FACSDiva software v.8.0.3) for assessment of the initial CytoTrack intensity on day 0. For the next 4 days, one well of each sample was dissociated each day and analysed by flow cytometry. The CytoTrack intensity values (negative log2) from the EGFP+ cells expressing the KD construct were normalized to those of the EGFP− cells from the same well to control for well-to-well variability in proliferation rates. They were further normalized to their initial CytoTrack intensity from day 0 to control for variability in dye incorporation. The results for each KD are displayed relative to the mean value from the NTC samples, plotted as the mean ± s.e.m. Statistical analysis was performed using ordinary two-way ANOVA with Dunnett’s multiple comparisons test (GraphPad Prism). Full details of sample numbers and statistical results are available in Supplementary Table 10.

Induction of FLICK recombination

To induce expression of Cre and initiate recombination of the FLICK construct, cells or organoids were treated with 2 μg ml−1 doxycycline (Sigma, D9891-1G) for 24 h.

Assessment of FLICK KD efficiency

To determine KD efficiency using FLICK constructs, we analysed three distinct samples per condition. Control 280-iPS cells were nucleofected with constructs encoding the NTC1 control gRNA array or a gRNA array targeting the gene ENSG00000261840, with a plasmid expressing piggyBac transposase69 to genomically integrate the FLICK constructs. For induction of KD, the cells were treated with 2 μg ml−1 doxycycline (Sigma, D9891-1G) for 24 h. Cells that had recombined the FLICK construct were then isolated through FACS (BD FACSymphony S6) for EGFP+ cells, which were replated in triplicate and grown for a further 2 days. RNA was then isolated using a Direct-zol RNA Microprep Kit (Zymo Research, R2062). To generate cDNA, we reverse transcribed 500 ng of RNA using SuperScript IV VILO Master Mix (Thermo Fisher Scientific, 11756050). We then performed quantitative PCR (qPCR) by combining cDNA at an RNA-equivalent concentration of 1 ng μl−1 (10 ng per 10-μl reaction), primers at a final concentration of 0.5 μM and PowerUp SYBR Green Master Mix for qPCR (Thermo Fisher Scientific, A25742). The qPCR reactions were run in technical triplicate for each sample on a CFX384 Touch Real-Time PCR Detection System (Bio-Rad, 1855484) using the ‘fast cycling mode’ conditions described in the Master Mix protocol. Expression of the KD target gene (ENSG00000261840) was assessed using the ΔΔCt method, with expression first normalized to that of housekeeping gene HPRT1 for each sample and then displayed relative to the average of the NTC control samples, plotted as mean ± s.e.m. Statistical significance was assessed using a two-tailed t-test (GraphPad Prism).

Processing of cerebral organoids for histological analysis

Cerebral organoids were fixed in 4% paraformaldehyde, diluted in PBS from 16% paraformaldehyde (Electron Microscopy Sciences, 15710), at 4 °C overnight. Following fixation, cerebral organoids were rinsed with PBS and then equilibrated in 30% sucrose (w/v in PBS) at least overnight at 4 °C. For large organoids, the 30% sucrose solution was replaced at least once following overnight incubation to ensure complete cryoprotection. For cryosectioning, cerebral organoids were frozen in Tissue-Tek O.C.T. Compound (Sakura, 4583) on dry ice. A Leica CM3050 S cryostat was used to obtain 30-μm sections, which were collected on Fisherbrand Superfrost Plus Microscope Slides (Fisher Scientific, 12-550-15) and stored at −80 °C.

For immunohistochemical analysis, we prepared a blocking buffer by combining 0.3 M glycine (Sigma, G7403-1KG), 1% bovine serum albumin (Sigma, A4503-50G), 0.3% Triton X-100 (Sigma, T8787-50ML), 10% normal goat serum (Fisher Scientific, NC9660079) and PBS. Slides were rinsed in PBS to remove O.C.T. and then incubated in blocking buffer for 1 h at room temperature. Blocking buffer was removed and replaced with fresh blocking buffer containing the primary antibodies of interest, and the slides were incubated in a slide moisture chamber (Avantor, 76278-848) at 4 °C overnight. Slides were then rinsed with PBS three times for 5 mins each time. Secondary antibodies and nuclear stain DAPI (Thermo Fisher Scientific, 62248) were diluted in blocking buffer (1:500 and 1:1,000, respectively) and added to the slides for 30 min at room temperature; the slides were protected from light from this point forward. The slides were then rinsed in PBS three times for 5 min each time. Cover slips (Epredia, 152455) were mounted using Fluoromount-G (Fisher Scientific, OB100-01), which was allowed to dry at least overnight at room temperature. After drying, slides were stored at 4 °C.

Antibodies for immunohistochemistry and immunocytochemistry

The following antibodies were used: PAX6 (BD Pharmingen, 561462, clone O18-1330), diluted 1:300; Nestin (Millipore Sigma, HPA026111-25UL), diluted 1:500; TBR2 (Millipore Sigma, HPA028896-100UL), diluted 1:300; CTIP2 (Abcam, ab18465, clone 25B6), diluted 1:300; EGFP (Aves, GFP-1010), diluted 1:1,000; Goat anti-chicken IgY, Alexa Fluor 488 (Thermo Fisher, A-11039), diluted 1:500; Goat anti-rabbit IgG, Alexa Fluor 488 (Thermo Fisher, A-11008), diluted 1:500; Goat anti-rat IgG, Alexa Fluor 488 (Thermo Fisher, A-11006), diluted 1:500; Goat anti-mouse IgG, Alexa Fluor 555 (Thermo Fisher, A-21424), diluted 1:500; Goat anti-rat IgG, Alexa Fluor 647 (Thermo Fisher, A-21247), diluted 1:500; Goat anti-rabbit IgG, Alexa Fluor 647 (Thermo Fisher, A-21245), diluted 1:500; Goat anti-rabbit IgG, Alexa Fluor 750 (Thermo Fisher, A-21039), diluted 1:500.

Confocal imaging and processing

Cerebral organoid sections were imaged using a Zeiss LSM 980 confocal microscope. Standard image processing was performed using Adobe Photoshop.

Cerebral organoid scRNA-seq

For scRNA-seq, cells were collected from:

  • 1 month: 10 organoids, pooled together and collected as a single batch;

  • 2 months: 14 organoids collected in 2 batches, with 7 organoids collected together on day 60 and a further 7 organoids collected together on day 61;

  • acute KD: 8 organoids, pooled together and collected as a single batch.

For collection of cells, mosaic cerebral organoids were cut into pieces using spring scissors, rinsed twice with PBS to remove debris, and then dissociated using a Neural Tissue Dissociation Kit – Postnatal Neurons (Miltenyi Biotec, 130-094-802) with the enzyme mixes described for brain tissue. Gentle dissociation was performed by trituration with pipette tips of decreasing diameter, starting with wide-orifice tips (Rainin, 30389218) and progressing to standard tips. Cells with recombined FLICK constructs were isolated through FACS (BD FACSymphony S6) for EGFP+ cells. Transcriptome-wide scRNA-seq libraries were prepared using Chromium Next GEM Single Cell 3′ Reagent Kits v.3.1 (Dual Index) from 10x Genomics, following user guide CG000316 Rev. A.

Construction of FLICK barcode sequencing libraries

Full details of protocols are available from protocols.io (https://doi.org/10.17504/protocols.io.5qpvoj5pzg4o/v1). Briefly, for quantification of FLICK constructs from the initial pool of iPS cells used to generate mosaic organoids (day 0), RNA was extracted from triplicate samples of 42,000 cells using a Direct-zol RNA Microprep Kit (Zymo Research, R2062). For generation of cDNA, 50–250 ng of RNA was reverse transcribed using a SuperScript IV First-Strand Synthesis System (Invitrogen, 18091050) with an oligo(dT)VN primer containing UMIs. Sequencing libraries were then generated using Q5 Hot Start High-Fidelity 2X Master Mix (NEB, M0494L) and custom primers.

