Published on 16 September 2026
Suicidal behaviour: what does a multivariate GWAS of more than 1.7 million people really show?
In brief
Five major genomic resources brought together: four biobanks or population programmes and one research consortium. The suicide attempt analysis covers 1,721,665 people, the suicidal ideation analysis 1,193,781. The authors run genome-wide association meta-analyses for ideation and for attempt, then a multivariate genome-wide analysis of the suicidality spectrum. Eighty-seven independent lead variants reach the significance threshold, forty-nine of them not previously described, and three conditionally independent signals are added, which gives the ninety associations announced in the abstract. Thirty-three genes obtain a prioritisation score of at least five points among the 1052 genes retained. SNP-based heritability is 11.5% for suicide attempt, 5.0% for the spectrum and 4.0% for ideation, but all three values concern European ancestry only: in the other ancestries, heritability could not be estimated significantly. It remains significant after conditioning on six psychiatric disorders, while collapsing, for example from 11.5% to 3.8% for attempt after conditioning on anxiety. Three factors emerge as putative causal factors through the convergence of three methods, each on a given phenotype: loneliness for ideation, medical abortion for attempt, age at first sexual intercourse for the spectrum. None of this translates today into a tool or a prescription, and the document remains a preprint that has not been peer reviewed.
The context
Genome-wide association studies ordinarily analyse one phenotype at a time. When several phenotypes are strongly correlated, as suicidal ideation, suicide attempt and suicidal behaviour in the broad sense are, that separation wastes information. The multivariate approach exploits the correlation to gain power.
The gain has a price. Here the authors report both per-phenotype meta-analyses, for ideation and for attempt, and a multivariate analysis of the suicidality spectrum: the multivariate estimates bear on a shared latent factor and not on each phenotype taken on its own. Their interpretation depends on how the phenotypes were harmonised across the five resources. That point is documented: the definitions used in the British biobank and in the United States population research programme are listed item by item in the supplement, the sources for the other cohorts are detailed, and the authors themselves acknowledge that definitions remain heterogeneous from one cohort to another.
The study at a glance
| Population | |
| Suicidal ideation, 179,881 cases against 1,013,900 controls, that is 1,193,781 people drawn from a British national biobank, a United States population research programme and a United States programme conducted among military veterans. Suicide attempt, 66,867 cases against 1,654,798 controls, that is 1,721,665 people drawn from the British biobank, the United States population programme, a Finnish biobank and the psychiatric genomics consortium. Ancestries are European, African, Admixed American, East Asian, Central and South Asian, and Middle Eastern, with a threshold of more than 100 cases per cohort. Recounting sample sizes by ancestry gives 74.2% of participants of European ancestry for ideation and 85.0% for attempt. | |
| Analysis | |
| Logistic regression with Firth correction, adjusted for sex, age and the first ten within-ancestry principal components. Genomic structural equation modelling to combine the European results, sample-size-weighted meta-analysis for the other ancestries, then cross-ancestry meta-analysis. Identification of lead variants, conditional and joint analysis, multi-ancestry fine-mapping. Five gene-discovery approaches: gene-based analysis, transcriptome-wide association, isoform-level transcriptome-wide association, proteome-wide association and Mendelian randomisation on methylation data. Tissue and gene ontology enrichment, drug repurposing analysis. Linkage disequilibrium score regression for heritability and genetic correlations, multi-trait conditioning on six psychiatric disorders, then three causal inference methods used jointly. | |
| Comparator | |
| Controls drawn from electronic health records and questionnaires, with two explicit independence rules: people with a history of attempt are excluded from the ideation analysis, and ideation cases are excluded from the controls of the attempt analysis. The authors point out that the suicidality status of some record-defined controls remains unknown, which underestimates prevalence and attenuates effect sizes. | |
| Outcomes | |
| Number of significant variants, SNP-based heritability, genetic correlations between phenotypes, prioritised genes, candidate causal factors, candidate molecules for repurposing. | |
| Design | |
