Shared genetic architecture between autism spectrum disorder, loneliness, and social isolation reveals novel genetic loci

Abstract

Objective

Deficits in social communication and social interaction are core features of autism spectrum disorder (ASD), and studies suggest that loneliness and social isolation are common. ASD has a strong genetic basis, but the genetic architecture and overlap with social phenotypes are not clear.

Methods

We analyzed summary statistics from genome-wide association studies on ASD (46 350), loneliness (452 302), and social isolation (288 950), using linkage disequilibrium score regression, local analysis of covariant annotation (LAVA), bivariate causal mixture model (MiXeR), and the conditional/conjunctional false discovery rate (cond/conjFDR).

Results

For ASD and social isolation, we found nonsignificant global genetic correlation (r_g = 0.02, P = 0.8), but LAVA identified 72 genomic regions with bidirectional correlations, and MiXeR estimated that 8.7 k of 13.1 k variants (81%) were shared, of which 53% had concordant effect directions. For ASD and loneliness, we found a positive genetic correlation (r_g = 0.26, P = 2e-10), LAVA identified 80 genomic regions with bidirectional genetic correlations, and MiXeR suggested that at least 3.8 k variants were shared. We identified nine specific shared genetic loci between ASD and loneliness and eight between ASD and social isolation (conjFDR < 0.05). Of these, 12 loci were novel for ASD. Genes mapped to these loci are involved in γ-aminobutyric acid (GABA), glutamate, calcium, and stress hormone signaling, cerebral glucose transport, TAU-accumulation, and immune function.

Conclusion

We found extensive overlap in genetic architecture between ASD, loneliness, and social isolation, with bidirectional effects. By leveraging data for ASD and social traits, we identified 12 novel ASD related genetic loci implicating several genes, thereby elucidating potential pathways underlying their shared genetic architecture.

Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental disorder affecting 1–2% of the population (Edition, 2013; Yuan et al., 2021). The core symptoms are persistent deficits in social communication and social interaction as well as restrictive, repetitive patterns of behavior, interests, or activities (Delobel-Ayoub et al., 2020). However, it is a complex condition with substantial phenotypic heterogeneity. As a group, persons with ASD are less likely to participate in social interactions (Kaale et al., 2018), have a higher frequency of comorbid mental disorders, and increased risk of somatic diseases (Koukouriki and Soulis, 2020; Pan et al., 2020).

ASD has a strong genetic basis, with estimated heritability of ~80% (Sandin et al., 2017). The genetic architecture of ASD is complex, comprising both rare pathogenic mutations and common variants (Sebat et al., 2007; Iossifov et al., 2014). These common variants each have a small effect, but are suggested to account for a large proportion of the genetic liability for ASD (Gaugler et al., 2014), and by acting additively, they may be clinically relevant (Torske et al., 2020; LaBianca et al., 2021). Although several common genetic variants have been discovered for ASD, sample sizes have been too small to detect a large proportion of the heritability (Grove et al., 2019). The largest genome-wide association study (GWAS) so far, including 18 381 cases, identified five genome-wide-significant loci, and by leveraging GWAS results from three phenotypes with significantly overlapping genetic architectures (schizophrenia, major depression, and educational attainment), seven additional loci were discovered (Grove et al., 2019). Thus, only 12 significant ASD-associated loci have been identified to date (Grove et al., 2019), and a large fraction of the polygenic architecture for ASD remains unknown. While rare pathogenic mutations can inform biology, the common genetic risk variants for ASD is of relevance for understanding shared genetic architecture (Gaugler et al., 2014).

Studies has reported that genetic liability to social symptoms of ASD is inherited largely independent from the liability to nonsocial symptoms (Robinson et al., 2012; Warrier et al., 2019). This is line with the fractionation theory of autism, which suggests that social, communicative, and rigid/repetitive symptom domains are independent traits with distinct causes (Happe and Ronald, 2008). Furthermore, it has become increasingly clear that autistic symptoms are not only present in persons diagnosed with ASD but also in persons with subclinical symptoms in the general population. This suggests that genetic liability for ASD might have similarities with genetic liability underlying social traits in the general population.

Loneliness and social isolation are social traits with somewhat different phenomenon (Peplau, 2022). ‘Loneliness’ reflects a subjective feeling of being alone, resulting from a discrepancy between the desired and actual level of social relationships, regardless of the amount of social interaction (Peplau, 2022), while ‘social isolation’ refers to an objective absence or paucity of social interactions (Poscia et al., 2018). Both loneliness and social isolation have profound negative consequences on mental and somatic health, and are associated with increased mortality and severe mental disorders (Hodgson et al., 2020; Rødevand et al., 2021; Elovainio et al., 2022) and are recognized as public health concerns (O’Sullivan et al., 2022). Studies suggest that there is a broad variation among people with ASD with respect to loneliness and isolation. Some people may choose to separate themselves from others due to reduced social motivation and do seldom feel lonely (Deckers et al., 2017; Bennett et al., 2018, 2019), while others possess a desire for social connections and may often feel lonely (Jaswal and Akhtar, 2019; Stice and Lavner, 2019; McGhee Hassrick et al., 2020).

The heritability of loneliness is estimated to be 40–50% in twin- and family-based studies (Devlin et al., 2013; Gao et al., 2017), and a recent GWAS from the UK Biobank identified genetic loci associated with both loneliness and social isolation (Day et al., 2018). Thus, how lonely someone feels (subjective) and how often one socializes (objective) are partly influenced by genetic variants, and those variants may also possibly influence social traits in ASD. We therefore hypothesized that there might be genetic overlap between ASD and social traits.

