AI-driven multi-omics and genetic causal inference implicate immune proteins in major depression

Background

Depression is biologically heterogeneous and lacks validated molecular targets. Immune and inflammatory processes have been implicated in depression in observational and genetic studies, yet it remains unclear which factors play a causal role which reflect downstream disease processes. While multi-omics and machine learning (ML) can identify predictive biomarkers, whether these reflect causal mechanisms remains unclear. To identify multi-omics biomarkers predictive of depression and evaluate their causal relevance using Mendelian randomisation and colocalisation analyses, with a focus on distinguishing upstream risk factors from downstream disease processes.

Methods

We analysed plasma proteomic (N=53,130, 2,924 proteins) and metabolomic data (N=291,000, 326 metabolites​) separately and combined from UK Biobank to predict three depression phenotypes: ICD-defined (cases N=3,371), probable depression (cases N=33,851), and broad depression (cases N=166,474). We used eight ML algorithms with nested cross-validation (10×5 folds), with models run with and without adjustment for key confounders (sex, BMI, smoking, medication use, and deprivation index). Model performance was evaluated using AUC, sensitivity, specificity, and balanced accuracy, and feature importance was derived from penalised regression coefficients and SHAP values for black-box models. To prioritise causal candidates, we integrated ML with two-sample Mendelian randomisation (MR). Forward MR used cis-pQTLs (N=54,219) as instruments for circulating proteins [1] and a large GWAS of major depression (~357k cases, ~1.28M controls, European ancestry) as the outcome [2]. Reverse MR assessed the downstream effects of genetic liability to depression. Colocalisation analyses were performed to identify shared causal variants.

Results

Regularised linear models consistently outperformed other approaches, with proteomic models achieving the highest AUC for ICD-defined depression (~0.82). Model performance varied substantially by phenotype definition, with AUC increasing from ~0.62 (broad) and ~0.63 (probable) to ~0.82 for ICD-defined depression. Multi-omics integration (proteomics + metabolomics) did not substantially improve depression prediction beyond proteomics alone. ML prioritisation consistently highlighted immune and inflammatory pathways, including cytokine signalling and TNF/NFκB-related processes. Forward MR identified five circulating proteins associated with depression risk after FDR correction. BTN3A2 (IVW β=0.029, FDR-p=3.3×10⁻⁹), LYPD3 (β=0.031, FDR-p=0.002), and CEACAM21 (β=0.011, FDR-p=0.029) increased risk, while PILRA (β=−0.014, FDR-p=0.005) and CD302 (β=−0.041, FDR-p=0.01) showed protective associations. These results were concordant across sensitivity analyses, and there was no evidence of directional pleiotropy. Colocalisation analyses provided strong evidence of shared causal variants at select loci, particularly RELT (PP.H4 > 0.80), strengthening causal inference for specific protein–depression relationships, while other signals showed weaker or independent genetic architectures. In contrast, reverse MR identified widespread associations between genetic liability to depression and circulating proteins across immune, metabolic, and neuroendocrine systems, but with limited colocalisation support, suggesting these predominantly reflect downstream or pleiotropic effects.

Discussion

The marked improvement in predictive performance for ICD-defined depression suggests that a biologically informed signal is diluted in broader phenotypes. Additionally, integrating ML with genetic causal inference prioritises immune-related proteins as putative contributors to depression risk while distinguishing upstream causal mechanisms from downstream biological consequences. These findings highlight immune pathways as promising targets for precision psychiatry and demonstrate the value of combining multi-omics prediction with Mendelian randomisation in distinguishing association from causation.

References

1. Sun BB, Chiou J, Traylor M, Benner C. Plasma proteomic associations with genetics and health in the UK Biobank. Nature. 2023 Oct 12;622(7982):329-38.
2. Adams MJ, Streit F, Meng X, Awasthi S, Gerring ZF. Trans-ancestry genome-wide study of depression identifies 697 associations implicating cell types and pharmacotherapies. Cell. 2025 Feb 6;188(3):640-52.