Main

Half of the population will meet criteria for at least one psychiatric disorder during their lifetime7, with many meeting criteria for multiple disorders1. High levels of psychiatric comorbidity complicate efforts to differentiate among psychiatric disorders. These challenges are heightened because psychiatric disorders are defined by signs and symptoms, as the underlying pathophysiologies remain largely unclear. Rapid progress in psychiatric genomics has identified hundreds of associated loci (genetic variants), many of which exhibit pleiotropic (shared) associations across disorders, and revealed high correlations in genetic liability across disorders8.

The present analyses represent the third major study from the Psychiatric Genomics Consortium Cross-Disorder working group9 (CDG3). Here we examined the shared and unique influences of common genetic variants across 14 psychiatric disorders. Triangulating across multiple, complementary analytic approaches, we dissected the genetic architecture across disorders at the genome-wide, regional, functional and individual genetic variant levels. Our results have implications for refining clinical nosology and repurposing and developing novel treatments.

GWAS data for 14 psychiatric disorders

A summary of the datasets is provided in Extended Data Table 1. Psychiatric disorders were included if described in a psychiatric diagnostic manual10,11 and power was sufficient to interpret genetic correlations4. This reflects a major update relative to previous CDG1 (ref. 12) and CDG2 (ref. 5) analyses (average case increase of around 165% above CDG2; Supplementary Fig. 1), with new genome-wide association studies (GWASs) for all eight disorders from CDG2: attention-deficit/hyperactivity disorder (ADHD), anorexia nervosa (AN), autism spectrum disorder (ASD), bipolar disorder (BIP), major depression (MD), obsessive–compulsive disorder (OCD), schizophrenia (SCZ) and Tourette’s syndrome (TS)13,14,15,16,17,18,19,20. We added six additional disorders: alcohol-use disorder (AUD)21, anxiety disorders (ANX)22, post-traumatic stress disorder (PTSD)23, nicotine dependence assessed using the Fagerström test for nicotine dependence (NIC)24, opioid-use disorder (OUD)25 and cannabis-use disorder (CUD)26. The three substance-use disorders (SUDs) are novel relative to a more recent cross-disorder analysis27, and sample size increases were significant for previously included disorders (average case increase of around 287%). The sample sizes, and therefore the power of the disorder GWAS, differed (Extended Data Table 1 (Neffective)).

Owing to an uneven representation of ancestral groups, the full set of cross-disorder analyses was restricted to GWAS summary statistics from a single genetic ancestry group—European-like (EUR-like)—defined on the basis of genetic similarity to European descent in global reference panels28. We also report bivariate results for MD29 and SCZ30 in East-Asian-like (EAS-like) genetic ancestry groups and AUD31, CUD26, OUD25 and PTSD23 in African-like (AFR-like) genetic ancestry groups similarly defined based on reference panels.

Genome-wide genetic correlations

Genetic correlations (rgs) estimated using linkage disequilibrium (LD) score regression (LDSC)4 revealed pervasive genetic overlap across disorders at the genome-wide level, with clusters of disorders demonstrating particularly high genetic overlap in individuals of EUR-like genetic ancestry (Fig. 1; Supplementary Table 1; see Supplementary Figs. 2–4 for consideration of high rg across PTSD and MD). The LDSC estimates within AFR-like participants were not significant, due to limited power (Supplementary Table 4). The rg between MD and SCZ in EAS-like participants (rg = 0.45, s.e. = 0.09) was double that observed in EUR-like participants (rg = 0.22, s.e. = 0.04), which has been shown29 to be driven by a single cohort of severe and recurrent MD32.

Fig. 1: Genome-wide structural models.

a, Heatmap of rgs across the 14 disorders as estimated using LDSC on the lower diagonal and the correlations among the psychiatric factors as estimated using GenomicSEM above the diagonal. Two-sided P values were derived from the Z-statistics, calculated as the point estimate of the rg divided by its s.e. Cells depicted with an asterisk reflect values that were significant at a Bonferroni-corrected threshold for multiple comparisons. Exact values are reported in Supplementary Table 1. Disorders that load on the same factor are shown in the same colour. Per the legend at the bottom of the panel, darker blue shading indicates larger, positive rgs. LDSC estimates were used as the input to genomic SEM to produce the results in b and c.** b**, Estimates from the five-factor model along with standard errors in parentheses. Estimates are standardized relative to SNP-based heritabilities, where this is equal to the sum of the squared factor loading (the single-headed arrow(s) from the factor to the disorder) and the residual variance (the values on the double-headed arrows on the single-colour circles with text labels that begin with u). Disorders are shown as pie charts; the proportion of residual variance is shaded in purple and the variance explained by the psychiatric factors is shaded in the colour of the corresponding factor. c, Standardized estimates from the p-factor model. The disorders are colour coded as in b, and the first-order factors (F1–F5) are also colour coded to show variance explained by the second-order p-factor in yellow.

