TL;DR
Researchers at Southeast University and Nanjing Medical University ran a generative AI model over resting-state brain scans from 802 people and found that suicide-related brain wiring in bipolar depression splits into two reproducible subtypes — one rooted in visual-sensory circuits, one in rumination networks. The patterns held across two independent replication cohorts and tracked changes in suicide risk over time. The genetic and treatment signals are preliminary, but the framework points toward making suicide-risk assessment a biological science instead of a questionnaire alone.
Suicide risk is not one thing
Clinicians assess suicide risk with self-report and clinical interviews — tools that are vulnerable to concealment and fluctuation. There is no reliable biological test, and candidate biomarkers have repeatedly failed to replicate.
A team led by Ting Wang, Xinruo Wei and Qing Lu at Southeast University, with colleagues at Nanjing Medical University, asked whether the brain signatures of suicidal thoughts and behaviors in bipolar depression might be more than one phenomenon. Their answer, published September 24 in BMC Medicine: at least two, each with its own wiring pattern, genetic background, and cognitive profile.
The two subtypes
The first is visual cortex-predominant: hyperconnectivity in the visual processing circuits at the back of the brain, tied to greater anxiety and poorer cognitive flexibility and working memory. In this group, a three-variant serotonergic genetic risk score — rs1631327 and rs6320 in HTR5A, rs8066602 in SLC6A4 (the serotonin transporter targeted by many antidepressants) — tracked current-episode suicidal thoughts and behaviors, and 11.4% of that genetic effect ran through disrupted right peripheral visual wiring.
The second is DMN-CEN predominant: hyperconnectivity between the default mode network, active during self-referential thought, and the central executive network, which governs goal-directed attention — a pairing linked to brooding rumination. In exploratory analysis, among rs8066602 TC/TT carriers in this subtype, patients with suicidal thoughts and behaviors showed greater symptom improvement after two weeks of pharmacotherapy plus repetitive transcranial magnetic stimulation (β = 31.01, p < .01).
The AI that found them
The scale is unusual for psychiatric neuroimaging: 802 individuals total — resting-state fMRI from 657 bipolar depression patients, 405 with current suicidal thoughts and behaviors and 252 without, plus 103 healthy controls. One discovery cohort, two independent replication cohorts, and 42 patients followed longitudinally.
The model was Smile-GAN, a semi-supervised clustering-generative adversarial network. The semi-supervised part matters: it did not sort patients into arbitrary clusters — it searched specifically for wiring patterns that deviate from patients without suicidal thoughts and behaviors. The subtypes replicated at r = 0.80/0.64 and 0.65/0.73 on spatial-permutation testing (adjusted p < 0.001), and in the longitudinal arm, subtype-specific dysconnectivity rose and fell with individual suicide-risk fluctuations.
The honest caveats
The data are retrospective. The genetic score rests on three variants drawn from targeted sequencing of 61 suicide-related SNPs, and the treatment-response signal was exploratory, small, and two weeks long. The authors say these signals are preliminary and require prospective replication before anything reaches a clinic.
Still, the framework is the point: stratify into biologically coherent groups first, then hunt for predictors — instead of averaging everyone together and washing out the signal. Two gene-brain-behavior profiles that survived independent testing and tracked clinical change over time is, for this field, a real step.