Research Outputs

Authors

Bianca Branco, Sam Bentwood, Esme Elsden, Matthew Iveson, Mark Adams, Andrew McIntosh, Peggy Seriès, Heather Whalley, Alex Kwong

Abstract

Prediction and prevention of depression remain challenging due to heterogeneity in how we define the disease and the complex aetiology of genetic and environmental factors that underpin it. These challenges limit the development of robust and generalisable predictive models needed for precision psychiatry. We evaluated key predictive factors for four depression phenotypes within the UK Biobank (max N = 153,491). These included one measure of current depressive symptoms during the previous two weeks (Patient Health Questionnaire-9 [PHQ-9]), and three lifetime measures reflecting history of depression (self-reported diagnosis, symptom-based screening using the Composite International Diagnostic Interview [CIDI], and a composite measure incorporating self-report, CIDI and PHQ-9). We examined the prediction performance using various genetic and commonly ascertained environmental predictors. Among the five machine models evaluated, CatBoost achieved the best performance across the four phenotypes, with the composite definition providing the most stable results (AUC-ROC: 0.75 and AUC-PR: 0.60). Relative SHAP importance of predictors was highly consistent across lifetime depression definitions (Pearson’s ρ = 0.92–0.99), but less consistent between lifetime depression and PHQ-9 (ρ = 0.24–0.44). Across lifetime phenotypes, the most influential predictors included sex, age, insomnia, self-rated health and depression polygenic risk scores (PRS), while current depressive symptoms were more strongly associated with modifiable factors including social connectedness and financial difficulties. Candidate PRS-by-environment interactions involving age, sleep, and health were identified by both SHAP interaction and regression analyses. Our findings demonstrate how depression phenotype definitions shape predictive performance and predictor importance rankings, reinforcing the need to consider phenotype heterogeneity and gene-by-environment interactions when developing depression prediction models.

Artur Miralles Méharon, from our 2024 cohort, was a co-author on a high impact paper “Replicated blood-based biomarkers for myalgic encephalomyelitis not explicable by inactivity”, published in June 2025.

Beentjes, S. V., Miralles Méharon, A., Kaczmarczyk, J., Cassar, A., Samms, G. L., Hejazi, N. S., Khamseh, A., & Ponting, C. P. (2025). Replicated blood-based biomarkers for myalgic encephalomyelitis not explicable by inactivity. EMBO Molecular Medicine, 17(7), 1868–1891.

https://doi.org/10.1038/s44321-025-00258-8

It concerns the largest ever biological study of ME/CFS (myalgic encephalomyelitis/chronic fatigue syndrome) that has identified consistent blood differences associated with chronic inflammation, insulin resistance and liver disease.

Significantly, the results were mostly unaffected by patients’ activity levels, as low activity levels can sometimes hide the biological signs of illness, researchers say.

The volume and consistency of the blood differences support the long-term goal of developing a blood test to help diagnose ME/CFS. ME/CFS’ key feature, called post-exertional malaise, is a delayed dramatic worsening of symptoms following minor physical effort. 

Other symptoms include pain, brain fog and extreme energy limitation that does not improve with rest. Causes are unknown and there is currently no diagnostic test or cure.

Researchers at the Institute of Genetics and Cancer worked with colleagues in the University of Edinburgh’s Schools of Mathematics and Informatics to better understand the biology that underpins the condition.

They used data from the UK Biobank – a health database of over half a million people – to compare 1,455 ME/CFS patients with 131,000 healthy individuals. They examined more than 3,000 blood-based biomarkers and used advanced models to account for differences associated with age, sex, and activity levels. 

The results, which were replicated afterwards using data from the US, showed that hundreds of biomarkers differed between ME/CFS patients and healthy people. 

Some 116 significant differences were found in both men and women, a key finding as ME/CFS can affect sexes differently. The consistent results across both groups strengthens the reliability of the biomarkers.  The strongest biomarker differences were found in people who reported symptoms consistent with post-exertional malaise, highlighting its central role in the illness. 

Researchers believe these biomarker changes are more likely a result of ME/CFS, rather than the initial trigger of the illness.

University of Edinburgh researchers were supported by partners from the Harvard T.H. Chan School of Public Health.

• BBC news story

In September 2026, our 2024 cohort students Emilia Agasi, Elisa Castagnari and Nuria Fabrega, together with colleagues, developed and delivered a full-day tutorial at the European Conference on Computational Biology (ECCB 2026) in Geneva.

The tutorial, From Multi-Omics to Gene–Disease Discovery: Knowledge Graphs and LLM-Augmented Analysis, explored how multi-omics data, knowledge graphs and large language models can be combined to support gene–disease discovery. The programme covered gene expression networks and knowledge graph construction, approaches for multi-omics integration, and the development of agentic LLM workflows for biomedical knowledge graphs.

Participants worked through practical sessions using real biomedical datasets, including cancer data from The Cancer Genome Atlas, gene expression data related to autism, and data from the Generation Scotland study. The hands-on exercises covered phenotype knowledge graph construction, multi-omic factor analysis (MOFA), querying LLM agents with molecular profiles, and using LLMs to query knowledge graphs and identify gene–disease relationships.

The tutorial was designed as a practical, reproducible resource, with Python notebooks, accompanying datasets and a dedicated JupyterHub environment for live coding. The tutorial materials and code have been made openly available through a public Jupyter Book and GitHub repository.

