Researchers will randomly allocate half of the participants to the ketogenic diet and the other half to the nutritionally balanced approach set out in the NHS EatWell Guide. Dietitians will support participants in following their assigned plan, while investigators monitor changes in depressive symptoms, including differences in mood and energy. The study will also examine how easy each diet is to follow and assess potential cost implications for the National Health Service. It will initially recruit in Scotland’s central belt before expanding to the Birmingham area in early 2027.

ENERGISE-BD follows a smaller University of Edinburgh study that suggested a ketogenic diet could be helpful for some people with bipolar disorder. The larger trial was funded by Wellcome and designed in partnership with people who have lived experience of bipolar depression. A Lived Experience Advisory Panel will remain involved in the research so that the study reflects patient priorities. Delivery and academic partners include the Edinburgh Clinical Trials Unit, NHS Lothian, and Birmingham and Solihull Mental Health NHS Foundation Trust, while Bipolar Scotland and Bipolar UK are among the supporting patient organisations.

About half of the participants will undergo brain scans intended to help researchers examine whether changes in brain chemistry are connected with any improvement in mood. The trial is testing both clinical outcomes and possible biological mechanisms associated with the two dietary approaches. A ketogenic diet restricts carbohydrates, which are the body’s usual source of glucose for energy. In the absence of carbohydrates, the body produces chemicals called ketones, which may be a more efficient fuel source for brain tissue. Whether this metabolic process affects depressive symptoms, and whether any observed changes are associated with brain-chemistry differences, remains an open question for the planned trial.

Daniel Smith, ENERGISE-BD co-lead from the University of Edinburgh’s Institute for Neuroscience and Cardiovascular Research, described the project as the world’s first large-scale randomized controlled trial of nutrition as a potential new treatment for bipolar depression. Steven Marwaha, co-lead from the University of Birmingham, said the study will evaluate whether a ketogenic diet or the NHS Eatwell plan can treat depressive symptoms in bipolar disorder. He said the trial would include a socioeconomically and ethnically diverse sample from the Midlands, intended to represent a wider range of people with the condition. The researchers characterized nutritional interventions for severe mental illness as novel and promising but cautioned that the research remains at an early stage.

Subject of Research: A randomized trial of ketogenic and NHS healthy-eating diets for depressive symptoms in bipolar disorder

Article Title: Innovative UK trial to explore impact of diet on bipolar depression

Article References: Innovative UK trial to explore impact of diet on bipolar depression. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: bipolar depression, ketogenic diet, randomized trial, clinical nutrition, depressive symptoms, brain metabolism, brain scans, mental health, Wellcome, NHS, Innovative, trial

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