As the majority of analyses were restricted to participants of EUR-like genetic ancestry, we sought to gauge how generalizable our findings were across ancestral groups. We achieved this using Popcorn33, which can estimate rgs for the same trait across ancestral groups. We estimated the genetic impact correlation (ρgi), which considers different allele frequencies across populations by calculating the correlation between the population-specific, allele-variance-normalized single-nucleotide polymorphism (SNP) effect sizes. The results were underpowered for many comparisons, but included a strong EAS–EUR correlation for SCZ (ρgi = 0.85, s.e. = 0.04), followed by lower correlations between EAS-like and EUR-like for MD (ρgi = 0.67, s.e. = 0.16) and for AFR-like and EUR-like PTSD (ρgi = 0.59, s.e. = 0.27; Supplementary Table 4). While these results suggest that the findings that follow for EUR-like ancestry groups may generalize better for some disorders (such as SCZ) than for others (for example, PTSD and MD), that conclusion awaits replication in more highly powered analyses.

MiXeR reveals pervasive genetic overlap

Genome-wide rgs from LDSC indicate shared genetic risk across psychiatric disorders. However, LDSC may underestimate the extent of genetic overlap if shared causal variants reflect a mixture of directionally concordant and discordant associations. We applied bivariate causal mixture modelling (MiXeR) to quantify the degree of genome-wide polygenic overlap reflecting the total number of shared causal variants regardless of magnitude or directionality6. Cross-trait analyses were limited to MD, SCZ, BIP, ANX, ADHD, PTSD, AUD and AN, because other disorders were underpowered (Methods; results for univariate MiXeR are reported in Supplementary Table 5 and Extended Data Fig. 1). Supplementary Fig. 5 displays cross-trait MiXeR results for pairwise overlap across four particularly well-powered disorder samples: ADHD, SCZ, BIP and MD (complete results are shown in Supplementary Figs. 6–9 and Supplementary Table 6). There was greater polygenic overlap across psychiatric disorders than suggested by the rgs from LDSC. Overall, MiXeR results suggested that the shared genetic signal for psychiatric disorders primarily reflects variants with concordant effects across disorders, while differentiation in genetic risk is driven by fewer shared discordant or unique variants.

Genomic SEM identifies five factors

We used genomic structural equation modelling (genomic SEM)27,34 in the EUR-like genetic ancestry datasets to model genetic overlap from LDSC across 14 disorders as latent factors representing dimensions of shared genetic risk (Methods). A five-factor model (Supplementary Tables 2 and 3) provided the best fit to the data (comparative fix index (CFI) = 0.971, standard root mean square residual (SRMR) = 0.063). These five latent genomic factors (capitalized throughout, to distinguish them from the psychiatric disorders that define them) (Fig. 1) comprised: F1, a Compulsive disorders factor defined by AN, OCD and, more weakly, TS and ANX; F2, a SB factor defined by SCZ and BIP; F3, a Neurodevelopmental factor defined by ASD, ADHD and, more weakly, TS; F4, an Internalizing disorders factor defined by PTSD, MD and ANX; and F5, a SUD factor defined by OUD, CUD, AUD, NIC and, to a lesser extent, ADHD.

Within this five-factor model, Internalizing disorders and SUD factors displayed the highest interfactor correlation (rg = 0.60; s.e. = 0.02). The median residual genetic variance unexplained by the latent factors was 33.5%, indicating that most genetic risk was shared among disorder subsets. TS displayed the most unique genetic signal, with 87% of its genetic variance unexplained by the factors. The structure of the first four factors was similar to that found by genomic SEM applied to subsets of these disorders in previous work5,27, indicating stability in the underlying factor structure, even as sample sizes and the number of disorders have increased. The newly added SUD traits formed the fifth factor.

