Spatial architecture of autism pathogenesis reveals mosaic structural disarray during early development Authors: Li Lin, Tzuen Yih Saw, Nigel Chou, Jie Lin Jolene Goh, Jing Eugene Kwa, Wan Kee Chock, Vipul Singhal, Zheng Li, Mike J. Huang, Huck Hui Ng, Chiea Chuen Khor, Hwee Lee Kuan, Kok Hao Chen, Shyam Prabhakar, Jinyue Liu Published: 2026 ABSTRACT Autism spectrum disorder (ASD) has been associated with diverse genetic factors and molecular changes. Yet, how pathophysiology emerges during development and contributes to clinical heterogeneity remains unclear. Here we use spatial and single-cell transcriptomics of genetically diverse patient-derived organoids of the prenatal cortex to map the spatial architecture of ASD pathogenesis. We find abnormal partitioning of progenitor and neuron zones alongside local patches of disorganized neurons that vary in distribution among patients. Such spatially mosaic disarray persists into neuronal maturation and is consistent with impaired adhesion between progenitors. Our findings suggest that the spatial landscape of biological processes leading to ASD may be more heterogeneous than previously thought. We propose a model in which different patterns of scattered abnormalities, arising from spatially mosaic pathogenesis during prenatal brain development, contribute to the wide range of symptoms and brain structure variations observed in ASD patients. Introduction Autism spectrum disorder (ASD) is a complex lifelong condition marked by repetitive behaviors, restricted interests, and challenges in social communication1. It is known as a spectrum disorder because its core symptoms vary widely in both presentation and severity among affected individuals. Despite being heritable, ASD is highly polygenic, with associated mutations spanning over 1000 genes2. Currently, 75% of all cases remain idiopathic3. Such clinical and etiological heterogeneity poses significant challenges in defining the pathogenic processes underlying ASD. Consequently, how ASD pathophysiology emerges during neural development remains unclear. One major hypothesis — the convergence hypothesis — posits that ASD individuals across the spectrum may share common underlying mechanisms despite diverse etiologies and clinical presentations. Indeed, known ASD risk genes are co-expressed during early-mid fetal and postnatal brain development, indicating shared molecular pathways4,5. However, mutations in these genes account for at most 20% of the entire ASD cohort6. Structurally, the cortex is a major site of pathology and displays abnormal regional patterning and disorganized patches that vary in the cortical layers and cell types affected7–11. Pathogenic convergence over space and time thus remains to be clarified. Genetic studies, architectural abnormalities in patients’ brains, and environmental risk factors such as maternal immune activation suggest that ASD pathogenesis begins in utero7,12,13. Since diagnosis can only be made postnatally, it is not possible to directly investigate early pathophysiological events in primary human fetal tissue. Postmortem brain tissues, on the other hand, display endpoint phenotypes that are extensively confounded by life history7,14–16. Thus, patient-derived organoid models offer a promising approach for studying ASD pathophysiology. They recapitulate essential features of prenatal human brain development while retaining each patient’s genetic load. In recent years, bulk and single-cell approaches on ASD organoids have revealed cell proportion changes such as overproduction of GABAergic neurons, depletion of intermediate progenitors, and imbalance of cortical neuron subtypes17–20. These alterations to cell composition implicated excitation-inhibition imbalance and asynchronous development of neuronal classes, often in a manner that is dependent on the individual’s genomic context17,19,20. However, no studies have addressed whether the spatial organization of the developing neural tissue is impacted, despite its critical role in circuit assembly and brain function. Furthermore, ASD organoid studies to date either focus on a small subset of ASD that is associated with macrocephaly or interrogate specific high-risk genes17–20. Together, they account for the most extreme forms of ASD that constitute at most 20% of all cases6. Since ASD is clinically and etiologically diverse, a diverse patient pool for organoid generation is essential for gaining representative and generalizable insights. Here, we address these unmet needs by generating brain organoids from neurotypical controls and a genetically heterogeneous patient cohort, and profiling these using spatial and single-cell transcriptomics and whole-exome sequencing (Fig. 1a). These data enabled a detailed investigation into the spatial architecture of ASD pathogenesis. We found that organoids derived from different patients varied in their extent of structural disarray and developmental timing. Our findings indicate that the spatial landscape of pathogenic processes in ASD is more heterogeneous than previously thought. We observed that pathogenesis emerges in variegated spatial patterns that, together with developmental mistiming, may give rise to the scattered distribution of disorganized patches seen in the postmortem cortex of ASD persons7–10. Such spatially mosaic pathogenesis may help explain how molecular convergence among ASD-associated genetic factors could lead to the diverse symptomology seen clinically.Fig. 1Single-cell RNA sequencing uncovered altered gene programs in ASD organoids.a Study design. Day 56 cortical organoids derived from 10 neurotypical control and 14 ASD individuals were profiled by single-cell RNA sequencing and imaging-based spatial transcriptomics (MERFISH). b Schematic illustrating tissue organization in each neuroepithelial bud within a human cortical organoid, representing typical architecture and key cell identities in the developing human cortex. c Left, UMAP representation of single-cell transcriptomes of progenitors from single-cell RNA sequencing of day 56 cortical organoids. Right, dot plot of the average single-cell RNA sequencing expression (dot color) for each marker gene per progenitor identity, and the fraction (dot size) of cells in which it was expressed. d Same as c, but for neurons. e Cell type proportions in brain organoids derived from each individual. Color per cell type is as indicated on the UMAPs in Fig. 1c, d. Source data are provided as a Source Data file. f UMAP representation of the odds of finding an ASD cell among 30 nearest neighbors in the transcriptomic space of neurons. UMAP regions enriched with ASD cells are magenta, while regions enriched with control cells would be green. Mid-color i.e., white, represents equal contributions of ASD and control cells. g Dysregulated GO processes in ASD organoids. Dot plot showing the top 5 significantly activated (NES > 0) and suppressed (NES < 0) pathways in each cell identity, as determined by Gene Set Enrichment Analysis (GSEA). NES, normalized enrichment score; padj, Benjamini-Hochberg adjusted p values, DESeq2 Wald test. Only padj ≤ 0.05 are shown. If significant GO terms share ≥70% gene set overlap, only the most significant GO term is shown. Red font indicates cell adhesion-related terms in RG cell types. Single-cell RNA sequencing uncovered altered gene programs in ASD organoids. a Study design. Day 56 cortical organoids derived from 10 neurotypical control and 14 ASD individuals were profiled by single-cell RNA sequencing and imaging-based spatial transcriptomics (MERFISH). b Schematic illustrating tissue organization in each neuroepithelial bud within a human cortical organoid, representing typical architecture and key cell identities in the developing human cortex. c Left, UMAP representation of single-cell transcriptomes of progenitors from single-cell RNA sequencing of day 56 cortical organoids. Right, dot plot of the average single-cell RNA sequencing expression (dot color) for each marker gene per progenitor identity, and the fraction (dot size) of cells in which it was expressed. d Same as c, but for neurons. e Cell type proportions in brain organoids derived from each individual. Color per cell type is as indicated on the UMAPs in Fig. 1c, d. Source data are provided as a Source Data file. f UMAP representation of the odds of finding an ASD cell among 30 nearest neighbors in the transcriptomic space of neurons. UMAP regions enriched with ASD cells are magenta, while regions enriched with control cells would be green. Mid-color i.e., white, represents equal contributions of ASD and control cells. g Dysregulated GO processes in ASD organoids. Dot plot showing the top 5 significantly activated (NES > 0) and suppressed (NES < 0) pathways in each cell identity, as determined by Gene Set Enrichment Analysis (GSEA). NES, normalized enrichment score; padj, Benjamini-Hochberg adjusted p values, DESeq2 Wald test. Only padj ≤ 0.05 are shown. If significant GO terms share ≥70% gene set overlap, only the most significant GO term is shown. Red font indicates cell adhesion-related terms in RG cell types. Results Clinical cohort characteristics We obtained induced pluripotent stem cells (iPSCs) for 10 neurotypical controls and 14 ASD individuals from 13 families (Supplementary Table 1). Among patients, 9 had a family history of ASD and 5 presented dysmorphologies. None had epilepsy, seizures, cerebral palsy, or a family history of these conditions. Both sexes were represented in the patient cohort (5 female, 9 male). One patient carries a de novo 15q11-13 duplication, a copy number variant associated with ASD. We further identified by exome sequencing 72 candidate genetic variants that are predicted to impair protein function in 10 patients. Only one variant, a stop-gain in GTF2I (7:74114630:C:T), was found in SFARI Gene (a database of ASD-linked genes and variants2) (Supplementary Table 1, Type 1). Other variants either matched a gene in the SFARI database but not the specific variant(s) previously reported to be found in persons with ASD (Supplementary Table 1, Type 2), or matched neither (Supplementary Table 1, Type 3). There was no overlap in candidate genetic variants across the patient cohort. Our patient cohort thus reflects diverse genetic etiology. Organoids from ASD individuals exhibit altered gene programs Since the cortex is a major site of ASD pathology11, we differentiated each iPSC line into organoids pre-patterned towards a cortical fate21 (Supplementary Fig. 1a). We validated their cortical identity by immunostaining for the forebrain marker FOXG122 and confirmed the emergence of BCL11B-expressing