You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). Rachel Fieldhouse is a reporter for Nature in Sydney, Australia. There have been cases in which people have had their lungs removed and been connected to an external device to maintain oxygen levels. But, the devices used in these cases don't count as artificial lungs because they do not maintain blood flow across the heart, meaning it cannot function normally, says Ankit Bharat, a thoracic surgeon at Northwestern University Feinberg School of Medicine in Chicago, Illinois, who helped to develop the artificial system. Bharat says his team's design is unique because it maintains a balanced and continuous flow of blood to the heart, reducing the risk of blood clots that could trigger a heart attack. It is difficult to maintain normal heart function in the absence of lungs, she says. The team's artificial-lung system could be used in other critically unwell people while they become healthy enough to receive lung transplants, she adds. The infection caused parts of his lungs to fill with pus, and he went into septic shock, at which point his heart and kidneys began to fail. At that point, the man received a double lung transplant and has showed no signs of organ rejection or impaired lung function years later. World's first whole-eye transplant: the innovations that made it possible First pig kidney transplant in a person: what it means for the future The surprisingly big health benefits of just a little exercise Funding cuts could put research into emerging threats to lung health at risk Shanghai, Beijing, Nanjing, Pune or New Delhi – hybrid working model World's first whole-eye transplant: the innovations that made it possible First pig kidney transplant in a person: what it means for the future An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Svalbard's polar bears are showing remarkable resilience to climate change These polar bears appear to be maintaining their physical health despite the loss of sea ice—their preferred hunting grounds A sedated polar bear lies on the ice with its cub huddled against it. Polar bears are the poster children of climate change—and for good reason. These giant bears hunt, mate and spend their days hanging out on Arctic sea ice, which is rapidly disappearing as the climate warms. But some polar bears, it seems, are far more resilient than we realized: new research suggests that in one region, the bears are adapting to the declining sea ice. Researchers took more than 1,000 body measurements from 770 polar bears over 24 years around Svalbard, a Norwegian archipelago in the Barents Sea—an area that has seen an especially acute decline in sea ice over the decades. By 2019, the region's annual sea ice season shriveled to more than two months shorter than what it was 24 years prior. If you're enjoying this article, consider supporting our award-winning journalism by subscribing. By purchasing a subscription you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today. Researchers record the mouth measurements of an anesthetized polar bear. Recapture of bears marked in previous years is important for obtaining accurate data on survival and reproduction. To Aars's surprise, the polar bears were in “fairly good” condition, even recovering from a drop in BCI from the 1990s to 2000—the opposite of what he and his team had anticipated. That's not to say the bears were totally unaffected by the loss of sea ice; some had relocated den areas or shifted their habitats north to follow the receding ice, Aars says. And of course, it's hard to know whether the polar bears would be in even better condition if it weren't for climate change. Aars and his colleagues aren't sure, but they think it may have something to do with changes in the polar bears' diet. It could be that, with less ice, seals “aggregate” on the ice that remains and are thus easier to hunt or that the bears are relying more on walrus carcasses or reindeer for food. Three polar bear cubs gather around their tranquilized mother. She had a litter of three cubs — an unusual brood size — and the smallest cub only weighed five kilograms. Lori Quakenbush, a researcher at the Alaska Department of Fish and Game's Arctic Marine Mammal Program, who was not involved in the study, says that polar bears are “part of a ‘complex relationship' that includes habitat, multiple prey species, their ability as predators and their ability to store energy.” The new findings are similar to separate observations of a group of bears off Alaska and Russia, the Chukchi Sea subpopulation, which has similarly held strong despite “substantial” sea ice decline, she says. “With continued monitoring of these and other subpopulations, we will likely learn more about the capabilities of polar bears as a species,” Quakenbush says. It's unclear what Svalbard will look like in five, 10 or 20 years, Aars says. “But the good news is that we are not there yet,” he says. A polar bear pictured standing on sea ice in the golden hour. Jackie Flynn Mogensen is a breaking news reporter at Scientific American. Before joining SciAm, she was a science reporter at Mother Jones, where she received a National Academies Eric and Wendy Schmidt Award for Excellence in Science Communications in 2024. If you enjoyed this article, I'd like to ask for your support. Scientific American has served as an advocate for science and industry for 180 years, and right now may be the most critical moment in that two-century history. If you subscribe to Scientific American, you help ensure that our coverage is centered on meaningful research and discovery; that we have the resources to report on the decisions that threaten labs across the U.S.; and that we support both budding and working scientists at a time when the value of science itself too often goes unrecognized. In return, you get essential news, captivating podcasts, brilliant infographics, can't-miss newsletters, must-watch videos, challenging games, and the science world's best writing and reporting. There has never been a more important time for us to stand up and show why science matters.

You are using a browser version with limited support for CSS. To make their promising compounds, they must trawl through millions of known chemical reactions, with hundreds of thousands more added annually, and then test whether it is possible to synthesize them. Now, researchers have created an artificial-intelligence system that vastly simplifies and accelerates the process of chemical synthesis. “The synthesis of small molecules is the slow step in drug discovery and a number of other important areas,” says study co-author Timothy Newhouse, a chemist at Yale University in New Haven, Connecticut. This translates chemical 3D structures into letters, numbers and punctuation, which are better suited to a system that recognizes language. By contrast, LLMs such as ChemCrow are trained for chemistry tasks using natural-language data2. AI chatbot shows surprising talent for predicting chemical properties and reactions AI chatbot shows surprising talent for predicting chemical properties and reactions The SMILES approach makes it easier to process chemical information such as starting materials and solvents. “Our goal was to build a general model that could read chemistry the way chemists write it [by] listening to the language of experimental procedures and quickly turning that collective voice into a practical suggestion,” says Newhouse. The researchers used an AI system they had developed previously3 to cluster a database of around one million reactions extracted from patents into 2,285 subsets. Using the subsets, the team trained Meta's partially open-source Llama LLM to create 2,498 separate expert models, each specialized in one combination of chemical transformation starting from one type of molecule. This approach can run on computers locally because it uses fewer parameters than do the major LLMs. “Each specialized model is more accurate within its domain,” Seifrid says. Read the paper: Connecting chemical and protein sequence space to predict biocatalytic reactions Read the paper: Bayesian reaction optimization as a tool for chemical synthesis AI chatbot shows surprising talent for predicting chemical properties and reactions Read the paper: Connecting chemical and protein sequence space to predict biocatalytic reactions Read the paper: Robot-assisted mapping of chemical reaction hyperspaces and networks Read the paper: Bayesian reaction optimization as a tool for chemical synthesis AI chatbot shows surprising talent for predicting chemical properties and reactions An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

People who stop taking prescription weight loss medications often regain weight and lose many of the health improvements they achieved, according to a study published by The BMJ. The analysis found that benefits tied to heart and metabolic health, including improvements in cholesterol levels and blood pressure, tend to fade after treatment ends. On average, individuals regained about 0.4 kilograms per month after stopping medication. Based on these trends, body weight and key risk markers for diabetes and heart disease were projected to return to pre-treatment levels in under two years. "This evidence suggests that despite their success in achieving initial weight loss, these drugs alone may not be sufficient for long term weight control," the researchers write. However, about half of people with obesity stop using GLP-1 drugs within a year. To examine these outcomes, researchers from the University of Oxford reviewed clinical trials and observational studies involving adults who used medications approved for weight loss. These results were compared with outcomes from non-drug approaches such as behavioral weight management programs or placebo treatments. In total, 37 studies published through February 2025 were included, covering 9,341 participants. Across the studies, participants who stopped taking weight loss drugs regained an average of 0.4 kilograms per month. Based on this rate, researchers estimated that body weight would return to pre-treatment levels within about 1.7 years. This pattern held true regardless of how much weight was lost during treatment. Only eight of the included studies examined newer GLP-1 medications, and none followed participants for more than 12 months after stopping treatment. In addition, relatively few studies were rated as having a low risk of bias. Even so, the researchers emphasize that they used three different analytical approaches, all of which produced similar results. Scientists Found a Surprising Way to Track Falling Space Junk A Strange Ice Process May Be Making Europa's Ocean Habitable Game-Changing Blood Pressure Drug Works for Patients Not Controlled by Standard Treatments New Study Shows Air Fryers Produce Far Fewer Toxic Particles Than Frying Stay informed with ScienceDaily's free email newsletter, updated daily and weekly. Keep up to date with the latest news from ScienceDaily via social networks: Tell us what you think of ScienceDaily -- we welcome both positive and negative comments.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript. Two decades of genetic studies in autism spectrum disorder (ASD) have identified more than 100 genes harbouring rare risk mutations1,2,3,4,5,6,7,8,9,10,11,12,13. Despite this substantial heterogeneity, transcriptomic and epigenetic analyses have identified convergent patterns of dysregulation across the ASD postmortem brain14,15,16,17. To identify shared and distinct mechanisms of ASD-linked mutations, we assembled a large patient collection of human induced pluripotent stem (hiPS) cells, consisting of 70 hiPS cell lines after stringent quality control representing 8 ASD-associated mutations, idiopathic ASD, and 20 lines from non-affected control individuals. Here we used these hiPS cell lines to generate human cortical organoids, profiling by RNA sequencing at four distinct time points up to 100 days after in vitro differentiation. Early time points harboured the largest mutation-specific changes, but different mutations converged on shared transcriptional changes as development progressed. We identified a shared RNA and protein interaction network, which was enriched in ASD risk genes and predicted to drive the observed downstream changes in gene expression. CRISPR–Cas9 screening of these candidate transcriptional regulators in induced human neural progenitors validated their downstream convergent molecular effects. These data illustrate how risk associated with genetically defined forms of ASD can propagate by means of transcriptional regulation to affect convergently dysregulated pathways, providing new insight into the convergent impact of ASD genetic risk on human neurodevelopment. Autism spectrum disorder (ASD) is a common neurodevelopmental disorder (NDD), with a childhood prevalence of close to 2%18. The last decade of genetic studies has yielded hundreds of risk genes, consistent with extraordinary aetiological heterogeneity1,2,3,4. More than 100 high-confidence mutations have been associated with ASD in genetic studies2,5,6,7,8,9,10,11,12,13. These rare, usually de novo mutations with large effect sizes are expected to account for 10–15% of ASD cases19,20, whereas common genetic variation is predicted to explain at least 50% of genetic risk21,22,23. Overall, ASD shows a complex genetic architecture, with a substantial component derived from a collection of distinct, rare disorders with overlapping clinical features. Despite its genetic heterogeneity, postmortem transcriptome analysis has revealed consistent changes in most individuals with idiopathic ASD, as well as individuals with a specific syndromic form of ASD, (dup)15q11–13 (refs. However, the mechanisms by which distinct mutations can lead to convergent molecular pathology, and whether convergence occurs across rare forms of ASD remains unknown. Understanding how these processes develop is complicated by the fact that the expression of most ASD risk genes peaks during fetal development, yet gene expression studies in human brain are conducted after this critical developmental window has ended16,24,25,26,27. Several lines of evidence, including genetic28, genomic29,30, neuroimaging31 and neuropathology32, indicate that early neurodevelopment has an essential role in the development of ASD. The advent of stem cell-based in vitro systems enables high-fidelity modelling of human brain development in NDDs33,34,35,36,37,38,39,40,41. Most studies have investigated relatively small numbers of lines from individuals with idiopathic ASD42, or focused on individual mutations, which have demonstrated the utility of human induced pluripotent stem (hiPS) cell-based systems to study the impact of ASD genetic risk on neurodevelopment43,44,45,46. In addition, recent work has highlighted the power of studying many genes in parallel, identifying evidence for molecular convergence using CRISPR-based perturbations of ASD risk genes on control genetic backgrounds47,48,49. However, studies of lines derived from affected individuals are a major gap in the field. Here we profile a large cohort of hiPS cell lines, starting from 96 hiPS cell lines ascertained from individuals with 8 different mutations associated with ASD and 11 individuals with idiopathic ASD and 30 lines derived from 25 matched control participants. From each of these lines, we derived neural organoids using a guided differentiation approach to create human cortical organoids (hCOs)50,51. Using orthogonal analytic approaches, we find evidence for convergence during early neuronal differentiation in those with genetically defined forms, but no significant shared signal across the idiopathic cases. We identify and characterize a downregulated, chromatin and transcriptional network that contains several ASD risk genes, including members of the SWI–SNF complex. This