Neurobiological signatures of stressor reactivity in early psychotic and affective disorders

Resilience to stress and adversity varies substantially across individuals and is particularly relevant in psychosis risk, where identifying protective and vulnerability-related mechanisms may inform early detection and prevention. Stressor reactivity (SR), defined as the discrepancy between expected and observed mental health response given exposure to stressful life events, may provide a clinically meaningful marker of individual adaptation to environmental adversity. However, the neurobiological mechanisms underlying interindividual differences in SR remain unclear. Here, we aimed to use machine learning to identify brain regions and functional patterns underlying individual differences in SR across healthy individuals and patients with early affective and psychotic disorders.

We analyzed data from 460 healthy controls, 285 individuals at clinical high risk for psychosis, 293 patients with recent-onset depression, and 296 patients with recent-onset psychosis from the Personalised Prognostic Tools for Early Psychosis Management (PRONIA) study. SR was computed as the difference between expected and observed mental health response, indexed by the Beck Depression Inventory-II, conditional on stressor exposure assessed with the Cologne Life Events scale. Linear support vector regression models were trained to predict SR using grey matter volume derived from T1-weighted magnetic resonance imaging (MIR), as well as fractional amplitude of low-frequency fluctuations (fALFF) and functional connectivity derived from resting-state functional MRI (rsfMRI). Additional models included Brain Age Gap Estimation (BrainAGE) scores combined with demographic and neurocognitive features, as well as polygenic risk scores indexing genetic liability for common psychiatric disorders. Models were evaluated in a repeated nested cross-validation framework with 5×5 inner and outer folds and permutations. While all imaging modalities were corrected for age, sex, and site, grey matter volume was additionally corrected for total intracranial volume, and rsfMRI measures for framewise displacement.

SR was significantly predictable from all neuroimaging modalities with the strongest predictive performance for fALFF, yielding an explained variance of R² = 8.31% (r = 0.29, mean absolute error [MAE] = 1.14, p < 0.001). Functional connectivity also predicted SR significantly, with R² = 5.41% (r = 0.23, MAE = 1.18, p < 0.001). Grey matter volume showed comparable predictive performance with R² = 5.81% (r = 0.24, MAE = 1.28, p < 0.001). In contrast, BrainAGE-demographic-neurocognitive models showed only weak predictive performance, with the best-performing model explaining R² = 0.73% of variance (r = −0.09, MAE = 1.79, p = 0.006). Similarly, PRS-based models also showed weak but significant prediction of SR, with the best-performing model explaining R² = 0.50% of variance (r = 0.07, MAE = 1.82, p = 0.036).

These findings suggest that individual differences in SR can be predicted from multimodal features, with structural and particularly resting-state functional neuroimaging providing the strongest predictive signal. Compared with BrainAGE-demographic-neurocognitive and genetic risk models, neuroimaging features showed substantially higher predictive relevance, supporting their potential utility for identifying brain mechanisms of stress resilience and vulnerability in early affective and psychotic disorders.

References

No references