The bivariate causal mixture model (MiXeR) is a statistical method for estimating the total number of shared genetic variants, irrespective of genetic correlations between traits (Frei et al., 2019). It allows for the detection of genetic overlap, despite a mixture of effect directions, that would otherwise be missed with analysis of genetic correlation using well-known methods such as linkage disequilibrium score regression (LDSR) (Bulik-Sullivan et al., 2015). Applying these tools has previously shown significant polygenic overlap, despite low genetic correlation, between a series of mental traits and disorders (Bahrami et al., 2021). Furthermore, genetic correlation can be further characterized by a local analysis of covariant annotation (LAVA). LAVA is a novel tool that enables genome-wide local genetic correlation analysis, which takes into account genetic correlations confined to specific genomic regions across the genome, and to correlations with opposing directions at different loci. Finally, if genetic overlap is present between two phenotypes, combining GWASs of these traits will increase the statistical power that can be leveraged by the conditional/ and conjunctional false discovery rate (cond/conjFDR) approach. The conditional FDR (condFDR) method uses the increased power to detect novel genetic variants associated with one trait, while the conjunctional FDR (conjFDR) identifies loci jointly associated with two traits.

Here, we aimed to investigate the genetic relationship between ASD, loneliness, and social isolation within a comprehensive analytical framework to gain more knowledge of the genetic architecture of ASD (Frei et al., 2019; Smeland et al., 2020; Bahrami et al., 2021; Werme et al., 2022). We anticipated that there would be overlapping genetic architecture between ASD, loneliness, and social isolation, and aimed to characterize overlapping genetic architecture by novel statistical genetics methods as MiXeR and LAVA. Further, we intended to leverage the genetic overlap to boost the statistical power to discover more common genetic variants associated with ASD (condFDR), and variants jointly associated with ASD and loneliness and social isolation (conjFDR). Finally, by applying functional analyses of the associated loci (Watanabe et al., 2019), we expected to gain new insight into the molecular pathways of ASD, which may form the basis for a better understanding of molecular functions and potential new treatment targets.

Methods

Samples

Autism spectrum disorder

We acquired GWAS summary statistics on ASD from the Psychiatric Genomics Consortium (Sullivan et al., 2018). The ASD dataset included 18 381 cases and 27 969 controls (Grove et al., 2019), see Table 1.

Table 1.

Genome-wide association studies characteristic

Sample	Sample size			Age group	Reference
	Total N	Cases	Controls
ASD	46 350	18 381	27 969	Adults and children	Grove et al. (2019)
Loneliness	445 024	80 134	364 890	Adults	Day et al. (2018)
Social isolation	288 950	2426	286 524	Adults	Day et al. (2018)

ASD, autism spectrum disorder.

Loneliness and social isolation

The GWAS data were analyzed using Multi-Trait Analysis of GWAS (MTAG) (Turley et al., 2018). This method accounts for sample overlap, and the effect direction of the primary trait will not be influenced by the effect direction of the related traits.

Our loneliness variable was based on summary statistics from the GWAS on loneliness published in the study of Day et al (2018) and was downloaded from the UK Biobank. The trait was based on self-reported answers (N = 452 302) to questions regarding (1) perceived feeling of loneliness, (2) ability to confide in someone close, and (3) frequency of social interactions with friends and family. Using MTAG, a loneliness composite variable was computed using question (1) (perceived feelings of loneliness) as the primary trait and using questions (2) and (3) as related, supporting traits. See Supplementary Methods, Supplemental digital content 1, https://links.lww.com/PG/A334, for more detailed information. Those who answered that they perceived feelings of loneliness were defined as cases, and those who answered otherwise were defined as controls (see Table 1). The effect direction for loneliness was set as positive for those who experienced loneliness.

Our social isolation variable was based on GWAS regarding social interactions and engagement in social activities (Day et al., 2018). The GWAS regarding engagement in social activities (social, sports, or religious groups) were downloaded from the UK Biobank, while the GWAS on social interaction was obtained by contacting the first author of the study (Day et al., 2018). Using MTAG, we computed a social isolation variable by defining social interaction as the primary trait and engagement in social activities as related, supporting traits. Those who answered that they lived alone and never or very seldom had social interactions with friends or family were recorded as cases, and those who answered otherwise were recorded as controls (Table 1). The effect direction for social isolation was set as positive for those who never or very seldom had social interactions. See Supplementary Methods, Supplemental digital content 1, https://links.lww.com/PG/A334 for more detailed information.

Data analysis

Global and local genetic correlations

Based on the suggested shared genetic architecture and social traits, we applied condFDR, an approach that reranks the test statistics of the primary phenotype (ASD) based on the strength of the association with the secondary phenotype (loneliness or social isolation). We interpreted conditional Q–Q plots and reversed conditional Q–Q where crosstrait enrichment is visualized as successive leftward shifts from the null line.

We then investigated the global genome-wide genetic correlation (r_g) using LDSR, which is a global summary of the correlation of all single-nucleotide polymorphisms (SNPs) after controlling for LD (Bulik-Sullivan et al., 2015). LDSR is unable to capture genetic overlap in the presence of a mixture of concordant and discordant effect directions, as they may cancel each other out (Frei et al., 2019).