network is predicted to drive the observed changes in downstream gene expression associated with these ASD susceptibility mutations. We use CRISPRi to validate the effects of many putative network drivers on downstream gene expression, which includes downregulation of important neurodevelopmental pathways including many ASD susceptibility genes. We reprogrammed somatic cells (Methods) to generate hiPS cells from a cohort of individuals with eight different mutations associated with ASD: (1) 22q11.2 deletion, a 1.5–3-megabase (Mb) deletion that leads to a constellation of variably present symptoms including heart defects, craniofacial features, intellectual disability, ASD (roughly 20%) and psychosis (roughly 25%)52,53 (n = 18)44; (2) 22q13.3 deletion, known as Phelan–McDermid syndrome, a deletion spanning 130 kilobases (kb) to 9 Mb that includes SHANK3, among other genes and presents with developmental delay, hypotonia and impaired social interactions54 (n = 4); (3) 15q13.3 deletion, a 1.5-Mb deletion that presents variably with intellectual disability and epilepsy55 (n = 3); (4) 16p11.2 deletion, a roughly 600-kb deletion presenting variably with intellectual disability, motor impairments, communication deficits and ASD56 (n = 4); (5) 16p11.2 duplication, which can share intellectual disability and ASD phenotypes with the reciprocal deletion57 (n = 4); (6) Timothy syndrome, which is characterized by variants within the CACNA1C gene and presents with syndactyly, prolonged QT interval, ASD and intellectual disability58,59 (n = 2)45,60,61; (7) PCDH19-related disorder, which is associated with epilepsy and can also include intellectual disability and ASD62,63 (n = 2) and (8) the SHANK3 R522W mutation (n = 1), a point mutation associated with neurodevelopmental risk64; mutations in SHANK3 share behavioural phenotypes with the larger 22q13.3 deletion61,65,66. We also profiled individuals with idiopathic ASD with no known pathogenic variants (n = 11) and unaffected individuals (n = 25) (Fig. 1a and Supplementary Table 1) for a total of 74 individuals. For several individuals we used several hiPS cell lines to assess reproducibility, amounting to a total of 96 hiPS cell lines (Fig. 1a,b and Extended Data Fig. a, Schematic workflow going from hiPS cells to cortical organoids to sequencing data. The number of hiPS cell lines and individuals (in parentheses) for each form of ASD and controls is indicated. b, Schematic representation of hCO differentiations derived from each hiPS cell line for two forms of ASD (16p11.2 deletion and 22q13.3 deletion). The other forms of ASD can be found in Extended Data Fig. 1. c, Spearman's correlation of gene expression between samples from the same time point and form of ASD that were derived either from different lines (red) or from the same line (blue). Boxplots show: centre, median; lower hinge, 25% quantile; upper hinge, 75% quantile; lower whisker, smallest observation greater than or equal to lower hinge −1.5× interquartile range; upper whisker, largest observation less than or equal to upper hinge +1.5× interquartile range. d, Genes within the CNVs are downregulated in deletions and upregulated in duplications, as exemplified by 16p11.2 deletion (del) and duplication (dup) and 15q13.3 deletion. *dreamlet P values of less than 0.005. Top illustration adapted from ref. 51, Springer Nature America; illustrations in the IP–MS and CRISPRi panels created in BioRender. We differentiated hiPS cell lines into hCOs for 100 days51, which yields hCOs that contain developing radial glia and ultimately give rise to functional glutamatergic neurons in a process that closely resembles cerebral cortical development36,51. We performed whole-genome sequencing and removed any lines with large copy number variations (CNVs), viral integration or in which the mutation of interest was absent or further mutations were present, which yielded 70 verified lines from 55 individuals passing quality control for downstream functional genomic analyses (Methods, Extended Data Figs. We next analysed the transcriptomes of pooled hCOs at days 25, 50, 75 and 100 of hCO differentiation (Fig. 1a,b and Extended Data Fig. 1), which provides a quantitative, unbiased metric67,68. Day 25 corresponds to an early period in cortical development36,50, consisting primarily of cycling progenitors and radial glia50,69. As differentiation progresses, the cell composition evolves to consist of intermediate progenitors, deep layer postmitotic neurons and astroglial lineage cells44,50. After stringent RNA sequencing (RNA-seq) quality control and outlier analysis to remove low-quality samples (Fig. 1–3 and Methods), 464 samples from 55 individuals (70 lines) were retained (Supplementary Table 1). We observed high reproducibility both between (overall mean Spearman correlation 0.96, range 0.88–0.99) and within (overall mean Spearman correlation 0.97, range 0.92–0.98) individuals, similar to published metrics36,44,51 (Fig. The largest driver of variation was stage of differentiation (Extended Data Fig. Genes within the CNV boundaries were downregulated in the deletion carriers and upregulated in the duplication carriers as expected (Fig. 1d and Extended Data Fig. Although several of the genes affected by point mutations showed a trend towards downregulation (PCDH19, log fold change (logFC) of −0.77, false discovery rate (FDR) of 0.45; CACNA1C: logFC = −0.22, FDR = 0.78), these changes were not significant, as expected70,71,72 (Extended Data Figs. We observed several reliable, distinct, gene-expression clusters across forms (Methods and Supplementary Table 3), with participants harbouring the 16p11.2 deletion clustering with those with PCDH19-related disorder, 16p11.2 duplication clustering with Timothy syndrome and 22q13.3 and SHANK3 mutation clustering together (Fig. We performed bootstrapping to show that these clusters were robust (Extended Data Fig. Despite the strong effect of differentiation day on transcriptional profiles, most samples from individual genetic mutations clustered together across time points, showing a consistent effect of mutation over developmental progression (Fig. a, Heatmap of hierarchical clustering based on correlation of logFC between different conditions (ASD form and differentiation day). Annotation bars (shades of blue) represent differentiation day and colourful boxes represent ASD form. b, The number of DEGs in each ASD form during hCO differentiation. Colours and sizes of the circles represent the number of DEGs. c, Overlap in DEGs between forms of ASD at day 25. Number and percentage of total genes are indicated. Statistics were derived from Fisher's exact test (two-sided). ***FDR < 0.005. d, Gene ontology terms significantly enriched in unique and intersecting DEGs. Statistics were derived from clusterProfiler enrichGO. e, Overlaps between DEGs at day 25 and ASD2,8, NDD and intellectual disability risk genes91. Only positive significant overlaps (OR > 1 and FDR < 0.05) are shown. Statistics were derived from Fisher's exact test (two-sided). f, Network representation of the correlation (Spearman's rho) between logFC of different ASD forms at days 25 and 100. Colour corresponds to rho value and line thickness corresponds to the rho absolute value. g, All correlations (Spearman's rho) between logFC of the different forms. Boxplots show: centre, median; lower hinge, 25% quantile; upper hinge, 75% quantile; lower whisker, smallest observation greater than or equal to lower hinge −1.5× interquartile range; upper whisker, largest observation less than or equal to upper hinge +1.5× interquartile range. Statistics were derived from n = 36 correlations per time point. ***Tukey HSD for all contrasts less than 0.001. h, The number of differentially expressed meta-significant genes comparing all affected forms with control participants. Colour and size of the circle represent the number of DEGs. BMP, bone morphogenetic protein; ID, intellectual disability. Next, we found that the largest changes in gene expression for individual forms were observed at early stages of hCO differentiation, with the largest number of differentially expressed genes (DEGs) found at day 25 in four forms (16p11.2 deletion, 979 genes; 16p11.2 duplication, 1,556 genes; PCDH19, 514 genes and Timothy syndrome, 832 genes), or at day 50 in 2 of the remaining forms (15q13.3 deletion, 616 genes and 22q13.3 deletion, 373 genes; Fig. Significant similarity was seen between genes differentially expressed at day 25 in samples harbouring deletion and duplication of 16p11.2 (375 out of 2,160 genes, P < 10−16, hypergeometric test) (Fig. The intersecting genes were significantly enriched for chromatin remodelling related terms (Fig. 2d and Supplementary Table 4), which have previously been associated with ASD risk genes8,24,73. Overall, genes downregulated at 25 days in these four forms of ASD were enriched for known high-confidence ASD and NDD risk genes (16p11.2del: SFARI genes odds ratio (OR) 3.0, FDR = 3.8 × 10−9; 16p11.2dup: SFARI genes OR = 3.3, FDR = 1.4 × 10−13; PCDH19: SFARI genes OR = 3.1, FDR = 1.5 × 10−5; Timothy syndrome: SFARI genes OR = 2.0, FDR = 4.3 × 10−2; 16p11.2del: brain disorder genes OR = 2.4, FDR = 6.6 × 10−3; 16p11.2dup: brain disorder genes OR = 2.2, FDR = 1.2 × 10−2; PCDH19: brain disorder genes OR = 3.2, FDR = 2.2 × 10−3; Fisher's exact test, Fig. Gene set enrichment analysis (GSEA) found downregulation of terms related to neural precursor proliferation (for example, ‘Beta catenin binding', ‘Neural precursor cell proliferation' and ‘Canonical WNT signalling pathway') in all 4 of these forms (Extended Data Fig. We detected only two significantly differentially expressed genes in the individuals with idiopathic ASD (PRRC2C at day 50 and the long non-coding RNA RP11-114H21.2 at day 100; FDR < 0.05). Although the largest mutation-specific changes were observed early, the shared differential expression signature of mutational effects increased with maturation. Examination of the correlation of log2FCs over time revealed that the correlations between hCO harbouring different genetic events were significantly stronger at day 100 (analysis of variance (ANOVA) F3,140 = 11, P = 1.6 × 10−6; Tukey's honestly significant difference (HSD): day 100 to day 25 P = 2.7 × 10−5, day 100 to day 50 P = 8.0 × 10−6, day 100 to day 75 P = 7.0 × 10−4; Fig. 2f,g) suggesting that the changes in gene expression converge as cortical differentiation progressed in vitro, with distinct mutational forms showing more consistent differences from control participants and less mutational specificity. This increase in convergence of gene-expression changes could not be explained by alterations in variability in control participants or the distinct genetic forms across development (Extended Data Fig. We observed that six different forms of ASD profiled were enriched for upregulated genes related to neuronal cell fate and protein translation-related terms by day 75 (ref. We also observed broad downregulation of synapse and ion channel related terms such as ‘Synaptic membrane' and ‘Gated channel activity', across many time points and all the main mutational forms and in meta-analyses, particularly at day 75 and 100 (Fig. The early disruption in pathways related to neurogenesis, neural progenitors and WNT signalling described above (for example, Fig. 2), coupled to the later emergence (day 75 and 100) of enrichment in synaptic pathways, provides a potential link between early mutational effects on neurogenesis that converge on signalling and synaptic biology during cortical neurogenesis across different genetic forms of ASD. a, Select gene ontology terms enriched in upregulated (red) or downregulated (blue) genes in each ASD form. Colour shows normalized enrichment score (NES); point size shows −log10(FDR). Statistics were derived from GSEA. b, Cell proportion changes of ASD forms. Top annotation, differentiation day (shades of blue); lower annotation, ASD form (colourful). Statistics were derived from linear mixed-effects models and contrasts comparing each form with control participants with two-sided Wald tests. *FDR < 0.05. c, Adjusted R of WGCNA modules associated with either a form of ASD (Form) or with the presence of ASD (Dx) at day 25 and day 100 (left). Changes (linear model beta) in module eigengene in each ASD form (right). +FDR < 0.1, *FDR < 0.05, **FDR < 0.01, ***FDR < 0.005. d, Cell-type enrichment of modules using EWCE (red), genome-wide association study (GWAS) enrichment using a stratified linkage disequilibrium score (blue) and ASD-related gene set enrichment using a two-sided Fisher's test (green). Only FDR < 0.05 are shown. e, Select gene ontology terms enriched in modules. Black lines represent FDR = 0.05. Statistics were derived from clusterProfiler enrichGO. f, The number of modules associated with either ASD forms (specific mutations) or with Dx. Significance was tested using two-sided Fisher's exact test. g, Magnitude of module association with ASD forms or ASD broadly (Dx). Significance was tested using Welch's two sample t-test (sample size: n = 22 modules per time point). Boxplots show: centre, median; lower hinge, 25% quantile; upper hinge, 75% quantile; lower whisker, smallest observation greater than or equal to lower hinge −1.5× interquartile range; upper whisker, largest observation less than or equal to upper hinge +1.5× interquartile range. ADHD, attention-deficit/hyperactivity disorder; BD, bipolar disorder; CycPro, cycling progenitor; Dx, diagnosis; ExNeu-1, excitatory neurons 1 (early); ExNeu-2, excitatory neurons 2 (late); InNeu, interneurons; IPC, intermediate progenitor cells; MDD, major depressive disorder; ME, module eigengene; ORG, outer radial glia; RG-1, radial glia 1 (early); RG-2, radial glia 2 (late); SCZ, schizophrenia. To further explore the relationships of gene-expression changes to specific cell types, we deconvolved bulk RNA-seq into cell-type transcriptomes using single-cell RNA-seq (scRNA-seq) from hCOs44 (Extended Data Fig. Predicted changes in cell-type proportions were subtle and largely seen at one or two time points per form (Fig. 3b and Extended Data Fig. The least mature progenitors (CycPros) and more mature neuronal cell types (ExNeu-2) showed predicted reductions in 16p11.2del and 22q13.3del, suggesting changes in the timing of neuronal maturation (ExNeu-2 decreased at day 25 and CycPro decreased at day 75), consistent with recent observations in a 16p11.2 organoid model75. Intermediate progenitors showed changes in 15q13.3 deletion, 16p11.2 duplication and Timothy syndrome, suggesting convergent effects on this cell type during development (Fig. 3b and Extended Data Fig. Correlation of deconvoluted cell-type compositional shifts increased at day 100 compared with day 25, demonstrating convergence in cell-type changes, similar to what we observed with differential expression and changes in biological process (Extended Data Fig. By day 100, although the correlation across forms increased, only InNeu showed any significant shifts from control participants in any given form (22q13.3del, 15q13.3del and PCDH19). This indicates that observed transcriptional convergence is probably related to both altered timing of differentiation of cell types and altered gene expression signatures within cell types, the latter appearing more prominent here. We used an orthogonal method, independent components analysis to determine whether pathway changes were robustly identified; we observed high concordance between independent components analysis and expression-based clustering (Rand index 0.58, Extended Data Fig. Upregulated independent components across 16p11.2 deletion and duplication were downregulated for processes involved in neuronal precursor renewal and differentiation, such as ‘Beta catenin-TCF complex assembly' and ‘Neuron fate specification'76 (Extended Data Fig. The broad role of these pathways in ASD has previously been suggested42,43,77, highlighting their potential role as a point of convergence. Independent component 31 (IC31), enriched for synaptic terms, was downregulated in 16p11.2 duplication (IC31 beta = −0.13, FDR = 1.11 × 10−3) and Timothy syndrome (IC31 beta = −0.14, FDR = 2.9 × 10−3; Extended Data Fig. 9 and Supplementary Table 6), similar to what was observed in the differential expression analysis. As transcripts do not act independently, but as part of highly regulated transcriptional networks78, we constructed co-expression networks using weighted gene co-expression network analysis (WGCNA)78,79,80. We tested whether modules derived from the day 25 samples changed in their association with individual forms and with all ASD-associated forms (diagnosis) compared with control participants over time (Fig. Notably, we observed that the 22 identified co-expression modules from day 25 data were preserved at all other time points, suggesting they represent robust, continuing biological processes over this 100-day period. 