To generate FLICK barcode libraries from scRNA-seq libraries, we used 5 μl of amplified and cleaned cDNA (produced by following the 10x User Guide CG000316 Rev. A to the end of step 2). Barcode-containing cDNA was amplified for sequencing using Q5 Hot Start High-Fidelity 2X Master Mix (NEB, M0494L) and custom primers. Libraries were cleaned using SPRIselect DNA Size Selection Reagent (Beckman Coulter, B23317) and assessed using TapeStation High Sensitivity D1000 reagents (Agilent, 5067-5603 and 5067-5584).

Alignment and quantification of organoid scRNA-seq data

We aligned the scRNA-seq data to the prebuilt Cell Ranger (v.9.0.0) hg38 reference using cellranger-count in feature-barcode mode to perform simultaneous gene expression and guide abundance quantification. Introns were not included in gene expression quantification (--include-introns=false) to avoid quantification of partially degraded transcripts that had been targeted by CRISPR–CasRx. CRISPR guide sequencing library FASTQs for 2-month samples were downsampled to 1.25% of reads using seqtk (v.1.4) to match CRISPR library sequencing depth for 1-month samples and to minimize background for guide assignment using the cellranger-CRISPR guide quantification pipeline. All other parameters were set to default values. Cell Ranger guide assignments were used to filter to only cells with exactly one assigned guide.

Preprocessing and quality control of organoid scRNA-seq data

Long-term and acute KD samples were processed separately but equivalently. CellBender (v.0.3.2) was used for detection and filtering of ambient RNA contamination, whereas scrublet (v.0.2.3) was used for automatic doublet detection and filtering. All other downstream analysis was performed with Seurat (v.5.0.0). High-quality cells were retained according to the following quality control metrics: (1) percent mitochondrial (percent.mt) gene expression < 20% and (2) 500 < number of expressed genes (nFeature_RNA) < 6,000.

Cell type annotation of organoid scRNA-seq data

Cell type annotation was first performed on only NTC cells. NTC cells were extracted to create a control-only Seurat object. This object subsequently underwent normalization (NormalizeData), variable feature identification (FindVariableFeatures), scaling (ScaleData; regressed covariates of percent.mt, percent_ribo, nCount_RNA and nFeature_RNA) and dimensionality reduction (RunPCA and RunUMAP). Integration was performed using harmony (v.1.2.0) across organoid batches, and the integrated object was used for unsupervised clustering (FindNeighbors and FindClusters, resolution values of 0.6 and 1.2). A small number of unsupervised clusters (fewer than 1,000 cells) were filtered owing to low cluster-wide quality control metrics (exceptionally high or low nCount_RNA, high percent.mt). All subsequent two-dimensional visualization of this control-only Seurat object was performed using a UMAP representation of the integrated data. Cell type annotation was performed using canonical brain marker genes. A Seurat object was next constructed for all cells, following the processing pipeline (for instance, normalization, scaling, dimensionality reduction and integration) described above. For cell type annotation, label transfer was performed from the control-only Seurat object (TransferData) to the full Seurat object. Expression patterns of canonical brain marker genes were then used to refine and finalize these cell type annotations.

Analysis of relative abundance per KD in cerebral organoids

The relative abundance of cells for each KD target was determined for each time point (day 0, day 30, day 60 and day 61). For day 0, relative abundance was determined from bulk RNA-seq libraries (‘Construction of FLICK barcode sequencing libraries’) using a custom pipeline (‘Code availability’). For days 30, 60 and 61, relative abundance was determined from scRNA-seq data using the CRISPR guide assignments generated by Cell Ranger (‘Alignment and quantification of organoid scRNA-seq data’). For each sample, results were normalized to the starting proportion at day 0 and displayed relative to NTCs from the same time point. Results were plotted as the mean ± s.e.m. Statistical analysis was performed using repeated measures one-way ANOVA with Dunnett’s multiple comparisons test (GraphPad Prism), considering all long-term samples (days 30, 60 and 61).

Analysis of relative cell type proportions in cerebral organoids

The relative ExN/RG proportion was determined for each KD target at day 30, day 60 and day 61. Results were normalized to the NTC sample from the same time point. Statistical analysis was performed using ordinary one-way ANOVA with Dunnett’s multiple comparisons test (GraphPad Prism), considering all long-term samples (days 30, 60 and 61).

Differential expression analysis of cerebral organoid scRNA-seq data

Differential expression analysis was performed between perturbed cells per guide and cells that received NTCs, stratified by main cell types (radial glia, intermediate progenitor cells, ExNs and inhibitory neurons). Seurat FindMarkers was used to perform differential expression with logfc.threshold = 0, min.pct = 0.10, min.cells.group = 3 and all other parameters set to default values. Enrichment analyses of DEGs were performed using one-sided Fisher’s exact tests, and P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. The universe was defined as all genes expressed in both comparisons of interest.

Visualization of shared DEGs between organoid cell types

We visualized shared DEGs between organoid cell types (Fig. 3c) using R package igraph (v.2.1.4). The Fruchterman–Reingold layout algorithm (layout_with_fr) was used to determine final node placement.

MAGIC analyses

MAGIC91 (v.1.1) was used to identify transcription factors and cofactors whose targets were enriched in the DEGs from individual perturbations. The strategy of ‘gene body (promoter to the end of the last exon) plus 1Kb flanking sequence either side of the gene body’ was used. For NPC analyses, the universe was set as all genes with baseMean > 10 in NTC samples. For organoid analyses, the universe was set as all genes expressed in at least 10% of cells of that cell type (Seurat min.pct = 0.10; no log2 fold-change cutoff), as these were the genes assessed for differential expression.

Re-analysis of REST ChIP–seq data

Data was extracted from Table S2 from ref. 92. Mouse genes were mapped to human genes using Orthology.eg.db (v.3.20.0), org.Mm.eg.db (v.3.20.0) and org.Hs.eg.db (v.3.20.0). Enrichment analysis was performed using a one-sided Fisher’s exact test with all BrainSpan genes with normalized expression ≥ 1 as the universe (version ‘RNA-Seq Gencode v10 summarized to genes’; https://brainspan.org; accessed 6 January 2024). P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. Results are displayed as bar plots, with main bars indicating the OR from a one-sided Fisher’s exact test and error bars indicating the lower bound of the 95% confidence interval on the OR.

GRN construction and benchmarking

A GRN was constructed using bioinformatic tool ARACNe-AP93. The input comprised the RNA-seq data from our 3 NTCs and the successful KD samples with at least 50 DEGs. We used curated lists of transcription factors94 and epigenetic regulators95 and selected genes expressed in NPCs (baseMean > 10 in NTC samples) as the candidate genetic regulators. To create the GRN, we used default parameters and 100 bootstraps that were then aggregated to form the final network. We then used VIPER96 (v.1.36.0) to identify differentially active regulons following standard operating procedures. In brief, we identified regulons with significant differential expression between NTCs and KDs using a t-test, with a null distribution defined by sample permutations.

To benchmark the predictive power of the ARACNe-derived GRNs, we generated further ARACNe networks after individually holding out perturbation samples for 5 target genes that were identified as regulators of at least 250 genes (range: 478–1,219 genes) in their second-degree regulons (that is, their direct targets, which constituted their first-degree regulon, as well as the direct targets of those genes). For each held-out regulator, we assessed the overlap between the predicted positively regulated genes within its second-degree regulon and the set of downregulated DEGs resulting from perturbation of the same regulator, using a one-sided Fisher’s exact test. P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. P values across the five held-out perturbations were combined using Fisher’s method.