| Meta-analyses of genome-wide association studies and multivariate analysis, observational design. The Oxford levels of evidence scales are built for questions of diagnosis, prognosis or treatment and contain no category corresponding to a meta-analysis of genome-wide association studies. No level has therefore been assigned. Preprint posted on medRxiv on 16 December 2025, version 1, not peer reviewed at the date of verification: the server page lists no review, neither professional nor community. |
Quality control
| Item checked | Verdict |
|---|---|
| Cohort sample size | Considerable sample size |
| FindingFive of the most widely used research resources in the field, brought together on a single set of phenotypes, for 1,721,665 participants in the suicide attempt analysis and 1,193,781 in the ideation analysis. Earlier work is identified in the reference list, but its sample sizes are not restated, so the gap with it is measured in number of loci and not in number of participants. | |
| Genotype quality control | Standard, explicit criteria |
| FindingMinor allele frequency above 1%, Hardy-Weinberg equilibrium with p above 10⁻⁶, imputation quality above 0.8, missingness below 5% per variant and per individual. Ancestries genetically inferred and relatedness removed. In the combined results, only variants present in several cohorts, with a total sample size above 10,000, and passing the heterogeneity test, are retained. Nothing unusual here, but everything is written down. | |
| Editorial status | Preprint |
| FindingThe document has not been peer reviewed. The server itself displays a warning stating that it reports medical research not yet evaluated and that it should not be used to guide clinical practice. The values reported may be revised on publication. | |
| Convergence between methods | Convergence partly across phenotypes |
| FindingThe prioritisation score awards one point per approach and per phenotype, then adds up across five approaches and three phenotypes, for a maximum of fifteen. The 33 genes retained have a score of at least five. Such a score can therefore come from five approach-phenotype combinations and not from five distinct approaches, all the more so as the three phenotypes are highly correlated with one another, with genetic correlations from 0.639 to 0.960. The convergence is real, it is less independent than the statement suggests. Only three genes reach the maximum observed score of eight: UGGT2, GMPPB and BRWD1. | |
| Assumptions of causal inference | Sensitivity analyses run, bidirectional effects |
| FindingThree frameworks are used jointly: a latent causal variable model, a Mendelian randomisation corrected for sample overlap, weak instrument bias and winner’s curse, and a generalised Mendelian randomisation on non-overlapping datasets with a filter for instrument heterogeneity. The set-up is serious. Still, the three factors retained pass on a nominal criterion for the latent causal variable model, with genetic causality proportions that are low in absolute value, from 0.126 to 0.294, and the Mendelian randomisation is significant in both directions for loneliness as for medical abortion. Reverse causation is therefore not ruled out. | |
| Funding and competing interests | Declared, one industry tie |
| FindingFunding from the American Foundation for Suicide Prevention (PDF-0-065-23), the National Institute of Mental Health (RF1MH132337) and MQ: Transforming Mental Health (UFA21\100014). The authors write that the funders had no role in design, data collection, analysis, interpretation or writing. A single competing interest is declared: the senior author is paid for his editorial work for the journal Complex Psychiatry and has received a research grant from the company Alkermes, outside the scope of this work. The other authors declare no conflict. Nothing abnormal, but the repurposing analysis concerns antipsychotics, and this declaration deserves to be read alongside it. | |
The findings
| Result | What the data show |
|---|---|
| Significant variants | 87 lead variants, 49 of them new, plus 3 conditional signals |
| ReadingRecounting the supplementary table confirms 87 lead variants spread over 80 loci, 49 labelled new and 38 already reported. The breakdown by analysis is given: 61 variants for the spectrum in European ancestry, 20 for ideation and 37 for attempt, overlaps deducted. The authors set this contribution against earlier work that identified four loci for ideation, twelve for attempt, and 77 for suicidality in a recent preprint (our analysis of that preprint). The three signals from the conditional and joint analysis are independent of this study’s lead variants, but the authors specify that they are not independent of loci already published. | |
| SNP-based heritability | Attempt 11.5% (± 0.5), spectrum 5.0% (± 0.2), ideation 4.0% (± 0.2), European ancestry only |