10) further demonstrates that these modules represent generalizable human neurodevelopmental processes. At the earliest time point, there were 22 modules that differentiated between any of the forms of ASD and control participants. However, no single module differentiated diagnosis broadly from control participants (Fig. Eleven of the modules were significantly associated with at least one form of ASD, and five were associated with several (Fig. By contrast, at day 100, which corresponds to early mid-gestation36,50,85 and contains a large pool of postmitotic neurons and remaining radial glia, we identified only six modules that were associated with specific forms of ASD (Methods and Fig. However, at day 100, significantly more modules differentiated those participants who were affected (diagnosis) broadly from control participants (6 at day 100 compared with 0 at day 25; Fisher test P = 0.02; Fig. 3f) and the association of the modules with diagnosis broadly was significantly higher compared with day 25 (day 25 = 0.03, day 100 = 0.11, Welch two sample t30.69 = −3.0, P = 5.0 × 10−3; Fig. 3g), again demonstrating that the different genetic forms become more similar as differentiation progresses (Fig. We found that two modules, M5 and M19, were significantly enriched for ASD risk genes (Fig. This indicates that known ASD risk genes are strongly co-expressed within networks affected by perturbation of neural development in several rare forms of ASD, similar to observations based on in vivo data17,24,25,29,86. Most notable in this regard is M5, which was significantly downregulated across several forms of ASD at day 25 (adjusted R = −0.76, FDR = 5.2 × 10−15; 16p11.2 deletion beta = −0.06, FDR = 2.3 × 10−6; 16p11.2 duplication beta = −0.05, FDR = 5.6 × 10−4; PCDH19 beta = −0.08, FDR = 1.7 × 10−5) (Fig. 3c) and across differentiation (16p11.2 deletion: t = −3.5, FDR = 2.0 × 10−3; 16p11.2 duplication: t = −3.6, FDR = 2.0 × 10−3; PCDH19 t = −2.6, FDR = 2.0 × 10−2; SHANK3 t = −2.3, FDR = 3.8 × 10−2; Methods and Extended Data Fig. 11) suggesting that it has a central role downstream of these four ASD-associated mutations. M5's relationship to in vivo human biology is supported by its significant overlap with two co-expression modules that peak during human fetal brain around mid-gestation that were previously identified as significantly enriched for rare ASD risk mutations24 (Fig. We observed that M5 was also enriched for genes co-expressed within a neuronal module that is downregulated in ASD postmortem cortex14 (Fig. 3d: ASD M4, OR = 3.5, FDR = 3.3 × 10−5), providing further evidence that it represents biological processes occurring in vivo in ASD. Functionally, M5 was enriched for terms related to regulation of gene expression such as ‘Histone modification' and ‘DNA-binding transcription factor binding' (Fig. 3e and Supplementary Table 8) and was enriched in developing cell types: early radial glia (RG-1; FC = 1.1, FDR = 4.5 × 10−6; Fig. 3d), intermediate progenitor cells (FC = 1.1, FDR = 4.5 × 10−6; Fig. 3d) and early cortical neurons (ExNeu-1; FC = 1.1, FDR = 3.0 × 10−5; Fig. M19, the other module that was enriched for ASD risk genes, was downregulated across several ASD forms (day 25 adjusted R = −0.35, FDR = 0.01; Fig. M19 also overlapped with two later expressed, neuronal, human fetal cortical co-expression modules enriched for ASD risk24 (Prenatal M16 OR = 4.1, FDR = 1.7 × 10−5; Prenatal M17 OR = 2.7, FDR = 1.2 × 10−3; Fig. Functionally, M19 was enriched for terms related to synaptic function such as ‘Synapse organization' and ‘Neurotransmitter secretion', and neuronal cell types (ExNeu-1 FC = 1.5, FDR = 3.0 × 10−6; ExNeu-2 FC = 1.6, FDR = 3.0 × 10−6; InNeu FC = 1.5, FDR = 3.0 × 10−6; Fig. As M5 met several criteria indicating it might be a factor in mediating convergence, including enrichment of early expressed ASD risk genes, significant differential expression across many forms, and enrichment for terms related to regulation of gene expression (Fig. 3c–e), we next tested the regulatory relationships between M5 and other modules associated with various forms of ASD (Fig. We determined the hierarchical regulatory relationships among modules (Fig. 4a, Supplementary Table 9, Methods and ref. 87), finding that M5 and M1 contained the highest levels of predicted upstream transcriptional regulators of other modules (Fig. 4b,c and Extended Data Fig. Transcriptional targets of M5 were enriched in ten of the ASD-associated modules, among which was M19 (Figs. This highlighted M5's potential role as a causal driver of the changes seen in the genetically defined forms of ASD, which is consistent with its expression at the earliest stages (Extended Data Fig. a, Workflow schema showing identification of enriched motifs (RcisTarget). Enriched motifs upstream of module genes are annotated to high-confidence DNA-binding genes. TSS, transcription start site. b, Modules predicted to be regulated by M5. Line width represents kME weighted score, circle size represents proportion of genes regulated by M5. c, Heatmap of regulatory relationships between modules using a weighted kME score to determine the strength of regulatory relationships between modules (x axis, regulating module; y axis, regulated module). d, Correlation of module 5 eigengene with downstream modules across differentiation, showing negative correlations with most downstream modules across time. e, Heatmap of logFC of gene regulators in M5 at day 25 shows strong decreases across several forms of ASD. Statistics are derived from dreamlet, * FDR < 0.05, ** FDR < 0.01, *** FDR < 0.005. f, Correlations of transcriptional regulators in M5 with ASD risk genes found in modules downstream of M5. M5 genes are in bold with a white ring. Most correlations are negative (blue) whereas genes within modules are positively correlated (red). g, Peak expression day of the regulators in M5 and the ASD genes that they regulate. Colour bars represent the percent of genes whose mean expression peak in the corresponding differentiation day. h, PPI of M5 regulator genes. These genes form a significant PPI network (P = 0.001, DAPPLE's permutation test for the direct-edges network). i, Enrichment of regulators in M5 as well as their indirect network for ASD genes2,8 as well as NDD and intellectual disability risk genes91. Significance was tested using a two-sided Fisher's exact test. Several modules downstream of M5 were associated with various processes known to be dysregulated in NDDs (Figs. We observed a relationship between M5 downregulation and the upregulated expression of genes in its downstream putative target modules, suggesting that M5 may be negatively regulating these modules (Fig. As genes in a module can vary in their degree of module membership88, we next asked whether the transcriptional regulators within M5 followed the same downregulation pattern. Indeed 65% (17 out of 26) of the annotated transcriptional regulators (Methods) in M5 were significantly downregulated in at least one of the forms of ASD with most (8 out of 9) of the other regulators trending in the same direction (Fig. To further elaborate the relationship between M5 regulatory targets and ASD, we asked whether genes regulated by M5 were enriched for ASD genetic risk factors. Indeed, even though the individual co-expression modules containing these genes were not enriched for ASD risk genes, we found that the M5 regulated genes within M5 regulated modules were enriched for ASD risk genes (OR = 2.0, P = 3.6 × 10−9). A total of 261 SFARI 1, 2 or S genes within enriched downstream modules were predicted to be regulated by at least one M5 regulator (Fig. Noticeably, these ASD risk genes were mostly (68%) negatively correlated with the M5 regulators (Fig. Consistent with their role as transcriptional regulators, the driver genes within M5 peak in expression earlier than the ASD genes they are regulating (Fig. These M5 regulated ASD risk genes included many high-confidence risk genes associated with synapse-related function, such as CNTNAP2 and NLGN1 (M16), forebrain neuron differentiation, such as FOXG1 and PAX6 (M15), and epigenetic remodellers including SET, DYRK1A, KMT2E and CHD2 (M2) (Fig. 4f), which are all processes previously related to ASD2,24,89. These genes also included PCDH19 (M15) and CACNA1C, the gene mutated in Timothy syndrome (M10) (Fig. 4f), further reflecting this transcriptional network's role as a point of convergence in ASD. Taken together, these findings suggest that a complex of transcriptional regulators orchestrates distinct biological pathways related to neurogenesis and neuronal maturation altered across genetically defined forms of ASD through regulatory relationships occurring during early stages of brain development. To determine whether these transcriptional regulators represent known biological processes, we asked whether they interact at the protein level. We found that they are predicted to form a highly significant protein–protein interaction (PPI) network (Methods; P = 0.001; Fig. 4h) enriched for transcription factor regulation and chromatin binding related gene ontology terms (Supplementary Table 10). Further examination of this PPI network indicates significant enrichment in known ASD and NDD risk genes (direct PPI, SFARI 1, 2, S genes OR = 15, FDR = 2.0 × 10−4; NDD OR = 8.4, FDR = 3.7 × 10−4; ID OR = 9.7, FDR = 2.0 × 10−4; Indirect PPI, SFARI genes OR = 3.7, FDR = 4.6 × 10−11; ASD risk genes OR = 2.4, FDR = 3.0 × 10−2; NDD genes OR = 2.1, FDR = 1.7 × 10−7; intellectual disability genes OR = 2.4, FDR = 1.9 × 10−9; brain disorder genes OR = 2.4, FDR = 3.6 × 10−3; Fig. Moreover, similar to the transcriptome-based M5, the PPI network strongly overlapped with the early to mid-gestation brain co-expression networks that have previously been associated with ASD24 (Direct PPI, Prenatal M2: OR = 12.3, FDR = 2.5 × 10−4; Fig. 4i), including common variant-enriched modules29 (Direct PPI, Developing Yellow: OR = 43, FDR = 7.1 × 10−10; Fig. We next tested whether the predicted PPI interactions among M5 genes based on public databases (Fig. 4h–i) were observed in human neural progenitor cells (hNPCs), given that protein interactions can be highly cell-type specific and most available data are not from neural cell-types90. We conducted immunoprecipitation followed by mass spectrometry (IP–MS) on proteins with reliable antibody signal (Fig. 5a,b, Extended Data Fig. 12, Supplementary Table 11 and Methods). We then examined connections between M5 proteins (Methods), observing that M5 regulators do indeed form a significant PPI with each other and with a subset of the predicted interactors (Fig. 5c–f), including SFARI genes (score 1, 2 and S combined and score S; Fig. 5c,f and Extended Data Fig. 12) and genes enriched in developmental delay as well as NDD broadly92, validating database predictions (Fig. Of the five M5 regulators that passed quality control (Methods), we found that most were connected to three out of four of their possible connections, except for POLR2A (Fig. This PPI includes members of the BAF complex, which has a critical role during neurogenesis and mammalian cortical development93,94,95,96,97, SMARCB1 (BAF47) and SMARCA4 (BRG1), EP300, an important coactivator known to interact with BAF complex members98 and preinitiation complex member, TBP. These findings show that M5 regulators interact with a network of genes essential for early neurodevelopment, both in terms of predicted transcriptional control and direct PPI in neural progenitors. a, Western blots (run once per antibody) confirm antibody binding for proteins of interest. b, Example volcano plot shows significantly bound proteins. Highlighted proteins are SFARI risk genes. Statistics were derived from three replicates using a two-tailed one-sample moderated t-test with FDR correction and a total of 632 variables. c, Z-score connectedness of gene lists show that M5 proteins detected in IP–MS experiments form a significant PPI and that this network is significantly connected with other neurodevelopmental risk such as SFARI genes (1, 2, S), NDD genes2,8,91 and genes enriched for developmental delay (DD), ASD or NDD broadly92. *FDR < 0.05, **FDR < 0.01, ***FDR < 0.005, z-score computed against a permutation-derived null; FDR-corrected P values obtained from standard normal approximation. d, Experimentally determined PPI network of M5 proteins tested in IP–MS experiements. e–f, Experimentally determined PPI networks of the five M5 baits used showing their interactions with SFARI genes (e) and risk genes as defined in Fu et al.92 (f). Size of circle represents the number of connections, edge width represents logFC between nodes from one direction of the IP for each UniProt accession number. Colour represents the gene list. Grey nodes are network baits not in the featured gene list. a, Schematic of experimental approach: 96 gRNAs corresponding to 26 genes and 18 non-targeting controls (NTCs) were designed and transduced into NPCs. Seven days posttransduction, single cells were sorted by fluorescence-activated cell sorting (FACS) to capture dCAS9–gRNA positive cells and scRNA-seq was performed (n = 6 replicates). NPCs were separated through single-cell clustering into cNPCs and NPCs. b, Ratio of target gene expression (y axis) remaining in target gRNA-expressing cells (x axis) compared to NTC. Boxplots show: centre, median; lower hinge, 25% quantile; upper hinge, 75% quantile; lower whisker, smallest observation greater than or equal to lower hinge −1.5× interquartile range; upper whisker, largest observation less than or equal to upper hinge +1.5× interquartile range. Dots show n = 6 replicate libraries. c, The log counts of DEGs (Padjusted  < 0.05, edgeR pseudobulk derived) in each target. d, Enrichment of predicted regulated genes in DEG dataset. e, Enrichment of DEGs in different groups of developmental disorder risk genes: SFARI genes (1, 2, S), ASD risk genes from refs. DP, double positive; GFP, green fluorescent protein; KD, knockdown; RFP, red fluorescent protein; sgRNA, single-guide RNA. Having validated components of the PPI network in developing neurons, we next sought to experimentally validate the regulatory role of M5 transcription factors and chromatin regulators on their predicted downstream targets. We leveraged the ability to perform a pooled, CRISPR–Cas9 inhibition (CRISPRi) screen to assess downstream gene expression on a single-cell basis99 (Methods, Fig. 6a and Extended Data Fig. We applied this to neural cells derived by NGN2 induction in hiPS cells (hNPCs, Methods), targeting the promoters of 26 M5 putative transcriptional regulators using 3 guides per gene100 (Methods and Supplementary Tables 14 and 15). We achieved consistent downregulation (Fig. 6b and Extended Data Fig. 13), ranging from roughly 25% (POLR2A) to 90% decrease (MBTPS2) (Fig. hNPCs separated