neurons by day 56 as reported previously21 (Supplementary Fig. 1b). To gain insight into the cellular makeup and transcriptomic landscape of our ASD organoids, we performed single-cell RNA sequencing on day 56, pooling three organoids per individual to mitigate inter-organoid variability (Fig. 1a). A total of 144,247 cells were profiled, (Supplementary Fig. 1c; Supplementary Table 3, tab 1) of which 90.5 ± 2.29% expressed FOXG1, further confirming that the organoids were predominantly cortical (Supplementary Fig. 1d, e). We also benchmarked our single-cell transcriptomes against the BrainSpan Atlas’ developmental transcriptome23,24 and confirmed that they best resembled human neocortex at 12-13 post-conception weeks (pcw) (Supplementary Fig. 1f). To resolve cellular subtypes within each major cell class, we subclustered progenitors and neurons separately (Fig. 1c, d, Supplementary Fig 1g) and recapitulated key cell identities of the human fetal cortex25 (Fig. 1b), many of which were enriched for ASD-linked genes (Supplementary Fig. 1h, adjusted p-value < 0.001, Fisher’s exact test; Supplementary Table 2). Among progenitors were radial glia (RG) marked by PAX6, SOX2 and SLC1A3, as well as intermediate progenitors (IP) expressing EOMES and HES625–27 (Fig. 1c). Excitatory neuron (ExN) subpopulations included transient identities such as Cajal-Retzius, pioneer ExN and subplate ExN, marked by RELN, CALB2 and LMO3 respectively28–31 (Fig. 1d). We further identified transitional neuronal states including migrating ExN marked by DPY19L1 and maturing ExN marked by NEUROD625,32 (Fig. 1d). Deep-layer ExN, which form cortical layers V/VI as early as 12pcw33, were also present and expressed TBR1, BCL11B and SOX525,27. Certain neuronal clusters shared gene signatures for the same identity but were segregated in transcriptomic UMAP space (Fig. 1d). We numbered such clusters in each identity to differentiate them, yielding deep-layer ExN1, deep-layer ExN2, migrating ExN1, and migrating ExN2. To understand the broader changes occurring in our ASD organoids, we compared cell type composition across lines. All major cell identities were present in every organoid line (Fig. 1e). Consistent with previous reports, committed IPs were diminished18 while interneurons were enriched in ASD organoids17,20 (Fig. 1e and Supplementary Fig. 1i; p = 0.077 and 0.0099, respectively, two-sample Kolmogorov-Smirnov test). Deep-layer ExN2 and migrating ExN2 were detected in organoids from 5 of 13 ASD individuals and rarely in controls (Fig. 1d–f, and Supplementary Fig. 1i; p = 0.024 for both). These ExN2 states were observed across three independent differentiation batches and co-occurred with ExN1 within the same differentiation batch and sequencing run (Supplementary Table 3, tab 1). The ExN2-presenting lines spanned both sexes (3 male, 2 female), included individuals with or without dysmorphologies (2 with, 3 without), and represented four distinct genetic backgrounds (family IDs: CIRM00056, CIRM00090, CIRM00082, CIRM00080) (Supplementary Fig. 1j; Supplementary Table 3, tab 1). The 5 ASD individuals that contributed to these states either carried Type 2 and/or Type 3 candidate variants (CW60171, CW60172, CW20081), or had no candidate variants identified (CW60245, CW20002) (Supplementary Table 1). The remaining 8 ASD individuals and all 10 controls were ExN1-dominant. We next compared migrating ExN2 with migrating ExN1, and deep-layer ExN2 with deep-layer ExN1, to identify transcriptomic features of the ExN2 states. Both migrating ExN2 and deep-layer ExN2 were upregulated in genes involved in histone modification, chromatin remodeling and transcriptional regulation (PTMS, CHD3, GNAS, PTMA, H1FX, C1orf61, RPL6, RPS26) (Supplementary Fig. 1k, common DEGs). Migrating ExN2 were selectively enriched for genes involved in neuronal development and organization, while deep-layer ExN2 were enriched for genes that mediate cell morphogenesis, adhesion and projection (Supplementary Fig. 1k, l). To investigate if cell states similar to ExN2 exist in vivo, we examined the expression of their enriched genes in primary human cortex from control and ASD individuals16. In postmortem ASD cortex, 65 out of 73 genes downregulated in ExN2 states presented negative fold changes while 85 out of 159 genes upregulated in ExN2 states presented positive fold changes (Supplementary Fig. 1m). Thus, most genes marking ExN2 states in organoids were also altered in postmortem ASD brains, predominantly changing in a similar direction. These results suggest that excitatory neurons with altered gene expression may persist after birth in certain ASD individuals. We further examined potential molecular changes that may occur in other cell types within our ASD organoids. Using pseudobulk differential gene expression analysis, we identified downregulated GO processes related to signal transduction, cell adhesion, neuronal differentiation, and tissue morphogenesis (Fig. 1g; Supplementary Data 1, 2). In contrast, upregulated processes were related to chromatin organization, DNA replication, and cilia assembly. Notably, dysregulation of GO processes related to cell adhesion and cytoskeleton was largely confined to radial glia populations (Fig. 1g, highlighted in red). These altered gene programs point towards changes in tissue organization during ASD pathogenesis. Neurons are improperly segregated from progenitors in ASD organoids To investigate this possibility, we mapped spatial transcriptomes in 36 tissue sections from 33 organoids covering all 24 individuals using Multiplexed Error-Robust Fluorescence In Situ Hybridization (MERFISH)34 (Fig. 2a; Supplementary Table 3, tabs 2–4). We assayed ~558 genes reported to be involved in neurodevelopment, cellular function or disease properties related to ASD2,25,26,35–37. As validation, 35 tissue sections from another 14 organoids across 8 individuals were profiled using three separate panels comprising 94 to 126 genes from the 558-gene panel (Supplementary Table 3, tabs 2–4). Genes shared across panels showed similar spatial distributions, such as enrichment at the lumen (ASPM), inner-radial regions (SOX2, NES, EOMES) or outer-radial regions (BCL11B, TBR1) (Supplementary Fig. 2a). We further benchmarked MERFISH RNA counts against sample-matched bulk RNA sequencing and obtained good correlation (r = 0.639 to 0.765, Supplementary Fig. 2b). In total, 68 million RNA molecules in 1.3 million cells across 71 tissue sections from 47 organoids were assayed.Fig. 2MERFISH showed abnormal zonation of progenitors and neurons in ASD organoids.RNA spot locations for 6 out of 558 genes assayed by MERFISH in a whole tissue section of a control (a) and an ASD (b) cortical organoid. Each colored spot represents a RNA molecule for MKI67 (orange), SFRP1 (green), EOMES (cyan), NHLH1 (blue), SOX11 (pink) or STMN2 (magenta). Scale: 500um on whole section, 100um on inset. 16 sections from 14 organoids of 10 control individuals and 20 sections from 19 organoids of 14 individuals were assayed with the same 558-gene panel (see Supplementary Table 3, tab 4). c Progenitor and neuron zones identified with spatially aware clustering in which the zone identity of a cell is determined by the average gene expression of its physical neighbors (see Methods). CW60040 is a control line, while the rest are ASD lines. d Equation for the calculation of the solidity of progenitor zones in each organoid section. Elliptical zones have high solidity, while zones with irregular boundaries have low solidity. e Solidity of progenitor zones in control (n = 88 zones from 13 organoids of 10 control individuals) and ASD organoids (n = 115 zones from 16 organoids of 11 ASD individuals) (p-value = 0.038, F(20,182) = 1.978, nested two-way ANOVA). Each dot represents an individual. Each box plot ranges from the 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Source data are provided as a Source Data file. MERFISH showed abnormal zonation of progenitors and neurons in ASD organoids. RNA spot locations for 6 out of 558 genes assayed by MERFISH in a whole tissue section of a control (a) and an ASD (b) cortical organoid. Each colored spot represents a RNA molecule for MKI67 (orange), SFRP1 (green), EOMES (cyan), NHLH1 (blue), SOX11 (pink) or STMN2 (magenta). Scale: 500um on whole section, 100um on inset. 16 sections from 14 organoids of 10 control individuals and 20 sections from 19 organoids of 14 individuals were assayed with the same 558-gene panel (see Supplementary Table 3, tab 4). c Progenitor and neuron zones identified with spatially aware clustering in which the zone identity of a cell is determined by the average gene expression of its physical neighbors (see Methods). CW60040 is a control line, while the rest are ASD lines. d Equation for the calculation of the solidity of progenitor zones in each organoid section. Elliptical zones have high solidity, while zones with irregular boundaries have low solidity. e Solidity of progenitor zones in control (n = 88 zones from 13 organoids of 10 control individuals) and ASD organoids (n = 115 zones from 16 organoids of 11 ASD individuals) (p-value = 0.038, F(20,182) = 1.978, nested two-way ANOVA). Each dot represents an individual. Each box plot ranges from the 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Source data are provided as a Source Data file. To evaluate reproducibility of our MERFISH data, we profiled 2-5 organoids per individual for 11 individuals (Supplementary Table 3, tab 4) and found RNA counts between different organoids from the same individual to highly correlate (r = 0.998, Supplementary Fig. 2c). We further profiled 2 tissue sections per organoid for 6 individuals (Supplementary Table 3, tab 4) and found that RNA counts between different tissue sections from the same organoid also highly correlate (r = 0.998, Supplementary Fig. 2c). These analyses confirmed that RNA counts obtained by MERFISH were highly reproducible. We thus performed all subsequent analyses on tissues profiled by the ~558-gene panel, as this panel contained the largest number of relevant genes. In control organoids, markers for dividing cells (MKI67), RG (SFRP1), IP (EOMES) and various ExN types (NHLH1, SOX11, STMN2) displayed radial spatial localization in each neuroepithelial bud (Fig. 2a), consistent with known anatomy (Fig. 1b). In contrast, spatial distribution for the same set of markers appeared less defined in ASD organoids (Fig. 2b). RNA spots for NHLH1, EOMES and SOX11 that were normally radially restricted became dispersed within each neuroepithelial bud. Neuronal markers such as STMN2 invaginated tissue regions