| ReadingEstimates by linkage disequilibrium score regression, bearing on the factors derived from genomic structural equation modelling, and therefore on the observed scale with effective sample sizes, not on the liability scale. The per-cohort estimates, by contrast, are on the liability scale with population prevalences of 9% for ideation and 2% for attempt, and range from 4.6% to 14.0% depending on the cohort. Two important reservations. First, these values hold only for European ancestry: in the other ancestries, the authors write that the estimates are not statistically significant. Second, conditioning on each of the six psychiatric disorders leaves a heritability that is significant but sharply reduced: 11.5% to 3.8% for attempt after conditioning on anxiety, 5.0% to 1.6% for the spectrum after conditioning on depression, 4.0% to 2.2% for ideation in the same case. What remains outside psychiatric diagnosis is therefore real and modest. | |
| Genetic correlations | Attempt and spectrum 0.960, ideation and spectrum 0.865, ideation and attempt 0.639 |
| ReadingThe announced range of 0.639 to 0.960 is bounded by the ideation-attempt pair, at 0.639 (p = 6.81 × 10⁻¹³⁵), and by the attempt-spectrum pair, at 0.960. Ideation and attempt are therefore only partly the same genetic object, which is confirmed by 10 variants specific to ideation and 16 specific to attempt beyond the 61 of the spectrum. Between cohorts, before combination, the correlations are markedly lower, from 0.238 to 0.862, which directly reflects the heterogeneity of the definitions. | |
| Prioritised genes | 1052 genes, 33 with a score of at least 5, 16 shared by all three phenotypes |
| ReadingRecounting the supplementary table gives exactly 1052 genes, 33 of them with a score of at least five and 3 at the maximum score of eight. Sixteen genes are supported for all three phenotypes at once, among them DCC, RBFOX1, GRIA1 and ANKK1. Conversely, 65 genes are specific to ideation, 114 to attempt and 537 to the spectrum, on the strength of at least one analysis. The great majority of the 1052 genes therefore rests on only one or two analyses. | |
| Causal inference | Loneliness for ideation, medical abortion for attempt, age at first intercourse for the spectrum |
| ReadingThree factors, and three only, pass all three methods, and each is tied to a specific phenotype, not to suicidality in general. Loneliness is in line with the epidemiological literature on social isolation. The other two are more delicate: the authors themselves recall that the link between a history of abortion and suicide attempt is disputed, a Danish registry study attributing it to confounding factors rather than to a direct effect. In the reverse direction, 23 traits emerge as putative consequences, among them hospital admissions, substance use disorders, difficulty falling asleep and various physical conditions. No consequence is identified for ideation alone. | |
| Candidate molecules | Four molecules for attempt, two for the spectrum, none for ideation |
| ReadingCariprazine, droperidol, molindone and paliperidone for attempt, cariprazine and chlorprothixene for the spectrum, that is five distinct molecules, all of them antipsychotics. The signal rests on a competitive analysis of drug target gene sets drawn from drug-gene interaction databases, with Bonferroni correction over 993 sets tested. A match between a flagged gene and a molecule listed in a target database is not an efficacy signal. None of these molecules was evaluated against a suicidality outcome in this work. We did not consult any regulatory source: nothing is asserted here about their regulatory status, marketing or reimbursement. | |
Critical appraisal
| Domain | Judgement |
|---|---|
| Design and level of evidence | Observational |
| FindingAn observational design at population scale. The genetically informed causal analyses strengthen the causal argument for a given factor, they do not turn the whole into experimental evidence. The authors themselves speak of natural experiments, a phrase that describes an aid to reasoning and not a randomisation. | |
| Traceability of the method | Usable methods and supplements |
| FindingA complete methods section, twenty supplementary tables and seventeen supplementary figures. Sample sizes by cohort and by ancestry, heritabilities before and after conditioning, genetic correlations, the list of lead variants and the count of prioritised genes can be recounted and agree with the text. Only two points call for vigilance: the abstract announces 90 lead variants where the results count 87 plus three conditional signals, and the word loci is used in the summary box to refer to the 33 best-prioritised genes. | |
| Phenotype harmonisation | Heterogeneous, acknowledged by the authors |
| FindingDefinitions range from self-report to hospital coding, by way of questionnaire items and health record concepts. The authors say so: in the Finnish biobank, the codes X60 to X84 used to define attempt include non-suicidal self-harm; some consortium cohorts were recruited to study specific psychiatric disorders, which limits transfer to the general population; sample overlap persists between two United States cohorts and could not be excluded. The cross-cohort genetic correlations, as low as 0.238 before combination, put a figure on this heterogeneity better than any commentary. | |