Evidence of moderate rg between factors suggests that a higher-order factor may explain common variance across the correlated factors. Consistent with this observation, a hierarchical model also fit the data well (CFI = 0.959, SRMR = 0.074). We refer to this as the *p-*factor model, which included a higher-order general psychopathology factor defined by the five lower-order psychiatric factors (such as SUD). Internalizing loaded most strongly on p (0.95), with the other 4 factors having moderate loadings (0.50–0.63).

As some of the underlying data were obtained using brief, self-reported diagnoses, we performed a sensitivity analysis in which those data were excluded (Supplementary Note 1, Supplementary Tables 7–11 and Supplementary Figs. 10–18). The rg matrix was largely unchanged; the five-factor model identified in the full sample continued to provide good fit to the data and produced similar point estimates, and downstream GWAS analyses (detailed below) identified similar loci.

Genetic correlations with factors

We estimated rgs between the five correlated factors, hierarchical p-factor and 31 complex traits (Supplementary Table 12) to place shared genetic liability indexed by the factors in a broader clinical context. These factors vary in their use for capturing shared genetic signal; accordingly, we used the QTrait heterogeneity statistic to assess this use at the genome-wide level. When QTrait is significant, this indicates a trait’s rg deviates from the factor structure. For example, if trait* X* is negatively correlated with SCZ but unrelated to BIP, QTrait would probably be significant, suggesting that trait X lies outside the shared signal captured by the factor. Significant correlations were defined at a Bonferroni-corrected threshold of P < 2.68 × 10−4, while not significant for QTrait at this same threshold. This QTrait exclusion criteria was relaxed for the p-factor if that trait was significantly associated with the majority (≥3) of the five correlated factors, as this indicates the trait is capturing transdiagnostic associations the *p-*factor is intended to index.

The Internalizing disorders and SUD factors were the only factors associated with household income (rg_Internalizing = −0.40, s.e. = 0.02; rg_SUD = −0.41, s.e. = 0.03; Fig. 2) and were the most pervasively associated with different cognitive outcomes, including childhood intelligence (rg_Internalizing = −0.27, s.e. = 0.05; rg_SUD = −0.40, s.e. = 0.07). Only the SUD factor was associated with adult intelligence (rg_SUD = −0.40, s.e. = 0.03) and verbal numerical reasoning (rg_SUD = −0.41, s.e. = 0.03). This was compared to more circumscribed cognitive associations for the Compulsive disorders and SB factors, including a large negative correlation with the pairs matching test (potentially indexing memory; rg_Compulsive = −0.33, s.e. = 0.03; rg_SB = −0.34, s.e. = 0.03). The SB and SUD factors were the only ones associated with risk tolerance (rg_SB = 0.31, s.e. = 0.03; rg_SUD = 0.38, s.e. = 0.03). The Neurodevelopmental factor was uniquely associated with childhood BMI (rg_Neurodevelopmental = 0.26, s.e. = 0.06) and showed high genetic overlap with childhood aggression (rg_Neurodevelopmental = 0.94, s.e. = 0.10). As would be expected, the five traits significantly associated with all five correlated factors were also among the top correlations for the p-factor, reflecting stress sensitivity (rg_p = 0.50, s.e. = 0.02), loneliness (rg_p = 0.62, s.e. = 0.02), neuroticism (rg_p = 0.64, s.e. = 0.02), self-harm (rg_p = 0.74, s.e. = 0.04) and suicide attempts (rg_p = 0.87, s.e. = 0.03). The full set of correlations is shown in Supplementary Table 13; comparison across factors is shown in Extended Data Fig. 2; and comparison across traits within each factor is shown in Extended Data Fig. 3.

Fig. 2: External trait genetic correlations for psychiatric factors.

Point estimates for the rgs between 14 external traits and the 5 psychiatric factors from the correlated factors model and the p-factor from the hierarchical model. These traits were selected as they were significantly correlated with at least one factor at >0.35 or <−0.35. Bars depicted with a dashed outline were significant for the QTrait heterogeneity statistic, which indicates that the pattern of rgs for that trait did not fit the factor structure. Bars depicted with an asterisk reflect values that were significant at a Bonferroni-corrected threshold for multiple comparisons, that were also not significant at this same Bonferroni corrected threshold for* Q*Trait. This is with the exception that the *p-*factor is depicted with an asterisk even if it is significant for the QTrait, provided that the same trait was significantly correlated with the majority (at least three) of the five other factors. The two-sided P values used to evaluate significance were derived from the Z-statistics, calculated as the point estimate of the rg divided by its s.e. Error bars are ±1.96 s.e., centred around the point estimate of the rgs. Traits are ordered by the point estimate for the *p-*factor. The implied sample size for the psychiatric factors was: Compulsive ((\hat{n}) = 54,100), SB ((\hat{n}) = 127,202), Neurodevelopmental ((\hat{n}) = 84,760), Internalizing ((\hat{n}) = 1,637,337), SUD ((\hat{n}) = 313,395) and p-factor ((\hat{n}) = 2,168,621). Sample sizes for the external traits are reported in Supplementary Table 12 and exact P values are reported in Supplementary Table 13.