into mitotically active (NPCs showing consistent G2 mitosis markers (termed cycling NPCs or cNPCs), Extended Data Fig. 13), or those not expressing markers of active cell division (termed NPCs). For most targets, knockdown did not have a major impact on survival or proliferation, with 25 out of 26 knockdown targets showing no difference in proportion of cells compared to control gRNA-expressing cells, except for TP53, which showed increased proportions of cNPCs and NPCs (Extended Data Fig. 13), consistent with its well-known role as a tumour suppressor101. We observed a wide spectrum of changes in gene expression following pooled CRISPRi, with RBBP5 knockdown leading to more than 1,250 downregulated and more than 500 upregulated genes in NPCs to genes such as RELA and BACH2 in which we observed no significant differential expression (Methods, Fig. RBBP5 is part of the WRAD complex, which supports H3K4 methylation for SET-domain containing methyltransferases, many of which are NDD risk genes102. RBBP5 knockdown led to massive changes in gene expression, including SWI–SNFI member SMARCA1 and synapse genes DLGAP1, NRXN3 and CACNA1C, the gene mutated in Timothy syndrome. Not every M5 regulator showed strong evidence for recapitulating phenotypes seen in hCOs, but knockdown of many M5 members produced a transcriptional response that was correlated with changes seen in organoids (Extended Data Fig. Overall, knockdown of 11 of the 26 M5 regulatory drivers showed significant differential expression enrichment of their predicted targets: 9 out of 26 in cNPC and 6 out of 26 in NPCs, confirming their regulatory relationships (Methods, Fig. 6d and Extended Data Fig. Parallel with the earlier bioinformatic predictions, we found that genes associated with NDDs and intellectual disability in addition to ASD risk genes2,8,91 (SFARI levels 1 and 2) were enriched in DEGs (Fig. Pathways downstream of several of the individual M5 regulators, including RBBP5, EP300 and SOX9 lead to disruptions of overlapping neurodevelopmental pathways (Extended Data Fig. To determine the shared downstream effects of the M5 regulators, we performed a meta-analysis of single-cell differential expression from the pooled CRISPRi experiments. We observed that the genes significant in meta-analysis were enriched in several modules predicted to be regulated by M5 genes, including M2, M4 and M5 itself, suggesting regulation of other genes within the module by M5 regulators in NPCs (Extended Data Fig. In summary, this examination of molecular responses to each M5 regulator perturbation, albeit in a single-cell type, was important to experimentally validate the bioinformatic predictions of regulatory relationships. These results indicate that further examination of these regulators in organoid models will help inform how changes in these genes alter cellular transitions during brain development. We characterized the biological impact of several different genetic forms of ASD modelled in a large cohort of hCO, based on quantitative analyses rooted in genome-wide transcriptomic profiling. We identified robust clusters of different ASD forms spanning across stages of cortical development, suggesting several distinct developmental trajectories. Orthogonal analytic approaches showed that as time progressed these trajectories converged, consistent with canalization of development in which changes in the early stages of brain development are buffered as development progresses103. Overall, these trajectories converge on diminished neuronal maturation across time points, with most forms showing changing trajectories starting from day 25 and all, except for idiopathic ASD, showing changes by day 100. Although the alterations in maturation are convergent, the path to these changes seems to be distinct depending on different mutations, consistent with previous observations from a cohort of three other genetic forms of ASD43. Our findings suggest that convergence represents the impact of shared early developmental changes in co-expression networks enriched in ASD risk genes, whose downstream effects represent a form of developmental canalization on neuronal and synapse maturation. Our findings point towards known ASD risk genes associated with chromatin remodelling as a bridge linking early mutational specificity with convergence during cortical neurogenesis across different genetic forms of ASD. We describe a core transcriptional network of driver genes acting at early stages and identified their downstream target biological processes. This core set of convergent genes is changed across several distinct genetic forms and is enriched in rare and common genetic risk for ASD. M5 dysregulation may not be limited to ASD, as it was also enriched for common variation associated with several psychiatric disorders104,105, and formed a PPI with neurodevelopmental risk genes broadly, reflecting the role of early neuronal differentiation in broad risk for psychiatric and NDDs. These gene regulatory relationships, which we show form interacting protein complexes, connect ASD risk genes across different pathways, including WNT signalling, translation and mTOR pathways, and synapse formation into a convergent network. It is also worth noting that we identified few consistent changes in lines from idiopathic ASD. This is probably due to the substantial polygenicity in ASD cases not harbouring major risk mutations106,107, and supports the need for much larger sample sizes in studies focused on idiopathic forms of ASD. This work and other recent work43,47 demonstrates the utility of studying many mutations in parallel across developmental time to identify shared and distinct pathways in ASD. We note that the absence of effects in idiopathic ASD tempers our ability to identify broad, core ASD phenotypes that overlap across polygenic and major gene forms. Despite the comparatively large number of idiopathic lines profiled here, our data indicate that even larger samples will be needed to find robust shared phenotypes. However, the known overlap between genes associated with NDD and with ASD and our experimental strategy to examine several genetically defined forms of ASD in parallel, supports the relevance of these findings to ASD. By describing a core transcriptional network of driver genes, enriched in causal ASD genes and identifying their downstream targets, which are also enriched in ASD risk genes, these data connect ASD risk in genetically defined forms to convergent biological processes that occur during early cortical development. Control hiPS cell lines in this study were reported and have been extensively characterized using standardized methods44,50,51. hiPS cell lines from 11 idiopathic ASD individuals were obtained from Coriell. The 22q11.2 deletion44, 22q13.3 deletion108 and Timothy syndrome45 hiPS cell lines were reported and characterized before. To confirm the various genetic mutations, whole-genome sequencing was performed. Cultures were regularly tested for mycoplasma and maintained mycoplasma free. Approval for reprogramming and differentiation of these cell lines was obtained from the Stanford Institutional Review Board panel, and informed consent was obtained from all individuals. Human fibroblast cells were reprogrammed using the Sendai virus-based CytoTune iPS 2.0 kit (CytoTune 2.0, Invitrogen, A16517). Fibroblasts were plated in six-well plates 2 days before transduction to achieve a density of 2 × 105 to 3 × 105 cells per well. Cells were then transduced with the indicated set of Sendai vectors at a multiplicity of infection of 5–5–3 (KOS–Myc–Klf4) as we previously described in ref. Virus was removed after 24 h and cells were cultured in fibroblast medium with 10% FBS (Gibco, 16000-044) in Dulbecco's modified Eagle medium (DMEM) (Gibco, 10569-010) for a total of 7 days, with media changes every other day. On day 7 after transduction, cells were collected with TrypLE (Gibco, 16563-029) and plated in the fibroblast medium onto truncated recombinant human vitronectin (VTN-N, Gibco, A14700)-coated plates. The following day, fibroblast medium was changed to Essential 8 medium (Gibco, A1517001) and cells were fed daily until days 20–22. For further culture, individual colonies were manually dissected and transferred to fresh vitronectin-coated dishes. Once hiPS cell clones were established, they were routinely cultured on vitronectin-coated dishes in Essential 8 medium, and cells were split every 4–5 days using 0.5 mM EDTA (Invitrogen, 15575-020). hiPS cell clones and lines were also generated using retroviruses or non-integrating episomal vectors from fibroblasts109. We did not observe any systematic differences between cells reprogrammed by the two methods. hiPS cells were cultured as previously described in ref. Briefly, hiPS cells were cultured on six-well plates coated with recombinant human vitronectin (VTN-N, Gibco, A14700) in Essential 8 medium (Gibco, A1517001) and passaged with 0.5 mM EDTA (Gibco, 15575). To coat the six-well plates, 1 ml of vitronectin (diluted at a 1:100 ratio with Dulbecco's phosphate-buffered saline (DPBS); Gibco, 14190) was added per well and then incubated at room temperature for 1 h. To passage hiPS cells with 80–90% confluency, cells were rinsed with 3–4 ml of DPBS per well, and then 1 ml of 0.5 mM EDTA (Gibco, 15575) was added and incubated for 7 min at room temperature. After the EDTA was removed, 2 ml of prewarmed complete Essential 8 medium was added to collect cells. The cell suspension was then diluted in Essential 8 medium (1:6–1:20 depending on the hiPS cell line) and distributed on vitronectin-coated wells. For the generation of 3D spheroids, hiPS cells were incubated with Accutase (Innovate Cell Technologies, AT-104) at 37 °C for 7 min and dissociated into single cells. To obtain uniformly sized spheroids, we used AggreWell 800 plates (STEMCELL Technologies, 34811) containing 300 microwells. There were roughly 3 × 106 single cells per well in Essential 8 medium supplemented with the ROCK inhibitor Y-27632 (10 μM, Selleckchem, S1049), centrifuged at 100g for 3 min to capture the cells in the microwells and incubated at 37 °C with 5% CO2. After 24 h, day 0 of differentiation, spheroids were collected from each microwell by firmly pipetting (with a cut end of a P1000 tip) medium in the well up and down, and transferring it into ultra-low-attachment plastic dishes (Corning, 3262) in Essential 6 medium (Gibco, A1516401) supplemented with two SMAD pathway inhibitors: dorsomorphin (2.5 μM, Sigma-Aldrich, P5499) and SB431542 (10 μM, Tocris, 1614) together with Wnt pathway inhibitor XAV939 (1.25 μM, Tocris, 3748). Media changes were performed daily, except for day 1. On the sixth day in suspension, neural spheroids were transferred to neural medium containing Neurobasal-A (Gibco, 10888), B-27 supplement without vitamin A (Gibco, 12587), GlutaMax (1:100, Gibco, 35050) and penicillin and streptomycin (1:100, Gibco, 15140). The neural medium was supplemented with epidermal growth factor (20 ng ml−1, R&D Systems, 236-EG) and basic fibroblast growth factor (20 ng ml−1, R&D Systems, 233-FB) until day 24. From days 25 to 42, to promote differentiation of the neural progenitors into neurons, the neural medium was supplemented with brain-derived neurotrophic factor (20 ng ml−1, Peprotech, 450-02) and neurotrophin 3 (20 ng ml−1, Peprotech, 450-03), with medium changes every other day. From day 43 onwards, only neural medium without growth factors was used for medium changes every 4–5 days. Sequencing libraries were prepared using TruSeq stranded RNA RiboZero Gold (Illumina) on ribosomal RNA-depleted (RiboZero Gold, Illumina) RNA. The libraries were sequenced with 100-base-pair paired-end reads on an Illumina HiSeq 4000. The reads were mapped to the human genome (hg38) with Gencode v.25 annotations using STAR (v.2.5.2b)110 and gene expression was quantified using RSEM (v.1.3.0)111. Genes that were expressed at very low levels (fewer than 10 reads in 30% of samples from a given day) were removed from the analysis. Samples with standardized sample network connectivity Z scores below −2 in each mutation were defined as outliers and removed112. To control for technical variation due to the sequencing and library prep we calculated the principal components of the Picard sequencing metrics (http://broadinstitute.github.io/picard/) using the CollectAlignmentSummaryMetrics, CollectRnaSeqMetrics and MarkDuplicates modules, and included them in our model. To infer genetic ancestry, we called single-nucleotide polymorphisms (SNPs) from the aligned reads using the GATK (v.3.3) Haplotype caller113. Sites with more than 5% missing samples, with rare minor allele frequency (less than 0.05) and Hardy–Weinberg disequilibrium (less than 1 × 10−6) were removed using plink (v.1.09)114. We then used the remaining high-quality SNPs to run MDS together with HapMap3.3 (hg38)115. The first two MDS values, referred to as genetic ancestry principal components 1 and 2 (PC1, PC2), were then included in our model. Sample identity was verified using the identity by descent algorithm from PLINK (v.1.09)114. Identity by descent was calculated for each pair of samples based on genotypes derived from RNA-seq analysis as well as for all pairs of RNA-seq and DNA sequencing samples. Samples with a \(\hat{\pi }\) < 0.8 from other samples derived from the same individual were removed. Sex was verified for each sample by detection of genes expressed by the Y chromosome. Sample with high levels of duplication (Picard tools, PERCENT_DUPLICATION > 65), high levels of intergenic mapping (Picard tools, PCT_INTERGENIC_BASES > 60) or with low levels of messenger RNA (RNA) (Picard tools, PCT_MRNA_BASES < 0.5) were removed. All RNA-seq samples passing quality control are listed in Supplementary Table 1. Variance partitioning was performed using the variancePartition package (v.1.20.0) with default parameters. Gene-expression reproducibility was measured using Spearman's correlation between every pair of samples in a given time point and mutation both within and between individuals. Libraries were prepared using TruSeq DNA PCR-Free (Illumina) and were sequenced with 150-base-pair paired-end reads on an Illumina HiSeq 4000. Lines that underwent whole-genome sequencing are in Supplementary Table 2: further lines were quality controlled previously44. Sequencing data were processed as previously described in ref. The sequencing reads were mapped to the human genome (hg38) using Burrows–Wheeler Aligner (bwa-mem, v.0.7.17)116. The resulting BAM files were merged using bamtools (v.2.5.1)117 resulting in a single BAM file per sample. Duplicate reads were marked using the Picard MarkDuplicates tool (v.2.5.0). Reads around indels were locally realigned using GATK's IndelRealigner (v.3.3). Base quality score was recalibrated using GATK113. Variant and non-variant bases in the genome were identified using GATK's HaplotypeCaller (v.3.3). Resulting variant cal formats were combined using GATK's combineGVCFs (v.3.3) and variants were called per chromosome using GATK's GenotypeGVCFs (v.3.3). Well-calibrated quality scores were generated using GATK's Variant Quality Score Recalibration (v.3.3). Large chromosomal changes, resulting from iPS cell chromosomal instability, were identified using DELLY (v.0.8.7)118. Lines with large indels (more than 750 kb) that were confirmed using allele frequency in the RNA-seq data were removed from the analysis. This included five individuals in whom all lines were dropped (four because of incorrect or absent mutations, one that did not pass quality control). Twenty-six lines in total (19 individuals) were dropped due to quality control failures. These include seven lines that were dropped due to failure to confirm mutation (four individuals), nine lines dropped due to large CNVs (seven individuals), two lines dropped due to leftover Sendai virus (two individuals), four