PageRank analysis of GRN

To identify central regulators in our GRN, we converted the GRN to a graph using R package igraph (v.2.1.4) and ran the PageRank algorithm97 using igraph method page_rank. We then ordered genes by their PageRank score and ran gene set enrichment analysis98 using R package fgsea (v.1.32.4), testing the following ontologies: SFARI genes, HC-SFARI genes, intellectual disability genes (exclusive of SFARI genes) and the 50 Molecular Signature Database Hallmark Pathways99 (obtained from https://www.gsea-msigdb.org/gsea/msigdb).

Identification of top candidate regulators for driving convergence in autism

VIPER96 (v.1.36.0) was used to identify GRN regulators whose regulons were significantly differentially active across the KDs compared with NTC controls. From these, we prioritized 70 regulators whose regulons were significantly enriched for both SFARI genes (one-sided Fisher’s exact test with Benjamini–Hochberg correction for multiple hypothesis testing) and high PageRank scores (gene set enrichment analysis with Benjamini–Hochberg correction for multiple hypothesis testing).

Further GRN analyses

We calculated the mean normalized degree for the indicated set of regulators and their first-degree regulons (Fig. 4d) using R package igraph (v.2.1.4) by running the function mean(degree(graph, normalized = TRUE)). We then created an empirical null distribution of mean normalized degrees for each indicated set of regulators by creating 1 million subnetworks of the same set size of randomly selected GRN regulators. The relative mean normalized degree was calculated as the observed mean normalized degree for a given set of regulators divided by the mean of the values for the corresponding empirical null distribution. A one-sided empirical P value was then calculated as the proportion of null values greater than or equal to the observed mean normalized degree, with a detection limit of P = 1 × 10−6.

To evaluate known or predicted protein–protein interactions between the top 70 regulators, we used the STRING database GUI100. For the universe, we used the full set of genes that were considered as candidate genetic regulators in the GRN: transcription factors94 and epigenetic regulators95 that were expressed in NPCs (baseMean > 10 in NTC samples). All other parameters were set to default values. For clustering of our 70 regulators into modules, we performed iterative walktrap clustering using R package igraph (v.2.1.4) by running the function cluster_walktrap with default parameters.

Re-analysis of CHD8 ChIP–seq data

For enrichment analysis of CHD8 targets, we used Supplementary Data 1 from ref. 45. Enrichment of the indicated gene sets was tested using one-sided Fisher’s exact tests, with the universe set to NPC-expressed genes (baseMean > 10 in NTCs). P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. Results were displayed using bar plots, with main bars indicating the OR from a one-sided Fisher’s exact test and error bars indicating the lower bound of the 95% confidence interval on the OR.

Identification of female-biased genes

To identify genes that exhibited female-biased expression in the neurotypical brain, we combined DEG call sets from the following work: Table S4 from ref. 101 (identified as significantly female-biased in either the UCLA cohort, BrainVar cohort or through meta-analysis); ref. 102 (https://storage.cloud.google.com/adult-gtex/bulk-qtl/v8/sb-eqtl/GTEx_Analysis_v8_sbgenes.tar.gz, identified as significantly female-biased in either BRNCTXA or BRNCTXB); Table S4 from ref. 103 (identified as significantly female-biased in the brains of all assessed species (human, mice, rats, dogs, and macaques)); and Table S13 from ref. 60 (identified as ubiquitously female-biased).

Analysis of snRNA-seq from autistic male individuals

Processed snRNA-seq data from the human frontal cortex was obtained from ref. 64. The sex of individual samples was verified on the basis of expression of XIST and Y-linked genes. Only a small number of female samples (one female control sample, one sample from a female participant with idiopathic autism and one sample from a female participant with 15q-duplication-associated autism) were included in this cohort owing to the known male bias of autism. Thus, all downstream analyses focused on the 17 male samples included in this cohort. Differential expression analysis was performed using Seurat function FindMarkers on a cell-type-specific basis between (1) male controls and samples only from male individuals with idiopathic autism; (2) male controls and samples only from male individuals with 15q-duplication-associated autism; and (3) male controls and all samples from male individuals with autism, either idiopathic or 15q-duplication-associated. To control for donor effects, we ran FindMarkers using MAST104. Significant genes were defined as those with absolute log2[fold change] > 0.1 and transcriptome-wide Padj < 0.05.

Plots of sex-differential expression of ZFX

Data were used from refs. 60,101,105,106.

Re-analysis of ZFX targets from refs. 53 and 65

ZFX ChIP–seq targets were obtained from Table S3E of ref. 65. ZFX targets were also obtained from Table S10 of ref. 53, by San Roman et al., who previously performed integrative analysis of ZFX ChIP–seq data (ENCODE), and ZFX KD RNA-seq data from ref. 65. Enrichment was analysed using one-sided Fisher’s exact tests, and P values were adjusted using Benjamini–Hochberg correction for multiple hypothesis testing. For analysis of common ZFX targets from ref. 65, the universe was defined as genes expressed in both of the cell lines used for ChIP–seq for which RNA-seq data were also available (C4-2B and MCF-7). For analysis of ZFX targets from ref. 53, the universe was defined as genes expressed in the corresponding cell type from ref. 65. Results are displayed using bar plots, with main bars indicating the OR from a one-sided Fisher’s exact test and error bars indicating the lower bound of the 95% confidence interval on the OR.

Software summary

Software used in this work included R (v.4.3.1) and R packages DESeq2 (v.1.42.0), Seurat (v.5.0.0), igraph (v.2.1.4), scrublet (v.0.2.3), harmony (v.1.2.0), Orthology.eg.db (v.3.20.0), org.Mm.eg.db (v.3.20.0), org.Hs.eg.db (v.3.20.0), VIPER (v.1.36.0) and fgsea (v.1.32.4), as well as Cell Ranger (v.9.0.0), CellBender (v.0.3.2), MAGIC (v.1.1), salmon (v.1.9), seqtk (v.1.4) and BD FACSDiva v.8.0.3.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

All raw data and processed count matrices are available at GEO under accession GSE332574. Differential gene expression results can also be found in Supplementary Tables 16–65. This study also used the following publicly available datasets and websites (copies of these datasets are also available at GitHub (https://github.com/MayaTalukdar/convergence_in_autism_ms)): GENCODE v.45 (reference transcriptome annotation; https://www.gencodegenes.org); SFARI Gene database (3 April 2025 release; https://gene.sfari.org); SysNDD database (accessed 23 August 2025; https://sysndd.dbmr.unibe.ch); RNA-seq expression data from CRISPR-edited iPS cell-derived neurons and iPS cells (ref. 20; GEO accession GSE107878); gnomAD v.2 (LOEUF/oe_lof_upper constraint scores), Table S11 (ref. 82); ASD sequencing study, Table S5G (ref. 2); ASD sequencing study, Tables S2 and S9 (ref. 52); ASD sequencing study, Table S6 (ref. 83); ASD sequencing study, Table S9 (ref. 84); ASD sequencing study, Table S6 (ref. 85); ASD sequencing study, Table S5 (ref. 86); ENCODE ChIP–seq compendium, accessed through MAGIC (ref. 91); https://www.encodeproject.org); human and mouse REST/NRSF ChIP–seq binding sites, Table S2 (ref. 92); BrainSpan RNA-Seq data, ‘Gencode v10 summarized to genes’ (https://www.brainspan.org; accessed 6 January 2024); MSigDB Hallmark gene set collection (https://www.gsea-msigdb.org/gsea/msigdb); STRING protein–protein interaction database: https://string-db.org; CHD8 ChIP–seq target genes, Table S1 (ref. 45); sex-biased expression data, Table S4 (ref. 101); sex-biased expression data (ref. 102; https://storage.cloud.google.com/adult-gtex/bulk-qtl/v8/sb-eqtl/GTEx_Analysis_v8_sbgenes.tar.gz); sex-biased expression data, Table S4 (ref. 103); sex-biased expression data, Table S13 (ref. 60); sex-biased expression data, Table S2 (ref. 105); sex-biased expression data, Tables S1 and S2 (ref. 106); snRNA-seq data from frontal cortex of autistic and control male participants (ref. 64); ZFX ChIP–seq target genes, Table S3E (ref. 65); and integrated ZFX target gene analysis, Table S10 (ref. 53). Source data are provided with this paper.

