Multivariate Brain-Blood Signatures in Early-Stage Depression and Psychosis

Abstract

Importance: Inflammation is increasingly implicated in the pathophysiology of mood and psychotic disorders. Integrating blood biomarkers and brain imaging may help uncover mechanistic pathways and guide targeted interventions.

Objective: To identify shared and distinct multivariate patterns of peripheral inflammation and gray matter volume (GMV) in early-stage depressive and psychotic disorders using a transdiagnostic machine learning approach.

Design, setting, and participants: The naturalistic multicenter PRONIA study was conducted between February 2014 and May 2019 with a follow-up period of up to 36 months; baseline data were analyzed between August 2021 and April 2024. Eight sites, including inpatient and outpatient facilities, in 5 European countries (Germany, Italy, Switzerland, Finland, and the United Kingdom) were included. The study included individuals with recent-onset depression (ROD, n = 163) or psychosis (ROP, n = 177) or clinical high-risk states for psychosis (CHR-P, n = 172), all with minimal medication exposure, and healthy control (HC) individuals (n = 166).

Exposures: Structural magnetic resonance imaging (MRI), peripheral assays of cytokines (eg, interleukin [IL] 6, IL-1β, tumor necrosis factor [TNF] α, C-reactive protein [CRP], brain-derived neurotrophic factor [BDNF], S100 calcium-binding protein B [S100B]); clinical assessments; neurocognitive testing.

Main outcomes and measures: After data collection, sparse partial least squares was used to identify latent brain-blood signatures. Support vector machine classification evaluated psychosocial and neurocognitive predictors of signature expression using repeated nested cross-validation.

Results: A total of 678 participants (346 [51.0%] female; median [IQR] age, 24.0 [20.9-28.9] years) were included. Four signatures were identified. A psychosis signature (ρ = 0.27; P = .002) differentiated ROP from CHR-P with elevated IL-6, TNF-α, and reduced CRP, alongside GMV shifts in corticothalamic circuits. A depression signature (ρ = 0.19; P = .02) differentiated ROD from HC individuals with elevated IL-1β, IL-2, IL-4, S100B, and BDNF and GMV reductions in limbic regions. Additional signatures reflected age (ρ = 0.67) and sex or MRI quality (ρ = 0.53). Psychosocial features, including a differential childhood trauma pattern, predicted both the psychosis (balanced accuracy [BAC] = 67.2%) and depression (BAC = 78.0%) signatures. Cognitive performance predicted only the psychosis signature (BAC = 65.1%).

Conclusions and relevance: In this study, early-stage depression and psychosis exhibited distinct neurobiological signatures involving immune and neuroanatomical markers, challenging fully dimensional disease models. These signatures are shaped by childhood trauma and cognition and may support biologically informed early interventions.

Figure 1.. Psychosis Signature

A, The bar plot displays both direction and magnitude of feature weights of the blood parameter pattern of latent variable 3 (LV3). If 2 feature weights had the same sign (ie, both positive or both negative), the respective features covaried positively with each other; an opposite direction of feature weights represents a negative covariation. Zero weights indicate that there was no significant contribution of the respective features to the covariance signature. Positive weights were assigned to recent-onset psychosis (ROP) status, age, body mass index (BMI), and levels of interleukin 6 (IL-6), tumor necrosis factor α (TNF-α), clinical high risk for psychosis (CHR-P), and C-reactive protein (CRP) status were negatively weighted. B, The brain pattern of LV3 was mapped onto the MNI152 standard space via the open-source 3-dimensional rendering software Connectome Workbench version 1.4.2. The spider plots highlight the top neuroanatomic brain regions (C; derived from the Brainnetome and Diedrichsen atlases) and corresponding functional networks (D; derived from an adapted, 8-network solution of the Yeo and Buckner atlases) according to the percentage of positive and negative voxels in these regions. Amyg indicates amygdala; BG, basal ganglia; CerH, cerebellum hemisphere; CG, cingulate gyrus; FuG, fusiform gyrus; Hipp, hippocampus; IFG, inferior frontal gyrus; INS, insular gyrus; IPL, inferior parietal lobule; ITG, inferior temporal gyrus; LOcC, lateral occipital cortex; MFG, middle frontal gyrus; MTG, middle temporal gyrus; MVOcC, medioventral occipital cortex; OrG, orbital gyrus; PCun, precuneus; PCL, paracentral lobule; PhG, parahippocampal gyrus; PoG, postcentral gyrus; PrG, precentral gyrus; pSTS, posterior superior temporal sulcus; SFG, superior frontal gyrus; SPL, superior parietal lobule; STG, superior temporal gyrus; Tha, thalamus; TNF, tumor necrosis factor.

Figure 2.. Depression Signature

A, The bar plot displays both direction and magnitude of feature weights of the blood parameter pattern of latent variable 4 (LV4). If 2 feature weights had the same sign (ie, both positive or both negative), the respective features covaried positively with each other; an opposite direction of feature weights represents a negative covariation. Zero weights indicate that there was no significant contribution of the respective features to the covariance signature. Positive weights were assigned to recent-onset depression (ROD) status, interleukin 1 receptor antagonist (IL-1RA), interleukin 4 (IL-4), S100 calcium-binding protein B (S100B), interleukin 1β (IL-1β), interleukin 2 (IL-2), and brain-derived neurotrophic factor (BDNF). Healthy control status was negatively weighted. B, The brain pattern of LV4 was mapped onto the MNI152 standard space via the open-source 3-dimensional rendering software Connectome Workbench version 1.4.2. The spider plots highlight the top neuroanatomic brain regions (C; derived from the Brainnetome and Diedrichsen atlases) and corresponding functional networks (D; derived from an adapted, 8-network solution of the Yeo and Buckner atlases) according to the percentage of positive and negative voxels in these regions. Amyg indicates amygdala; BG, basal ganglia; CerH, cerebellum hemisphere; CG, cingulate gyrus; FuG, fusiform gyrus; Hipp, hippocampus; IFG, inferior frontal gyrus; INS, insular gyrus; IPL, inferior parietal lobule; ITG, inferior temporal gyrus; LOcC, lateral occipital cortex; MFG, middle frontal gyrus; MTG, middle temporal gyrus; MVOcC, medioventral occipital cortex; OrG, orbital gyrus; PCun, precuneus; PCL, paracentral lobule; PhG, parahippocampal gyrus; PoG, postcentral gyrus; PrG, precentral gyrus; pSTS, posterior superior temporal sulcus; SFG, superior frontal gyrus; SPL, superior parietal lobule; STG, superior temporal gyrus; Tha, thalamus.

Figure 3.. Post Hoc Exploration of the Psychosis and Depression Signature

Prediction of latent variables (LVs) 3 (A) and 4 (B) high and low scorers using Support Vector Machine Classification with psychosocial (ie, Childhood Trauma Questionnaire [CTQ], Global Assessment of Functioning, Disability and Impairment [GAF:D/I], Global Assessment of Functioning, Symptoms [GAF:S], Global Functioning, Role [GF:R], Global Functioning, Social [GF:S], Premorbid Adjustment Scale [PAS], NEO Five-Factor Inventory [NEO-FFI], World Health Organization Quality of Life–Brief Version [WHOQOL-BREF]) (LV3: balanced accuracy [BAC], 67.97%; LV4: BAC, 78.00%), and neurocognitive data (ie, social cognition, verbal learning, speed of processing, global cognition, attention, reasoning, and working memory) (LV3: BAC, 65.06%; LV4: BAC, 55.73%). The significance of the predictive features was assessed by means of sign-based consistency. FDR indicates false discovery rate.