due to further or improper deletions (four individuals) and four that were not fully characterized or did not pass quality control (two individuals). Point mutations in SHANK3, PCDH19 and CACNA1C were visually confirmed using Integrative Genomics Viewer (v.2.9.4)119. All whole-genome sequencing samples are listed in Supplementary Table 2. Gene counts were normalized using trimmed mean of M values121 as implemented in the calcNormFactors function from the edgeR package (v.3.26.8)122. The mixed linear model used included differentiation day, sex, batch, the first 2 principal components calculated for genetic ancestry and 15 sequencing principal components (SeqPCs) as fixed effects, and accounted for many samples coming from the same individual, hiPS cell line and differentiation, and for samples that were sequenced several times as random effects. Genes with FDR < 0.05 were considered to be differentially expressed. To conduct meta-analyses on mutational differential gene expression, we used the metafor (v.4.6.0) package by fitting a fixed-effect model with the rma function for each time point124. FDR < 0.05 genes were considered to be differentially expressed. GSEA was run on metafor analysis output to determine enriched pathways using the fgsea package (v.1.3.0). Independent component analysis was performed using the JADE algorithm as implemented in the MineICA package (v.1.30.0)125. The number of components was set to 36 with 10,000 maximum iterations. Association of the independent component with the forms of ASD was tested using a mixed linear model. Contrasts were used to compare each form of ASD at each time point to the control at the same time point using the glht function from the multcomp package (v.1.4.18). Like other component analyses, independent component directions are random and so enrichment was performed for both positively and negatively contributing genes. Unlike principal components, independent components are not inherently ordered and were therefore ordered by the amount of kurtosis in each independent component. GSEA was performed using the fgsea package (v.1.10.1)126 on all genes with 1,000,000 permutations and a set size ranging from 30 to 500. For differential expression genes were ranked by log2FC and for independent component analysed genes were ranked by the projection values of the genes on each component. Enrichment for high-confidence ASD genes (gene score of two or less, or syndromic genes) from the SFARI database (https://gene.sfari.org/database/gene-scoring/), from large scale ASD sequencing studies2,8, and risk genes for NDDs, intellectual disability and brain disorders91 was performed using Fisher's exact test. Gene counts were normalized using conditional quantile normalization (v.1.36.0)127 and the covariates sex, tissue of origin, genetic ancestry PC1 and PC2 and the first 15 technical sequencing principal components were regressed out. The network was constructed using the WGCNA package (v.1.70.3)88 using a signed network for each time point separately. A modified version of robust WGCNA was used to reduce the influence of potential outlier samples on network architecture26. Samples were resampled from within each form of ASD ensuring that each iteration contained at least four samples from each form. In total, 100 networks were constructed followed by consensus network analysis that takes the median of the topological overlap dissimilarity matrix across the 100 networks88. Soft-threshold powers of 13, 9, 13 and 14 were used for days 25, 50, 75 and 100, respectively, to achieve approximate scale-free topology (R2 > 0.8). Modules were constructed using a minimal module size of 160, deep split of 4, cut height for creation of modules of 0.9999 and cut height for merging modules of 0.1. Each module eigengene (first principal component of the module) was tested for association with the different forms of ASD as well as for association with ASD overall using linear models. Module gene-ontology term enrichment was performed using clusterProfiler (v.4.0.5)128 with default parameters. Enrichment for common variation associated with ASD21, schizophrenia129, attention-deficit/hyperactivity disorder130, major depressive disorder131 and bipolar disorder132 was calculated using a stratified LDscore regression (v.1.0.0)133. SNPs with 10 kb of a gene were assigned to that gene and the enrichment was calculated as the proportion of SNP heritability accounted for by the genes in the module divided by the proportion of total SNPs within the module. Cell-type enrichment was performed using the bootstrap enrichment method from the EWCE package (v.1.0.1)134 using control samples from previously published single-cell data from a human cortical spheroids set44 with 100,000 permutations. As the background gene set, all genes expressed in both the current dataset and the single-cell dataset were used. Enrichment for high-confidence ASD genes (gene score of two or less or syndromic genes) from the SFARI database (https://gene.sfari.org/database/gene-scoring/) and ASD postmortem modules14 was performed using Fisher's exact test. Module trajectories were modelled using polynomial splines with two degrees of freedom and two degrees of the piecewise polynomial and tested for significant differences using a linear model. Module preservation was performed using the modulePreservation function from the WGCNA package with 100 permutations. Preservation was tested (1) at other time points within the current data set (2) in cortical samples from early developmental stages (8–16 weeks postconception) from the BrainSpan in vivo data set and (3) in other stem cell-based neuronal models81,82,83,84. To test the association of each ASD form with WGCNA modules, we ran a linear regression comparing the module eigengene of each module by form (as factor variable with control participants as the base level) or by diagnosis (affected versus control) using the lm() function. Adjusted R values of the regression are multiplied by the sign of the sum of model coefficients to determine direction of effect. Estimates for individual forms were derived from the model coefficients. We show FDR-corrected P values in the figures. The different conditions (ASD form/differentiation day) were clustered based on the Spearman correlation of log2FC of all expressed genes using the Ward.D2 clustering method on Euclidean distances as implemented in the ComplexHeatmap package (v.2.9.3)135. Cluster significance was tested using multiscale bootstrap resampling as implemented in the pvclust function in the pvclust package (v.2.2.0)136 using the Ward.D2 clustering method on Euclidean distances with 10,000 bootstraps. The bootstrap probability measure is the frequency that a cluster appears in the bootstrap replicates. To test for clustering similarity across clustering methods, clustering was performed using the Ward.D, Ward.D2, complete, average and Mcquitty methods on each of the Euclidean, maximum, Manhattan and Minkowski distances. For each pair of clustering and distance methods the cophenetic correlation was measured using the cor.dendlist function from the dendextend package (v.1.14.0)137. Cophenetic correlation is a measure of how well the pair-wise distances of the unmodeled data are preserved in the dendrogram138. In addition, the Rand index was measured for each combination of these pairs using the rand.index function from the fossil package (v.0.4.0)139. Rand index is a measure of the similarity between two trials of clustering using the same data140. To verify that the clusters were not driven by a small subset of genes, we subset the genes used for clustering to 10,000, 5,000 or 2,000 random genes 100 times. The resulting genes were correlated using Spearman correlation and clustered using the Ward.D2 clustering method on the basis of Euclidean distances. Cophenetic correlation and Rand indexes were calculated as above. Bulk RNA-seq counts were deconvolved using the control cell profiles from previously published single-cell data44 from day 75–85 organoids to deconvolve day 50, 75 and 100 time points, and new single-cell libraries were created with standard 10X methods from day 25 organoids to deconvolve the day 25 time point. Single-cell datasets were processed and integrated separately using standard Seurat v.5 workflows. Changes in cell-type proportion were tested using a logit mixed linear model. The logit transformation was performed using the logit function from the car package (v.3.0.11) with a 0.001 adjustment. Contrasts were used to compare each form of ASD at each time point to the control at the same time point using the glht function from the multicomp package (v.1.4.18). To identify transcription factor binding motifs over-represented in each module we used RcisTarget (v.1.6.0)142. The transcription factor data base used was hg38 and included 10 kb upstream and downstream of the transcription start site (version 9; hg38__refseq-r80__10kb_up_and_down_tss.mc9nr). Only high-confidence annotated transcription factors were used for downstream analysis. The regulatory relationship between modules was calculated as a kME (which is the correlation of each gene in the module with the module eigengene) weighted proportion of the number of genes regulated by the upstream module in the downstream module such that: \(\mathrm{score}=\frac{\sum k\mathrm{ME}\,\mathrm{of}\,\mathrm{regulated}\,\mathrm{genes}\,\mathrm{in}\,\mathrm{the}\,\mathrm{module}}{\mathrm{Total}\,\mathrm{number}\,\mathrm{of}\,\mathrm{genes}\,\mathrm{in}\,\mathrm{the}\,\mathrm{module}}\). This weighted proportion puts larger weight on genes that are highly connected within the module. A module was considered to be highly regulated by another module if the phosphastaseME weighted score was above 1 standard deviation above the mean (which corresponded to a weighted score of 0.341). PPIs were mapped using the Dapple143 module (v.0.19) in GenePattern (http://genepattern.broadinstitute.org) using the hg19 genome build with 1,000 permutations. To validate predicted PPIs, we generated mass spectrometry data from key regulator proteins predicted to form a PPI. We tested antibodies against RELA (Active motif 40916), TP53 (Abcam ab26), CNOT6 (CST no. All antibodies were run at 1:1,000 concentrations. Antibodies that generated western blot bands at the appropriate size were selected for continued mass spectrometry analysis (run once; Fig. For western blot analysis of the proteins of interest, cell lysates were diluted in 6× SMASH buffer (5 mM Tis HCL pH 6.8, 10% glycerol, 2% SDS, 0.02% bromophenol blue and 1% beta-mercaptoethanol) and boiled for 10 min at 95 °C. Samples were then separated on a NuPAGE 4–12% Bis-Tris Protein Gel (Invitrogen) and transferred to a polyvinyl difluoride membrane (Life Technologies, 100 V, 2 h). Membranes were incubated at room temperature for 1 h in tris-buffered saline with Tween (TBST) (0.1% Tween) with 5% BioRad Blotting-grade Blocker and then incubated overnight with antibodies listed above at 4 °C. After 3× 10-min washes in TBST, blots were incubated with horseradish peroxidase (HRP)-conjugated secondary antibodies for 45 min and again washed three times in TBST. The following secondary antibodies were used: anti-mouse IgG HRP-linked (NA9310V, Sigma-Aldrich) and anti-rabbit IgG HRP-linked (GENA934, Sigma-Aldrich) at a concentration of 1:5,000. Western blots were imaged using SuperSignal West Femto Maximum Sensitivity Substrate (Thermo Scientific). Western blots were considered successful if the band was visualized at the expected molecular weight (Fig. For the 6 proteins that showed successful western blots, 1–2 mg of protein extract was incubated with primary antibody (1–2 µg) at 4 °C overnight (1 µg of primary antibody to 1 mg of protein). Following primary antibody incubation 100 µl of Protein A/G beads (Pierce) were added and incubated for 4 h at 4 °C, followed by 1 wash with lysis buffer (Pierce) with added Halt protease and phosphatase inhibitors (Thermo Scientific), 2 washes with PBS and then resuspended in PBS for mass spectrometry. Samples were processed and quantified for mass spectrometry as in ref. 90 for the 6 proteins (TBP, POLR2A, SOX9, SMARCB1, SMARCA4 and EP300) with a western blot signal in NGN2-derived NPCs at day 4 (refs. Briefly, NPCs were collected at day 4 following NGN2 expression and dual SMAD/WNT inhibition in the presence of doxycycline (rtTA, tetON-NGN2). Cells were pelleted, washed with PBS and resuspended in 10× packed cell volume immunoprecipitation lysis buffer (Thermo Scientific) with added Halt protease and phosphatase inhibitors (Thermo Scientific) for 20 min at 4 °C. Cells were then centrifuged (16,200g, 20 min, 4 °C and resuspended in 3× packed cell volume. Samples were quantified using the Thermo BCA protein assay. Immunoprecipitations were eluted using SDS and subject to electrophoresis for gel extraction. Bands were excised and cut into roughly 1 mm3 pieces. Gel pieces were then subjected to a modified in-gel trypsin digestion procedure, washed and dehydrated with acetonitrile for 10 min, followed by removal of acetonitrile. Pieces were then completely dried in a speed-vac. Rehydration of the gel pieces was with 50 mM ammonium bicarbonate solution containing 12.5 ng µl−1 modified sequencing-grade trypsin (Promega) at 4 °C. After 45 min, the excess trypsin solution was removed and replaced with 50 mM ammonium bicarbonate solution to just cover the gel pieces. Samples were then placed in a 37 °C room overnight. Peptides were later extracted by removing the ammonium bicarbonate solution, followed by 1 wash with a solution containing 50% acetonitrile and 1% formic acid. The extracts were then dried in a speed-vacuum (roughly 1 h). The samples were then stored at 4 °C until analysis. On the day of analysis, the samples were reconstituted in 5–10 µl of high-performance liquid chromatography solvent A (2.5% acetonitrile, 0.1% formic acid). A nano-scale reverse-phase high-performance liquid chromatography capillary column was created by packing 2.6-µm C18 spherical silica beads into a fused silica capillary (100 µm inner diameter by roughly 30 cm in length) with a flame-drawn tip. After equilibrating the column each sample was loaded by using a Famos auto sampler (LC Packings) onto the column. A gradient was formed, and peptides were eluted with increasing concentrations of solvent B (97.5% acetonitrile, 0.1% formic acid). As peptides eluted, they were subjected to electrospray ionization and then entered an LTQ Orbitrap Velos Pro ion-trap mass spectrometer (Thermo Fisher Scientific). Peptides were detected, isolated and fragmented to produce a tandem mass spectrum of specific fragment ions for each peptide. For all IP–MS experiments, the following isotype control antibodies were used: rabbit IgG monoclonal (EPR25A) (Abcam, no. ab172730), rabbit IgG polyclonal (Abcam, no. ab37415) and mouse IgG1 κ monoclonal [MOPC-21] (Abcam, no. Mass spectrometry experiments that yielded significance for the protein bait over IgG control samples were then included in analyses (all proteins except SOX9). Peptide sequences (and hence protein identity) were determined by matching the UniProt human protein database (release 2023_01) with the acquired fragmentation pattern using Sequest (Thermo Fisher Scientific)146. The database included a reversed version of all the sequences and the data were filtered to between 1% and 2% peptide FDR. Protein quantification was performed using GFY Core Version 3.8 (Harvard University). IP–MS data analysis is described in ref. To determine whether the M5 proteins formed a significant interaction network, we compared the logFCs between each ‘bait' and the other interactors predicted to be in the M5 network and found as a protein in any of the mass spectrometry experiments (943 unique proteins identified across all IP–MS experiments). We used the mean logFC across baits as a measure