LAVA finds regional hotspots of overlap

Global estimates of pleiotropy, such as the genome-wide rgs from LDSC, provide an average of the degree of shared signal across the genome. However, as genetic overlap is unlikely to be constant across genomic regions, we segmented the genome into 1,093 LD-independent regions, and applied local analysis of (co)variant association (LAVA35; Methods) to assess the rg between disorders within these regions. In addition to capturing heterogeneity in genetic overlap and pinpointing relevant regions, LAVA identifies potential rg hotspots shared among several disorders, thereby providing further insight into genetic architecture.

We restricted analyses to loci with sufficient SNP-based heritability for the disorders analysed (P < 4.6 × 10−5 = 0.05/1,093; Methods). Correcting for the number of bivariate tests performed across all regions and disorder pairs, we detected 458 significant pairwise local rgs (P < 2.1 × 10−6 = 0.05/24,273). The pairs of disorders with the greatest number of local rg hits were MD and ANX (113 regions), MD and PTSD (88 regions), and BIP and SCZ (40 regions), accounting for over half of all significant local rgs detected (Fig. 3a). This is consistent with the genome-wide levels of overlap indicated through the LDSC global rg (Fig. 1), the polygenic overlap estimated with MiXeR (Supplementary Figs. 5–9), and the multivariate genetic structure identified by genomic SEM. Both global and local rgs tended to be positive, with significant negative rgs identified in only three instances (Supplementary Fig. 19). This indicates that the genetic risk for one disorder typically increases the risk for another (Supplementary Fig. 20).

Fig. 3: Local genetic correlations.

a, An overview of the average patterns of local rgs across the genome for all pairs of disorders, shown as a heatmap (below diagonal) and a network plot (above diagonal). The colours of the heatmap represent the average local rgs across all evaluated loci, with darker red and blue shading indicating more negative and positive rg, respectively; the dot size reflects the strengths of average associations; and the numbers indicate how many of the local rgs were significant. These results are mirrored in the network plot, where the width or the edges reflect the number of significant associations, meaning that only disorders with at least one significant local rg are connected, and the edge opacity reflects the strength of the average local rg across tested loci. Note that label colours are concordant with the genomic SEM factor structure from Fig. 1 and, as shown, disorders of similar colours also tend to be proximally located within the network. b, The local rg structure within the top rg hotspot on chromosome (chr.) 11 (112755447–114742317, GRCh37 reference genome), that is, the region where the greatest number of significant rgs were found across all disorder pairs. Here, the network plot illustrates all significant rgs detected in this region, with both edge width and opacity reflecting the strength of the association. The region plot in the middle displays the genes contained within the hotspot, and the table below shows the rg estimates (Rho), 95% confidence intervals (CIlower, CIupper), variance explained (R2) and P values for all significant pairwise local rgs in this region. Label colours are again concordant with those used for the genomic SEM factor structure in Fig. 1.

We detected 101 regions that contained significant local rgs between several disorder pairs, which we call rg hotspots (see Supplementary Tables 14–23 for local rgs across disorders in the top 10 hotspots). The most pleiotropic of these hotspots was on chromosome 11, which contained 17 positive and significant local rgs involving 8 of the 14 analysed disorders (Fig. 3b). This region also stands out as the most significantly associated with 8 of these 17 disorder pairs, while ranking in the top 25% of associated loci for 12 of them (Supplementary Fig. 21). Notably, this region contains the NCAM1–TTC12–ANKK1–DRD2 gene cluster that has been frequently associated with psychiatric phenotypes36,37,38,39, and flagged as a likely pleiotropy hotspot for a range of cognitive and behavioural outcomes related to, for example, intelligence, personality, substance use and sleep35,40,41,42.

