What if Severe Mental Illness (SMI) leaves a metabolic/chemical fingerprint?

Author: Stephen Binaansim

Image by Javaistan from Pixabay, Sept. 2026

Imagine walking into a clinic feeling sick, but instead of relying on a straightforward lab result, your diagnosis depends almost entirely on how you describe your feelings. Unlike most medical fields, where clear biological tests instantly unveil what is wrong inside the body, psychiatric diagnosis has long operated in the shadows of biological uncertainty. For many diseases, diagnosis is well supported by simple biological measurements: a blood sample and glucose test for diabetes, a CT scan for a tumor, or a genetic test for a genetic disease, but this is not the case for psychiatric conditions. Severe Mental Illness (SMI), which include conditions such as bipolar disorder, schizophrenia, and severe depression, have relied heavily on a patient’s symptoms and the experience and expertise of psychiatric professionals. These evaluations are subjective, increasing the risk of misdiagnosis, mistreatment, or delayed preventive interventions (Zhang et al., 2026).

SMIs are highly influenced by many factors including genetics, environmental factors, lifestyle, medications, and interactions, which are primarily rooted in brain-body connections via biochemical reactions (metabolism), leaving a trail of biological clues that remain to be investigated. Therefore, two people with the same mental health diagnosis can have very different underlying biology, though with similar symptoms (Shi et al., 2024)(Cavaleri et al., 2026) . This has led to fundamental questions that researchers keep asking: beyond their symptoms, can we find measurable biological signatures of SMIs that reveal what mere observations of symptoms cannot?

One way to find out is METABOLOMICS, measuring many of the small molecules found in a sample of blood or another body fluid. These molecules, known as metabolites, are produced as the body breaks down food, makes energy, and responds to its surroundings. To understand the advantage of this approach, we first need to look at what metabolism is and what it can reveal.

Reading The Body’s Chemistry

This brings to mind the famous assertion that ‘Chemistry is life’—a quote made by the fictional 1960s protagonist character, Elizabeth Zott in Bonnie Garmus’s novel Lessons in Chemistry (Garmus, 2022), which explains the chemistry of the human body. The human body is constantly undergoing chemical reactions through interactions among small molecules. It produces, consumes, and transforms thousands of these metabolites, including lipids, amino acids, sugars, and other molecules involved in energy production (Shih, 2019).

A collection of these molecules provides a biochemical fingerprint that offers insight into what is happening in the body at a given moment. This can be measured using technologies such as mass spectrometry, which measures large quantities of these molecules at once. Instead of looking at one molecule at a time, researchers can examine patterns across hundreds or thousands of metabolites (Cavaleri et al., 2026).

This is important to metabolic psychiatry, since metabolism sits at the intersection of many influences. Genes provide instructions, but metabolism reflects what is happening in the cell, capturing influences such as diet, medication, sleep, exercise, and the environment (Shih, 2019)(Cavaleri et al., 2026). This has opened many doors, and current research indicates that brain energy metabolism contributes significantly to the core symptoms of SMIs, suggesting that altered metabolism may be related to psychiatric symptoms and affect the biological patterns researchers observe (Farran et al., 2026).

Brain-body metabolism is shaped by multiple factors, including exercise, diet, sleep, and genetics (Created by Stephen Binaansim with Canva, Sept. 2026)

For example, a recent 2025 review of schizophrenia metabolomics identified recurring metabolic patterns, including amino acid and fatty acid metabolism, that have been shown to be associated with schizophrenia, offering diagnostic biomarkers—measurable clues that can define the disease state (Yao et al., 2025). But finding the metabolic differences in disease states is only the beginning.

A molecule that differs between two people does not automatically become a biomarker, since it may be greatly influenced by factors such as medication, diet, environmental exposure, or other underlying causes of the illness rather than being the cause itself. A biomarker then becomes genuinely useful when researchers can demonstrate that it is reliable, reproducible, and useful in human subjects beyond the original study.

There is another challenge: metabolomics produces far more information than a person can easily interpret by eye.