Therefore, the LAVA analysis (Werme et al., 2022) was applied following the protocol described in the original article, using the LD reference panel based on the 1000 Genomes phase 3 genotype data for European samples (Auton et al., 2015). The genome was partitioned into 2495 nonoverlapping semiindependent regions, defined based on LD, with an average size of 1 Mb, as described when LAVA was introduced (Werme et al., 2022).

To adjust for multiple comparisons, we used FDR with a limit of FDR < 0.05, which is in line with common procedures for controlling for type I errors in GWAS analysis (Kim et al., 2021).

Estimating total genetic overlap

MiXeR is a statistical tool developed for quantifying polygenic overlap irrespective of genetic correlation. As for Q–Q plots, which are suitable for visual assessment of the genetic, they do not quantify it. For this reason, we use the MiXeR method, as it allows us to quantify the shared genetic architecture while efficiently accounting for subthreshold genetic signals. This MiXeR method estimates the total number of shared and ‘trait-influencing’ variants (i.e. variants with nonzero additive genetic effects on a trait) (Frei et al., 2019). More details about MiXeR are presented in Supplementary Methods, Supplemental digital content 1, https://links.lww.com/PG/A334. The Akaike Information Criterion (AIC) for the best-fitting MiXeR model compared to a reference model was used to evaluate the ability of the MiXeR to predict the actual GWAS data. A positive AIC value indicates that the polygenic overlap is beyond the minimal level needed to explain genetic correlation, while a negative AIC indicates that MiXeR cannot accurately distinguish the best-fitting estimates from the reference model.

snp2gene

To identify genetic loci shared between ASD and social traits, we employed the condFDR/conjFDR statistical framework (Smeland et al., 2020).

This approach reranked the test statistics of a primary phenotype (e.g. ASD) based on the strength of the association with a secondary phenotype (e.g. loneliness or social isolation). In the conditional Q–Q plot, this crosstrait enrichment is visualized as successive leftward shifts from the null line. Further, the condFDR method is a statistical approach that builds on an Empirical Bayesian statistical framework by combining GWAS summary data to increase the statistical power to detect shared genetic risk variants (Andreassen et al., 2013). The condFDR method leverages systematic colocalization of SNP associations to prioritize likely pleiotropic SNPs (Schork et al., 2016). Additionally, the conjunctional FDR (conjFDR) is an extension of the condFDR, which allows for the discovery of SNPs significantly associated with two phenotypes jointly. Together, these approaches improve the discovery rate, allow for the detection of a mixture of effect directions, and enable the detection of shared specific genomic loci that do not reach significance threshold in traditional GWAS analyses (Frei et al., 2019; Smeland et al., 2020). We apply condFDR < 0.01 for loci associated with ASD conditional on loneliness or social isolation. Further, we search for shared genetic loci between traits at conjFDR < 0.05, which is defined as the maximum of two condFDR values after repeating the condFDR analysis for both phenotypes (Frei et al., 2019; Smeland et al., 2020). This approach identifies the localization of specific shared variants. The model fit analyses were performed after excluding the major histocompatibility region (chr6:25000000–34000000), the 8p23 inversion (chr8:7200000–12500000), and the MAPT region (chr17:40000000–47000000) to avoid inflated enrichment. See Supplementary Methods, Supplemental digital content 1, https://links.lww.com/PG/A334 for more details of the analyses.

Genomic loci definition and functional annotation

We defined independent genomic loci using Functional Mapping and Annotation of geetic associations (FUMA) (Watanabe et al., 2019). Significant independent SNPs were defined as conjFDR < 0.05 and r² < 0.6. Lead SNPs were chosen if they were in approximate linkage equilibrium with each other (r² < 0.1). Due to the technical limitation of the FUMA web interface, which allows setting the level of genome-wide significance to 1E-5 or smaller, we had to scale the condFDR/conjFDR levels in our summary statistics before uploading them to FUMA. Specifically, we kept the default threshold for genome-wide significance at 5E-8 and divided the conjFDR and condFDR values by 1 000 000 and 200 000, respectively, to match our significance thresholds of conjFDR = 0.05 (0.05/1 000 000 = 5E-8) and condFDR = 0.01 (0.01/200 000 = 5E-8).

Candidate SNPs, defined as any SNP within each jointly associated genomic locus with a conjFDR value <0.10 and an LD r² < 0.6 with an independent significant SNP, were functionally annotated using FUMA ‘SNP2Gene’ with default parameters (Watanabe et al., 2019). Lead SNPs were mapped to genes by three methods. We used (1) positional mapping to align SNPs to genes based on their physical proximity, (2) expression quantitative trait locus (eQTL) mapping to match cis-eQTL SNPs to genes whose expression is associated with allelic variation at the SNP level, and (3) chromatin interaction mapping to link SNPs to genes based on three-dimensional DNA–DNA interactions between each SNP’s genomic region and nearby or distant genes. Thus, a ‘mapped gene’ is defined as a gene that was linked with the lead SNP by at least one of three different gene-mapping analyses. The genes that were mapped by all three strategies were defined as a ‘credible gene’. All gene-mapping strategies were limited to brain tissues. Finally, we queried SNPs for known QTLs in brain tissues using the GTEx portal (GTEx, version 8) (Aguet et al., 2017). If the gene annotation of a specific SNP was marked as ‘NA’, we searched for information in the dbSNP database (Smigielski et al., 2000).