Risk loci for psychiatric factors

We used multivariate GWAS within genomic SEM34 to identify SNPs associated with the factors from the five-factor model or the p-factor in the hierarchical model. Similar to the* Q*Trait metric, we estimated factor-specific QSNP heterogeneity statistics. This indexes SNPs that deviate strongly from the factor structure, due to either disorder-specific or directionally discordant effects. We defined genomic hits for the factors as those that were significant after Bonferroni correction (P < 5 × 10−8/6 genomic factors) and did not overlap with QSNP hits for that factor (Methods). Most hits were identified for the SB (n = 102) and Internalizing (n = 150) factors. After merging overlapping loci across the five correlated factors, 238 unique hits remained, including 27 broadly pleiotropic loci associated with two or more factors. The hierarchical model identified 160 hits for the p-factor (Fig. 4, Supplementary Fig. 22 and Supplementary Tables 24–36), 57 of which were not identified in the five-factor model (295 unique hits across both models). Forty-eight hits were novel relative to the univariate GWAS, of which 38 have been described in previous GWAS for a broad range of outcomes, and 10 are entirely novel (Supplementary Table 37).

Fig. 4: Locus-level results.

a, Heatmap of CC-GWAS loci below the diagonal across pairwise combinations of disorders; the darker orange shading indicates a higher number of CC-GWAS hits. CC-GWAS results are not shown for the Internalizing disorders as their rgs were too high, or for nicotine dependence as this is a continuously measured trait. Genomic SEM results (number of hits and mean χ2 for each factor and factor-specific QSNP estimate) are reported above the diagonal. Results for the p-factor are shown above the plot along with a Venn diagram of the overlap between p-factor, p-factor QSNP and overall CC-GWAS hits. The disorders are ordered and coloured according to the genomic SEM factor structure from Fig. 1.** b**,c, The Miami and QQ-plots for the p-factor (b) and SBs factors (c), respectively. These panels show the results for the −log10-transformed two-tailed P values for the factor on the top half of the Miami plot and the log10-transformed one-tailed P values for QSNP on the bottom half. Factor hits that were within 100 kb of univariate hits are shown as black triangles, novel hits for the factors that were not within 100 kb of a univariate or QSNP hit are shown as red triangles and QSNP hits are shown as purple diamonds. d, The two-tailed −log10[P] in a Manhattan plot for the CC-GWAS comparison across MD and SCZ, which produced the most hits (orange diamonds), as well as the scatterplot of standardized case–control effect sizes of MD (x axis) versus SCZ (y axis), with CC-GWAS significant SNPs labelled in red. For b–d, the grey dashed lines indicate the significance threshold, which was defined using Bonferroni correction for multiple comparisons.

We identified 33 unique hits with significant QSNP effects across the factors from the five-factor model. By comparison, we identified 117 QSNP hits from the p-factor model that showed significantly divergent effects across the five, lower-order psychiatric factors (Supplementary Table 36). These p-factor QSNP hits also included the chromosome 11 LAVA hotspot, where this region was found not to confer transdiagnostic risk due to an absence of signal for the Neurodevelopmental factor. For the SUD factor, highly significant QSNP hits were driven by variants in the genes involved in biological pathways specific to particular psychoactive substances, including the alcohol dehydrogenase genes (ADH1A, ADH1B and ADH1C) for AUD and the CHRNA3–CHRNA5–CHRNB4 nicotinic receptor subunit gene cluster for NIC. More QSNP loci for the *p-*factor model relative to the five-factor model indicates that many shared genetic relationships are better captured by the five factors (Supplementary Figs. 23 and 24).

A phenome-wide association study conducted in the Mayo Clinic Biobank revealed that factor hits were associated with multiple psychiatric disorders, especially those that loaded on the factor (Supplementary Table 38 and Supplementary Fig. 25). The Internalizing disorders (Supplementary Fig. 25d) and *p-*factor (Supplementary Fig. 25f) loci were also associated with a range of medical outcomes (for example, chronic pain and hypertension).

Divergent loci across disorders

In more fine-grained analyses of disorder pairs, case–case GWAS (CC-GWAS)43 was used to identify loci with different allele frequencies across cases of different disorders. Such loci may reflect distinctive genetic effects across disorder pairs. CC-GWAS was applied to 75 disorder pairs, comparing 13 disorders. NIC was excluded because it is a continuous trait, and the pairs ANX–MD, ANX–PTSD and MD–PTSD were excluded because all had an rg estimate of >0.8, thereby risking an inflated type I error rate (Methods). The genome-wide significance threshold was defined at 5.5 × 10−10 (that is, 5 × 10−8/91 pairwise comparisons). An overview of CC-GWAS input parameters is provided in Supplementary Table 39.