| Fit between claim and evidence | Measured, final conclusion firmer |
| FindingThe molecules are presented as potential candidates, the causal factors as putative. The abstract writes that the findings may inform targeted suicide prevention strategies, in the conditional. The conclusion of the full text drops the conditional and writes that the findings inform strategies for suicide control and prevention. The gap is slight, it goes in the direction of a hardening, and it is the wording of the abstract that matches what the data allow. | |
| External validity | Diversity displayed, low non-European power |
| FindingThe diversity is real but unbalanced. Recounting gives 74.2% of participants of European ancestry for ideation and 85.0% for attempt, the Asian groups representing about 2% of the ideation analysis and 1.9% of the attempt analysis. In the per-ancestry analyses, European ancestry provides 12 lead variants for ideation and 31 for attempt, against 5 and 5 for all the other ancestries combined. Above all, the authors write that heritability could not be estimated significantly outside European ancestry, which prevented multivariate modelling from being applied there. The title announces diverse ancestries, the genetic architecture is described only in Europeans. | |
| Independence and transparency | Declared, data to come |
| FindingFunding, the role of the funders, author contributions and competing interests are declared. The study was judged by the ethics committee of the host university not to constitute human subjects research, each cohort also having its own approval. On the other hand, the summary statistics produced by this work are not yet accessible: the authors announce public release at the time of publication. The original datasets are referenced with their access addresses. No statement on sharing of the analysis code appears in the text consulted. | |
Level of evidence
Confidence is high in the existence of a large-scale polygenic genetic signal for suicidal behaviour. The 87 lead variants, 49 of them new, can be recounted in the supplements, the quality control criteria are explicit, and the genome-wide significance threshold is applied without adjustment. It is also high that ideation and attempt overlap only partly: a genetic correlation of 0.639 between them, 10 variants specific to one and 16 specific to the other, a single protein shared by both in the proteome-wide analysis.
It is moderate on heritability. The values are accurate and documented, but they concern a single ancestry, they bear on latent factors estimated on the observed scale, and they fall by two thirds for attempt as soon as one conditions on anxiety. It is moderate too on the 33 best-prioritised genes, whose score adds up approaches and phenotypes that are highly correlated with one another.
It is low on the three causal factors. The inference set-up is careful, but the genetic causality proportions are low, the threshold used for one of the three methods is only nominal, and the effect is significant in both directions for loneliness as for medical abortion. It is nil on any practical consequence: no tool, no test, no prescription follows from this work, and the drug repurposing lead remains upstream of any clinical evaluation. The preprint status adds a reservation that applies to all the figures cited.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ More than 1.7 million people commands respect, and this time the methods are on the table. What I take from it: the heritability of 11.5% for attempt holds only in Europeans, it drops to 3.8% when you condition on anxiety, and it could not even be estimated elsewhere. Loneliness comes out, which I already knew by other routes, but the effect is significant in both directions, so I still do not know which drives which. And I am waiting for the peer reviewed version. ”
What this means in practice: nothing changes in the consulting room. The paper reinforces, without demonstrating it on its own, the value of assessing social isolation in the evaluation of a patient at suicidal risk. The molecules named are not therapeutic options in this indication, and it would be wrong to present them as such to a patient.
What you can do with this
- Keep asking explicitly about social isolation when assessing suicide risk. The genetic argument does not ground this practice on its own, it joins an epidemiological literature that is already consistent.
- Do not present the molecules named as treatments for suicide risk. They are matches produced by a repurposing analysis, between a flagged gene and a pharmacological target database, with no clinical data whatsoever on a suicidality outcome. We did not consult any regulatory source: nothing is asserted here about their regulatory status.