Functional annotation were used to analyze genetic function by using FUMA ‘Gene2func’ (Watanabe et al., 2019), which provide four different outputs: (1) information about gene expression in different tissues and whether the gene set is expressed during different development periods; (2) information regarding whether the gene set has increased or reduced expression in a certain tissues (differently expressed genes); (3) information about gene set enrichment, that is, whether the gene set are overrepresented in other predefined gene sets in the GWAS catalog (Cerezo et al., 2025), MsigDB (Liberzon et al., 2015), or WikiPathways (Kutmon et al., 2016); and (4) analysis of molecular function, with links to external biological information of input genes.

Colocalization analysis of identified shared genetic loci

To assess whether the same or distinct variants are responsible for the GWAS signals within loci shared between ASD and social traits, we applied the coloc R package (Giambartolomei et al., 2014) For each shared locus identified in our conjFDR analyses, we extracted effect sizes and respective variances for all SNPs within the locus boundaries from the original GWASs of ASD and the corresponding social trait. We then applied the coloc.abf function to each locus and reported the corresponding probability of colocalization (PP4).

Gene-level analysis

In addition to mapping genes and gene sets to loci shared between ASD and the social traits identified in our conjFDR analyses, we deployed Multi-marker Analysis of GenoMic Annotation (MAGMA) v1.10 (de Leeuw et al., 2015) [and Gene Set Analysis (GSA)-MiXeR v2.2.1 (Frei et al., 2024)]tools to carry out gene-level analyses. These tools aggregate genetic signals across genes while accounting for SNPs that do not reach genome-wide significance in the analyzed GWASs.

Neither MAGMA nor GSA-MiXeR can be applied to conjFDR results, as they are designed to use P-values from standard GWAS analyses. Therefore, we applied these tools to the original GWAS summary statistics for ASD, loneliness, and social isolation, and compared the identified genes and gene sets. MAGMA gene and gene set analyses were performed using default settings. Specifically, for gene analysis, an SNP-wide mean model was applied to the GWAS summary statistics, taking gene boundaries without a surrounding window and using the 1000 Genomes Phase 3 European reference panel for LD estimation. For this analysis, we used a list of protein-coding genes (N = 19 427) provided on the MAGMA webpage (de Leeuw et al., 2015).

Gene set analysis was conducted by restricting the sets under investigation to those that are part of the Gene Ontology (GO) biological processes, GO molecular function, and GO cellular component subsets (total N = 10 529), as listed in the Molecular Signatures Database (MsigdB) v2023.1Hs (Subramanian et al., 2005).

Results were controlled for multiple testing using the Benjamini-Hochberg procedure with an FDR < 0.05 (Benjamini and Hochberg, 2018). FDR correction was applied separately for the analyses of ASD, loneliness, and social isolation. GSA-MiXeR was then used to quantify fold enrichments for the genes identified by MAGMA. AIC values were used to assess the reliability of the GSA-MiXeR estimates.

Results

Genetic enrichment in conditional Q–Q plots

We applied condFDR and got Q–Q plots for ASD conditioned on loneliness and for loneliness conditioned on ASD. These are shown in Figs. 1a, b. The plots demonstrate consistent genetic enrichment and indicate polygenic overlap between these traits. For ASD conditioned on social isolation, the Q–Q plot showed weak SNP enrichment (Fig. 1c), while the reverse Q–Q plot demonstrated enrichment for social isolation conditioned on ASD (Fig. 1d). Interpreted together, the two Q–Q plots suggest genetic overlap between social isolation and ASD.

Fig. 1.

Conditional quantile-quantile (Q–Q) plots and reversed Q–Q plots for ASD and loneliness and for ASD and social isolation. Conditional Q–Q plots of nominal vs. empirical −log10(P) values (corrected for inflation) in ASD below the standard GWAS threshold of P < 5 × 10⁻⁸ as a function of significance of association with loneliness (a) and (b) and with social isolation (c) and (d) at the significance levels of P < 0.1, P < 0.01, and P < 0.001, respectively. The blue lines indicate all SNPs. The dashed lines indicate the null hypothesis. The plots build on the condFDR method. ASD, autism spectrum disorder; GWAS, genome-wide association studies; SNPs, single-nucleotide polymorphisms.

Global and local genetic correlations

LDSR analyses showed a significant positive genetic correlation (r_g) between ASD and loneliness (r_g = 0.26, P = 2e-10) and a weak, nonsignificant genetic correlation between ASD and social isolation (r_g = 0.02, P = 0.8).

LAVA analysis for ASD and loneliness identified 72 regions of local r_g with nominal significance (P < 0.05), and 17 regions after FDR correction (Fig. 2a). As seen in Fig. 2a, there is a rightward shift for greater red dots in line with a positive r_g. For ASD and social isolation, LAVA identified 80 regions of local r_g with nominal significance (P < 0.05), and one region (chr3:20754843–21790867) was significant after FDR correction (Fig. 2b). Thus, LAVA indicated genetic overlap between ASD and isolation despite the nonsignificant global r_g.

Fig. 2.

Volcano plots illustrating genetic overlap between ASD, loneliness, and social isolation. The Volcano plots are based on LAVA analysis and display the local genetic correlation coefficients (rho) against –log10 (−P) per locus between ASD & loneliness and ASD & social isolation. Gray points are nonsignificant correlations. Blue points represent negative correlations, and red points (positive) correlated with nominal significance (P < 0.05). Larger red/blue points represent correlations that surpass FDR correction. (a) the figure regarding ASD and loneliness, (b) the one regarding ASD and social isolation. ASD, autism spectrum disorder; LAVA, local analysis of covariant annotation.