of average connectedness of the network. To assess the statistical significance of these networks, we ran 10,000 permutations by randomly selecting proteins of the same size as the network from the remaining set of 943 uniquely identified proteins. We compared the average logFC of each of these randomly generated networks to create a population curve. We then examined the z-scored connectedness of the network. Gene lists used to compare networks were SFARI genes (levels 1, 2, S), ASD risk genes2,8, intellectual disability risk genes91, NDD risk genes91 and developmental disorder enriched, ASD enriched and NDD non-specific92. Three genomic RNAs (gRNAs) per target as well as 18 total non-targeting controls (NTC) were designed using the CRISPick database (https://portals.broadinstitute.org/gppx/crispick/public) (Supplementary Table 15). NTCs were both gRNAs with no sequence in the human genome as well as gRNAs targeting random intergenic sequences. Sequences were designed as follows: a 5′ overhang (ATCTTGTGGAAAGGACGAAACACC) and a 3′ overhang (GTTTAAGAGCTATGCTGGAAACAGCATAGCAAGT) that is compatible with the CRISPR droplet sequencing (CROP-seq) vector were added on the outside of the gRNA sequence. A G nucleotide was also incorporated between the 5′ overhang and the gRNA sequence. All oligos were ordered as a pool (Integrated DNA Technologies). The CROP-seq-opti-dsRed vector backbone was obtained from Addgene (Addgene no. 201999) and digested using New England Biolabs' (NEB) BsmBI enzyme. The oligo pool was diluted at a concentration of 100 µM, PCR amplified, (Q5 Hot Start High-Fidelity 2X Master Mix) gel extracted, and cleaned. The vector and the pooled oligo insert was then assembled (NEBuilder HiFi DNA Assembly Reaction). To validate the insertion was present post assembly, a PCR amplification using the U6 outer forward primer (TTTCCCATGATTCCTTCATATTTGC) and gRNA end reverse primer (AGTACAAGCAAAAAGCAGTGTCTCAA) was run and visualized. All gRNA sequences used for M5 targets and control samples are supplied in Supplementary Tables 14 (targets) and 15 (control samples). The pooled vectors were then transformed into NEB Stable Competent E. coli (High Efficiency), and plated at various dilutions on Luria-Bertani medium:ampicillin 1:1,000 plates overnight at 30 °C. Colonies were picked and PCR amplification of the insert was done to insure its presence. All bacterial colonies were then scraped from the plate and grown using 200 ml of 2.5% Luria-Bertani medium + 1:1,000 ampicillin. The broth containing colonies was then shaken at 170 rpm at 30 °C and this was complete once the optical density reached 2.0 nm. DNA was then isolated by use of the Qiagen Maxi Plasmid Kit. Human embryonic kidney 293T (HEK293T) cells obtained from American Type Culture Collection (lines CRL-3216 and TIB-202) were expanded in DMEM, 10% FBS and 1% penicillin and streptomycin (HEK medium). Then 5 × 106 HEK293T were plated onto each of 2 poly-ornithine (5 µg ml−1) coated 10-cm plates for 24 h. Media from both plates was replaced with Opti-MEM before cells from 1 10-cm plate were transduced with a lentiviral vector (pLV-KRAB-dCas9 (Addgene no. 71236) or CROP-seq-opti-dsRed (Addgene no. 201999)) while the other plate of cells was transfected with 15 µg of the pooled plasmid library, 15 µg of psPAX2 and 1.5 µg of VSV-G with lipofectamine 3000 following the manufacturer's instructions. After 16 h, the Opti-MEM media was replaced with 10 ml of HEK medium, which was collected and replaced every 24 h twice. The HEK medium was then spun down with 12.5 ml of lenti-X concentrator (Takara Bio) and centrifuged at 1,500g for 45 min. The supernatant was discarded and the pellet was resuspended in 400 µl of PBS, aliquoted and stored at −80 °C. We derived NGN2-induced hNPCs according to ref. A six-well plate was coated with Matrigel (1:50 in Knockout DMEM) for 2 h at 37 °C. A dox-inducible NGN2 iPS cell line, WTC11 (from L. Gan, Weill Cornell, available at Coriell), was seeded at 75,000 cells per cm2 in MTeSR+ supplemented with 10 µM of Y-27632. After 24 h, cells were subjected to induction medium (1:100 glutamax, 1:66.7 20% glucose in DMEM:F12, N2 supplement 1:100, doxycycline 2 µg ml−1, LDN-193189 200 nM, SB431542 10 µM, XAV939 2 µM). Day 2 involved the induction and selection using the following media: 1:100 glutamax, 1:66.7 20% glucose in DMEM:F12, N2 supplement 1:100, doxycycline 2 µg ml−1, LDN-193189 100 nM, SB431542 5 µM, XAV939 1 µM and puromycin 5 µg ml−1. Then 24 h postselection, the media was changed fully with NES complete and selection media (DMEM:F12, glutamax 1:50, penicillin and streptomycin 1:100, MEM NEAA 1:100, B-27 without vitamin A 1:50, N2 supplement 1:100, epidermal growth factor 10 ng ml−1, basic fibroblast growth factor 10 ng ml−1, puromycin 5 µg ml−1 and Y-27632 1:1,000). One day later and for the duration of the culture, cells were put on NES complete medium, which is the same as listed above apart from the use of Y-27632 and puromycin. Once hNPCs were generated, 500,000 hNPCs were seeded in every well of a six-well plate with NPC maintenance media (n = 6 replicates). To each well, GFP-KRAB-dCas9 (40 µl per well of a six-well plate) plasmid and CROP-dsRed-gRNA (33 µl per well of a six-well plate) were added. Media was replaced daily for 7 days before cells were collected for fluorescence-activated cell sorting (FACS). For FACS, hNPCs were washed with PBS, dissociated into single-cell suspension using Accutase and resuspended in 500 µl of FACS buffer per replicate (7.5% BSA, 0.5 M EDTA, RNase inhibitor 1:200, DAPI 0.5 mg ml−1). A population of NPCs that were not transfected with plasmid were used as a negative control to establish FACS gating. Double positive GFP/TdTomato (dCAS9–single-guide RNA double expressing) cells were sorted for downstream single-cell analyses (roughly 30,000 cells per replicate after FACS, roughly 20% of live cell fraction). Sorted cells then underwent scRNA extraction using the 10X genomics Chromium Next GEM Single Cell 3′ GEM, Library & Gel Bead Kit v.3.1 (PN-1000121). The 10X protocol for library prep was adapted from ref. 147. gRNA libraries were sequenced separately from scRNA (GEX) libraries. scRNA and gRNA libraries were aligned to the human genome (GRCh38) with CellRanger (v.7.0.1). Roughly 10,000 aligned cells were recovered per replicate. Cells were filtered to include cells with more than 200 features and percentage of mitochondria less than 10%. After quality control and doublet filtering before selecting uniquely gRNA positive cells, the median nCounts per cell was 17,071 and the median nGenes was 4,675, from 55,216 singlets (Extended Data Fig. Single-cell data were then processed using standard best practices for Seurat v.5. Briefly, we log-normalized counts and performed integration using SCTransform followed by clustering. Canonical markers were used to determine cell-type classification including NPC markers SOX2, NES, VIM and HES1 and cNPCs expressing G2 mitotic markers TOP2A, HMGB2, CDK1 and NUSAP1 (Extended Data Fig. We then assigned gRNAs to cells and filtered cells as follows. BAM files generated from gRNA libraries were analysed as in ref. In short, a text file of gRNA sequences was supplied as well as the common sequence appearing before the gRNA sequence, and each of these reads was counted from the BAM file and paired to the cells within the scRNA libraries through cell unique molecular identifier matching. Only cells containing gRNA barcodes with a unique molecular identifier count of at least ten were kept in the downstream analysis (Extended Data Fig. This left a total of 23,672 high-quality singlet cells that had 1 and only 1 gRNA-target assigned (Extended Data Fig. This cut-off has also been previously used148 and was determined to be sufficient for significant decrease in expression of the target. All differential analysis was performed on cells that expressed the gRNA of a single target. A standard pseudobulk differential expression using the edgeR (v.3.40.2) Quasi-Likelihood F test was performed using the libra package (v.1.0.0) function, run_de(). All differential gene expression compared all cells expressing any control gRNAs with all cells expressing any target gRNAs, above the unique 10 unique-molecular-identifier cut-off (3,163 total cells considered as control cells). The range of cells containing each gene target was 171–2,423, median 775. The proportion of cells containing each unique barcode (summarized by gene target) per replicate is shown in Extended Data Fig. Each replicate sample, taken from different plate wells was considered one sample for psuedobulking. We have complied with all ethics guidelines and regulations in conducting this research. Informed consent was obtained from all individuals. Participants were recruited at Stanford University, UCLA, NIMH. Ethics oversight is conducted by Stanford University, UCLA and NIMH. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Processed datasets generated and analysed as a part of this current study are included as Supplementary Tables 1–15. Aligned RNA-seq counts data are available through the Gene Expression Omnibus at accession GSE271853. Owing to participant consent and privacy, we are not able to make raw data public, but they can be made available by the authors on reasonable request. Publicly available data used in this paper were as follows: datasets used to test module preservation: (1) from ref. 81, https://www.sciencedirect.com/science/article/pii/S0006322320317029#sec1 (Synapse ID, syn8118403); (2) ref. 84, https://www.ncbi.nlm.nih.gov/pubmed/30617258 (EMBL-EBI ArrayExpress with the accession code E-MTAB-6018) and (5) ref. Datasets used for single-cell deconvolution reference to deconvolute days 50–100 at accession no. Datasets used for enrichment testing: (1) SFARI genes https://gene.sfari.org/database/gene-scoring/; (2) ASD risk genes from ref. 2, https://pubmed.ncbi.nlm.nih.gov/31981491/ (Supplementary Table 2) and ref. 8, https://pubmed.ncbi.nlm.nih.gov/31398340/ (Supplementary Table 3); (3) intellectual disability risk genes and NDD risk genes from ref. 91, https://pubmed.ncbi.nlm.nih.gov/33932580/ (Supplementary Table 1) and (4) developmental disorder enriched, ASD enriched and NDD non-specific from ref. Datasets used for genome-wide association study enrichment are available through LDSC: for ASD from ref. 21 at https://www.nature.com/articles/s41588-019-0344-8, data available from iPSYCH at https://ipsych.dk/en/research/downloads/2; for schizophrenia from ref. 129 at https://www.nature.com/articles/s41588-018-0059-2, data available from the Walters Group at https://walters.psycm.cf.ac.uk/3; for attention-deficit/hyperactivity disorder from ref. 130 at https://www.nature.com/articles/s41588-018-0269-7, data through PGC at https://pgc.unc.edu/for-researchers/download-results/; for major depressive disorder from ref. 131, https://www.nature.com/articles/s41593-018-0326-7, data available from Edinburgh DataShare at https://datashare.ed.ac.uk/handle/10283/3203 and for bipolar disorder from ref. The hg38 genome for STAR alignment is available through GENCODE at https://www.gencodegenes.org/human/release_25.html. The Rcistarget databases are available from Stein Aerts laboratory at https://resources.aertslab.org/cistarget/databases/homo_sapiens/hg38/refseq_r80/mc9nr/gene_based/. The UniProt database is available at https://ftp.uniprot.org/pub/databases/uniprot/previous_releases/release-2023_01/. Code used to analyse and generate plots for this paper is available at Zenodo (https://doi.org/10.5281/zenodo.17645341)149. de la Torre-Ubieta, L., Won, H., Stein, J. L. & Geschwind, D. H. Advancing the understanding of autism disease mechanisms through genetics. Large-scale exome sequencing study implicates both developmental and functional changes in the neurobiology of autism. Huguet, G., Ey, E. & Bourgeron, T. The genetic landscapes of autism spectrum disorders. & State, M. W. Genomics, convergent neuroscience and progress in understanding autism spectrum disorder. Whole genome sequencing resource identifies 18 new candidate genes for autism spectrum disorder. Werling, D. M. et al. An analytical framework for whole-genome sequence association studies and its implications for autism spectrum disorder. Feliciano, P. et al. Exome sequencing of 457 autism families recruited online provides evidence for autism risk genes. Inherited and de novo genetic risk for autism impacts shared networks. De Rubeis, S. et al. Synaptic, transcriptional and chromatin genes disrupted in autism. The contribution of de novo coding mutations to autism spectrum disorder. Sebat, J. et al. Strong association of de novo copy number mutations with autism. Insights into autism spectrum disorder genomic architecture and biology from 71 risk loci. O'Roak, B. J. et al. Multiplex targeted sequencing identifies recurrently mutated genes in autism spectrum disorders. Parikshak, N. N. et al. Genome-wide changes in lncRNA, splicing, and regional gene expression patterns in autism. Velmeshev, D. et al. Single-cell genomics identifies cell type–specific molecular changes in autism. Gandal, M. J. et al. Broad transcriptomic dysregulation occurs across the cerebral cortex in ASD. Wamsley, B. et al. Molecular cascades and cell type-specific signatures in ASD revealed by single-cell genomics. Global prevalence of autism: a systematic review update. Havdahl, A. et al. Genetic contributions to autism spectrum disorder. Geschwind, D. H. & State, M. W. Gene hunting in autism spectrum disorder: on the path to precision medicine. Identification of common genetic risk variants for autism spectrum disorder. Weiner, D. J. et al. Polygenic transmission disequilibrium confirms that common and rare variation act additively to create risk for autism spectrum disorders. Most genetic risk for autism resides with common variation. Parikshak, N. N. et al. Integrative functional genomic analyses implicate specific molecular pathways and circuits in autism. Willsey, A. J. et al. Coexpression networks implicate human midfetal deep cortical projection neurons in the pathogenesis of autism. Shared molecular neuropathology across major psychiatric disorders parallels polygenic overlap. Gupta, S. et al. Transcriptome analysis reveals dysregulation of innate immune response genes and neuronal activity-dependent genes in autism. Ben-David, E. & Shifman, S. Combined analysis of exome sequencing points toward a major role for transcription regulation during brain development in autism. Walker, R. L. et al. Genetic control of expression and splicing in developing human brain informs disease mechanisms. Won, H., Huang, J., Opland, C. K., Hartl, C. L. & Geschwind, D. H. Human evolved regulatory elements modulate genes involved in cortical expansion and neurodevelopmental disease susceptibility. Early brain development in infants at high risk for autism spectrum disorder. A., Peñagarikano, O., Belgard, T. G., Swarup, V. & Geschwind, D. H. The emerging picture of autism spectrum disorder: genetics and pathology. Pașca, S. P. et al. A nomenclature consensus for nervous system organoids and assembloids. Seah, C., Huckins, L. M. & Brennand, K. J. Stem cell models for context-specific modeling in psychiatric disorders. Amin, N. D. & Paşca, S. P. Building models of brain disorders with three-dimensional organoids. Long-term maturation of human cortical organoids matches key early postnatal transitions. Arlotta, P. & Paşca, S. P. Cell diversity in the human cerebral cortex: from the embryo to brain organoids. Pașca, S. P. The rise of three-dimensional human brain cultures. Using stem cell models to explore the genetics underlying psychiatric disorders: linking risk variants, genes, and biology in brain disease. Wen, Z. et al. Synaptic dysregulation in a human iPS cell model of mental disorders. Zhang, Z., Wang, X., Park, S., Song, H. & Ming, G. L. Development and application of