dominated by progenitor markers such as SFRP1. MKI67 RNA, which typically lined the apical ventricular zone, also adopts irregular arrangements (Fig. 2b). These spatial gene patterns suggest that cortical tissue organization in ASD organoids is compromised. To better assess overall morphology, we mapped progenitor and neuron zones in organoids using spatially-aware clustering38. A total of 203 neuroepithelial buds in 31 tissue sections from 29 organoids across 21 individuals were identified (Supplementary Table 3, tab 5). Neuroepithelial buds in all controls were radially organized with inner progenitor and outer neuron zones clearly demarcated (Fig. 2c, CW60040). Progenitor and neuron zones were similarly partitioned in ASD organoids, but the progenitor-neuron divide was overall less defined compared to controls (Fig. 2c, CW60172, CW20012, CW20002). As orthogonal validation, we immunostained an independent set of organoids for the progenitor markers, phosphorylated VIM (pVIM) and SOX2, as well as the neuronal marker BCL11B, and similarly observed more irregular zone boundaries in ASD organoids (Supplementary Fig. 2d). We quantified this cytoarchitectural difference by calculating, for each progenitor zone identified by MERFISH, the ratio of its actual physical area to its convex area. This metric, termed solidity (Fig. 2d), was significantly lower in ASD compared to control (Fig. 2e; P = 0.038, F(20,182) = 1.978, nested two-way ANOVA). In contrast, the number and size of neuroepithelial buds were unaffected in ASD organoids (Supplementary Fig. 2e). Thus, cellular zonation during cortical development may be compromised in ASD individuals.Fig. 3Neurons were locally disorganized in a spatially mosaic fashion in ASD organoids.a Top, metagene expression (see Methods) of deep-layer ExN (blue) versus migrating ExN (red) for control and ASD organoids. Bottom, zoom-in of a cortical plate region boxed in the corresponding organoid above and oriented apicobasally. b Distribution of relative metagene expression along cortical plate depth for each corresponding organoid in (a). Negative values indicate expression for migrating ExN while positive values indicate expression for deep-layer ExN. 0% in cortical plate depth marks apical side while 100% marks the basal surface. c Cortical plate lamination index in control and ASD organoids, defined as the peak-to-trough range in the relative metagene expression distribution for each cell line. Each box plot ranges from 25th to 75th percentile. Center line is median and whiskers are 1.5x interquartile range. p = 8.61 × 10−4, n = 16 organoids from 10 control individuals and 20 organoids from 14 ASD individuals. F(1, 29) = 13.8, two-way ANOVA; p = 8.61 ×10−4 for ASD versus control, Tukey-Kramer test. Source data are provided as a Source Data file. d Left, models of neuronal organization in cortical organoids. Three possibilities are listed: control, in which neurons form layers throughout the organoid; local, in which neurons are layered in some neuroepithelial buds but disorganized in others, and global, in which neurons throughout all buds in the organoid are disorganized. Right, metagene expression map for CW60218, an ASD sample with a mosaic pattern of neuronal disorganization. Neuroepithelial bud 1 is well-layered while bud 2 is disorganized. e Proportion of cortical plate ROIs that are layered or disorganized in organoids derived from each individual. Source data are provided as a Source Data file. Neurons were locally disorganized in a spatially mosaic fashion in ASD organoids. a Top, metagene expression (see Methods) of deep-layer ExN (blue) versus migrating ExN (red) for control and ASD organoids. Bottom, zoom-in of a cortical plate region boxed in the corresponding organoid above and oriented apicobasally. b Distribution of relative metagene expression along cortical plate depth for each corresponding organoid in (a). Negative values indicate expression for migrating ExN while positive values indicate expression for deep-layer ExN. 0% in cortical plate depth marks apical side while 100% marks the basal surface. c Cortical plate lamination index in control and ASD organoids, defined as the peak-to-trough range in the relative metagene expression distribution for each cell line. Each box plot ranges from 25th to 75th percentile. Center line is median and whiskers are 1.5x interquartile range. p = 8.61 × 10−4, n = 16 organoids from 10 control individuals and 20 organoids from 14 ASD individuals. F(1, 29) = 13.8, two-way ANOVA; p = 8.61 ×10−4 for ASD versus control, Tukey-Kramer test. Source data are provided as a Source Data file. d Left, models of neuronal organization in cortical organoids. Three possibilities are listed: control, in which neurons form layers throughout the organoid; local, in which neurons are layered in some neuroepithelial buds but disorganized in others, and global, in which neurons throughout all buds in the organoid are disorganized. Right, metagene expression map for CW60218, an ASD sample with a mosaic pattern of neuronal disorganization. Neuroepithelial bud 1 is well-layered while bud 2 is disorganized. e Proportion of cortical plate ROIs that are layered or disorganized in organoids derived from each individual. Source data are provided as a Source Data file. Neuronal disorganization is spatially mosaic among ASD organoids We next examined the arrangement of ExN identities in our organoid sections, focusing on deep-layer ExN and migrating ExN to assess early patterns of cortical lamination (Fig. 1b). As a proxy for their locations, we mapped the average expression of each identity’s marker genes (metagene expression) across the apicobasal depth of each cortical plate39 (Fig. 3a). We quantified the spatial segregation of the two metagenes in each organoid as the lamination index, defined as the peak-to-trough range of the relative metagene expression distribution (Supplementary Fig. 3a, b; Fig. 3b). Lamination index was significantly lower in ASD organoids compared to controls (Fig. 3c, F(1, 29) = 13.8, two-way ANOVA; p = 8.61 ×10−4 for ASD versus control, Tukey-Kramer test). Only 21% of ASD organoids had a lamination index above median, while 80% of control organoids met this criterion (p = 0.00142, Chi-square test). In control organoids, the spatial distribution of the two metagenes matched their typical organization in vivo (Fig. 1b): migrating ExN occupied the inner cortical plate while deep-layer ExN occupied the outer cortical plate (Fig. 3a, CW60040; Supplementary Fig. 3a). In contrast, the metagenes were spatially intermixed in ASD organoids (Fig. 3a, CW60172, CW20012, CW20002; Supplementary Fig. 3b). Our findings thus suggest that neuronal organization during cortical development may be disrupted in ASD.Fig. 4ASD organoids were discordant in developmental timing.a Distribution of gene-shared latent time among cells of controls, and of ASD organoids that contain predominantly ExN1 or ExN2 states. Each box plot ranges from the 25th to the 75th percentile. Center line is median and whiskers are 1.5x interquartile range. *p < 0.05; ****p < 0.0001; ns, not significant. p = 0.0112 (neurons) and 0.0180 (progenitors) for ExN2 versus control and 7.09 × 10−5 (neurons) and 0.0177 (progenitors) for ExN2 vs ExN1. n = 10 neurotypical controls and 13 ASD individuals, two-way ANOVA and post-hoc Tukey’s test. Source data are provided as a Source Data file. b Spatial distributions of latent time in control and ASD organoids. Scale, 400um. c, Spatial gradient in latent time in control and ASD organoids. Each box plot ranges from 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Spatial gradient is taken as the maximum rate of change in latent time over apicobasal distance of each neuroepithelial bud. p = 0.028, n = 14 organoids from 9 control individuals and 15 organoids from 10 ASD individuals, one-tailed two-sample Wilcoxon test. Source data are provided as a Source Data file. ASD organoids were discordant in developmental timing. a Distribution of gene-shared latent time among cells of controls, and of ASD organoids that contain predominantly ExN1 or ExN2 states. Each box plot ranges from the 25th to the 75th percentile. Center line is median and whiskers are 1.5x interquartile range. *p < 0.05; ****p < 0.0001; ns, not significant. p = 0.0112 (neurons) and 0.0180 (progenitors) for ExN2 versus control and 7.09 × 10−5 (neurons) and 0.0177 (progenitors) for ExN2 vs ExN1. n = 10 neurotypical controls and 13 ASD individuals, two-way ANOVA and post-hoc Tukey’s test. Source data are provided as a Source Data file. b Spatial distributions of latent time in control and ASD organoids. Scale, 400um. c, Spatial gradient in latent time in control and ASD organoids. Each box plot ranges from 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Spatial gradient is taken as the maximum rate of change in latent time over apicobasal distance of each neuroepithelial bud. p = 0.028, n = 14 organoids from 9 control individuals and 15 organoids from 10 ASD individuals, one-tailed two-sample Wilcoxon test. Source data are provided as a Source Data file. To validate the neuronal disorganization in ASD, we stained organoids for markers of migrating ExN (BHLHE22) and deep-layer ExN (IGFBP5) that we had identified through single-cell RNA sequencing (Fig. 1d). In control organoids, IGFBP5 was enriched in the outer CP while BHLHE22 was relatively dispersed or enriched in the inner CP (Supplementary Fig. 4a). In contrast, IGFBP5 and BHLHE22 were significantly intermixed to varying extents across ASD organoids (Supplementary Fig. 4b), resulting in no overall enrichment of either protein along the depth of the CP (p = 4.46 × 10−3, linear model T-test, Supplementary Fig. 4c). These results are consistent with the MERFISH-derived distributions of metagene expression and confirm that ASD organoids were disorganized compared to controls. We noted that the lamination index of ASD organoids was not all-or-none but instead existed as a continuum (Fig. 3c). We reasoned that neuronal disorganization in each organoid may be localized to some neuroepithelial buds and not others, rather than being a global phenomenon (Fig. 3d, left). To test this hypothesis, we clustered metagene distributions from all organoids as either layered or disorganized and quantified the proportion of layered and disorganized ROIs for each individual. Indeed, proportions were non-binary, indicating that neuronal disorganization occurred locally—in some, but not all parts of the organoid (Fig. 3d, e). Control and ASD individuals differed significantly in their proportions (Fig. 3e, p = 1.29 × 