Bidirectional genetic overlap beyond genetic correlation using MiXer

For ASD and loneliness, MiXeR estimated that the number of shared variants is between 3.8 and 11.1 k. The lower boundary for this estimate is the minimum amount of the shared variants for a given level of genetic correlation (Frei et al., 2019), while the upper boundary is the total polygenicity of loneliness. The high uncertainty of the MiXeR model for ASD and loneliness is reflected in both best_versus_minimum AIC and best_versus_maximum_AIC values being negative (Supplementary Table 11, Supplemental digital content 2, https://links.lww.com/PG/A335).

For ASD and social isolation, MiXeR estimated that 8.7 k out of 13.1 k genetic variants influencing both phenotypes are shared, with a Dice similarity coefficient of 0.81 (Supplementary Fig. 2b, Supplemental digital content 3, https://links.lww.com/PG/A336). Of the shared genetic variants, 53% were estimated to have concordant effect directions. The conditional Q–Q plots and Log-likelihood plots from MiXeR analysis regarding ASD and social isolation are presented in Supplementary Fig. 1b, Supplemental digital content 3, https://links.lww.com/PG/A336, and as shown in Supplementary Table 11, Supplemental digital content 2, https://links.lww.com/PG/A335, the AIC value indicates that the MiXeR can accurately differentiate the estimated fit from the minimum possible overlap based on genetic correlation.

Genetic loci discovery

To leverage the increased statistical power from the genetic overlap, we applied cond/conjFDR to identify specific genetic loci.

CondFDR analysis identified seven genetic loci associated with ASD after conditioning on loneliness (condFDR < 0.01). These loci are presented in Supplementary Table 1, Supplemental digital content 4, https://links.lww.com/PG/A337 and Supplementary Fig. 3a, Supplemental digital content 3, https://links.lww.com/PG/A336. CondFDR also identified five genetic loci associated with ASD after conditioning on social isolation (condFDR < 0.01), and these loci are presented in Supplementary Table 2, Supplemental digital content 4, https://links.lww.com/PG/A337 and Supplementary Fig 3c, Supplemental digital content 3, https://links.lww.com/PG/A336. Three of the loci detected by these two condFDR analyses were not detected in the original ASD GWAS, and their nearest genes were: KANSL1, LINC00461, and SERPIN1a.

ConjFDR analysis revealed nine loci jointly associated with ASD and loneliness (conjFDR < 0.05). These are presented in Fig. 3a and in Supplementary Table 3, Supplemental digital content 4, https://links.lww.com/PG/A337. Among them, eight were not discovered in the original ASD GWAS (Grove et al., 2019), and eight of them had concordant allelic effect directions (Supplementary Table 3, Supplemental digital content 4, https://links.lww.com/PG/A337).

Fig. 3.

Manhattan plots showing common genetic variants jointly associated with ASD and social traits. Manhattan plots for ASD and loneliness (a) and for ASD and social isolation (b), showing the −log10 transformed conjFDR values for each SNP on the y-axis and chromosomal positions along the x-axis. Significant independent SNPs (based on conjunctional FDR < 0.05) are shown with enlarged blue dots. A black circle around the enlarged blue dot indicates the most significant independent SNP in its locus. The significant loci shown are new, except the loci at chromosome 5 for ASD & loneliness and the loci at chromosome 8 and 9 for ASD & social isolation, which were identified in the original ASD GWAS. Two loci at chromosome 17 are shared with both social isolation and loneliness. ASD, autism spectrum disorder; GWAS, genome-wide association studies; SNPs, single-nucleotide polymorphisms.

ConjFDR analysis of ASD and social isolation revealed eight shared loci, which are presented in Fig. 3b and in Supplementary Table 4, Supplemental digital content 4, https://links.lww.com/PG/A337. Of these loci, six were not identified in the original ASD GWAS (Grove et al., 2019), and five had concordant effect direction (Supplementary Table 4, Supplemental digital content 4, https://links.lww.com/PG/A337). Two loci were overlapping with both loneliness and social isolation, yielding a total of 15 independent loci. Of these, 12 were not identified in the original ASD GWAS (Grove et al., 2019) (Table 2).

Table 2.

Novel genetic loci and mapped genes for autism spectrum disorder, loneliness, and social isolation

