AI Reveals Two Distinct Brain Subtypes Behind Suicide Risk in Bipolar Depression

Suicide risk has long been treated as a single, monolithic danger that clinicians try to detect with questionnaires and clinical judgment. A new study published in BMC Medicine challenges that assumption by showing that the brain signatures associated with suicidal thoughts and behaviors in bipolar depression are not one phenomenon but at least two, each with its own wiring pattern, genetic background, cognitive profile, and even preliminary treatment response. The research, led by Ting Wang, Xinruo Wei, Qing Lu and colleagues at Southeast University and Nanjing Medical University, analyzed multimodal brain imaging and genetic data from 802 individuals and used a generative artificial intelligence model to uncover biologically coherent subtypes that had previously been hidden inside the statistical noise of group comparisons.

The scale and design of the study set it apart from most neuroimaging work in psychiatry. The team assembled resting-state functional MRI data from 657 patients with bipolar depression, of whom 405 were experiencing current suicidal thoughts and behaviors and 252 were not, along with 103 healthy controls. Crucially, the data came from one discovery cohort and two independent replication cohorts, allowing the researchers to test whether any patterns they found were stable across different groups of people rather than artifacts of a single sample. A longitudinal component followed 42 patients over time, providing a rare opportunity to see whether the brain-based subtypes tracked fluctuations in clinical suicide risk.

The analytical centerpiece of the study was a semi-supervised clustering-generative adversarial network, abbreviated Smile-GAN. Generative adversarial networks are machine learning architectures in which two neural networks compete: one generates candidate outputs while the other tries to distinguish them from real data. In this application, the framework was adapted to identify imaging-defined subtypes of dysconnectivity, using patients without suicidal thoughts and behaviors as the reference group against which anomalous connectivity patterns could be detected. This semi-supervised design means the model did not simply sort patients into arbitrary clusters; it specifically searched for patterns of brain connectivity that deviated from the non-suicidal reference state, a strategy well suited to a field where the boundaries between clinical categories are blurry and labels are often unreliable.

What emerged were two reproducible neurophysiological subtypes with strikingly different characteristics. The first was a visual cortex-predominant subtype, in which patients showed hyperconnectivity involving the visual cortex, the region at the back of the brain that processes incoming visual information. This pattern was associated with greater anxiety and poorer performance on tests of cognitive flexibility and working memory, including measures such as the Trail Making Test and the Digit Span Backward task. The second subtype was defined by hyperconnectivity between the default mode network and the central executive network, two large-scale brain systems that are normally engaged in a dynamic balance. The default mode network is active during self-referential thought and mind-wandering, while the central executive network governs goal-directed attention and control. Excessive coupling between these two networks was linked to brooding rumination, the repetitive, passive dwelling on negative feelings that is a well-established psychological risk factor for suicidal thinking.

The genetic findings added a layer of biological specificity to these imaging-defined groups. In the visual cortex-predominant subtype, the researchers constructed a genetic risk score from three single nucleotide polymorphisms in serotonergic genes: rs1631327 and rs6320 in HTR5A, which encodes the 5-hydroxytryptamine receptor 5A, and rs8066602 in SLC6A4, the gene for the serotonin transporter, a protein targeted by many antidepressant drugs. This three-variant score was associated with current-episode suicidal thoughts and behaviors within this subtype, and statistical mediation analysis showed that the effect was partially carried through right peripheral visual dysconnectivity. The mediated portion was 11.4 percent, with an effect size of 1.94 multiplied by ten to the power of minus two and a 95 percent confidence interval running from 1.08 multiplied by ten to the power of minus three to 0.05. In plain terms, a small but statistically detectable share of the genetic association with suicidal thinking appeared to operate by altering how visual brain circuits connect with the rest of the network.

The involvement of the serotonin system is scientifically coherent, since serotonergic signaling has been implicated in impulsivity, aggression, and suicide risk for decades, but the localization of the effect to visual cortex is more surprising and potentially important. It suggests that in some patients, inherited differences in serotonin biology may shape suicide-related vulnerability partly through sensory processing circuits rather than through the emotion and control networks that have traditionally dominated the literature. The researchers also examined transcriptomic and neurotransmitter maps from resources such as the Allen Human Brain Atlas to characterize the subtypes, situating the connectivity patterns within the broader molecular architecture of the cortex.