Code availability

Scripts, analyses and reference files used for this study are available at GitHub (https://github.com/MayaTalukdar/convergence_in_autism_ms).

References

  1. Satterstrom, F. K. et al. Large-scale exome sequencing study implicates both developmental and functional changes in the neurobiology of autism. Cell https://doi.org/10.1016/j.cell.2019.12.036 (2020).

  2. Trost, B. et al. Genomic architecture of autism from comprehensive whole-genome sequence annotation. Cell 185, 4409–4427 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  3. Ruzzo, E. K. et al. Inherited and de novo genetic risk for autism impacts shared networks. Cell 178, 850–866 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  4. Liao, C. et al. Convergent coexpression of autism-associated genes suggests some novel risk genes may not be detectable in large-scale genetic studies. Cell Genomics 3, 100277 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  5. Maenner, M. J. et al. Prevalence and characteristics of autism spectrum disorder among children aged 8 years — Autism and Developmental Disabilities Monitoring Network, 11 sites, United States, 2020. MMWR. Surveill. Summ. 72, 1–14 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  6. Jacquemont, S. et al. A higher mutational burden in females supports a ‘female protective model’ in neurodevelopmental disorders. Am. J. Hum. Genet. 94, 415–425 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  7. Antaki, D. et al. A phenotypic spectrum of autism is attributable to the combined effects of rare variants, polygenic risk and sex. Nat. Genet. 54, 1284–1292 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  8. Werling, D. M. The role of sex-differential biology in risk for autism spectrum disorder. Biol. Sex Differ. 7, 58 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

  9. Dougherty, J. D. et al. Can the ‘female protective effect’ liability threshold model explain sex differences in autism spectrum disorder? Neuron 110, 3243–3262 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  10. Schneider-Gädicke, A., Beer-Romero, P., Brown, L. G., Nussbaum, R. & Page, D. C. ZFX has a gene structure similar to ZFY, the putative human sex determinant, and escapes X inactivation. Cell 57, 1247–1258 (1989).

    Article  ADS  PubMed  Google Scholar 

  11. Iakoucheva, L. M., Muotri, A. R. & Sebat, J. Getting to the cores of autism. Cell 178, 1287–1298 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  12. Mendes, M. et al. Chromosome X-wide common variant association study in autism spectrum disorder. Am. J. Hum. Genet. 112, 135–153 (2024).

  13. Talukdar, M. & Page, D. C. The inactive X chromosome as a female protector in autism and beyond. Nat. Genet. 58, 687–694 (2026).

    Article  CAS  PubMed  Google Scholar 

  14. Abrahams, B. S. et al. SFARI Gene 2.0: a community-driven knowledgebase for the autism spectrum disorders (ASDs). Mol. Autism 4, 36 (2013).

    Article  PubMed  PubMed Central  Google Scholar 

  15. Ang, C. E. et al. The novel lncRNA lnc-NR2F1 is pro-neurogenic and mutated in human neurodevelopmental disorders. eLife 8, e41770 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  16. Andersen, R. E. et al. Chromosomal structural rearrangements implicate long non-coding RNAs in rare germline disorders. Hum. Genet. 143, 921–938 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  17. Konermann, S. et al. Transcriptome engineering with RNA-targeting type VI-D CRISPR effectors. Cell 173, 665–676 (2018).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  18. Haddad Derafshi, B. et al. The autism risk factor CHD8 is a chromatin activator in human neurons and functionally dependent on the ERK-MAPK pathway effector ELK1. Sci. Rep. 12, 22425 (2022).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  19. Hart, S. K. et al. Precise RNA targeting with CRISPR–Cas13d. Nat. Biotechnol. 44, 64–69 (2026).

    Article  CAS  PubMed  Google Scholar 

  20. Deneault, E. et al. Complete disruption of autism-susceptibility genes by gene editing predominantly reduces functional connectivity of isogenic human neurons. Stem Cell Rep. 11, 1211–1225 (2018).

    Article  CAS  Google Scholar 

  21. Fernandez Garcia, M. et al. Transcriptomic and phenotypic convergence of neurodevelopmental disorder risk genes in vitro and in vivo. Nat. Neurosci. 29, 1079–1094 (2026).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  22. Gordon, A. et al. Developmental convergence and divergence in human stem cell models of autism. Nature 651, 707–719 (2026).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  23. Weerts, M. J. A. et al. Delineating the molecular and phenotypic spectrum of the SETD1B-related syndrome. Genet. Med. 23, 2122–2137 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  24. Parra, A. et al. Clinical heterogeneity and different phenotypes in patients with SETD2 variants: 18 new patients and review of the literature. Genes 14, 1179 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  25. Crippa, M. et al. SETD5 gene haploinsufficiency in three patients with suspected KBG syndrome. Front. Neurol. 11, 631 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  26. O’Donnell-Luria, A. H. et al. Heterozygous variants in KMT2E cause a spectrum of neurodevelopmental disorders and epilepsy. Am. J. Hum. Genet. 104, 1210–1222 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

  27. Quental, R. et al. Congenital heart defects associated with pathogenic variants in WAC gene: Expanding the phenotypic and genotypic spectrum of DeSanto–Shinawi syndrome. Am. J. Med. Genet. A 188, 1311–1316 (2022).

    Article  CAS  PubMed  Google Scholar 

  28. Connacher, R. et al. Autism NPCs from both idiopathic and CNV 16p11.2 deletion patients exhibit dysregulation of proliferation and mitogenic responses. Stem Cell Rep. 17, 1380–1394 (2022).

    Article  CAS  Google Scholar 

  29. Fombonne, E., Rogé, B., Claverie, J., Courty, S. & Frémolle, J. Microcephaly and macrocephaly in autism. J. Autism Dev. Disord. 29, 113–119 (1999).

    Article  CAS  PubMed  Google Scholar 

  30. Nordahl, C. W. et al. Brain enlargement is associated with regression in preschool-age boys with autism spectrum disorders. Proc. Natl Acad. Sci. USA 108, 20195–20200 (2011).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  31. Alsafh, R. et al. Multiplex consanguineous family highlights CLASP1 as a candidate gene for lissencephaly. Neurol. Genet. 10, e200172 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  32. Rots, D. et al. Comprehensive EHMT1 variants analysis broadens genotype-phenotype associations and molecular mechanisms in Kleefstra syndrome. Am. J. Hum. Genet. 111, 1605–1625 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  33. Sewani, S. et al. Neurodevelopmental and other phenotypes recurrently associated with heterozygous BAZ2B loss-of-function variants. Am. J. Med. Genet. A 194, e63445 (2024).

    Article  CAS  PubMed  Google Scholar 

  34. Paulsen, B. et al. Autism genes converge on asynchronous development of shared neuron classes. Nature 602, 268–273 (2022).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  35. Jourdon, A. et al. Modeling idiopathic autism in forebrain organoids reveals an imbalance of excitatory cortical neuron subtypes during early neurogenesis. Nat. Neurosci. 26, 1505–1515 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  36. Voineagu, I. et al. Transcriptomic analysis of autistic brain reveals convergent molecular pathology. Nature 474, 380–384 (2011).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  37. Gandal, M. J. et al. Broad transcriptomic dysregulation occurs across the cerebral cortex in ASD. Nature 611, 532–539 (2022).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  38. Wamsley, B. et al. Molecular cascades and cell type-specific signatures in ASD revealed by single-cell genomics. Science 384, eadh2602 (2024).