Chr	Lead SNP	Sig	Mapped gene	Genetic function
A: ASD and loneliness
1p36.23a	rs10864367	1E-4	RERE e SLC45A1 SLC25A33	DNA-binding transcription factor Cerebral glucose transporter Mitochondrial transporter
2q34a	rs10932542	8E-5	VWC2L e	Part of AMPA glutamate receptor/calcium
3p24.3a	rs748832	2E-4	PLCL2 e	GABA receptor binding, calcium, and immune
6p24.1a	rs932345	2E-6	ADTRP e HIVEP1	Androgen-dependent TFPI regulation Immune, influence viral and cellular genes
10q25.3a	rs1885435	3E-5	NRAP e HABP2	Nebulin/muscle Serine protease, coagulation, inflammation
14q32.13b	rs28929474	2E-6	SERPINA1 e	Alpha-1 antitrypsin deficiency, lung/liver
17q21.31a,c	rs17652520	1E-5	MAPT d , e , f CRHR1 d , f WNT3 d KANSL1 d , f ARHGAP27 d PLEKHM1 d SPPL2C d STH d , f ARL17B d LRRC37A d	Encodes tau–protein involved in dementia Stress hormone release receptor Modulates neuronal calcium AMPA transmission and oxidative stress Activates GTP-metabolizing enzymes. Vesicular transport in osteoclast/bone Vesicular transport, acrosome formation MAPT-associated, Tau/dementia-related Enables GTP binding (energy transfer) Leucine-rich membrane protein
17q24.2a	rs56305452	3E-5	PSMD12 BPTF C17orf58 KPNA2	Proteasome and inflammation Cell division and immune defense Collagen-containing extracellular matrix Nuclear import and immune regulation
B ASD and social isolation
5q14.3a	rs4916723	2E-6	TMEM161B	Calcium-activated chloride channel
8p23.1a	rs278736	5E-5	MTMR9 d MSRA d FAM167A d	Glucose and insulin tolerance Converts methionine sulfoxide Increases autoinflammation
10p12b	rs1277731	7E-5	CACNB2	Neuronal calcium channel
17p13.3a	rs2447091	1E-4	SCARF1 SGSM2 MNT AC006435.1 CLUH	Receptor in endothelial cells GTPase activator (energy) Genetic transcription repressor Unknown Mitochondrial biogenesis
17q21.31b,c 17q24.2a	rs80028338 rs56305452	8E-6 2E-02	Similar genes as those associated with loneliness but opposite effect directiond Similar genes as those associated with loneliness

Only loci that were not discovered in the original ASD GWAS are presented (Grove et al, 2019). Complete lists of loci with more details are presented in Supplementary Tables 3 and 4, Supplemental digital content 4, https://links.lww.com/PG/A337 and complete lists of mapped genes are presented in Supplementary Tables 7 and 8, Supplemental digital content 2, https://links.lww.com/PG/A335. An extended table including references is presented in Supplementary Results, Table 13, Supplemental digital content 3, https://links.lww.com/PG/A336.

ASD, autism spectrum disorder; Chr, chromosome location; Sig, significance level for ASD according to conjFDR analysis.

Genetic function: A short presentation of description found at the GeneCards database of human genes (61).

^aConcordant effect direction with loneliness.

^bConcordant effect direction with social isolation.

^cFor 17q21.31 only genes that were credible by tree different methods are presented.

^dCredible by three methods.

^eBoth mapped gene and nearest gene.

^fSignificant for all three traits in our MAGMA analysis.

Gene-mapping

Using the three different strategies described in the methods, we mapped 36 genes from candidate SNPs within loci shared between ASD and loneliness (Supplementary Table 7, Supplemental digital content 2, https://links.lww.com/PG/A335). Of these, 10 were considered credible after being mapped by all three gene-mapping strategies. Further, we mapped 67 genes for ASD and social isolation (Supplementary Table 8, Supplemental digital content 2, https://links.lww.com/PG/A335). Of these, 13 were considered credible after being mapped by all three gene-mapping strategies. The mapped genes for the novel significant loci are presented in Table 2, and as shown, 11 of the novel genes were considered credible.

Gene expression in different tissues and development periods

The heatmaps of the gene sets derived from candidate SNPs are shown in Supplementary Fig. 4, Supplemental digital content 3, https://links.lww.com/PG/A336. The gene sets were not significantly upregulated in any of the 54 body tissues that were analyzed (Supplementary Fig. 5, Supplemental digital content 3, https://links.lww.com/PG/A336). Expression of mapped genes in different developmental periods is presented in Supplementary Fig. 6, Supplemental digital content 3, https://links.lww.com/PG/A336, and showed no significant differences of gene expression during the developmental periods assessed.

Gene-set enrichment

Mapped genes were enriched in several other GWAS phenotypes, both in somatic traits such as BMI and blood pressure, brain-related traits including handedness, Parkinson’s disease, and multiple system atrophy, neuroticism, alcohol use, number of sexual partners, and cognitive function (see Supplementary Tables 9 and 10, Supplemental digital content 2, https://links.lww.com/PG/A335).

Analysis of molecular function

Mapped genes were significantly enriched in four overlapping immune related reactomes/curated gene sets (BETA defensins, defensins, antimicrobial peptides, and Epstein–Barr virus nuclear antigen 1 anticorrelated) (Supplementary Table 10, Supplemental digital content 2, https://links.lww.com/PG/A335), and were involved in calcium-, glutamate- and GABA signaling, cerebral glucose transport, stress hormone, androgen receptor signaling, tau protein formation/dementia and immune activation, as presented in Table 2.

Colocalization analysis of identified shared genetic loci

Colocalization analysis of loci shared between ASD and social traits, identified in conjFDR analyses, has not revealed strong evidence of shared causal SNPs for most of these regions (Supplementary Tables 3 and 4, Supplemental digital content 4, https://links.lww.com/PG/A337), with notable exceptions being the LINC00461 gene locus for ASD and social isolation (locus 1 in Supplementary Table 4, Supplemental digital content 4, https://links.lww.com/PG/A337), which shows a colocalization probability of 91.3%, and the SERPINA1 gene locus for ASD and loneliness (locus 7 in Supplementary Table 3, Supplemental digital content 4, https://links.lww.com/PG/A337), with a colocalization probability of 85.9%. The latter locus is the only one in our analyses where the lead SNP (rs28929474) is located within the exon of the protein-coding gene (SERPINA1), while all other lead SNPs are in noncoding regions.