10−6, Pearson’s Chi-square test): most controls were uniformly layered, while organoids from ASD individuals were disorganized to different extents (e.g., CW60218; Fig. 3d, right). While both ExN1- and ExN2-presenting ASD organoids showed disorganization, ExN2-presenting ones had a higher proportion of disorganized ROIs (Fig. 3e, p = 0.00288, Pearson’s Chi-square test), suggesting that the transcriptomic state of excitatory neurons may correlate with the severity of architectural abnormalities across organoid lines. To investigate the contribution of any familial or genetic components to the observed disorganization, we fitted a linear regression model predicting the extent of disorganization from diagnosis and donor family. The greater disorganization in ASD remained significant even after adjusting for the effect of donor family on organoid appearance (p = 5.62 × 10−8, linear model T-test), suggesting that ASD status, and not family-specific mutations, primarily accounts for the disorganization. Together, these findings indicate that ASD pathogenesis during prenatal cortical development may occur in a scattered manner with varying severity, underlying the focal patches of cortical disorganization seen in postmortem ASD cortices7–10. Cortical organization is temporally asynchronous across ASD individuals Since spatial conformation of the developing cortex is normally tightly coordinated with temporal identity40, we next asked if the architectural disarray in ASD organoids reflected altered developmental progression of their cells. We performed RNA velocity analysis on our organoid single-cell RNA sequencing data and inferred latent time per cell as a representation of each cell’s temporal position along its developmental trajectory. This means that cells at a later latent time would represent a more advanced stage. Given that transcriptomic states correlate with the severity of structural derangement (Fig. 3e), we analyzed ASD individuals with and without ExN2 states separately. Developmental progression for both neurons and progenitors was delayed in ExN2-presenting ASD organoids when compared to controls (Fig. 4a, p = 0.0112 (neurons) and 0.0180 (progenitors), two-way ANOVA and post-hoc Tukey’s test). In contrast, both neurons and progenitors were unaffected in ExN1-dominant ASD organoids (Fig. 4a). Temporal progression during cortical development is thus predicted to be discordant among ASD organoids, highlighting heterogeneity that is consistent with previous reports of developmental asynchrony in ASD organoids19,20,41. To correlate temporal identity with tissue architecture, we imputed latent time for each cell in our spatially profiled organoids. As expected, progenitor zones displayed earlier latent times than neuron zones in all organoids (Fig. 4a). Within control organoids, cells in each neuroepithelial bud were layered by their temporal identity: latent time transitioned from progenitor stage (deep purple) through an intermediate stage at the inner cortical plate (off-white) to the most advanced stage at the outer cortical plate (orange) (Fig. 4b, CW60040). Temporal layering in the cortical plate mirrored the metagene distribution of migrating and deep-layer ExN (Fig. 3a), indicating that neurons of the same developmental stage are typically produced at the same time. In contrast, latent time in ASD organoids transitioned sharply from progenitors to neurons (Fig. 4b, CW20012, CW20002; Fig. 4c, p = 0.028, two-sample Wilcoxon test), bypassing the intermediate phase that was evident in controls. Neuronal identities in ASD organoids were thus no longer temporally layered, indicating that dysregulated developmental timing may impact the spatial organization of neuronal subtypes. Cell adhesion between ASD progenitors is dysregulated Since spatiotemporal coordination during cortical development depends on tightly regulated cell-cell interactions, we sought to identify signaling changes in ASD organoids. We integrated MERFISH with single-cell RNA sequencing and imputed transcriptome-wide gene expression for MERFISH-imaged cells42, revealing the in situ expression of all signaling molecules. Imputed gene expression recapitulated both cell identities (Supplementary Fig. 5a) and organoid phenotypes (Supplementary Fig. 5b) derived from measured RNA counts. We then mapped the spatial enrichment of ligand-receptor interactions between pairs of cell identities using a metric called Spatial Signaling Score (S3). S3 is the sum of the product of ligand and receptor expression between interacting cell pairs for each pair of cell identities, divided by the number of cells of the receiver cell identity (Fig. 5a). We restricted our analysis to cells in close physical proximity, where ligand concentration is expected to closely mirror its expression levels in interacting sender cells. As expected, S3 calculated from imputed expression correlated well with those from measured RNA counts (r = 0.7) (Supplementary Fig. 5c, d) and recapitulated known signaling events in the developing human cortex26,43–46 (Supplementary Fig. 5e).Fig. 5ASD organoids were dysregulated in cell adhesion signaling.a Schematic illustrating Spatial Signaling Score (S3) per ligand-receptor pair. Left, interactions between two contacting cells, or edge weights, quantified as the product of ligand (L) expression in sender cell and receptor (R) expression in receiver cell. Center, example cell-cell contact graph for 6 cells. Each node is a cell. Each arrow represents an edge defined by proximal soma contact. Right, S3 between two cell identities (sum of cell-cell contact edge weights normalized by number of all receiver cells in cell-cell contact graph). b Signaling perturbations in disorganized ASD organoids as defined by differential S3 (see Methods). c Correlation plot of absolute differential S3 values on log2 scale for the 651 perturbations shared between ExN1- and ExN2-presenting ASD organoids. d Left, number of shared signaling perturbations for each sender-receiver cell identity pair. Only affected cell identity pairs are shown. Right, number of shared signaling perturbations in each signaling class for progenitors (RG senders – RG receivers) and neurons (maturing ExN – maturing ExN). ECM, extracellular matrix; IgSF, immunoglobulin superfamily; LDL, low density lipoprotein; RTN, reticulon; TAM, TYRO3, AXL and MERTK; EGF, epidermal growth factor; FGF, fibroblast growth factor. Source data are provided as a Source Data file. e Dot plot of differential S3 between all ASD and all control datasets for each ligand-receptor pair in each cell identity pair, normalized across all cell identity pairs. Green circles indicate top 100 differential S3 (adjusted p-value < 0.05, absolute normalized differential S3 > 0.006, two-tailed two-sample Wilcoxon test, n = 14 organoids from 10 control individuals and 19 organoids from 14 ASD individuals). SFARI genes are in red. f Left, immunostaining for ITGA6, SOX2 (progenitor marker) and DAPI (all nuclei) in day 56 control and ASD neuroepithelial buds. White dotted lines demarcate SOX2-enriched (progenitor zone) and SOX2-depleted (neuron zone) regions. Scale:100 um. Right, ratio of ITGA6 immunoreactivity in progenitor zone versus that in neuron zone (n = 11 images from 3 control organoid lines and 18 images from 3 ASD organoid lines, two-tailed Welch’s T-test). Each box plot ranges from the 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Source data are provided as a Source Data file. ASD organoids were dysregulated in cell adhesion signaling. a Schematic illustrating Spatial Signaling Score (S3) per ligand-receptor pair. Left, interactions between two contacting cells, or edge weights, quantified as the product of ligand (L) expression in sender cell and receptor (R) expression in receiver cell. Center, example cell-cell contact graph for 6 cells. Each node is a cell. Each arrow represents an edge defined by proximal soma contact. Right, S3 between two cell identities (sum of cell-cell contact edge weights normalized by number of all receiver cells in cell-cell contact graph). b Signaling perturbations in disorganized ASD organoids as defined by differential S3 (see Methods). c Correlation plot of absolute differential S3 values on log2 scale for the 651 perturbations shared between ExN1- and ExN2-presenting ASD organoids. d Left, number of shared signaling perturbations for each sender-receiver cell identity pair. Only affected cell identity pairs are shown. Right, number of shared signaling perturbations in each signaling class for progenitors (RG senders – RG receivers) and neurons (maturing ExN – maturing ExN). ECM, extracellular matrix; IgSF, immunoglobulin superfamily; LDL, low density lipoprotein; RTN, reticulon; TAM, TYRO3, AXL and MERTK; EGF, epidermal growth factor; FGF, fibroblast growth factor. Source data are provided as a Source Data file. e Dot plot of differential S3 between all ASD and all control datasets for each ligand-receptor pair in each cell identity pair, normalized across all cell identity pairs. Green circles indicate top 100 differential S3 (adjusted p-value < 0.05, absolute normalized differential S3 > 0.006, two-tailed two-sample Wilcoxon test, n = 14 organoids from 10 control individuals and 19 organoids from 14 ASD individuals). SFARI genes are in red. f Left, immunostaining for ITGA6, SOX2 (progenitor marker) and DAPI (all nuclei) in day 56 control and ASD neuroepithelial buds. White dotted lines demarcate SOX2-enriched (progenitor zone) and SOX2-depleted (neuron zone) regions. Scale:100 um. Right, ratio of ITGA6 immunoreactivity in progenitor zone versus that in neuron zone (n = 11 images from 3 control organoid lines and 18 images from 3 ASD organoid lines, two-tailed Welch’s T-test). Each box plot ranges from the 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Source data are provided as a Source Data file. For each cell identity pair, we identified ligand-receptor pairs with differential S3 between ASD organoids versus all controls. Out of 79,303 interactions, 1,116 (1.4%) and 3,175 (4%) were altered in disorganized ASD organoids harboring ExN1 and ExN2 states, respectively (adjusted p-value < 0.05, two-sample Wilcoxon test). Of these, 465 and 2,524 were specific to each group, perhaps reflecting distinct microenvironments resulting from neuronal disarray and developmental asynchrony (Supplementary Fig. 5f, g). 