The default mode network-central executive network subtype carried its own genetic and clinical signal, this time in the treatment domain. In exploratory analyses, among carriers of the rs8066602 TC or TT genotypes who belonged to this subtype, patients with suicidal thoughts and behaviors showed greater symptom improvement after two weeks of treatment with pharmacotherapy and repetitive transcranial magnetic stimulation than comparable patients without suicidal thoughts and behaviors. The effect was substantial, with a beta coefficient of 31.01, a p value below 0.01, and a 95 percent confidence interval from 16.94 to 45.08. Repetitive transcranial magnetic stimulation, or rTMS, is a non-invasive brain stimulation technique that modulates cortical activity through magnetic pulses, and the finding hints that a patient’s imaging subtype and genetic background might eventually help predict who responds to which intervention.

Perhaps the most convincing evidence for the reality of these subtypes is their reproducibility. The subtype-specific dysconnectivity patterns replicated across the two independent cohorts, with correlation coefficients of 0.80 and 0.64 in the first replication sample and 0.65 and 0.73 in the second, all reaching adjusted p values below 0.001 on spatial permutation testing designed to account for the autocorrelated structure of brain maps. Moreover, in the longitudinal arm of the study, the strength of subtype-specific dysconnectivity covaried with fluctuations in individual suicide risk over time, suggesting that these are not static traits but dynamic markers that rise and fall with the clinical state. That combination of cross-cohort replication and within-person longitudinal tracking is exactly the kind of evidence that neuroimaging biomarkers have historically lacked.

The authors are careful to frame the genetic and treatment-related signals as preliminary. The genetic risk score rests on a small number of variants drawn from targeted sequencing of 61 suicide-related single nucleotide polymorphisms, and the treatment response analysis was exploratory, involving a limited number of patients followed for a short period. Prospective replication in larger, independently recruited samples will be needed before any of these findings can inform clinical decisions. The researchers also note that suicidal thoughts and behaviors in bipolar depression are profoundly heterogeneous, and that no single biomarker, genetic or neural, is likely to capture that complexity on its own. What this study offers instead is a framework: a way of stratifying patients into biologically meaningful groups before searching for predictors, rather than averaging across a mixed population and diluting the signals that matter.

If the findings hold up, the implications for suicide prevention could be significant. Clinicians currently lack reliable biological tools for assessing suicide risk, relying instead on self-report and clinical interview, both of which are vulnerable to concealment and fluctuation. A future in which a resting-state MRI scan, combined with a handful of genetic markers, could indicate whether a patient with bipolar depression falls into a visual-sensory risk profile or a rumination-linked network profile would open the door to subtype-informed stratification of risk and, eventually, to targeted interventions matched to each profile. The study also demonstrates the growing power of generative machine learning models in psychiatry, showing that adversarial architectures can extract reproducible structure from noisy, high-dimensional clinical data. For a field in which candidate biomarkers have repeatedly failed to replicate, the demonstration that two gene-brain-behavior profiles can survive independent testing and track clinical change over time is a noteworthy step toward making suicide risk assessment a genuinely biological science.

Subject of Research: Neurophysiological subtypes of suicide-related brain dysconnectivity in bipolar depression

Article Title: Gene-brain-behavior links revealing generative neurophysiological subtypes of suicide-related dysconnectivity in bipolar depression

Article References: Wang, T., Wei, X., Shen, N., Xia, Y., Dai, Z., Shao, J., Yan, R., Xiong, T., Tian, S., Yao, Z., Lu, Q., & Yao, Z. (2026). Gene-brain-behavior links revealing generative neurophysiological subtypes of suicide-related dysconnectivity in bipolar depression. BMC Medicine. https://doi.org/10.1186/s12916-026-05228-6

Image Credits: AI Generated

DOI: 10.1186/s12916-026-05228-6

Keywords: bipolar depression, suicidal thoughts and behaviors, resting-state fMRI, generative adversarial network, brain connectivity, serotonin genes, HTR5A, SLC6A4, default mode network, transcranial magnetic stimulation, genetic risk score, psychiatry