Gene-level analysis

Our MAGMA gene analyses identified five genes that were common to ASD, loneliness, and social isolation: KANSL1, STH, MAPT, CRHR1, and NSF. The FDR-corrected p-values from the MAGMA analysis, along with fold enrichments and corresponding reliability estimates (AIC values) assessed by GSA-MiXeR for these genes across all three analyzed phenotypes, are presented in Supplementary Table 12, Supplemental digital content 2, https://links.lww.com/PG/A335. All these genes are located within the 17q21.31 region, and all, except for NSF, were also mapped to the locus with the rs17652520 lead SNP identified in our conjFDR analysis (Table 2). Due to the moderate statistical power of the analyzed GWAS summary statistics, many GSA-MiXeR estimates of fold enrichment were not robust (AIC < 0), except for the MAPT gene, which had a fold enrichment of about 5.8 in ASD, and the STH gene, which was about 15 times more enriched in loneliness compared to the baseline model. The MAGMA gene-set analyses did not reveal any gene sets that were common to ASD and social traits.

Discussion

This is the first study to show an extensive genetic overlap between ASD, loneliness and social isolation, and to identify 12 genomic loci not discovered in the original ASD GWAS (Grove et al., 2019).

For both loneliness and social isolation, conditional Q–Q plots and MiXeR analyses suggested genetic overlap with ASD. For ASD and loneliness, we found a positive genome-wide genetic correlation (r_g = 0.26) and 72 local regions of genetic correlations. For ASD and social isolation, we identified 80 locally correlated genetic regions, and MiXeR indicated extensive genetic overlap with 81% shared variants, despite no significant genome-wide genetic correlation. Taken together, our findings show extensive overlap in genetic architecture between ASD and social traits, both subjective (loneliness) and objective (social isolation).

The current findings indicate differences in the overlapping genetic architecture between ASD and subjective and objective social traits, as the majority of the ASD loci shared with loneliness were different from those shared with social isolation. In other words, most of the variants in the ASD genetic architecture, which were also associated with being socially isolated, were not associated with feeling lonely and vice versa. Although the absence of a significant association does not necessarily mean that it does not exist, as it could be caused by a lack of statistical power, the result supports the idea that loneliness and social isolation are distinct social traits with relevance for social symptoms in ASD.

Our results also showed that the genetic overlap between ASD and social traits is bidirectional. For ASD and loneliness, LAVA analysis showed local genetic correlations with both concordant and discordant effect directions. Nearly half of the variants (47%) identified as shared by LAVA and MiXeR between ASD and social isolation had the opposite effect direction. This implies that many variants associated with ASD are also associated with more social interactions. This may be in line with findings of genetic association between the trait systemizing and the ASD symptoms ‘restricted and repetitive behavior’, as well as with prosocial traits (extroversion and openness) (Warrier et al., 2019). This finding of bidirectional overlap between ASD and social traits may be relevant for understanding the heterogeneity in social function among autistic persons. Furthermore, that some of the common genetic liability for ASD is shared with liability for prosocial traits is in line with a neurodiversity understanding of ASD (Pellicano and den Houting, 2022).

Although common variants individually have small effect sizes, they may collectively account for a substantial proportion of the genetic liability for ASD (Gaugler et al., 2014). Furthermore, polygenic liability for ASD (PRS) has previously been associated with poorer function and increased need for specialist care, more severe behavior regulation difficulties (Torske et al., 2020; LaBianca et al., 2021), and increased sensitivity to facial images (Gui et al., 2021), indicating that clinical relevance might be obtained from PRS (Gui et al., 2021). Furthermore, due to indices of different and independent genetic mechanisms underlying social and nonsocial aspects of ASD, it has been suggested that separate polygenic liabilities for social symptoms and for rigid/repetitive traits should be calculated (Happe and Frith, 2020). This is not possible yet, but hopefully our current results may be used in future studies for designing more specific PRS. Such traits informed PRSs of relevance to social traits might increase the clinical relevance of PRS, aiding the development of more personalized interventions and stratification (Frye et al., 2022).

By using data on genetic overlap between ASD and social traits, we got improved statistical power to identify 12 novel genetic loci associated with ASD, and follow-up analysis enabled us to map several involved genes jointly associated with ASD, loneliness, and social isolation. Our functional analysis showed that the mapped genes are expressed in many different brain regions and without any significant differences in neurodevelopmental stages. This could imply that these genes are not only expressed during early neurodevelopment but also later in life, and might be targeted for treatment (Modabbernia et al., 2017; Burris et al., 2020; Warrier and Baron-Cohen, 2021). Our molecular function analysis showed significant enrichment in a few gene sets, which were immune-related pathways for antimicrobial defensins. This is noteworthy as immune-related factors are implicated in the pathology of ASD (Modabbernia et al., 2017; Hadwin et al., 2019; Warrier and Baron-Cohen, 2021). Furthermore, antimicrobial defensins have also been found to mediate the negative effects of loneliness on risk of preterm birth (Burris et al., 2020), which is associated with increased risk of ASD (Crump et al., 2021). These results suggest that genetic factors may be involved in mechanisms underlying immune dysregulation in ASD (Robinson-Agramonte et al., 2022).