    Article  ADS  PubMed  Google Scholar 

  39. Li, C. et al. Single-cell brain organoid screening identifies developmental defects in autism. Nature 621, 373–380 (2023).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  40. Schoenherr, C. J. & Anderson, D. J. The neuron-restrictive silencer factor (NRSF): a coordinate repressor of multiple neuron-specific genes. Science 267, 1360–1363 (1995).

    Article  ADS  CAS  PubMed  Google Scholar 

  41. Chong, J. A. et al. REST: a mammalian silencer protein that restricts sodium channel gene expression to neurons. Cell 80, 949–957 (1995).

    Article  CAS  PubMed  Google Scholar 

  42. Katayama, Y. et al. CHD8 haploinsufficiency results in autistic-like phenotypes in mice. Nature 537, 675–679 (2016).

    Article  ADS  CAS  PubMed  Google Scholar 

  43. Abuhatzira, L., Makedonski, K., Kaufman, Y., Razin, A. & Shemer, R. MeCP2 deficiency in the brain decreases BDNF levels by REST/CoREST-mediated repression and increases TRKB production. Epigenetics 2, 214–222 (2007).

    Article  PubMed  Google Scholar 

  44. Tang, X. et al. KCC2 rescues functional deficits in human neurons derived from patients with Rett syndrome. Proc. Natl Acad. Sci. USA 113, 751–756 (2016).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  45. Sugathan, A. et al. CHD8 regulates neurodevelopmental pathways associated with autism spectrum disorder in neural progenitors. Proc. Natl Acad. Sci. USA 111, E4468–E4477 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  46. Cotney, J. et al. The autism-associated chromatin modifier CHD8 regulates other autism risk genes during human neurodevelopment. Nat. Commun. 6, 6404 (2015).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  47. Werling, D. M., Parikshak, N. N. & Geschwind, D. H. Gene expression in human brain implicates sexually dimorphic pathways in autism spectrum disorders. Nat. Commun. 7, 10717 (2016).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  48. Velmeshev, D. et al. Single-cell analysis of prenatal and postnatal human cortical development. Science 382, eadf0834 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  49. Tukiainen, T. et al. Landscape of X chromosome inactivation across human tissues. Nature 550, 244–248 (2017).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

  50. Shepherdson, J. L. et al. Variants in ZFX are associated with an X-linked neurodevelopmental disorder with recurrent facial gestalt. Am. J. Hum. Genet. 111, 487–508 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  51. Chahrour, M. H. et al. Whole-exome sequencing and homozygosity analysis implicate depolarization-regulated neuronal genes in autism. PLoS Genet. 8, e1002635 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  52. Zhou, X. et al. Integrating de novo and inherited variants in 42,607 autism cases identifies mutations in new moderate-risk genes. Nat. Genet. 54, 1305–1319 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  53. San Roman, A. K. et al. The human Y and inactive X chromosomes similarly modulate autosomal gene expression. Cell Genomics 4, 100462 (2024).

    Article  CAS  PubMed  Google Scholar 

  54. Czerminski, J. T., King, O. D. & Lawrence, J. B. Large-scale organoid study suggests effects of trisomy 21 on early fetal neurodevelopment are more subtle than variability between isogenic lines and experiments. Front. Neurosci. 16, 972201 (2022).

    Article  PubMed  Google Scholar 

  55. Zyla, J. et al. Evaluation of zero counts to better understand the discrepancies between bulk and single-cell RNA-seq platforms. Comput. Struct. Biotechnol. J. 21, 4663–4674 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  56. Chatterjee, S. et al. Enhancer variants synergistically drive dysfunction of a gene regulatory network in Hirschsprung disease. Cell 167, 355–368 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  57. Yim, K. M. et al. Cell-type-specific dysregulation of gene expression due to Chd8 haploinsufficiency during mouse cortical development. Cell Genomics 5, 100986 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  58. Durak, O. et al. Chd8 mediates cortical neurogenesis via transcriptional regulation of cell cycle and Wnt signaling. Nat. Neurosci. 19, 1477–1488 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  59. Willsey, H. R. et al. Parallel in vivo analysis of large-effect autism genes implicates cortical neurogenesis and estrogen in risk and resilience. Neuron 109, 788–804 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  60. DeCasien, A. R. et al. Sex effects on gene expression across the human cerebral cortex at cell type resolution. Science 392, eaea9063 (2026).

    Article  CAS  PubMed  Google Scholar 

  61. Ferrer, J. & Dimitrova, N. Transcription regulation by long non-coding RNAs: mechanisms and disease relevance. Nat. Rev. Mol. Cell Biol. https://doi.org/10.1038/s41580-023-00694-9 (2024).

  62. Wong, C. C. Y. et al. Genome-wide DNA methylation profiling identifies convergent molecular signatures associated with idiopathic and syndromic autism in post-mortem human brain tissue. Hum. Mol. Genet. 28, 2201–2211 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  63. Weiner, D. J. et al. Statistical and functional convergence of common and rare genetic influences on autism at chromosome 16p. Nat. Genet. 54, 1630–1639 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  64. Dias, C. et al. Cell-type-specific effects of autism-associated 15q duplication syndrome in the human brain. Am. J. Hum. Genet. 111, 1544–1558 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  65. Rhie, S. K. et al. ZFX acts as a transcriptional activator in multiple types of human tumors by binding downstream from transcription start sites at the majority of CpG island promoters. Genome Res. 28, 310–320 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  66. Wei, J. et al. Deep learning and CRISPR-Cas13d ortholog discovery for optimized RNA targeting. Cell Syst 14, 1087–1102 (2023).

    Article  CAS  PubMed  Google Scholar 

  67. Camacho, C. et al. BLAST+: architecture and applications. BMC Bioinformatics 10, 421 (2009).

    Article  PubMed  PubMed Central  Google Scholar 

  68. Ventura, A. et al. Cre-lox-regulated conditional RNA interference from transgenes. Proc. Natl Acad. Sci. USA 101, 10380–10385 (2004).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  69. Yusa, K., Zhou, L., Li, M. A., Bradley, A. & Craig, N. L. A hyperactive piggyBac transposase for mammalian applications. Proc. Natl Acad. Sci. USA 108, 1531–1536 (2011).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  70. Kumamoto, T. et al. Direct readout of neural stem cell transgenesis with an integration-coupled gene expression switch. Neuron 107, 617–630 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  71. Schnütgen, F. et al. A directional strategy for monitoring Cre-mediated recombination at the cellular level in the mouse. Nat. Biotechnol. 21, 562–565 (2003).

    Article  PubMed  Google Scholar 

  72. Kaczmarczyk, S. J. & Green, J. E. A single vector containing modified cre recombinase and LOX recombination sequences for inducible tissue-specific amplification of gene expression. Nucleic Acids Res. 29, E56–E56 (2001).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  73. Qian, K. et al. A simple and efficient system for regulating gene expression in human pluripotent stem cells and derivatives. Stem Cells 32, 1230–1238 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  74. Schlaeger, T. M. et al. A comparison of non-integrating reprogramming methods. Nat. Biotechnol. 33, 58–63 (2015).

    Article  CAS  PubMed  Google Scholar 

  75. de Leeuw, V. C. et al. An efficient neuron-astrocyte differentiation protocol from human embryonic stem cell-derived neural progenitors to assess chemical-induced developmental neurotoxicity. Reprod. Toxicol. 98, 107–116 (2020).

    Article  PubMed  Google Scholar 

  76. Giandomenico, S. L., Sutcliffe, M. & Lancaster, M. A. Generation and long-term culture of advanced cerebral organoids for studying later stages of neural development. Nat. Protoc. 16, 579–602 (2021).