651 signaling perturbations were shared between both groups (Fig. 5b). They largely comprised downregulated interactions and were highly concordant between ExN1- and ExN2-presenting ASD organoids (Fig. 5c, r = 0.98), indicating a common set of signaling changes across heterogeneous ASD phenotypes. Notably, 77.5% of genes participating in these shared signaling perturbations were either differentially expressed or contributed to dysregulated GO terms (Fig. 1i, Supplementary Data 1, 2), indicating that altered intercellular signaling aligns with the transcriptomic changes and functional pathways disrupted in ASD organoids. Among shared perturbations, 53% were involved in cell adhesion including integrin, immunoglobulin superfamily (IgSF) and ephrin signaling (Fig. 5d, right; Fig. 5e), consistent with our pseudobulk transcriptomic analysis in which the top dysregulated GO processes – particularly within RG identities – were dominated by programs related to cell adhesion and signal transduction (Fig. 1g). Indeed, 94% of shared perturbations occurred among RG identities, while only 6% involved neurons (Fig. 5d, left). Adhesion between RG soma and processes is known to be required for maintaining RG architecture, which supports neuronal migration and cortical layering47. The reduction in adhesion-related signatures between RG identities in ASD organoids thus suggests that RG fibers may be structurally compromised, resulting in improper ‘touch-and-go’ contact, which is consistent with the observed neuronal disorganization. Other dysregulated pathways included Notch, Wnt, and growth factor signaling, indicating altered neural progenitor dynamics26,48 that may relate to the observed abnormalities in cellular zonation and developmental timing (Fig. 2c–e). To validate the reduced cell adhesion signaling shared by ASD organoids, we focused on integrin-mediated signaling, which exhibited the highest number of altered interactions (Fig. 5d). We stained organoids for integrin subunit alpha 6 (ITGA6), an integrin receptor known to be expressed in the human fetal cortex25, and involved in GO processes and interactions that were dysregulated in our ASD organoids (Fig. 5e). In control organoids, ITGA6 protein was enriched in the progenitor zone and also detected in the neuron zone, particularly along the basal surface where RG fibers typically extend (Fig. 5f). In contrast, ASD organoids showed comparable ITGA6 enrichment in the progenitor zone but a marked reduction in the neuron zone, consistent with disrupted RG structure and contact (Fig. 5f; p = 6.17 × 10−3, two-tailed Welch’s T-test). Hence, the spatial disorganization in ASD organoids features dysregulated integrin-mediated adhesion. Dysregulated cell adhesion is associated with sustained disorganization To examine whether decreased cell adhesion signaling in ASD organoids (Fig. 5d) is associated with altered RG architecture, we assessed RG structure using Nestin immunostaining, which labels RG cytoplasm49 (Fig. 6a). Nestin immunoreactivity density in neuronal zones of ASD organoids was decreased relative to controls (Fig. 6b, p = 2.48 × 10−5, two-tailed Welch’s T-test) while that in progenitor zones was unaffected (Fig. 6c), despite comparable Nestin mRNA expression levels (Supplementary Fig. 6a) and radial glia abundances (Fig. 1e, Supplementary Fig. 1i) in control and ASD organoids. This result indicates that the integrity of RG fibers in ASD organoids was compromised, consistent with the decrease in cell adhesion signaling among RG. Since neurons migrate along RG processes, defective RG architecture is likely related to the observed neuronal disorganization.Fig. 6Dysregulated cell adhesion in ASD is associated with sustained disorganization.a Immunostaining for nestin (radial glia marker), MAP2 (neuronal marker) and DAPI (all nuclei) in day 56 control and ASD neuroepithelial buds. Scale: 50 um. b Quantification of nestin immunoreactivity density in neuron zones (n = 21 images from 3 control organoids and 70 images from 4 ASD organoids, two-tailed Welch’s T-test). Each box plot ranges from the 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Source data are provided as a Source Data file. c Same as b but for progenitor zones. Source data are provided as a Source Data file. d Day 80 organoids immunostained with migrating ExN marker BHLHE22, deep-layer ExN marker IGFBP5, and pan-neuronal marker MAP2. Three replicate images, each showing cortical plates from different organoids, are shown for each line. Composite images feature IGFBP5 (yellow) and BHLHE22 (blue). The corresponding individual channels for IGFBP5, BHLHE22, and MAP2 are shown below. Red dotted lines separate IGFBP5-enriched regions from IGFBP5-deficient and/or BHLHE22-enriched regions. Scale: 100 um. e Relative distribution of BHLHE22 and IGFBP5 immunoreactivity along the depth of the cortical plate in day 80 cortical organoids. n = 7 organoids from 3 control lines and 10 organoids from 4 ASD lines. Data are presented as mean values ± SEM. p = 2.12 × 10−6, linear model T-test. Source data are provided as a Source Data file. f Model illustrating the spatial molecular architecture of ASD pathogenesis as proposed by this study. Control individuals undergo typical cortical development as reflected by uniform layering of deep-layer and migrating ExN in control organoids. In contrast, ASD individuals likely harbor focal patches of pathogenesis during cortical development. These pathogenic sites are spatially mosaic and vary in distribution among patients, reminiscent of focal pathology observed in ASD postmortem cortex. They are characterized by neuronal disorganization, discordant developmental timing, and dysregulated cell adhesion. Dysregulated cell adhesion in ASD is associated with sustained disorganization. a Immunostaining for nestin (radial glia marker), MAP2 (neuronal marker) and DAPI (all nuclei) in day 56 control and ASD neuroepithelial buds. Scale: 50 um. b Quantification of nestin immunoreactivity density in neuron zones (n = 21 images from 3 control organoids and 70 images from 4 ASD organoids, two-tailed Welch’s T-test). Each box plot ranges from the 25th to 75th percentile. Center line is median, and whiskers are 1.5x interquartile range. Source data are provided as a Source Data file. c Same as b but for progenitor zones. Source data are provided as a Source Data file. d Day 80 organoids immunostained with migrating ExN marker BHLHE22, deep-layer ExN marker IGFBP5, and pan-neuronal marker MAP2. Three replicate images, each showing cortical plates from different organoids, are shown for each line. Composite images feature IGFBP5 (yellow) and BHLHE22 (blue). The corresponding individual channels for IGFBP5, BHLHE22, and MAP2 are shown below. Red dotted lines separate IGFBP5-enriched regions from IGFBP5-deficient and/or BHLHE22-enriched regions. Scale: 100 um. e Relative distribution of BHLHE22 and IGFBP5 immunoreactivity along the depth of the cortical plate in day 80 cortical organoids. n = 7 organoids from 3 control lines and 10 organoids from 4 ASD lines. Data are presented as mean values ± SEM. p = 2.12 × 10−6, linear model T-test. Source data are provided as a Source Data file. f Model illustrating the spatial molecular architecture of ASD pathogenesis as proposed by this study. Control individuals undergo typical cortical development as reflected by uniform layering of deep-layer and migrating ExN in control organoids. In contrast, ASD individuals likely harbor focal patches of pathogenesis during cortical development. These pathogenic sites are spatially mosaic and vary in distribution among patients, reminiscent of focal pathology observed in ASD postmortem cortex. They are characterized by neuronal disorganization, discordant developmental timing, and dysregulated cell adhesion. To assess if disrupted cell adhesion and RG architecture may have persistent effects, we examined neuronal organization in our organoids over time, immunostaining for migrating ExN marker BHLHE22 and deep-layer ExN marker IGFBP5. We focused on days 30, 56 and 80, as organoids at these time points manifest key hallmarks in cortical organoid development21: a consistent progenitor to neuron ratio at day 30, the emergence of neuronal layers at day 56 and the attainment of features of neuronal maturation at day 80 (Supplementary Fig. 1a). At day 30, all samples examined had a consistently thin layer of neurons marked by MAP2, and both BHLHE22 and IGFBP5 were not detectable (Supplementary Fig. 6b). By day 56, distribution of these two markers was evidently layered in controls and disorganized in ASD organoids (Supplementary Fig. 4a–c), indicating that neuronal disorganization is an emergent phenotype. At day 80, IGFBP5 remained localized to the outer CP, while BHLHE22 became restricted to the inner CP in control organoids. In contrast, the two markers remained variably intermixed across ASD organoids (Fig. 6d, e, p = 2.12 × 10−6, linear model T-test; Supplementary Fig. 6c). This observation indicates that neuronal disorganization occurring alongside adhesion defects persists into later development, consistent with the focal pathology seen in postmortem ASD brains7–10. Together, our work puts forth a model in which the diverse clinical symptoms and brain structure variations among ASD individuals may be explained by focal patches of pathogenesis that arise during prenatal development (Fig. 6f). These pathogenic loci are spatially mosaic, varying in distribution among patient-derived organoids. Neuronal organization is impacted, manifesting as structural derangement and altered gene expression. These phenotypes may relate to dysregulated ‘touch-and-go’ adhesive contacts between radial glia progenitors in the ASD state. Discussion Our study proposes that a spatially mosaic pattern of pathogenesis, together with discordant developmental timing, diversifies perturbations in neuronal progeny and contributes to clinical heterogeneity. Surprisingly, we observed greater heterogeneity than expected in the way patient-derived brain organoids develop, especially in how cellular layers were established. We thus present a parallel model alongside the ongoing convergence hypothesis. Our study further highlights the strength of extending investigations beyond well-defined subtypes of ASD — such as those associated with macrocephaly or single-gene mutations — in discovering disease-relevant phenotypes. We anticipate similar multi-modal approaches applied to cohorts of idiopathic forms of complex disorders to reveal susceptible ligand-receptor interactions that can serve as candidate therapeutic targets. The disorganization observed in our ASD organoids is notably more extensive than that found in patient postmortem brains, which tends to occur in relatively small regions of the cortex7–9. It is possible that early disorganization in vivo may be partially resolved