brain region-specific organoids for investigating psychiatric disorders. Modeling idiopathic autism in forebrain organoids reveals an imbalance of excitatory cortical neuron subtypes during early neurogenesis. Paulsen, B. et al. Autism genes converge on asynchronous development of shared neuron classes. Khan, T. A. et al. Neuronal defects in a human cellular model of 22q11.2 deletion syndrome. Birey, F. et al. Dissecting the molecular basis of human interneuron migration in forebrain assembloids from Timothy syndrome. Lancaster, M. A. et al. Cerebral organoids model human brain development and microcephaly. Li, C. et al. Single-cell brain organoid screening identifies developmental defects in autism. In vivo Perturb-Seq reveals neuronal and glial abnormalities associated with autism risk genes. Meng, X. et al. Assembloid CRISPR screens reveal impact of disease genes in human neurodevelopment. Paşca, A. M. et al. Functional cortical neurons and astrocytes from human pluripotent stem cells in 3D culture. Reliability of human cortical organoid generation. McDonald-McGinn, D. M. et al. 22q11.2 deletion syndrome. & McDonald-McGinn, D. Developmental trajectories in 22q11.2 deletion. Sharp, A. J. et al. A recurrent 15q13.3 microdeletion syndrome associated with mental retardation and seizures. Chung, W. K., Roberts, T. P. L., Sherr, E. H., Snyder, L. G. & Spiro, J. E. 16p11.2 deletion syndrome. Steinman, K. J. et al. 16p11.2 deletion and duplication: characterizing neurologic phenotypes in a large clinically ascertained cohort. Splawski, I. et al. Ca(V)1.2 calcium channel dysfunction causes a multisystem disorder including arrhythmia and autism. Bauer, R., Timothy, K. W. & Golden, A. Update on the molecular genetics of Timothy syndrome. Using iPSC-derived neurons to uncover cellular phenotypes associated with Timothy syndrome. Using induced pluripotent stem cells to investigate cardiac phenotypes in Timothy syndrome. Ryan, S. G. et al. Epilepsy and mental retardation limited to females: an X-linked dominant disorder with male sparing. Identification of PCDH19 gene mutations/deletions in patients with early onset epilepsy. Gauthier, J. et al. De novo mutations in the gene encoding the synaptic scaffolding protein SHANK3 in patients ascertained for schizophrenia. De Rubeis, S. et al. Delineation of the genetic and clinical spectrum of Phelan–McDermid syndrome caused by SHANK3 point mutations. Definition and clinical variability of SHANK3-related Phelan–McDermid syndrome. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Ferraro, N. M. et al. Transcriptomic signatures across human tissues identify functional rare genetic variation. Kanton, S. et al. Organoid single-cell genomic atlas uncovers human-specific features of brain development. Han, D., Xue, X., Yan, Y. & Li, G. Dysfunctional Cav1.2 channel in Timothy syndrome, from cell to bedside. Gerosa, L., Francolini, M., Bassani, S. & Passafaro, M. The role of protocadherin 19 (PCDH19) in neurodevelopment and in the pathophysiology of early infantile epileptic encephalopathy-9 (EIEE9). Leblond, C. S. et al. Meta-analysis of SHANK mutations in autism spectrum disorders: a gradient of severity in cognitive impairments. Mossink, B., Negwer, M., Schubert, D. & Nadif Kasri, N. The emerging role of chromatin remodelers in neurodevelopmental disorders: a developmental perspective. Mohammad, L., Wiseman, J., Erickson, S. & Yang, G. in The Oxford Handbook of Neuronal Protein Synthesis (ed. Urresti, J. et al. Cortical organoids model early brain development disrupted by 16p11.2 copy number variants in autism. Ohtaka-Maruyama, C. & Okado, H. Molecular pathways underlying projection neuron production and migration during cerebral cortical development. Impaired neurodevelopmental pathways in autism spectrum disorder: a review of signaling mechanisms and crosstalk. & Geschwind, D. H. Systems biology and gene networks in neurodevelopmental and neurodegenerative disorders. Oldham, M. C. et al. Functional organization of the transcriptome in human brain. Li, M. et al. Integrative functional genomic analysis of human brain development and neuropsychiatric risks. Adhya, D. et al. Atypical neurogenesis in induced pluripotent stem cells from autistic individuals. Flaherty, E. et al. Neuronal impact of patient-specific aberrant NRXN1alpha splicing. Lin, M. et al. Integrative transcriptome network analysis of iPSC-derived neurons from schizophrenia and schizoaffective disorder patients with 22q11.2 deletion. Schafer, S. T. et al. Pathological priming causes developmental gene network heterochronicity in autistic subject-derived neurons. Trevino, A. E. et al. Chromatin accessibility dynamics in a model of human forebrain development. Liao, C. et al. Convergent coexpression of autism-associated genes suggests some novel risk genes may not be detectable in large-scale genetic studies. Verfaillie, A., Imrichova, H., Janky, R. & Aerts, S. iRegulon and i-cisTarget: reconstructing regulatory networks using motif and track enrichment. Langfelder, P. & Horvath, S. WGCNA: an R package for weighted correlation network analysis. Fundamental elements in autism: from neurogenesis and neurite growth to synaptic plasticity. Protein interaction studies in human induced neurons indicate convergent biology underlying autism spectrum disorders. Leblond, C. S. et al. Operative list of genes associated with autism and neurodevelopmental disorders based on database review. Fu, J. M. et al. Rare coding variation provides insight into the genetic architecture and phenotypic context of autism. Braun, S. M. G. et al. BAF subunit switching regulates chromatin accessibility to control cell cycle exit in the developing mammalian cortex. & Crabtree, G. R. The role of BAF (mSWI/SNF) complexes in mammalian neural development. The BAF complex interacts with Pax6 in adult neural progenitors to establish a neurogenic cross-regulatory transcriptional network. Tuoc, T. C. et al. Chromatin regulation by BAF170 controls cerebral cortical size and thickness. Jin, Y. et al. Loss of BAF (mSWI/SNF) chromatin-remodeling ATPase Brg1 causes multiple malformations of cortical development in mice. Acute depletion of the ARID1A subunit of SWI/SNF complexes reveals distinct pathways for activation and repression of transcription. Datlinger, P. et al. Pooled CRISPR screening with single-cell transcriptome readout. Guss, E. J. et al. Protocol for neurogenin-2-mediated induction of human stem cell-derived neural progenitor cells. Tumor suppressor p53 restrains cancer cell dissemination by modulating mitochondrial dynamics. Wagner, G. P., Booth, G. & Bagheri-Chaichian, H. A population genetic theory of canalization. Schork, A. J. et al. A genome-wide association study of shared risk across psychiatric disorders implicates gene regulation during fetal neurodevelopment. Genomic relationships, novel loci, and pleiotropic mechanisms across eight psychiatric disorders. Eyring, K. W. & Geschwind, D. H. Three decades of ASD genetics: building a foundation for neurobiological understanding and treatment. Bourgeron, T. From the genetic architecture to synaptic plasticity in autism spectrum disorder. Miura, Y. et al. Generation of human striatal organoids and cortico-striatal assembloids from human pluripotent stem cells. Cell stress in cortical organoids impairs molecular subtype specification. Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. & Dewey, C. N. RSEM: accurate transcript quantification from RNA-Seq data with or without a reference genome. Oldham, M. C., Langfelder, P. & Horvath, S. Network methods for describing sample relationships in genomic datasets: application to Huntington's disease. The Genome Analysis Toolkit: a MapReduce framework for analyzing next-generation DNA sequencing data. Purcell, S. et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. International HapMap, C. et al. Integrating common and rare genetic variation in diverse human populations. Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler transform. Barnett, D. W., Garrison, E. K., Quinlan, A. R., Stromberg, M. P. & Marth, G. T. BamTools: a C + + API and toolkit for analyzing and managing BAM files. Rausch, T. et al. DELLY: structural variant discovery by integrated paired-end and split-read analysis. Robinson, J. T. et al. Integrative genomics viewer. Danecek, P. et al. Twelve years of SAMtools and BCFtools. A scaling normalization method for differential expression analysis of RNA-seq data. & Smyth, G. K. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Hoffman, G. E. & Roussos, P. Dream: powerful differential expression analysis for repeated measures designs. Viechtbauer, W. Conducting meta-analyses in R with the metafor package. Independent component analysis uncovers the landscape of the bladder tumor transcriptome and reveals insights into luminal and basal subtypes. Korotkevich, G., Sukhov, V. & Sergushichev, A. Fast gene set enrichment analysis. Preprint at bioRxiv https://doi.org/10.1101/060012 (2019). Removing technical variability in RNA-seq data using conditional quantile normalization. Yu, G., Wang, L. G., Han, Y. & He, Q. Y. clusterProfiler: an R package for comparing biological themes among gene clusters. Pardinas, A. F. et al. Common schizophrenia alleles are enriched in mutation-intolerant genes and in regions under strong background selection. Discovery of the first genome-wide significant risk loci for attention deficit/hyperactivity disorder. Howard, D. M. et al. Genome-wide meta-analysis of depression identifies 102 independent variants and highlights the importance of the prefrontal brain regions. Mullins, N. et al. Genome-wide association study of more than 40,000 bipolar disorder cases provides new insights into the underlying biology. Finucane, H. K. et al. Partitioning heritability by functional annotation using genome-wide association summary statistics. Skene, N. G. & Grant, S. G. Identification of vulnerable cell types in major brain disorders using single cell transcriptomes and expression weighted cell type enrichment. Gu, Z., Eils, R. & Schlesner, M. Complex heatmaps reveal patterns and correlations in multidimensional genomic data. Suzuki, R. & Shimodaira, H. Pvclust: an R package for assessing the uncertainty in hierarchical clustering. Galili, T. dendextend: an R package for visualizing, adjusting and comparing trees of hierarchical clustering. Comparison of hierarchical cluster analysis methods by cophenetic correlation. Vavrek, M. J. fossil: palaeoecological and palaeogeographical analysis tools. Rand, W. M. Objective criteria for the evaluation of clustering methods. Accurate estimation of cell composition in bulk expression through robust integration of single-cell information. Aibar, S. et al. SCENIC: single-cell regulatory network inference and clustering. Rossin, E. J. et al. Proteins encoded in genomic regions associated with immune-mediated disease physically interact and suggest underlying biology. Combining NGN2 programming with developmental patterning generates human excitatory neurons with NMDAR-mediated synaptic transmission. Pintacuda, G. et al. Genoppi is an open-source software for robust and standardized integration of proteomic and genetic data. Eng, J. K., McCormack, A. L. & Yates, J. R. An approach to correlate tandem mass spectral data of peptides with amino acid sequences in a protein database. Gasperini, M. et al. A genome-wide framework for mapping gene regulation via cellular genetic screens. Hill, A. J. et al. On the design of CRISPR-based single-cell molecular screens. Gordon, A. et al. Code for ‘Developmental convergence and divergence in human stem cell models of autism'. & Horvath, S. Defining clusters from a hierarchical cluster tree: the Dynamic Tree Cut package for R. Bioinformatics 24, 719–720 (2008). This work was supported by several grants from the National Institutes of Health, including grant nos. ), and the National Institute of Mental Health Convergent Neuroscience Consortium (grant no. ); by grants from the California Institute of Regenerative Medicine (to D.H.G. ); the Stanford Human Brain Organogenesis Program in the Wu Tsai Neuroscience Institute (to S.P.P. ); the Stanford Wu Tsai Neuroscience Institute Big Idea Grant (to S.P.P. ); the Kwan Funds (to S.P.P. ); the Senkut Research Fund (to S.P.P. ); the Autism Science Foundation (ASF) and the Brain and Behavior Research Foundation Young Investigator award (Brain & Behavior Research Foundation) (to A.G.). is a New York Stem Cell Foundation (NYSCF) Robertson Stem Cell Investigator and a Chan Zuckerberg Initiative (CZI) Ben Barres Investigator. was supported by the AP Giannini Foundation. was supported by the Stanley Center for Psychiatric Research, the Lundbeck Foundation (grant no. R350-2020-963), and the Novo Nordisk Foundation (grant no. is supported by the Simons Foundation Autism Research Initiative Fellows-to-Faculty Award (00002804). These authors contributed equally: Aaron Gordon, Se-Jin Yoon, Lucy K. Bicks Program in Neurogenetics, Department of Neurology, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, USA Aaron Gordon, Lucy K. Bicks, Jacqueline M. Martín, Stephanie Arteaga, Brie Wamsley, Qiuyu Guo & Daniel H. Geschwind Department of Psychiatry and Behavioral Sciences, Stanford University, Stanford, CA, USA Se-Jin Yoon, Lubayna Elahi, Ruth O'Hara, Joachim F. Hallmayer & Sergiu P. Pasca Stanford Brain Organogenesis Program, Wu Tsai Neurosciences Institute and Bio-X, Stanford University, Stanford, CA, USA Se-Jin Yoon, Lubayna Elahi & Sergiu P. Pasca Stanley Center for Psychiatric Research, Broad Institute, Cambridge, MA, USA Greta Pintacuda & Kasper Lage Tempero Bio, Boston, MA, USA Novo Nordisk Foundation Center for Genomic Mechanisms of Disease, Broad Institute of MIT and Harvard, Cambridge, MA, USA Department of Human Genetics, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, USA Program in Neurobehavioral Genetics, Semel Institute, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, USA Center for Autism Research and Treatment, Semel Institute, David Geffen School of Medicine, University of California Los Angeles, Los Angeles, CA, USA Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar Search author on:PubMed Google Scholar conceived, designed, supervised and provided funding for this work and the writing of the paper. carried out experiments, analysis and writing. designed, carried out and analysed CRISPRi experiments. carried out and provided data for IP–MS experiments and contributed to analysis. aided in participant recruitment and sample collection. All authors provided critical feedback in writing the paper. Correspondence to Sergiu P. Pasca or Daniel H. Geschwind. Stanford University holds patents for the generation of cortical organoids/spheroids (listing S.P.P. holds equity in Sidera and serves as a consultant for Sidera and ZS Associates and has equity in and serves as a consultant to Quantum Cell. Nature thanks Kristen Brennand and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Publisher's note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. (a) Schematic representation of hCO differentiations from each derived hiPS cell line from each individual. (b) Schema of dropped lines due to quality control metrics. Whole genome sequencing depth in (a) 16p11.2 (b) 15q13.3 (c) 22q11.2 and (d) 22q13.3 loci averaged over 10 kb windows. (e) Mutational forms not shown in main figure showing genes inside the CNV locus or genes carrying the point mutations (marked in blue). Genes significantly differentially expressed over all time points are denoted by asterisks (dreamlet p-values of p < 0.005). The percent of DE genes in the CNV across the 4 time points is as follows: 22q13del - 77.3% of genes, 22q11del - 88.9% of genes. (a) Point mutations in CACNA1C (G406R) identified in the 3 lines from individuals with Timothy syndrome. (b) Point mutations in PCDH19 (Y243* and Q201P) identified in