Biomarkers in psychiatric disorders: status quo, impediments and facilitators
Michael Berk
Abstract
As probes of the beating heart of a disorder, few research domains match both the promise and complexity of biomarkers. From monoamines to cortisol, inflammatory markers, neuroimaging and cognition, serial waves of enthusiasm have broken concerning biological markers in psychiatry only to dissipate feebly on the shores of research validation, but research is still very active in this area. Biomarkers have diverse potential roles: there may be biomarkers of risk, of diagnosis/trait, of state or acuity, of stage, of treatment response, and of prognosis1. This classification is not arcane; a marker might succeed in one domain but fail in others – there are multiple examples in general medicine that this is indeed the case. In this issue of the journal, Abi-Dargham et al2 explore the most promising candidate biomarkers in major mental disorders. They highlight an electroencephalographic event-related brain potential, the N170 signal, for autism spectrum disorder; striatal resting-state functional magnetic resonance imaging (fMRI) measures for schizophrenia; an electrophysiological metric, error-related negativity, for predicting the onset of generalized anxiety disorder; and resting-state and structural brain connectomics for social anxiety disorder. All of these candidate biomarkers await confirmation by definitive and replicated studies. There are multiple hurdles to be cleared in the race to the finishing line of clinical translation of biomarkers. One of the most significant ones is related to current diagnostic classifications. It is implausible that symptom-based classifications can cleave the biology of nature at its joints, yet they remain the reference point against which biomarkers are indexed. Most psychiatric disorders are extremely heterogeneous and at the same time overlap extensively with other disorders. Comorbidity, with other psychiatric disorders and with non-communicable physical disorders, is the rule, and both can influence any exploratory marker. There are also extensive interactions between any potential marker and a plethora of variables, including early life experiences, genetics and epigenetics, current stressors, medications and other therapies, environmental and lifestyle risk factors, stage of illness trajectory, age, as well as secondary biological adaptations to these variables. A frequent stumbling block is power, with most biomarker studies of relatively small sample size confounding the efforts to detect influences which generally are of similarly small effect size. Aggravating this is the selection of controls: many studies compare healthy “supernormal” controls with clinical populations, amplifying perceived differences. Methodological problems are also legion. Even within a singular disorder such as schizophrenia, there are large differences driven by stage, comorbidity, inpatient or community setting, background treatment, and many more. Several markers, such as cytokines, are highly sensitive to collection variables – including phase of menstrual cycle, fasting status, concomitant medications, and diurnal rhythms – and to environmental factors such as smoking, physical activity, nutritional status, as well as substance and alcohol abuse. Also, most biomarker studies use stored samples, and many analytes deteriorate significantly with storage. Most biomarker studies are cross-sectional in nature, a design which does not disentangle state and trait effects and cannot inform causal associations. Even if a biomarker is found, it may reflect another factor: for example, low vitamin D appears to be a marker of a sedentary lifestyle – a consequence rather than a cause. Additionally, there are huge commercial and personal interests in this area, with biotech companies and individuals incentivized to be overly optimistic and amplify promise. Against all these challenges, the bar for adoption by clinicians and funders remains extremely high: to achieve clinical utility, any marker needs to have very high sensitivity and specificity, as well as low complexity, low cost and easy integration into clinical care. The limitations of single marker studies, alongside the availability and increasing capacity of omics technologies, have catalyzed a series of studies using these latter technologies. Platforms exist for metabolomics, transcriptomics, genomics, proteomics and lipidomics amongst others3. This facilitates simultaneous dynamic assessment of multiple metabolites and can capitalize on systems biology approaches that appreciate the tight interconnection between multiple processes, including inflammation, oxidative biology, cell signalling pathways, lipid biology and cellular metabolism. Meta-analyses of omics studies have found several lipidomic abnormalities in mood disorders4. Multimodal neuroimaging approaches combining clinical and imaging data might predict treatment outcomes5. Combinations of different omics modalities among themselves and with data sources such as neuroimaging and cognition offer promise. Such large-scale data allow artificial intelligence analysis. The stratified medicine approach also provides a template to bypass the lack of gold standard pathophysiology. As an example, even though the pathophysiology of breast cancer is only partly understood, recognition of the overexpression of human epidermal growth factor receptor subtype 2 (HER2) in breast cancer facilitated enhanced prognostic models and the development of monoclonal antibodies against this receptor subtype. Similarly, even though the cause of colorectal cancer is unknown, the characterization of KRAS mutations enabled stratification and detection of those who may respond to epidermal growth factor receptor (EGFR) inhibitors such as cetuximab. In psychiatry, pharmacogenomics targeting P450 enzymes offers the possibility of detecting individuals who might need higher or lower doses of medication, which can increase the likelihood of response to therapy. It is critical to be mindful that failure or success in one domain – such as diagnosis or trait, state or stage, response or prognosis – does not imply outcomes in another domain. As an example, structural neuroimaging has not proved so helpful in informing the differential diagnosis of depression. But a multimodal imaging approach was more promising in predicting the clinical course and outcome of this condition6. While a definitive Google map to the destination is not available, the following road signs are likely to be useful. First, it needs to be emphasized that the bar for biomarker research should not be based on p values or effect sizes, but on sensitivity and specificity, or positive and negative predictive value in a clinical context. Any clinically impactful test must be both cost-effective and simple to implement. As the field moves from singular to aggregate and more complex markers, this barrier increases in height. Second, the field needs to adopt a consistent terminology of the different biomarker domains. Third, like clinical trials, biomarker studies need rigorous a priori power calculations and concomitantly adequate sample sizes. To increase methodological rigour, biomarker studies should ideally be pre-registered, with pre-specified primary outcomes, criteria for multiplicity, and assessment criteria. Fourth, rigorous guidelines for methodological standardization, for example guidelines for the collection and preparation of biological samples7, are needed. Finally, consistent reporting standards, such as the STARD (Standards for Reporting of Diagnostic Accuracy) framework, will enhance the field8. Notwithstanding the necessity of hypothesis-generating studies, markers with a plausible link to known pathophysiology should be prioritized. Complementing top-down disorder-based approaches, bottom-up symptom- and symptom cluster-based approaches might add value; exemplars include biomarker stratification based on typical vs. atypical symptoms in depression. Automated collection of selected measures by remote sensing using digital technologies offers promise, and such large-scale data are amenable to artificial intelligence methodologies. Large-scale and long-duration international longitudinal cohort studies with repeated multiple biomarkers that span peripheral blood measures, electrophysiology, neuroimaging and cognitive neuroscience, mirrored by deep clinical phenotyping, are likely to be a path forward – the planned BD2 Integrated Network for Bipolar Disorder is an exemplar9. In conclusion, considerable progress has been made in identifying a diverse suite of candidate biomarkers in psychiatry, but substantial challenges remain. Fortunately, multiple promising approaches are on the horizon. But, mindful of the road travelled so far, and the obstacles ahead, T. Bayes may still have the last word.