by compensatory mechanisms that occur during later developmental stages that are not typically recapitulated in our organoids. Alternatively, our organoids may be primed towards certain cortical subregion identities that are particularly susceptible to disorganization in autism, such as the prefrontal cortex21. Organoid culture conditions may also exacerbate the biological processes that allow the disorganization to manifest. These considerations, however, do not negate our key finding on the varied distribution of disorganization seen in ASD organoids, since organoids from all individuals were cultured under identical conditions. The ability to model pathological features in organoids opens new doors for investigating disease mechanisms and stratifying patients by their neurobiological patterns. Elucidating the molecular pathways that exacerbate phenotypic manifestations in organoids relative to in vivo may reveal new ways to mitigate the pathological phenotypes. Triggers of focal pathogenesis remain to be addressed. Lineage tracing of abnormal progeny may be the first step in defining such triggers, along with the affected brain circuits that contribute to the core symptoms of ASD, thereby enabling targeted therapeutic approaches. Prior to this study, ASD has been associated with defects in progenitor proliferation, while implicating neurons and synaptic adhesion as primary drivers of disease1,19. The use of spatial technologies empowered us to deep-dive into a new dimension: the physical relationship between cells. We found cell adhesion between progenitors to be majorly impaired across ASD organoids, alongside the structural and transcriptomic phenotypes identified in neurons. This result suggests that the way progenitor cells interact may have a profound way in determining developmental phenotypes in ASD. It is also possible that the disorganization observed in ASD organoids results from an altered maturation timeline. However, we note that ASD organoids were disorganized regardless of whether their developmental timing was delayed or not, suggesting that disorganization and timing of maturation are likely to act separately or in parallel in the pathogenesis of ASD. We note that a minority of ASD samples were non-phenotypic for the metrics assayed. It remains an open possibility that they contain latent phenotypes that emerge at later developmental stages. Indeed, an early and late phase of brain growth have been implicated in the pathogenesis of ASD4,50. Alternatively, the patients’ symptoms may be due to environmental factors that could not be recapitulated in an organoid or in vitro setting. On the other hand, a minority of control organoids also lacked clear layering organization. Though surprising, this observation is consistent with those in human postmortem cortex, where a minority of control individuals were reported to exhibit patches of cortical disorganization7,9. Effects due to variation in iPSC lines cannot be ruled out completely. Given that autism is a spectrum disorder, this possibility remains until many more organoid lines have been derived from individuals with autism across the spectrum. This is a mammoth effort unlikely to be achieved by any individual lab, and we acknowledge that we are playing a small pioneering part by contributing observations based on the autism lines in our hands. As with many similar studies, future research will provide more clarity. This study holds exciting implications for our understanding of ASD by modelling focal patches of pathology observed in postmortem brains in an in vitro organoid system. Our results suggest that such postmortem pathology may be developmentally derived and propose a model in which spatially mosaic pathogenesis contributes to ASD heterogeneity. Methods Inclusion & Ethics statement All experiments were carried out in accordance with Singapore’s Agency for Science, Technology and Research (A*STAR) guidelines. All hiPSC lines were obtained from the NIGMS Human Genetic Cell Repository at the Coriell Institute for Medical Research, and in compliance with research use regulations in each organization. Human induced pluripotent stem cells (hiPSC) lines The following hiPSC lines were obtained from the NIGMS Human Genetic Cell Repository at the Coriell Institute for Medical Research: CW20002A, CW20012E, CW60040F, CW20041A, CW20047D, CW20050D, CW20051C, CW20053A, CW20069F, CW20081F, CW60108A, CW60109C, CW60115F, CW60171E, CW60172A, CW60134F, CW60218E, CW60219E, CW60220F, CW60228D, CW60230E, CW60245L, CW20320C, CW20324C. Each cell line has a certificate of analysis from CIRM indicating that it passed quality control testing, namely single-nucleotide polymorphism profiling, qPCR for pluripotency genes, teratoma assays, and SNP assays (Illumina Infinium HumanCore BeadChips) to check chromosomal integrity and karyotype. All hiPSCs were cultured on hESC-qualified Matrigel (Corning, 354277) in mTeSR1 Complete Kit (Stemcell Technologies, 85850) and maintained at 5% CO2, 37 °C with daily media change and weekly passage. All hiPSCs were maintained below passage 50 and routinely tested negative for mycoplasma using EZ-PCR mycoplasma test kit (20-700-20; Bio-REV). They also expressed pluripotency markers, SOX2 (Abcam ab92494, 1:500), OCT3/4 (Santa Cruz Biotechnology sc-5279, 1:500) and NANOG (R&D systems, AF1997, 1:500). Generation of cortical organoids Cortical organoids were generated as previously described21 with minor modifications. On Day 0, hiPSC colonies were lifted using Collagenase type IV and transferred onto ultra-low attachment six-well plates containing DMEM/F12, 20% KOSR, 1x GlutaMAX, 1x MEM-NEAA, 0.1 mM 2-Mercaptoethanol, 1x Penicillin/Streptomycin, 2 µM Dorsomorphin, 2µM A-83. On Day 5 and 6, 50% medium was replaced with DMEM/F12, 1x N2 supplement, 1x GlutaMAX, 1x MEM-NEAA, 1x Penicillin/Streptomycin, 1 µM CHIR-99021, 1 µM SB-431542. On Day 7, organoids were embedded in growth factor-reduced (GFR)-Matrigel. On Day 14, organoids were mechanically dissociated from GFR-Matrigel and cultured with shaking in DMEM/F12, 1x N2 supplement, 1x B27 supplement, 1x GlutaMAX, 1x MEM-NEAA, 1x Penicillin/Streptomycin, 0.1 mM 2-Mercaptoethanol, 2.5 µg/ml Insulin. Medium was supplemented with 0.5% (v/v) GFR-Matrigel from Day 35 onwards. Bulk RNA sequencing RNA was extracted from three to five 33-day-old cortical organoids derived from a subset of control and patient iPSC lines using the DirectZol RNA miniprep Plus kit (Zymo Research, R2051) as per the manufacturer’s recommendations. These individuals are CW20012E, CW60040F, CW20047D, CW20053A, CW20069F, CW60219E, CW60220F, CW60230E (Supplementary Table 3, tab 6). Each individual was replicated twice. Library preparation and sequencing were performed at the Next-Generation Sequencing Platform (NGSP), Genome Institute of Singapore (GIS), A*STAR, Singapore. Sequencing reads were aligned to hg38 using STAR51 and transcript abundances quantified using RSEM52. Exome sequencing analysis All samples were sequenced in-house with the Illumina HiSeq 4000 platform, Exome were captured with either SeqCap EZ Human Exome Probes v3.0 (Roche) or KAPA HyperExome Prep Kit (Roche) due to the discontinuation of the former one, with 151 bp pair-end sequencing reads. Sequence data were processed from the de-multiplexed Fastq files, started with alignment against the GRCH37 human genome assembly with BWA53, followed by variants realignment, recalibration and both germline INDELs and SNVs calling were done with Genome Analysis Toolkit (GATK, https://gatk.broadinstitute.org/hc/en-us). QC is started by Variant Quality Score Recalibration, with variant quality score log-odds of 99.4%. Genotypes with GQ < 25, DP < 10X, and allelic balance > 0.8 or <0.2 were set to missing. Qualified Samples were selected as they must have 20X or above among 80% of the captured region. All post-QCed genetic variants were annotated using the GRCh37 (equivalent to hg19) database with ANNOVAR54, and compatible modes of inheritance of variants were done by Jannovar (https://github.com/charite/jannovar). Subsequently, disease-causing variants were filtered by matched inheritance pattern, very low allelic frequency ( < 0.5%) or undetected in ExAc and 1000Genome, and predicted outcome as deleterious by MutationTaster (https://www.mutationtaster.org). Immunohistochemistry Whole organoids were fixed in 4% paraformaldehyde in PBS for 15 minutes, sucrose-sunk overnight, frozen in O.C.T. compound, and cryosectioned at 10 µm. Tissue sections were collected on polylysine-coated glass slides, blocked with 5% donkey serum (or 1.5% BSA)/0.3% Triton X-100 /PBS for 30 minutes, and incubated with primary antibodies in 3% donkey serum/0.3% Triton X-100/PBS overnight. After two PBS washes, sections were incubated with secondary antibodies (Jackson Immunoresearch) and DAPI for 2 hours and mounted with FluorSave reagent (Calbiochem, 345789-20) or VECTASHIELD mounting medium (Vector Laboratories, H-1500). Primary antibodies used were MAP2 (Synaptic Systems 188004, 1:600), Nestin (R&D Systems MAB1259, 1:500), SOX2 (Abcam ab92494, 1:500), BHLHE22 (Sigma Aldrich HPA064872, 1:500), IGFBP5 (R&D Systems MAB875, 1:500), FOXG1 (Abcam ab18259, 1:500), pVIM (MBL D076-3S, 1:1000), BCL11B (Abcam ab18465, 1:500) and ITGA6 (eBioscience 12-0495-81, 1:500). Microscopy images in Figs. 5f, 6a and 6d as well as Supplementary Fig. 1b, 2d, 4a, b and 6b, c were acquired on Nikon Eclipse Ti-E Inverted Microscope at x-y resolution of 0.324 um x 0.324 um using NIS-Elements AR software. Images were analyzed in ImageJ. Gamma, contrast, and brightness settings of the images were adjusted to make dim features visible without eliminating the background signal. Quantification of Nestin immunoreactivity density was performed uniformly across all images using a customized CellProfiler pipeline. Quantification of ITGA6 immunoreactivity was done by obtaining the mean gray values of ITGA6 signal in the progenitor and neuron zones of each image using ImageJ’s Analyze -> Measure function. Single-cell RNA sequencing Three 56-day-old cortical organoids were collected per hiPSC line, rinsed with Earle’s Balanced Salt Solution (EBSS), chopped in pre-warmed papain solution (Worthington, LS003126), and incubated at 37 °C for 30 minutes with gentle shaking. Cell suspension was gently triturated, strained, and inactivated with ovomucoid solution (Worthington, LS003087). Cells were resuspended to ~1000 cells/ul. More than 90% of cell viability was obtained. All individuals were processed except for CW20050, whose sample was lost due to technical issues. Single-cell libraries were prepared using 10x Genomics Chromium Single Cell 3’ Kit v3 according to the manufacturer’s instructions at the Spatial and Single Cell Genomics Platform (S2GP), GIS, to obtain a targeted 8000 cells per sample and sequenced on Illumina NovaSeq 6000 at NGSP. At least 80,000 mean reads per cell were obtained. Pseudobulk differential gene expression and GO analyses Pseudobulks were generated from single-cell RNA sequencing data by summing gene expression counts per cell type per donor. Pseudobulks containing at least 10 cells per donor and genes with at least 10 counts in at least five pseudobulks were retained for differential expression analysis using DESeq2 (v1.30.1). Differential testing was performed using a design formula that included donor family as a covariate ( ~ family + diagnosis). P values were re-estimated using fdrtool (v1.2.18). Differential expression results for each cell type are provided in Supplementary Data 1. Genes were ranked by the product of log fold change and fdrtool-derived P values and used for Gene Set Enrichment Analysis (GSEA) with clusterProfiler (v4.8.3) and org.Hs.eg.db (v3.17.0). Gene Ontology results for each cell type are provided in Supplementary Data 2. MERFISH encoding library probe design and synthesis A Modified Hamming Distance 4 encoding scheme was adopted for MERFISH probe design. 