From Static to Dynamic Metabolomics: The Role of AI

This is where Artificial Intelligence comes in. Machine learning (ML) algorithms can learn patterns across multiple variables simultaneously. Take, for example, measuring 1000 metabolites in a sample collected from 500 people in a study. It becomes much more interesting when 30 samples are collected from each person. A researcher can compare groups manually, but the number of possible biological relationships becomes unbearable and overwhelming.

Machine learning models can instead learn from data and search for a combination of features that effectively distinguishes between groups or predicts each disease state.

A recent study by Zhang et al., (2026) on the diagnosis of mental health disorders analyzed blood samples collected from 173 healthy controls (HC) and 100 cases each of major depressive disorder (MDD), bipolar disorder (BD), schizophrenia (SCZ), and anxiety disorder (AD), identifying 634 metabolomic and 2,391 protein features. The study identified eight metabolic biomarkers and developed four machine learning (ML) models that distinguished healthy controls from individuals with mental disorders. The models achieved scores close to perfect separation between the groups. However, such results need to be replicated in larger, independent studies before they can support clinical diagnosis (Zhang et al., 2026).

Though this demonstrates the great success and practical usefulness of ML in the diagnosis of SMIs, it creates another problem: A model may correctly identify a pattern without explaining why that pattern is associated with a condition. This is often called the ‘black box’ problem.

Advances in metabolomics technology moving from static snapshots in time to dynamic metabolomics (created by Stephen Binaansim with Canva, Sept. 2026).

Traditional metabolomics measures a static snapshot of the metabolome, but biology is not static. Imagine taking a blood sample at 9am. It tells a person’s metabolism at 9am, but metabolism changes throughout the day. Our bodies follow an approximately 24-hour biological cycle known as a “Circadian Rhythm” where hormones, metabolism, sleep, body temperature, and many other biological processes fluctuate and repeat over time. Mental health has been shown to be closely linked to biological timing, with sleep disruptions and altered metabolic rhythms affecting several mental health conditions (Rakshasa-Loots et al., 2025).

Hence, perhaps the question is no longer, “what metabolites does a person have?” but rather “how does their metabolism change in time?” Though continuous serum glucose monitoring has made great progress in measuring glucose metabolites over time, measuring a large number of metabolic molecules over long periods remains challenging. This would require collecting blood samples at minute-, hourly-, or daily-sampling intervals, which would be inconvenient for subjects and is a research challenge currently being addressed by Upton et al., (2023). Thus, introducing another layer of complexity, for which machine learning may be a useful analytical tool.

A single metabolomic measurement is like a photograph. Repeated measurements can produce a film of metabolic changes over time.

From Diagnosis to True Biology!

In summary, these ideas suggest a shift from asking whether a molecule is present to asking how the body’s chemistry changes. Psychiatry has diagnosed mental illness mainly through what we can observe: mood swings, behaviour, cognition, and perception. Though these observations remain essential, metabolomics offers an extra layer that shows the biochemical processes occurring throughout the body, not just in a snapshot but over time. AI can navigate this enormous and complex body of information, providing valuable insights to researchers.

Neither of these technologies solves all our problems. Metabolomics cannot offer a simple chemical test for mental illness, whereas AI cannot independently tell why a person develops BD or SCZ from the data it analyses. But these two technologies can allow researchers to ask increasingly complex questions, such as “which specific metabolite can predict a mental illness?” AI may help researchers discover that the biological categories underlying severe mental illness are more nuanced — and potentially more measurable — than our current diagnostic labels suggest. A conventional metabolomic experiment gives us a snapshot of biology. High-resolution measurements collected over time could give us something closer to a film and artificial intelligence brings everything together, helping us read the entire metabolic psychiatry story.

Article written by Stephen Binaansim, A bioinformatician and UKRI-funded PhD researcher at the University of Edinburgh AI4BI CDT, specialising in artificial intelligence, metabolomics, and data-driven approaches to improving human health.

Article edited by Luka Daniel, a fourth-year MPhys Astrophysics student at the University of Edinburgh and an Online Editor for EUSci.