We were able to map several genes with well-known functions, which are relevant not only for ASD but may also be involved in the mechanism underlying loneliness and social isolation, and could potentially suggest new targets for treatment. Our MAGMA gene analyses identified five genes that were common to ASD, loneliness, and social isolation: KANSL1, STH, MAPT, CRHR1, and NSF. Of these, the MAPT and the genes are is important as they are encoding the Tau protein and an intron in the Tau protein, which are well-known treatment targets for frontotemporal dementia (Bouzigues et al., 2025). Tau protein accumulation has been found in ASD brains and has received attention as a possible treatment target for ASD (Villavicencio-Tejo et al., 1496; Grigg et al., 2020). We also mapped the stress hormone signaling gene CRHR1. The variant mapping to this gene was NS in either the original ASD (Grove et al., 2019) nor the loneliness GWAS (Day et al., 2018; Grove et al., 2019), but association with ASD was supported by MAGMA analysis (Grove et al., 2019). Several lines of evidence suggest increased vulnerability for mental stress in ASD (Courtemanche et al., 2021; Makris et al., 2021), and our study supports that cortisol signaling may be involved in this vulnerability (Chou et al., 2014). Several of our identified mapped genes are involved with calcium-related glutamate and GABA signaling (CACNB2, vW2CL, and PLCL2), supporting previous reports that altered calcium signaling may be involved in ASD. Thus, medications targeting GABA and glutamate may be beneficial (Marotta et al., 2020; Reilly et al., 2020; Manco et al., 2021). Furthermore, the genes that regulate cerebral glucose transportation (SLC45A1 and MTMR9) support reduced brain glucose metabolism in ASD (Manco et al., 2021).

Furthermore, the mapping of an androgen receptor signaling protein (ADTRP), supports the importance of regulating fetal exposure to testosterone for reducing the liability to ASD (Palm et al., 2022).

Some limitations of our study include that all samples consisted of persons of European ancestries. It is also likely that the questionnaire used to assess loneliness and social isolation has limitations with respect to sensitivity and specificity for these traits. The difference in statistical power between the different GWAS analyzed makes it difficult to make valid comparisons between the differences in genetic overlap and could affect the statistical power to identify shared genetic loci. Investigating genes with concordant and discordant effect directions in ASD and social traits separately might help to dissect the potential biological underpinnings of the observed phenotypic heterogeneity in the autism spectrum. However, with the currently available summary statistics, our conjFDR analysis identified nine loci shared between ASD and loneliness, and eight loci shared between ASD and social isolation, with eight and five of them being concordant, respectively. This limited number of loci precludes meaningful separate gene-set analyses of concordant and discordant genes.

Furthermore, functional analyses are based on statistical associations that annotate genes in a biological context (Watanabe et al., 2019), and not on experimental work to determine molecular mechanisms and should be considered with caution.

In summary, the current study revealed genetic overlap between ASD and both loneliness and social isolation. While ASD had a positive genetic correlation with loneliness, there was a pattern of mixed effect directions among the shared variants with social isolation. The results suggest that genetic overlap may underlie some of the observed association between ASD and these social traits, with different variants underlying subjective and objective aspects. This may have implications for our understanding of autistic traits and potential clinical impact. Further work is needed to delineate the molecular genetic mechanisms underlying these associations, which may contribute to progress in treatment options to improve the quality of life of people with ASD.

Acknowledgements

We thank the research participants contributing to the GWASs used for the present study. We also thank the Psychiatric Genomics Consortium and the UK Biobank, as well as Felix Day and John Perry for making their GWAS summary statistics available. This work was performed on Services for sensitive data (TSD), University of Oslo, Norway, with resources provided by UNINETT Sigma2 – the National Infrastructure for High Performance Computing and Data Storage in Norway.

This study used summary statistics from previously published GWAS studies. Regional Committees for Medical Research Ethics – South-East Norway has evaluated the current protocol and found that no additional institutional review board approval was necessary because no individual data were used.

All data are either publicly available [GWAS Catalog (Cerezo et al., 2025) (ebi.ac.uk)] or available on request. Analysis tools are available at GitHub – bulik/ldsc: LD Score Regression (LDSC), GitHub – precimed/mixer: Causal Mixture Model for GWAS summary statistics, and GitHub – precimed/pleiofdr: Pleiotropy-informed conditional and conjunctional false discovery rate.

This work was supported by the Research Council of Norway (#223273, #273291, #276082, #296030, and #300309), KG Jebsen Stiftelsen (SKGJ-MED-021), Norway Regional Health Authority (#2020060), and EU’s Horizon 2020 RIA grant #847776 CoMorMent. This work was performed on Services for sensitive data (TSD), University of Oslo, Norway, with resources provided by UNINETT Sigma2 – the National Infrastructure for High Performance Computing and Data Storage in Norway.

O.A.A. and T.N. conceived the study. I.S., W.C., N.P., and S.D. contributed to data acquisitions. A.L. and A.A.S. performed data analysis with assistance from S.B., O.F., K.O’.C., W.C., and N.P. A.L., L.R., S.J.H., S.B., W.C., N.P., K.S.O’.C., A.A.S., K.S., G.F.L.H., D.S.Q., and A.K. contributed to data interpretations. S.H., S.J.H., and A.S. drafted the manuscript. All authors revised and approved the final manuscript.

Conflicts of interest

O.A.A. is a consultant for HealthLytix and received speaker´s honorarium from Lundbeck and Sunovion. The remaining authors have no competing interest.

Supplementary Material

pg-36-13-s001.docx ^{(230.1KB, docx)}

pg-36-13-s002.docx ^{(131.5KB, docx)}

pg-36-13-s003.docx ^{(1.8MB, docx)}

pg-36-13-s004.docx ^{(41.7KB, docx)}