30-nucleotide long target regions were identified using a previously published algorithm34,55, with GRCh38 as reference, a specificity cut-off of 0.5 and an isospecificity cut-off of 0.4. Three readout sequences per encoding probe were used; they were either 20-nucleotide long as published in ref. 56, or 30-nucleotide long as published in ref. 34. smFISH genes were similarly detected, except that all encoding probes for the same gene carried the same readout sequence to allow spot detection in a single hybridization channel. Encoding probe libraries prepared as previously described55. Encoding probes were synthesized as an oligonucleotide pool (CustomArray or Twist Biosciences) and PCR-amplified (Phusion Hot Start Flex 2x Master Mix M0536L): 98 °C, 30secs; 98 °C, 5 secs, 72 °C, 15 secs, 72 °C, 18secs for 22 cycles; 98 °C, 5 mins. PCR products were then purified and concentrated before in vitro transcription (HiScribe T7 Quick High Yield RNA Synthesis Kit E2050) and reverse transcription (Maxima H Minus Reverse Transcriptase EP0752). RNA templates were removed by alkaline hydrolysis with 1 M NaOH and 0.5 M EDTA at 95 °C, 10 minutes. Probes were purified and eluted in nuclease-free water, evaporated in a vacuum concentrator, and re-suspended in hybridization buffer to 1 mM. Readout probes were ordered from IDT. MERFISH sample preparation Fresh organoids were OCT-embedded and cryosectioned at 10 µm thickness onto glass coverslips functionalized with 2% vol/vol (3-aminopropyl) triethoxysilane/acetone. Sections were fixed in 4% PFA/PBS for 15 minutes, permeabilized with 0.1% Triton X-100/PBS, and pre-incubated with hybridization buffer. Sections were then incubated with the encoding probe library at 37 °C overnight in humidified chambers. Probe concentration was optimized for each MERFISH panel library. The next day, sections were incubated with DAPI (1:1000) and 0.1 µm TetraSpeck fluorescent microspheres (1:1000, Invitrogen, T7279), rinsed with 2x SSC, and kept cold till imaging. MERFISH imaging Tiled imaging was performed at 60x magnification using a custom-built microscope comprising a Nikon Ti-E body and MS-2000 ASI X–Y stage. Samples were mounted on a flow chamber (Bioptechs, FCS2), which was secured to the microscope stage. Hybridization of readout probes and washes in between hybridization rounds were performed sequentially within the flow chamber by buffer exchange controlled by an automated fluidics system. Imaging buffer consisted of 2× SSC, 50 mM Tris-HCl pH 8, 10% glucose, 2 mM Trolox (Sigma, 238813), 0.5 mg ml–1 glucose oxidase (Sigma, G2133), and 40 μg ml–1 catalase (Sigma, C30). We performed four-color imaging using Cy3, Cy5, Alexa Fluor 594 and IRDye 800CW as dyes on our readout probes. Dyes were quenched by photobleaching after each imaging round. Focus was maintained using the Nikon Perfect Focus system. A single Z position was imaged per field of view (FOV). MERFISH spot decoding Raw MERFISH images were processed using a published image analysis pipeline57 modified to include adaptive thresholding. Images from each hybridization cycle were corrected for field and chromatic distortion and registered for translation relative to a selected frame by phase correlation using a subpixel registration algorithm from the Scikit-image (v0.16.2) package. All images were filtered with a high-pass filter and a low-pass filter to suppress cell background and camera noise, respectively. Images are further corrected for intensity variations across channels and hybridizations: we performed a global bit-wise normalization by pooling all pixels above the 99.9th percentile of intensity in each field-of-view (FOV), and then taking the 50th percentile of the intensities of pooled pixels as a normalization value for the bit. MERFISH spots were decoded with a three-step process. First, barcodes were matched by finding the closest barcode match against the pixel vector using the nearest neighbor search to the reference barcode using a KDTree. The distance is computed using the normalized Euclidean distance. Second, adjacent pixels matched to the same barcode are aggregated into spots. Third, spots are classified as true or false positives and filtered using the following three features: i. the minimum normalized Euclidean distance of the aggregated pixels of a spot, ii. the mean L2 norm intensity of a spot, iii. number of adjacent pixels aggregated into a spot. These three metrics are binned into a 3D histogram. Each bin is assigned a blank fraction score, which is the mean(blank_barcode_count(bin)/ mean(barcode_count(bin)). A blank fraction score threshold is set such that spots found in bins with a blank fraction score above that threshold are filtered out. This threshold is chosen to obtain a target misidentification rate of 0.04, with the misidentification rate defined as the mean(blank_barcode_count) / total(blank_barcode_count). Total counts per gene were compared to FPKM values from bulk RNAseq by Pearson’s correlation. smFISH spots were called out using methods adapted from ref. 58. Peaks were localized using local max intensity, and a threshold separating signal from noise was determined by the inflection point on the cumulative distribution function of the peak intensities, where the first derivative is at a minimum and the second derivative is at a maximum. Retained spots were visually inspected and thresholds adjusted accordingly. For mosaic stitching, adjacent FOV alignments were estimated using the phase-correlation algorithm from scikit-image that was modified to output a value for the phase-correlation peak magnitude, which is an indication of registration accuracy. We then generated a graph with FOVs as vertices and edges weighted by the negative of the phase-correlation peak value. The full mosaic was then stitched by calculating the minimum spanning tree (SciPy v1.5.2) and shifting each FOV accordingly. Overlapping regions were blended using maximum intensity projection. Power analysis We performed power analysis to determine if our sample sizes are sufficient for detecting meaningful effects. With 36 total observations (organoid sections), 2 groups (ASD and control), alpha = 0.05, and 80% power, the estimated effect size (Cohen’s d) is 0.48 (this means the group means differ by 0.48 standard deviations). This is considered a medium effect size59 and is reasonable given that our patient cohort is heterogeneous. Evaluation of cortical organoid single-cell RNA sequencing Cortical organoid single-cell transcriptomes were benchmarked against BrainSpan using the VoxHunt (v1.0.1) package24. ASD gene enrichment analysis was performed for each cell identity using Fisher’s Exact test of the GeneOverlap (v1.26.0) package. Gene ontology enrichment analysis was performed based on semantic similarity using simplifyEnrichment (v1.0.0) package60. Gene ontology terms with q-value < 0.05 were clustered using the Louvain method. Differences in proportions of cell identities were evaluated using the Kolmogorov-Smirnov test. Nuclei segmentation Organoid cells in DAPI images were segmented using a deep learning segmentation pipeline, Cellpose 2.061,62. For model training and evaluation, we manually annotated 31 images from 20 different organoid lines. All manual annotations are divided into a training set (70%) and a validation set (30%). Images from the same organoid lines were not separated in the training and validation sets to prevent information leakage. We used a computer with an 11th Gen Intel(R) Core(TM) i7-11800H @ 2.30 GHz processor and 32 GB memory to train our own segmentation model for 2000 epochs based on the pretrained model called cyto2. Then, we evaluated our model on the validation set. The precision and recall are 0.766 and 0.885, respectively, when the intersection over union between segmented cells and manually annotated ground truth cells is above 0.5. We used our trained model to segment nuclei for each dataset and obtained a nuclei mask, which is a matrix of pixel values that gives the location of every cell. MERFISH spot assignment To reduce the misassignment of RNA spots, we generated a mask from non-segmented regions of each DAPI image. The pixels outside the segmented nuclei mask and with intensity above the median intensity of the segmented regions were considered as non-segmented regions. RNA spots within the non-segmented region were discarded. Next, we dilated each segmented nucleus by 2.2 um using the Euclidean distance transform method to obtain a cell mask. If an RNA spot is located within a segmented cell, it is assigned to that cell. The percentage of spots assigned to cells is 84.2 ± 1.2%. Single-cell QC and clustering Feature-barcode matrices for single-cell RNA sequencing were generated using CellRanger (v7.0.1). Cells were discarded if they expressed 1800 genes or fewer, 11000 genes or more, 15% or more mitochondrial genes, or 1% or less mitochondrial genes. Genes were discarded if they were expressed in less than 1% of all cells. Gene expression was normalized using scTransform (v0.4.1)63. Gene-cell matrices across all datasets were concatenated, and clustering was performed with 500 top-ranked features and 15 PCs using Seurat (v4.0.2)64. Clusters marking signatureless cells, stressed cells, or transcriptomic multiplets as identified by DoubletFinder (v2.0.3)65 