    Article  CAS  PubMed  Google Scholar 

  77. Han, X. et al. Millisecond-timescale optical control of neural dynamics in the nonhuman primate brain. Neuron 62, 191–198 (2009).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  78. Lohman, B. K., Weber, J. N. & Bolnick, D. I. Evaluation of TagSeq, a reliable low-cost alternative for RNAseq. Mol. Ecol. Resour. 16, 1315–1321 (2016).

    Article  CAS  PubMed  Google Scholar 

  79. Rabe, B. A. Investigation of the Multiple Roles of Notch Signaling in Embryonic Retinal Development Using Novel High-Throughput Techniques (Harvard University Graduate School of Arts and Sciences, 2020).

  80. Ye, J. et al. Primer-BLAST: a tool to design target-specific primers for polymerase chain reaction. BMC Bioinformatics 13, 134 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  81. Kochinke, K. et al. Systematic phenomics analysis deconvolutes genes mutated in intellectual disability into biologically coherent modules. Am. J. Hum. Genet. 98, 149–164 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  82. Karczewski, K. J. et al. The mutational constraint spectrum quantified from variation in 141,456 humans. Nature 581, 434–443 (2020).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  83. Wang, T. et al. Integrated gene analyses of de novo variants from 46,612 trios with autism and developmental disorders. Proc. Natl Acad. Sci. USA 119, e2203491119 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  84. Doan, R. N. et al. Recessive gene disruptions in autism spectrum disorder. Nat. Genet. 51, 1092–1098 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  85. Rodin, R. E. et al. The landscape of somatic mutation in cerebral cortex of autistic and neurotypical individuals revealed by ultra-deep whole-genome sequencing. Nat. Neurosci. 24, 176–185 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  86. Wilfert, A. B. et al. Recent ultra-rare inherited variants implicate new autism candidate risk genes. Nat. Genet. 53, 1125–1134 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  87. Alcantara, D. et al. Mutations of AKT3 are associated with a wide spectrum of developmental disorders including extreme megalencephaly. Brain 140, 2610–2622 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

  88. Marcelis, A. & Van Reet, E. A boy with KIF11-associated disorder along with ADHD and ASD: collaboration between paediatrics and child. Case Rep. Psychiatry 2024, 5535830 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  89. Ruiz-Reig, N., Hakanen, J. & Tissir, F. Connecting neurodevelopment to neurodegeneration: a spotlight on the role of kinesin superfamily protein 2A (KIF2A). Neural Regen. Res. 19, 375–379 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  90. McGhee, C. A. et al. Genotype–phenotype correlations with autism spectrum disorder-related traits in Noonan syndrome and Noonan syndrome with multiple lentigines: a cross-sectional study. Mol. Autism 16, 51 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  91. Roopra, A. MAGIC: a tool for predicting transcription factors and cofactors driving gene sets using ENCODE data. PLoS Comput. Biol. 16, e1007800 (2020).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  92. Rockowitz, S. & Zheng, D. Significant expansion of the REST/NRSF cistrome in human versus mouse embryonic stem cells: potential implications for neural development. Nucleic Acids Res. 43, 5730–5743 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  93. Lachmann, A., Giorgi, F. M., Lopez, G. & Califano, A. ARACNe-AP: gene network reverse engineering through adaptive partitioning inference of mutual information. Bioinformatics 32, 2233–2235 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  94. Lambert, S. A. et al. The human transcription factors. Cell 172, 650–665 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  95. Boukas, L. et al. Coexpression patterns define epigenetic regulators associated with neurological dysfunction. Genome Res 29, 532–542 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  96. Alvarez, M. J. et al. Functional characterization of somatic mutations in cancer using network-based inference of protein activity. Nat. Genet. 48, 838–847 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  97. Page, L., Brin, S., Motwani, R. & Winograd, T. The PageRank citation ranking: bringing order to the web. The Web Conference https://www.semanticscholar.org/paper/The-PageRank-Citation-Ranking-%3A-Bringing-Order-to-Page-Brin/eb82d3035849cd23578096462ba419b53198a556 (1999).

  98. Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl Acad. Sci. USA 102, 15545–15550 (2005).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

  99. Liberzon, A. et al. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Syst 1, 417–425 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  100. Szklarczyk, D. et al. The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest. Nucleic Acids Res. 51, D638–D646 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  101. Kissel, L. T. et al. Sex-differential gene expression in developing human cortex and its intersection with autism risk pathways. Biol. Psychiatry Glob. Open Sci. 4, 100321 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  102. Oliva, M. et al. The impact of sex on gene expression across human tissues. Science 369, eaba3066 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  103. Naqvi, S. et al. Conservation, acquisition, and functional impact of sex-biased gene expression in mammals. Science 365, eaaw7317 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  104. Finak, G. et al. MAST: a flexible statistical framework for assessing transcriptional changes and characterizing heterogeneity in single-cell RNA sequencing data. Genome Biol. 16, 278 (2015).

    Article  PubMed  PubMed Central  Google Scholar 

  105. Benoit-Pilven, C. et al. Early establishment and life course stability of sex biases in the human brain transcriptome. Cell Genomics 5, 100890 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  106. Fass, S. B. et al. Relationship between sex biases in gene expression and sex biases in autism and Alzheimer’s disease. Biol. Sex Differ. 15, 47 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

Download references

Acknowledgements

We thank members of the Walsh laboratory for helpful discussions. We also thank. N. K. Hylton and E. M. Bushinsky of the Walsh laboratory for assistance with cerebral organoid culture; A. Y. Huang for technical feedback on statistical analyses; and A. K. San Roman and M. E. Tharp of the Page laboratory for insights with respect to ZFX.

Funding

This work was facilitated by the BCH Flow Cytometry Research Core (funded by 5U54DK110805), the BCH IDDRC Molecular Genetics Core Facility (funded by NIH P50 HD105351), the BCH Cellular Imaging Core (funded by NIH P50 HD105351) and L. Dong of the HMS Research Computing Core. This work was supported by Autism Speaks Postdoctoral Fellowship 13008 and NIGMS T32 GM007748 (R.E.A.); NIH T32 GM007753 and T32 GM144273 (M.T.); the Simons Foundation Autism Research Initiative, the Howard Hughes Medical Institute, The Brit Jepson d’Arbeloff Center on Women’s Health, Arthur W. and Carol Tobin Brill, Matthew Brill, Charles Ellis, Carla Knobloch, The Brett Barakett Foundation, the Howard P. Colhoun Family Foundation, and the Seedlings Foundation (M.T. and D.C.P.); the Harvard Program for Research in Science and Engineering and the Harvard College Research Program (T.S.); a Howard Hughes Medical Institute Fellowship from the Helen Hay Whitney Foundation and Grant K99MH136290 from NIMH/NIH (J.H.T.S.); K99 NS135123 from the NINDS/NIH (X.Q.); Simons Foundation fellowships AN-SURFiN-00008133, AN-SURFiN-00015215 and SFI-AN-SURFiN-00019272 (N.M.); the Howard Hughes Medical Institute and Grant K99EY034603 from NEI/NIH (R.N.D.); the Harvard-Amgen Scholars Program (G.E.); and grants from the Simons Foundation and Autism BrainNet (953759), NIMH (U01MH106883), Yang Tan Collective at Harvard University, Hock E. Tan and K. Lisa Yang Center for Autism Research at Harvard University, and the Allen Discovery Center program, a Paul G. Allen Frontiers Group advised program of the Paul G. Allen Family Foundation (C.A.W.). D.C.P. and C.A.W. are Investigators of the Howard Hughes Medical Institute.