In March 2026, our 2024 cohort researcher Melina Müller, alongside her supervisors Nina Kudryashova and Luke Bashford, organised and delivered the workshop Edge of Tomorrow: Theory and Practice of Closed-Loop Neural Control at COSYNE 2026 in Cascais, Portugal.

The workshop focused on the emerging field of closed-loop neural control: moving beyond the passive recording of neural activity towards actively probing and controlling neural dynamics. It brought together researchers working across control theory, computational neuroscience and experimental systems neuroscience to explore how theoretical methods for neural system identification, modelling and control can be translated into practical experimental settings.

A particular focus was on applications in human electrophysiology, including brain-computer interfaces and intracranial stimulation. Sessions covered data-driven models of brain and behaviour, optimal stimulation design, real-time learning of neural dynamics, and the number of recording and stimulation channels required to control latent neural dynamics. The programme featured invited talks from researchers working on human cortical microstimulation, implanted brain-computer interfaces, dynamical systems, optimal control and real-time neural modelling, followed by a panel discussion bringing together speakers and participants.

The workshop provided an opportunity to connect theoretical developments with the practical challenges of implementing closed-loop neural control in experimental settings, while highlighting the potential of these approaches to enable new ways of investigating and influencing neural computation.

In May 2026, Imperial College London hosted the annual UKRI Machine Intelligence for Health conference at the University of Warwick. It brought together PhD researchers and early-career scientists from across the UKRI AI for Healthcare Centres for Doctoral Training to share research, exchange ideas, and build new collaborations across AI, healthcare, and biomedical innovation.

Our 2024 cohort researcher, Bianca Branco, presented her work on machine learning prediction of depression across phenotype definitions integrating genetic and environmental factors, winning the runner-up prize for best oral presentation!

In July 2026, our 2025 cohort researcher Nikoo Moradi attended the FENS–Chen Institute Summer Program on AI-Accelerated Neuroscience Discovery and Translation, hosted by the Cambridge Centre for Data-Driven Discovery.

The experience involved a week full of lectures on AI and neuroscience, from spiking neural networks to computational psychiatry to AI for neurodegenerative disorders. It was an amazing opportunity to learn directly from some of the leading researchers in the field.

Nikoo presented a poster on her latest results in discovering an EEG-based biomarker for SYNGAP1 related disorder, and received valuable feedback and ideas from people across different areas of neuroscience. 

In May 2026, Imperial College London hosted the annual UKRI Machine Intelligence for Health conference at the University of Warwick. It brought together PhD researchers and early-career scientists from across the UKRI AI for Healthcare Centres for Doctoral Training to share research, exchange ideas, and build new collaborations across AI, healthcare, and biomedical innovation.

Across keynote talks, oral presentations, and poster sessions, attendees showcased cutting-edge doctoral research spanning:

• foundation models and multimodal AI
• medical imaging and diagnostics
• clinical decision support
• digital health and wearable technologies
• causal and explainable AI
• translational healthcare applications

It was great to see such a broad range of interdisciplinary research and to create space for researchers from different centres and backgrounds to connect and exchange ideas.

In January 2026, our 2024 cohort researchers, Núria Fàbrega Ribas and Emilia Agasi presented their work at FOG26 in London, accompanied by their supervisor and our Director, Ian Simpson.

A highlight for Núria was the workshop on “Making Spatial Biology FAIR” organised by the Initiative for FAIR Spatial Data IO-FAST. Beyond the session itself, the opportunity to ask questions and speak directly with people working on metadata standards was especially valuable and led to several insightful conversations.

For Emilia, some highlights were Ketan Patel from Bioptimus talking about their new multimodal foundation model building on H-optimus-1 and Greg Slabaugh on the interpretable fusion of histopathology and transcriptomics data, as well as Paula Cunnea on the spatial heterogeneity of high grade serous ovarian carcinoma and Karen Sayal on the integration of AI into early phase oncology trials.

What FOG does really well is provide a collaborative space for conversations on innovative research and their practical implementation in clinical settings.

In December 2025, Artur Miralles Méharon attended EurIPS 2025 in Copenhagen. The conference brought together researchers from academia and industry, with a strong emphasis on methodological foundations, sustainability, uncertainty, and the translation of AI research into real-world applications.

Artur attended the Causality for Impact workshop, focused on the challenges in applying causal methods to real-world problems, particularly in health, earth and social sciences. Here he presented his first academic poster, entitled “Replicated blood-based biomarkers for Myalgic Encephalomyelitis not explainable by inactivity”, which enabled constructive discussion with both method developers and applied researchers.

In October 2025, our 2024 cohort researcher Bianca Branco presented her poster and a lightning talk about using ML methods for depression prediction at the UKRI Mental Health Platform Research Summit. The event perfectly coincided with World Mental Health Day, providing an important opportunity to reflect on mental health and raise awareness.

In June 2025, our students Elisa Castagnari and Núria Fàbrega Ribas had a fantastic experience representing AI4BI at HealTAC2025 in Glasgow. 

For Elisa, it was a milestone — her first poster presentation, accompanied by a lightning talk, both of which sparked insightful discussions around her research on improving digital healthcare solutions with data interoperability and LLMs.