were discarded. The resulting filtered gene-cell matrices were re-clustered using 100 top-ranked features and 10 PCs. Progenitor and neuron identities were assigned to each cluster. All forebrain neuron cells were subclustered with 500 top-ranked neuron features and 20 PCs, while all forebrain progenitor cells were subclustered with 1000 top-ranked progenitor features and 20 PCs. MERFISH feature-barcode matrices were generated using the spot assignment pipeline described above. Cells were discarded if they meet the following criteria: cells (i) smaller than value at half-peak of nuclei area distribution for all datasets combined, (ii) above 99.9th quantile of nuclei area distribution for all datasets combined, (iii) contained 10 counts or less per cell, (iv) have a blank fraction per cell (blank counts per cell divided by total spots counts per cell) greater than 0.1, (v) expressed fewer than 10 genes or (vi) contained 70% or more counts from smFISH. For gene panels 2-4, cells with fewer than 3 genes were discarded. Genes present in less than 1% of all cells in each diagnosis were discarded. Counts were normalized using SCTransform63. Stressed and non-forebrain cells identified by BANKSY (v0.99.0)38 at lambda = 1, k_geometric = 36 were discarded. Progenitor and neuron zones were defined by BANKSY at lambda = 1, k_geometric = 6. Seurat (v4.0.2)64 was used for transcriptomic cell typing. All filtered cells were first clustered using the top 316 expressed genes and 10 PCs and annotated as progenitors or neurons. Progenitor cells were subclustered using the top 53 variable features and 7 PCs, while neuron cells were subclustered using the top 82 variable features and 6 PCs. De novo cell identities were assigned to each progenitor and neuron cluster based on the markers each cluster expressed. Progenitor zone analysis To delineate the progenitor zones, we initially mapped all progenitor cells onto images based on their spatial coordinates, representing each cell with a 3 × 3 cross. Subsequently, we computed the distances between the nearest neighbors and selected the dilated value as the 90th percentile of these distances, ranging between 7 and 10 μm among different datasets. Using this value, we expanded each progenitor cell to define the progenitor zones. After this expansion, we filled holes to address any internal voids within the zones. In instances where tears were observed in the dataset, the dilated value was adjusted to 23 μm to encompass these tears. Following this adaptation, we eroded the zones by 13 μm post the fill hole operation to mitigate over-dilation effects at the zone boundary. Each maximum connected region was then identified to represent each progenitor zone, with a filtering process excluding regions with areas less than 5000 μm², or those at the edges, or containing significant tears. Finally, the area, convex hull area, and solidity of each zone were computed, alongside a count of the cells contained within. A nested two-way ANOVA was performed to evaluate ASD-control differences. Imputation of gene expression Gene expression was imputed from single-cell RNA sequencing (reference) to MERFISH (query) using Seurat (v4.0.2)64. Imputation was done separately for progenitors and neurons. Transfer anchors between single-cell RNA sequencing and MERFISH were identified by mutual nearest methods using the FindTransferAnchors function with k.anchor = 8, k.filter = 30, reduction = rpca, dims1 = 1:29 (progenitors) or 1:6 (neurons), k.score = 15 (progenitors) or 40 (neurons), and shared features between the MERFISH and single-cell RNA sequencing objects. Single-cell RNA sequencing gene expression was imputed onto MERFISH cells by calculating the anchor-weighted sum using the TransferData function with weight.reduction = MERFISH[‘pca’], dims2 = 1:7 (progenitors) or 1:6 (neurons) and k.weight = 300 (progenitors) or 200 (neurons). MERFISH cells were subsequently clustered by their imputed expression using 15 PCs and expressed MERFISH genes with a log2 fold change greater than 0.8 in single-cell RNA sequencing data (for progenitors) or all expressed MERFISH genes (for neurons) as variable features. Cell identities were then assigned based on cluster markers. All parameters were determined by optimizing the correspondence between de novo identities obtained from unbiased clustering of MERFISH cells using measured counts and those using imputed expression. Quantification of metagene expression Metagene per cell identity is defined as the mean Pearson residuals of cell identity markers from single-cell RNA sequencing that are present in the MERFISH gene panel. If imputed expression is used to calculate the metagene per cell identity, min-max-normalized imputed expression is used instead of Pearson residuals. Lamination index We used a well-established method39 to quantify the spatial enrichment of metagene expression along the depth of each cortical plate (CP). Line scans were drawn across the CP of every neuroepithelial bud. Incomplete regions (e.g., edge of imaging area, tears, and folded tissue) were excluded. Metagene intensity along the depth of each CP was obtained using the ImageJ/Analyze/Plot Profile function. Depth values were normalized such that 0% indicates the apical border and 100%, the basal border. Intensity values were normalized to the maximum intensity within each line scan. Intensity values, in arbitrary units, were binned for every 5% CP depth and averaged across line scans in each dataset. The lamination index is the difference in metagene expression at the peak and trough of each dataset’s intensity distribution. A higher lamination index indicates better layering. Datasets with a lamination index less than 80 were considered to present neuronal mislayering. A linear regression model was used to evaluate differences in the distribution of lamination indices in controls versus ASD. Estimation of latent time using RNA velocity Spliced and unspliced count matrices were constructed using velocyto (v0.6)66. Gene-shared latent time was calculated using the dynamical model in scVelo (v0.2.5)67. Root cells and terminal states were selected using default settings. Spatial gradient in latent time was calculated as such: line scans were drawn across the CP of every neuroepithelial bud. Incomplete regions (e.g., edge of imaging area, tears, and folded tissue) were excluded. Metagene intensity along the depth of each progenitor zone and its corresponding neuron zone was obtained using the ImageJ/Analyze/Plot Profile function. Depth values were normalized such that the first bin indicates the apical border and the last bin indicates the basal border. The maximum gradient within ± 20bins from the neuron-progenitor border was calculated. Gradients obtained from organoids lacking neuroepithelial buds were beyond 1.5x interquartile range; for consistency, gradients beyond 1.5x interquartile range and/or from budless organoids were excluded from analysis. A two-sample ANOVA test followed by post-hoc Tukey’s test was performed to identify ASD-control differences. Spatial Signaling Score (S3) A spatial network between contacting cells (cell-cell contact graph) was first constructed: each cell is dilated by 1 pixel, and the resulting cell masks that overlap form an edge in the spatial network. Cell identity pairs in which 20% or less of the receiver cell identity interacts in both edge directions (A-B and B-A for heterotypic pairs, A-A or B-B for homotypic pairs) are removed. We obtained a list of literature-supported ligand-receptor pairs from connectomeDB202068 and shortlisted ligand and receptor genes with the highest mean expression per cell identity across all cell identities greater than 0.09 (value at the valley of the bimodal distribution in mean expression per cell identity). For each kept ligand-receptor pair, we calculated cell-cell contact edge weight, S3 and scaled S3. Cell-cell contact edge weight is the product of ligand expression in sender cell and receptor expression in the receiver cell. To avoid spurious interactions, ligand-receptor pairs with edge weight less than a threshold were set to 0. This threshold was determined as the value at the intersection of the edge weight distributions for kept and discarded ligand-receptor pairs. All edges are bi-directional. S3 for each sender-receiver cell identity pair is the sum of all cell-cell contact edge weights belonging to this sender-receiver cell identity pair, divided by the total number of receiver cells for that receiver cell identity in the cell-cell contact graph. Scaled S3 is the S3 for that sender-receiver cell identity pair divided by the sum of all S3 across all cell identity pairs. To identify affected signaling pathways and cell identities, differential S3 was calculated by subtracting, for each ligand-receptor pair in each cell identity pair, the mean S3 of all ASD samples (or just disorganized ASD samples presenting either ExN1 or ExN2 states) from the mean S3 of all control samples. For each ligand-receptor pair, differential S3 was normalized by the per-diagnosis average of the sum of mean S3 for all datasets within each diagnosis across all cell identity pairs. The Wilcoxon test was used to assess significance in normalized differential S3. For each cell identity pair, p-values from all ligand-receptor interactions were used as input to fdrtool (v1.2.18)69 to re-estimate p-values and obtain adjusted p-values. Normalized differential S3 between all ASD samples (or just disorganized ASD samples presenting either ExN1 or ExN2 states) and all controls was considered significant if it had an adjusted p-value less than 0.05 and a value greater than the 99th percentile of the normalized differential S3 distributions of 50 randomized control-to-control comparisons. Data analysis and visualization tools In addition to the packages listed above, the following were used for data analysis and visualization: on R (v4.1.0): patchwork (v1.1.1), Rmisc (v1.5), reticulate (v1.22), data.table v1.14.2), readxl (v1.3.1), reshape2 (v1.4.4), pheatmap (v1.0.12), ggplot2 (v3.3.6), igraph (v1.4.2), dplyr (v1.1.1), ggsignif (v0.6.4) and stats (v4.2.3); on Python (v3.6.5): numpy (v1.19.1), pandas (v1.0.3), matplotlib (v3.3.2), seaborn (v0.10.0), scikit-learn (v0.22.1), opencv-python (v4.4.0.44), h5py (v2.10.0), PIL (v8.0.1), tifffile (v2020.6.3), json (v2.0.9) and mpmath (v1.1.0). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Reporting Summary Transparent Peer Review file Supplementary Information Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Reporting Summary Transparent Peer Review file Source data Source Data Source Data