Author information

Author notes

  1. Janet H. T. Song

    Present address: Department of Human Evolutionary Biology, Harvard University, Cambridge, MA, USA

  2. Xuyu Qian

    Present address: Department of Pediatrics, Division of Neurology, Children’s Hospital of Philadelphia, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA

  3. These authors contributed equally: Rebecca E. Andersen, Maya Talukdar

Authors and Affiliations

  1. Division of Genetics and Genomics and Manton Center for Orphan Diseases, Boston Children’s Hospital, Boston, MA, USA

    Rebecca E. Andersen, Maya Talukdar, Tyler Sakamoto, David Exposito-Alonso, Janet H. T. Song, Xuyu Qian, Seungil Lee, Nila Murugan, Sijing Zhao, Gwenyth Eichfeld, Julia Harms & Christopher A. Walsh

  2. Departments of Pediatrics and Neurology, Harvard Medical School, Boston, MA, USA

    Rebecca E. Andersen, Maya Talukdar, Tyler Sakamoto, David Exposito-Alonso, Janet H. T. Song, Xuyu Qian, Seungil Lee, Nila Murugan, Sijing Zhao, Gwenyth Eichfeld, Julia Harms & Christopher A. Walsh

  3. Howard Hughes Medical Institute, Boston Children’s Hospital, Boston, MA, USA

    Rebecca E. Andersen, Maya Talukdar, Tyler Sakamoto, David Exposito-Alonso, Janet H. T. Song, Xuyu Qian, Seungil Lee, Nila Murugan, Sijing Zhao, Gwenyth Eichfeld, Julia Harms & Christopher A. Walsh

  4. Broad Institute of MIT and Harvard, Cambridge, MA, USA

    Rebecca E. Andersen, Maya Talukdar & Christopher A. Walsh

  5. Allen Discovery Center for Human Brain Evolution, Boston, MA, USA

    Rebecca E. Andersen, David Exposito-Alonso, Janet H. T. Song, Xuyu Qian & Christopher A. Walsh

  6. Harvard-MIT MD/PhD Program and Harvard PhD Program in Biomedical Informatics, Harvard Medical School, Boston, MA, USA

    Maya Talukdar

  7. Whitehead Institute, Cambridge, MA, USA

    Maya Talukdar & David C. Page

  8. Harvard College, Cambridge, MA, USA

    Tyler Sakamoto & Seungil Lee

  9. University of Massachusetts Boston, Boston, MA, USA

    Nila Murugan

  10. Department of Genetics, Blavatnik Institute and Howard Hughes Medical Institute, Harvard Medical School, Boston, MA, USA

    Ryan N. Delgado

  11. Harvard BBS PhD Program, Harvard Medical School, Boston, MA, USA

    Sijing Zhao

  12. Colgate University, Hamilton, NY, USA

    Gwenyth Eichfeld

  13. University of California Berkeley, Berkeley, CA, USA

    Julia Harms

  14. Department of Biology, Massachusetts Institute of Technology, Cambridge, MA, USA

    David C. Page

  15. Howard Hughes Medical Institute, Whitehead Institute, Cambridge, MA, USA

    David C. Page

Authors

  1. Rebecca E. Andersen
  2. Maya Talukdar
  3. Tyler Sakamoto
  4. David Exposito-Alonso
  5. Janet H. T. Song
  6. Xuyu Qian
  7. Seungil Lee
  8. Nila Murugan
  9. Ryan N. Delgado
  10. Sijing Zhao
  11. Gwenyth Eichfeld
  12. Julia Harms
  13. David C. Page
  14. Christopher A. Walsh

Contributions

R.E.A. and M.T. conceived the experiments and led the analyses, with supervision from C.A.W. M.T. led all bioinformatic analyses. R.E.A. designed all plasmid constructs with support from T.S. and S.L. Molecular cloning and lentivirus production were performed by R.E.A., T.S. and S.L. The CasRx knock-in iPS cell line was generated by J.H.T.S. Cell culture was performed by R.E.A., T.S., S.L. and N.M., with contributions from S.Z., G.E. and J.H. Bulk RNA-seq libraries were prepared by R.E.A. and S.L. Analysis of bulk RNA-seq results was performed by M.T., with code contributions from R.N.D. NPC proliferation assays were performed by R.E.A., S.L., S.Z., G.E. and J.H. Cerebral organoids were cultured by R.E.A., T.S. and N.M., with advice and support from X.Q. Cerebral organoid scRNA-seq libraries were prepared by R.E.A. and N.M. Analysis of scRNA-seq libraries was performed by M.T. Cryosectioning and immunohistochemical labelling of cerebral organoids was performed by R.E.A. and N.M. Confocal imaging and histological analyses of cerebral organoids were performed by D.E.-.A. D.C.P. contributed to the interpretation of ZFX as a contributor to the FPE in ASD. R.E.A., M.T. and C.A.W. wrote the manuscript with contributions from all of the authors. All authors read and approved the final version.

Corresponding authors

Correspondence to Rebecca E. Andersen or Christopher A. Walsh.

Ethics declarations

Competing interests

C.A.W. consults for Maze Therapeutics, Mosaica Medicines and Bioskryb Genomics (cash, equity). R.E.A., M.T., D.C.P. and C.A.W. are planning a patent application with respect to ZFX and ASD. The other authors declare no competing interests.

Peer review

Peer review information

Nature thanks the anonymous reviewers for their contribution to the peer review of this work. Peer reviewer reports are available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended Data Fig. 1 CasRx-based knockdown of ASD genes in NPCs.

a, Relative expression of the targeted gene within each KD sample compared to NTC samples. Error bars: mean +/− 95% confidence interval. n = three biologically independent samples per condition. Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini-Hochberg correction for multiple hypothesis testing. b, Analysis of protein levels by auto-western blot from NTC and WAC-KD NPCs, from three biologically independent samples per condition.

Extended Data Fig. 2 Differential gene expression upon CasRx-based KD of ASD genes in NPCs.

Volcano plots depicting transcriptome-wide effects on gene expression following CasRx-based KD of ASD genes. Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini-Hochberg correction for multiple hypothesis testing. Statistically significantly downregulated DEGs are shown in blue, while statistically significantly upregulated DEGs are shown in red. The gene targeted for KD (listed above each plot) is highlighted in green. SFARI DEGs are indicated by a thick border.

Extended Data Fig. 3 Transcriptomic convergence on shared differentially expressed genes.

a, Circos plots depicting DEGs shared between different KD conditions. For direct comparison, the Circos plot for high confidence SFARI genes (shown in Fig. 1d) is also shown here. b, Heatmap showing the relative expression of non-SFARI genes that are DEGs in at least 5 KD conditions. c, Heatmap showing the relative expression of high confidence SFARI genes that are DEGs in at least 5 KD conditions. Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini-Hochberg correction for multiple hypothesis testing, and are indicated with *. Genes that were identified amongst the 26 strongest ASD risk genes in ref. 1 are bolded.

Extended Data Fig. 4 Shared DEGs are not dependent on CasRx-based KD.

a, Scatter plots depicting the number of total DEGs (left plot) or SFARI DEGs (right plot) on the y-axis and the expression of the KD target gene on the x-axis. Linear regression is shown in red. b, Flow cytometry analysis of fluorescence intensity in NPCs expressing the fluorophores mKO2, mScarlet, and iRFP670. CasRx was used to individually KD each of the fluorophores. Results depicted relative to the NTC control average. **** adj. p = 1.44e-25 for mKO2, 5.74e-24 for mScarlet, and 1.45e-24 for iRFP670 (ordinary two-way ANOVA with Bonferroni’s multiple comparisons test). n = three biologically independent samples per condition. c, Volcano plots depicting transcriptome-wide effects on gene expression following shRNA-based KD of ASD genes. Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini-Hochberg correction for multiple hypothesis testing. Statistically significantly downregulated DEGs are shown in blue, while statistically significantly upregulated DEGs are shown in red. The gene targeted for KD (listed above each plot) is highlighted in green. SFARI DEGs are indicated by a thick border. d, Bar plot depicting enrichment of the same downregulated DEGs following KD of the same target gene by CasRx vs shRNA. Bars indicate the odds ratio (OR) from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR (no upper bound is defined for a one-sided test). p-values were adjusted using Benjamini-Hochberg correction for multiple hypothesis testing.