- Remember that ideation and attempt are not the same object. This is the most solid and most transferable finding of this work: a genetic correlation of 0.639, largely distinct variants and proteins, and a heritability nearly three times higher for attempt.
- Do not quote the heritability of 11.5% without saying that it concerns European ancestry only and that it falls sharply once psychiatric disorders are taken into account.
- Be able to tell a patient who has read a press article on the subject that this work leads to no genetic test of suicide risk, and that its authors propose none.
- Wait for the peer reviewed version before citing these results in a formal setting, all the more so as publication could change the estimates.
Frequently asked questions
What does a SNP-based heritability of 11.5% for suicide attempt mean?
That 11.5% of the variability of the phenotype, in this sample and with this definition, is explained by the common genetic variants measured. It is not an individual probability, it is not a familial risk, and it is not a biological constant. Here the value bears on a latent factor estimated on the observed scale, in people of European ancestry only, and it falls to 3.8% once anxiety is taken into account.
Should the drugs flagged by this study be prescribed to patients at suicide risk?
No. None was evaluated against a suicidality outcome in this work: they come out of a repurposing analysis based on matches between flagged genes and pharmacological targets. Four come out for attempt, two for the spectrum, none for ideation. We did not consult any regulatory source, and nothing is asserted here about their authorisation, marketing or reimbursement status.
Does loneliness cause suicide, according to this genetic study?
The analysis designates it as a putative causal risk factor for suicidal ideation, through the convergence of three methods. Two reservations matter. The estimated genetic causality proportion is low, about 0.29 in absolute value, far from what would be expected of a fully causal link. And the Mendelian randomisation finds a significant effect in both directions, so loneliness that follows suicidal experience is not distinguished from loneliness that precedes it. It is one more argument, consistent with the epidemiological data, not a demonstration.
Do these genetic findings apply to patients of every ancestry?
Not equally. The authors include participants of European, African, Admixed American, East Asian, Central and South Asian, and Middle Eastern ancestry, and report significant variants in several of these groups, which is new. But European ancestry accounts for 74.2% of the ideation analysis and 85.0% of the attempt analysis, only ten lead variants come from the other ancestries, and heritability could not be estimated significantly in them. The diversity is displayed, transferability is not demonstrated.
How does a multivariate GWAS differ from a standard GWAS?
A standard analysis treats one phenotype at a time. The multivariate approach analyses them jointly and gains power when they are correlated. In return, the estimates bear on a shared latent factor and not on each phenotype separately. Here the multivariate analysis covers the suicidality spectrum, ideation and attempt each being the subject of a separate meta-analysis, and it is the spectrum that yields the most variants, 61 of 87.
Annotated bibliography
Source study. He J, Cabrera-Mendoza B, Qiu D, Davtian D, Mao Z, Chen Q, Penichet EN, Zhang Q, Karaca S, Polimanti R. Multivariate genome-wide association study of suicidal behaviors in >1.7 million individuals of diverse population descents. medRxiv [Preprint]. 2025 Dec 16:2025.12.15.25342298 (version 1). DOI 10.64898/2025.12.15.25342298.
Status of the document and material consulted. Preprint posted on medRxiv on 16 December 2025, version 1, not peer reviewed at the date of verification. It carries PMID 41445618 and PMCID PMC12723771. Consulted: the full text of the preprint, abstract, summary box, introduction, methods, results, discussion, declarations and list of 120 references; the file of supplementary figures and tables, containing seventeen supplementary figures; and the supplementary material spreadsheet, containing thirty-five sheets corresponding to the twenty supplementary tables and their sub-tables. Sample sizes by cohort and by ancestry, heritabilities before and after conditioning, genetic correlations, lead variants, prioritised genes, repurposing candidates and causal inference results were recounted programmatically in these tables and agree with the text. Not consulted: the PubMed Central record gives no access to the full text, the licence terms chosen by the authors not allowing it to be archived, and the genome-wide summary statistics produced by the study are not yet public, the authors announcing them for publication. No check therefore bore on individual data or on the summary statistics themselves.