the 2 lines from individuals with PCDH19 syndrome. (c) Point mutations in SHANK3 (R522W) identified in the 2 lines from the individual with SHANK3 mutation. All mutations were visualized with the Integrative Genomics Viewer (IGV). (a) Hierarchical clustering of all samples based on gene expression showing that samples cluster by differentiation day and the genetic form of ASD. Annotation colour bars represent day (shades of blue) and form (colourful). (b) Association of the top 20 gene expression principal components (PC) with the different covariates. Numbers in brackets on the y-axis are the percent of variance explained by these PCs. Differentiation time is highly associated with the 1st PC (Adjusted R2 = 0.92). (c) Variance explained by each of the covariates for each gene (n = 17,963 genes). Distributions represent the density of the percent of variance explained. The median value of percent of explained variance for each variable are denoted below the plots. Boxplots in c and e show: centre, median, lower hinge – 25% quantile, upper hinge – 75% quantile, lower whisker – smallest observation greater than or equal to lower hinge –1.5× interquartile range, upper whisker – largest observation less than or equal to upper hinge +1.5× interquartile range. (d) Scatter plot of the first two gene expression PCs from each form of ASD. Colour represents the differentiation day (top) or form (bottom). PC1 explains 24% of the variation and is highly associated with differentiation day. PC2 explains 7.9% of the variation. Sample numbers: Day 25: Controls – 46, 15q13.3del – 7, 16p11.2del – 5, 16p11.2dup – 4, 22q11.2del – 13, 22q13.3del – 11, Idiopathic – 12, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 3. Day 50: Controls – 53, 15q13.3del – 6, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 15, 22q13.3del – 12, Idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 3. Day 75: Controls – 54, 15q13.3del – 7, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 23, 22q13.3del – 12, Idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 2. Day 100: Controls – 50, 15q13.3del – 6, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 15, 22q13.3del – 12, Idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 2. (e) Correlation of gene expression with control samples and forms (Affected) shows no difference in within-control and within-form correlation across development. Correlation by age shows a slight effect (Anova Eta2 = 0.04) with a slight decrease in correlation across development. Total number of pair-wise correlations: Day 25: Control – 1035, Affected – 241, Day 50: Control – 1378, Affected – 317, Day 75: Control – 1431, Affected – 469, Day 100: Control – 1225, Affected – 315. (f) Pseudotime testing of forms using linear mixed model and contrasts with Wald test shows no significant FDR corrected differences in pseudotime (PC-1). (a) Cophenetic correlation and Rand index across multiple clustering and distance methods (left) and gene subsampling (right). Cophenetic correlation measures how faithfully the pairwise distances of the unmodeled data are preserved in the dendrogram. Rand index measures the similarity between two different clusterings of the same data. Number of individual correlations plotted: Clustering methods – 190. Boxplots show: centre, median, lower hinge – 25% quantile, upper hinge – 75% quantile, lower whisker – smallest observation greater than or equal to lower hinge –1.5× interquartile range, upper whisker – largest observation less than or equal to upper hinge +1.5× interquartile range. (b) Cluster significance using multiscale bootstrap resampling. P values are indicated in red for each branching point. Bootstrap probability measures the frequency that a cluster appears in bootstrap replicates. (a) Overlaps between differentially expressed genes at all time points and risk genes for ASD from either SFARI or from large scale whole exome (WES) sequencing studies2,8 as well as with neurodevelopmental disorders (NDDs) and intellectual disability (ID) risk genes91. Only positive significant overlaps (OR > 1 and FDR < 0.05) are shown. (b) Select gene ontology (GO) terms enriched in upregulated (red) or downregulated (blue) genes in each of the ASD forms at day 25 (top) or day 100 (bottom). Statistics derived from GSEA. (c) Distribution of normalized expression for meta analysis significant example genes at day 100. Boxplots show: centre, median, lower hinge – 25% quantile, upper hinge – 75% quantile, lower whisker – smallest observation greater than or equal to lower hinge –1.5× interquartile range, upper whisker – largest observation less than or equal to upper hinge +1.5× interquartile range. Number of samples (differentiations) for each form: Control – 50, 15q13.3del – 6, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 15, 22q13.3del – 12, Idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 2. (d) Select GO terms enriched in upregulated or downregulated meta-analysis significant genes Colour corresponds to normalized enrichment score (NES). Point size corresponds to the level of significance (−log10(FDR)). Statistics derived from GSEA. (a) Clustering of d25 organoid samples (left) and Clusters identified using samples from Khan et al. 202044 (right). For Khan et al. samples, cluster characterization was based on control samples only. (b) Selected markers for Cycling progenitor (MKI67, PIF1; CycPro), Radial glia (PAX6, SOX2, PTN, SAT1; RG-1, RG-2, and RG-3), intermediate progenitors (EOMES, NHLH1; IPC), Interneurons (GAD1, DLX5; InNeu) and Excitatory neurons (TBR1, BCL11B; ExNeu-1 and ExNeu-2). (a) Cell proportion in each ASD form across hCO differentiation. Significant changes are marked by asterisk. CycPro - cycling progenitor, RG-1 - radial glia 1 (early), RG-2 - radial glia 2 (late), IPC - intermediate progenitor cells, ORG - outer radial glia, ExNeu-1 - excitatory neurons 1 (early), ExNeu-2 - excitatory neurons 2 (late), InNeu - interneurons. Grey bar indicates proportion of cell type in reference used to deconvolute. Stastics were derived from linear mixed-effects models and contrasts comparing each form to controls with two-sided Wald tests, *FDR < 0.05. Sample Numbers: Day 25: Controls – 46, 15q13.3del – 7, 16p11.2del – 5, 16p11.2dup – 4, 22q11.2del – 13, 22q13.3del – 11, Idiopathic – 12, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 3. Day 50: Controls – 53, 15q13.3del – 6, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 15, 22q13.3del – 12, Idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 3. Day 75: Controls – 54, 15q13.3del – 7, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 23, 22q13.3del – 12, Idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 2. Day 100: Controls – 50, 15q13.3del – 6, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 15, 22q13.3del – 12, Idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 2. (b) Correlation of deconvoluted celltype beta across time points. Day 100 cell type deconvolution effect is more correlated than day 25 (p-adj = 0.01). Stastics derived ANOVA followed by TukeyHSD post-hoc tests. (c) Normalized expression of IPC marker genes at day 50. Number of samples for each form: Control – 53, 15q13.3del – 6, 16p11.2del – 6, 16p11.2dup – 4, 22q11.2del – 15, 22q13.3del – 12, idiopathic – 15, PCDH19 – 2, SHANK3 – 2, Timothy Syndrome – 3. Boxplots a-c show: centre, median, lower hinge – 25% quantile, upper hinge – 75% quantile, lower whisker – smallest observation greater than or equal to lower hinge –1.5× interquartile range, upper whisker – largest observation less than or equal to upper hinge +1.5× interquartile range. (a) Association of independent components (IC) with the different forms of ASD across time points. Stastics were derived from linear mixed-effects models and contrasts comparing each form to controls with two-sided Wald tests followed by FDR correction. (b) Select gene ontology (GO) terms enriched for the positively or negatively contributing genes for the indicated ICs. Negatively enriched terms are downregulated in ICs that are positively associated with forms of ASD and are upregulated in ICs that are negatively associated with forms of ASD. Stastistics were derived from GSEA. (c) Developmental expression of independent components (IC) for each ASD form. IC trajectory significance was tested using a linear model with polynomial splines with 2 degrees of freedom and 2 degrees of the piecewise polynomial. (a) Dendrogram from which the WGCNA modules at day 25 were derived using the dynamic tree cut algorithm150. Genes associations (adjusted R2) with covariates is also indicated. (b) Preservation of day 25 modules at other time points (left), in vivo (8-16 weeks post conception)80 and in other stem cell derived neuronal data sets: Adhya202081, Flahrety201982, Lin201683 and Schafer 201984. The Zsummary preservation score is indicated. (c) Number of TF-regulators (significantly enriched as regulators for at least 1 downstream module) within each module. (d) Number of TF-regulators within each module (x-axis) significantly regulating each downstream module (bar graph). (a) Module eigengene trajectories of day 25 modules across differentiation. Significance was tested using a linear model with polynomial splines with 2 degrees of freedom and 2 degrees of the piecewise polynomial. (b) Module eigengene trajectories of the regulators in M5 (blue) and the ASD risk genes that are regulated by these regulators (grey). (a) Volcano plots show logFC of proteins compared to IGG controls. Highlighted are SFARI genes (1,2,S). Statistics were derived from 3 replicates using a two-tailed one-sample moderated t-test to calculate P values and FDR. (b) Significance of connectedness of networks is determined by comparing the network mean logFC to 10000 randomly selected networks of the same size across each of the protein baits. Red dotted line shows the network mean logFC and black line shows the 95th percentile. (c-d) Connectivity networks from gene lists of interest. Size of circle represents number of significant proteins for each bait. d) NDD and ASD risk genes2,8,91. (a) Schema of CROP-Seq vector. (b) Example of gating for flow cytometry. Left shows a set of hNPCs not expressing virus used for gating compared with hNPCs expressing both dCas9-Krab (GFP) and Crop-Seq-opti-dsRed (TdTomato). Box in the upper right shows double positive cells selected for downstream analysis (~20% of live cell fraction). (c) QC measures for CRISPRi single cell libraries organized by gRNA negative (no gRNA detected in sequencing), Control gRNA positive, or Target gRNA positive. (d) Number of gRNA UMIs per cell. X-axis shows cut-off of 10 UMIs used to consider a cell gRNA positive (left). Number of of unique gRNAs expressed in each cell (x-axis). Only cells expressing gRNA for a single gene-target are retained for downstream analyses. (e) Single cell UMAPs of markers used to differentiate cycling and non-cyling NPCs (f) Single cell UMAP coloured by Target-gene presence (g) Proportion of cells uniquely expressing gRNAs barcodes for each target within each experiment (n = 18 per Gene comprised of 6 technical replicates, 3 gRNAs barcodes per target, for Controls n = 108: 18 gRNAs across 6 technical replicates). TP53 show increased proportion compared to non-targeting controls (NTCs). Boxplots a-c show: centre, median, lower hinge – 25% quantile, upper hinge – 75% quantile, lower whisker – smallest observation greater than or equal to lower hinge –1.5× interquartile range, upper whisker – largest observation less than or equal to upper hinge +1.5× interquartile range. ***Family wise corrected p < 0.005, Dunnets post-hoc contrast of linear model. (a) Broad positive correlation between DEGs across targets in NPCs (right) and cNPCs (left) and DEGs in patient-derived organoids across developmental stages. (b) Enrichment of rare genetic variation gene-lists in the differentially expressed genes of each target. Significance derived from two-sided Fishers exact tests. (c) Jaccard Index of DEG overlaps for each target. (d) Meta-analysis significant genes across targets show enrichment in several hCO-derived regulatory modules. Significance derived from two-sided Fishers exact tests. Gene Set Enrichment Analysis (GSEA) shows down-regulation of neuronal-related terms across M5 transcription factor regulators, particularly in RBBP5, SOX9, and EP300. Statstics were derived from fgsea R package. 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/. et al. Developmental convergence and divergence in human stem cell models of autism. Version of record: 29 January 2026 Anyone you share the following link with will be able to read this content: Sorry, a shareable link is not currently available for this article. Provided by the Springer Nature SharedIt content-sharing initiative Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). Rachel Fieldhouse is a reporter for Nature in Sydney, Australia. Posts on social-media platform X that are critical of scientific research can act as early warning signs of problematic articles, according to two large studies. The findings reflect how post-publication commentary can help to identify errors or fraudulent results, say scientists. Previous research has shown that potentially problematic articles get substantial attention on social media before being formally retracted1, and receive more attention than do similar articles that do not get retracted2. Er-Te Zheng, a PhD student studying computational social science at the University of Sheffield, UK, wanted to find out whether social-media platforms such as X could be used to identify integrity issues in articles. Zheng and his colleagues defined critical tweets as posts that contained sarcasm, criticism, accusations or doubt about the article. The researchers found that negative sentiment in tweets and the inclusion of any of 95 ‘red flag' words — such as fraud, retract, hoax or flawed — were associated with an increased risk that the paper would be retracted. Sotudeh and her team found that tweets containing red-flag words were associated with papers being retracted faster than those that had posts without these terms. The team noted, however, that they could not show causality between negative X posts and retractions. These studies shows that social media platforms are an important place for post-publication commentary, says Virginia Barbour, co-chair of the Declaration on Research Assessment initiative, which aims to reform how scientific outputs are evaluated, and a researcher in academic publishing at the Queensland University of Technology in Brisbane, Australia. Post-publication commentary is needed because conventional peer review can miss errors or fraud, says Barbour. Sotudeh says researchers and publishers are paying more attention to commentary online. But there is no guidance yet for how publishers can translate social media critiques into evidence-based editorial decisions. “If publishers are to give more weight to such discourse, they will need systems for screening, verifying, and contextualizing claims before taking action,” Sotudeh adds. How to speed-up peer review: make applicants mark one another The early days of peer review: five insights from historical reports ArXiv says submissions must be in English: are AI translators up for the job? Wikipedia is needed now more than ever, 25 years on ‘A serious problem': peer reviews created using AI can avoid detection More than half of researchers now use AI for peer review — often against guidance ArXiv says submissions must be in English: are AI translators up for the job? Shanghai, Beijing, Nanjing, Pune or New Delhi – hybrid working model How to speed-up peer review: make applicants mark one another The early days of peer review: five insights from historical reports An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday. Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.