In total, 412 loci showed significantly different effects across the 75 disorder pairs (Supplementary Tables 40 and 41); most (294 out of 412) were in comparisons that included SCZ, possibly reflecting either greater power for the SCZ GWAS or more distinctive biology for this disorder. Owing to overlap among the hits, the 412 loci comprised 109 LD-independent loci (Supplementary Table 42). Five of these were CC-GWAS specific, implying that they were not significantly associated with case–control status in either of the disorders in the respective disorder pair. CC-GWAS also computes a genome-wide genetic distance between the cases of two disorders (FST,causal), indicating how genetically dissimilar the cases are on average. As expected, these genetic distances were inversely correlated (r = −0.79, s.e. = 0.07) with rg (Supplementary Table 43). In support of the five-factor model, >99% of the CC-GWAS hits were identified for disorder pairs that loaded on separate factors (Supplementary Tables 44 and 45). Disorders that cluster on the same factor from the five-factor model are, apparently, largely indistinguishable at the level of individual genetic variants.

Functional annotation

Enrichment analyses

To understand biological functions influenced by the risk loci, we prioritized candidate risk genes implicated by the multivariate GWAS loci using expression quantitative trait loci (eQTL)44,45 and Hi-C44,46 datasets collected from fetal and adult brain samples (Methods and Supplementary Tables 46 and 47). Owing to the limited number of variants associated with other factors, analyses were restricted to the *p-*factor, the SB and Internalizing disorders factors and QSNP for these latter two factors. We first compared the target gene expression along the temporal trajectory of human brain development, finding that genes associated with the three factors were expressed at higher levels than QSNP target genes across the lifespan, with the largest difference observed at fetal stages and early life (Fig. 5 and Supplementary Fig. 26). This suggests that pleiotropic variants are involved in early, fundamental neurodevelopmental processes. We next examined biological processes using Gene Ontology (GO) enrichment analysis47. The target genes of the *p-*factor were primarily enriched in broader biological processes related to gene regulation (Fig. 5). To enhance the specificity of the gene sets, we removed Internalizing disorders and SB target genes that also appeared for the *p-*factor. SB (minus p-factor) target genes were enriched in more specific terms related to neuron development. No significant results were identified for the Internalizing disorders factor, probably reflecting the large proportion of target genes overlapping with the *p-*factor. Results from MAGMA48 (Supplementary Methods) provided convergent support for the role of early neurodevelopmental processes in transdiagnostic psychiatric risk. Specifically, genetic signal for the five correlated factors and *p-*factor showed enrichment in genes identified from rare variant studies of ASD49,50,51, neurodevelopmental delay49 or both (Supplementary Fig. 27).

Fig. 5: Functional annotation of factor variants.

a, GO enrichment analysis of predicted target genes with transdiagnostic associations (that is, variants associated with the *p-*factor), or those target genes associated with the SB factor that were not overlapping with *p-*factor target genes. Depicted −log10-transformed P values are one-sided, calculated using a χ2 test; false-discovery rate (FDR) correction was applied for multiple comparisons. b, The averaged and normalized expression levels of target genes of the indicated classes along the temporal trajectory of human brain development. Shading around the lines reflects 95% CIs. pcw10, post-conception week 10. c,d, Average log10[P] values across EWCE and MAGMA enrichment for genes associated with the indicated factors in fetal brain cell types using two independent single-cell RNA-sequencing (scRNA-seq) datasets53,54 (c) or adult brain cell types using three independent single-nucleus RNA-seq (snRNA-seq) datasets55,56,57 (d). The P values from EWCE and MAGMA were two-sided and each had an FDR correction applied for multiple comparisons before averaging the two sets of results. EWCE P values were empirically derived using a permutation test; MAGMA P values were calculated using an F-test. Int, Internalizing disorders factor. The implied sample size for the three depicted psychiatric factors was: SB ((\hat{n}) = 127,202), Internalizing ((\hat{n}) = 1,637,337) and p-factor ((\hat{n}) = 2,168,621). CycProg, cycling progenitor; Endo/BBB, endothelial/blood brain barrier; ExNeu, excitatory neuron; InNeu, interneurons; IP, intermediate progenitor; OPC, oligodendrocyte progenitor cell; RG, radial glia; Astro, astrocyte; MSN, medium spiny neuron; ODC/Oligo, oligodendrocyte.

Averaged results across expression-weighted cell type