Funding and competing interests. American Foundation for Suicide Prevention (PDF-0-065-23); National Institute of Mental Health (RF1MH132337); MQ: Transforming Mental Health (UFA21\100014). The authors write that the funders had no role in study design, data collection, analysis or interpretation, or in writing the report. Declaration of interests: R. Polimanti is paid for his editorial work for the journal Complex Psychiatry and has received a research grant, outside the scope of this study, from the company Alkermes; the other authors declare no competing interests.
Background references. All the references below are cited and identified in the reference list of the source publication.
Ashley-Koch AE, Kimbrel NA, Qin XJ, et al. Genome-wide association study identifies four pan-ancestry loci for suicidal ideation in the Million Veteran Program. PLoS Genet 2023; 19(3): e1010623. Contribution: the source of the four suicidal ideation loci used as a point of comparison, and the source of the ideation data from the programme conducted among military veterans that are reused here. Limitation: a population of United States veterans, whose age, sex and comorbidity structure does not represent the general population.
Docherty AR, Mullins N, Ashley-Koch AE, et al. GWAS Meta-Analysis of Suicide Attempt: Identification of 12 Genome-Wide Significant Loci and Implication of Genetic Risks for Specific Health Factors. Am J Psychiatry 2023; 180(10): 723-38. Contribution: these are the twelve suicide attempt loci used as a comparator. Limitation: it is not an independent comparator, since this same meta-analysis supplies the consortium data used as input to the present work, amounting to 41,438 cases and 580,259 controls.
Colbert SMC, Group tPGCSW, Ruderfer D, Docherty AR, Mullins N. Genome-wide association studies identify 77 loci for suicidality and provide novel biological insights. medRxiv 2025: 2025.10.22.25338076. Contribution: the closest comparator, the 77 suicidality loci and the single ideation locus reported in East Asian ancestry, and it is against it that the 49 new associations are defined. Limitation: a preprint that has not been peer reviewed either, so the claimed gain is measured against an unconsolidated reference.
Kimbrel NA, Ashley-Koch AE, Qin XJ, et al. A genome-wide association study of suicide attempts in the million veterans program identifies evidence of pan-ancestry and ancestry-specific risk loci. Mol Psychiatry 2022; 27(4): 2264-72. Contribution: the first work cited to report suicide attempt loci in African, Admixed American and Asian ancestries, which qualifies the novelty of the multi-ancestry component. Limitation: the same population of United States veterans.
Motillon-Toudic C, Walter M, Seguin M, Carrier JD, Berrouiguet S, Lemey C. Social isolation and suicide risk: Literature review and perspectives. Eur Psychiatry 2022; 65(1): e65. Contribution: the reference on which the authors base their interpretation of the loneliness finding, with the idea of mediation by depression and of a protective effect of social support. Limitation: a literature review, not a meta-analysis, and therefore without a pooled effect estimate.
Steinberg JR, Laursen TM, Adler NE, Gasse C, Agerbo E, Munk-Olsen T. The association between first abortion and first-time non-fatal suicide attempt: a longitudinal cohort study of Danish population registries. Lancet Psychiatry 2019; 6(12): 1031-8. Contribution: the reference the authors set against their own finding on medical abortion, indicating that it attributes the association to confounding factors rather than to a direct causal link. Limitation: Danish national registries, whose transferability to other health systems is not discussed in the source publication.
Ren J, Qi X, Cao W, et al. Early Sexual Initiation Is Associated with Suicide Attempts among Chinese Young People. Int J Environ Res Public Health 2022; 19(7). Contribution: the reference cited in support of the third causal factor, age at first sexual intercourse. Limitation: the authors themselves describe it as a cross-sectional study, and therefore without an argument from temporality.
GBD 2023 Causes of Death Collaborators. Global burden of 292 causes of death in 204 countries and territories and 660 subnational locations, 1990-2023: a systematic analysis for the Global Burden of Disease Study 2023. Lancet 2025; 406(10513): 1811-72. Contribution: the source of the epidemiological framing of the introduction, 766,700 deaths by suicide worldwide in 2023 and an age-standardised mortality rate of 9.0 per 100,000. Limitation: an estimate from a systematic analysis for which the source publication restates neither the uncertainty intervals nor the quality of cause-of-death registration by country.