Source data

Extended Data Fig. 5 Transcriptomic convergence occurs across cell types and experimental contexts.

Analysis of transcriptomic convergence across multiple cell types and experimental contexts. Statistically significant cross-enrichment of downregulated DEGs is shown in green (one-sided Fisher’s exact tests with Benjamini-Hochberg correction for multiple hypothesis testing). Data from iPSCs and ExNs were re-analysed from ref. 20.

Extended Data Fig. 6 Effects on proliferation and cell abundance following ASD gene KD.

a, Relative CytoTrack dye intensity (−log2) over time. Only KD conditions that exhibited non-significant or inconsistent effects are displayed. Plot depicts mean +/− standard error of the mean (SEM). Statistical significance was determined using ordinary two-way ANOVA with Dunnett’s multiple comparisons test. At least four biologically independent samples per condition were analysed for each timepoint. See also Fig. 1f, and see Supplementary Table 10 for full statistical results. b, Scatter plot depicting the relative abundance of KD cells on the y-axis and the expression of the KD target gene on the x-axis. Linear regression is shown in red. Genes that have been implicated in microcephaly and whose KD resulted in decreased proliferation (Fig. 1f) are highlighted.

Extended Data Fig. 7 KD of ASD genes in mosaic cerebral organoids reveals transcriptomic convergence.

a, KD efficiency of FLICK construct in iPSCs. Normalized expression of the KD target gene ENSG00000261840 shown relative to NTC samples, as mean +/− SEM. **** p = 3.87e-5, two-tailed t test. n = three biologically independent samples per condition. b, Immunohistochemistry of a FLICK organoid at day 15, representative of three independent analyses with similar results. c, UMAP of organoid scRNA-seq colored by timepoint. d, UMAP of organoid scRNA-seq data coloured by targeted gene. e, FLICK efficiency in RGs as quantified by scRNA-seq from one-month organoids. Statistically significant DEGs were identified using Seurat (non-parametric wilcoxon rank test) with Bonferroni correction for multiple hypothesis testing. We note that BAZ2B quantification may be impacted by low cell number, and WDFY3 capture in scRNA-seq may be impacted by the length (>14 kb) of the transcript. f, Cell type composition per KD target in two-month organoids. Dotted lines indicate the NTC proportions of major cell types. Samples for which the ExN/RG proportion was significantly different than NTC are indicated with * (adj. p = 0.021 for BAZ2B, 0.018 for CLASP1, and 0.012 for EHMT1), ordinary one-way ANOVA with Dunnett’s multiple comparisons test, considering all long-term samples from one month (see Fig. 2f) and two months (shown here). g, Cross-enrichment of DEGs from IPCs (one-sided Fisher’s exact tests with Benjamini-Hochberg correction for multiple hypothesis testing). h, Bar plot depicting cross-enrichment of DEGs from RGs from one-month organoids versus DEGs from previous NPC cultures upon KD of the same target gene. Only perturbations with at least 20 DEGs were tested for cross-enrichment. Bars indicate the odds ratio (OR) from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR (no upper bound is defined for a one-sided test). p-values were adjusted using Benjamini-Hochberg correction for multiple hypothesis testing. i, UMAP of scRNA-seq data from acute KD organoids, coloured by KD target gene as in panel d. j, Expression of marker genes used for identifying cell types. k, Cell type composition across different perturbations in acute KD organoids. l, Bar plot depicting cross-enrichment of DEGs from excitatory neurons (ExNs) from acute KD mosaic organoids versus DEGs from previous NPC cultures upon KD of the same target gene. Bars indicate the odds ratio (OR) from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR (no upper bound is defined for a one-sided test). p-values were adjusted using Benjamini-Hochberg correction for multiple hypothesis testing.

Source data

Extended Data Fig. 8 Key transcription factors, including REST, target recurrently dysregulated genes.

a, MAGIC analysis (Kolmogorov-Smirnov test with Benjamini-Hochberg correction for multiple hypothesis testing) identifying transcriptional regulators whose targets were enriched within the DEGs from different cell types following KD of ASD genes. For each population, MAGIC regulators were scored by the number of ASD gene KDs that implicated the regulator. The 25 top scoring regulators are included here, along with any additional regulators with a tied score. b, Quantitative Euler diagrams depicting the overlapping MAGIC transcriptional regulators identified in different conditions. c, Bar plot depicting enrichment of SFARI, HC-SFARI, or ID-only genes within REST targets from human and mouse embryonic stem cells. REST target data from ref. 92. Bars indicate the odds ratio (OR) from a one-sided Fisher’s exact test; error bars indicate the lower bound of the 95% confidence interval on the OR (no upper bound is defined for a one-sided test). p-values were adjusted using Benjamini-Hochberg correction for multiple hypothesis testing.

Source data

Extended Data Fig. 9 Top drivers of convergence identified through reconstruction of an NPC gene regulatory network.

a, Volcano plots depicting differential expression upon KD of the regulator shown above the plot. Genes predicted by GRN analysis to be positively regulated by the indicated regulator are highlighted, with statistically significantly decreased DEGs shown in blue, increased DEGs shown in red, and targets with no significant change shown in yellow. Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini-Hochberg correction for multiple hypothesis testing. Statistically significant enrichment of 2nd degree regulon genes within the downregulated DEGs is indicated by the adjusted p-value (one-sided Fisher’s exact tests with Benjamini-Hochberg correction for multiple hypothesis testing). b, Heatmap showing the relative expression of the top GRN drivers of convergence across the different ASD gene KD samples. Statistically significant DEGs were identified using DESeq2 (Wald test) with Benjamini-Hochberg correction for multiple hypothesis testing, and are indicated with *. Genes that were identified amongst the 26 strongest ASD risk genes in ref. 1 are bolded.

Extended Data Fig. 10 ZFX exhibits female-biased expression in the brain and is recurrently downregulated upon ASD gene KD.

a, Heatmaps showing the relative expression of candidate contributors to the FPE in ASD. Left: expression in NPCs across KD conditions relative to NTCs, with statistically significant DEGs (indicated with *) identified using DESeq2 (Wald test) with Benjamini-Hochberg correction for multiple hypothesis testing. Right: expression in ExNs from males with ASD relative to male idiopathic controls, re-analysed from ref. 64, with statistically significant DEGs (indicated with *) identified using MAST (likelihood ratio test) with Bonferroni correction for multiple hypothesis testing. Lack of detection is indicated with X. b, Bar plot depicting the significantly female-biased expression of ZFX across several tissues of the human central nervous system (CNS), including prenatal and adult cortex. Distinct regions indicated by different colours. The depicted expression levels and statistical results were taken directly from the following references: [1] ref. 105, [2] ref. 101, and [3] ref. 106. For details regarding the statistical analyses, see the corresponding reference. c, Heatmap depicting expression of ZFX across cortical areas and cell types, relative to expression in males. Cell types and areas are abbreviated as in the source60.

Supplementary information

Source data

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

[Check for updates. Verify currency and authenticity via CrossMark]

Cite this article

Andersen, R.E., Talukdar, M., Sakamoto, T. et al. Transcriptomic convergence and the female protective effect in autism. Nature (2026). https://doi.org/10.1038/s41586-026-11236-6

Download citation

  • Received: 03 September 2025

  • Accepted: 01 October 2026

  • Published: 04 November 2026

  • Version of record: 04 November 2026

  • DOI: https://doi.org/10.1038/s41586-026-11236-6