2026 ECNP Congress

PS02-2037

Integrating transcriptome and metabolome data for treatment outcome prediction in depression through multi-omics factor analysis

Lea Steinbach — T. Lange, S.M. Wissing, C. Hohoff, L.J. Schmidt, E. Van Assche, B.T. Baune · University of Münster, Department of Psychiatry, Münster, Germany

2026 ECNP Congress

Background

There is a strong need for improved prediction of treatment outcome for major depressive disorder (MDD). Multi-omics-based machine learning (ML) has potential to detect the molecular patterns underlying treatment response. However, integration and dimensionality reduction of multi-omics data while preserving intra- and inter-modal patterns and biological interpretability remain challenging. We aim to explore the use of multi-omics factor analysis (MOFA) for unsupervised integration of transcriptome and metabolome data and prediction of treatment response in MDD.

Methods

The ongoing Esketamine multi-omic biomarker evaluation in MDD (EMBER-MDD) study aims to identify biosignatures of treatment resistance in patients treated with esketamine or treatment as usual. We present work based on 140 patients (Mage=38, 45% women) from EMBER-MDD. Baseline transcriptome and metabolome data were obtained from peripheral blood. Following standard procedures and QC, data on 23,163 transcripts (mRNA-seq: NovaSeq-6000/-X, Illumina) and 582 metabolites (mass spectrometry: MxP500 XL, Biocrates) were available for MOFA. Moreover, genomic variants (Infinium GSA v3.0+MD, Illumina) were used to calculate the polygenic load for MDD [1] for the final ML model. MOFA was performed in mofapy2 using transcriptome and metabolome data (15 initial factors, variance threshold=1%). Initially, MOFA was run to explore inferred latent factors independent from treatment outcome. The best MOFA model from 25 random initializations was selected for downstream analysis in MOFA2, investigating variance explained, factor loadings, enriched pathways and associations with clinical variables. Afterwards, a ML model predicting treatment response was built. Response was defined as a 50% reduction in depression severity (MADRS or HAM-D) from begin to end of treatment. A XGBoost classifier was trained using 5-fold cross-validation based on area under the curve (AUC) scores. Input data included six clinical variables, polygenic MDD load and latent factors of transcriptome and metabolome data from MOFA. Factors of test data were estimated based on weights learned on train data only.

Results

MOFA identified 10 factors which explained 27% of variance in transcriptome data and 62% of variance in metabolome data. Factor 1 was most active in metabolome data (49% variance explained) and factor 2 was most active in transcriptome data (11% variance explained). Loadings for factor 1 were driven by triglycerides. Enrichment analysis of factor 2 revealed enriched immune response and protein synthesis pathways. Factor 1 was associated with sex (Wilcoxon rank-sum test, p<0.001) and positively correlated with age (Spearman’s correlation, p=0.004). The ML model for predicting treatment response showed poor performance with just above random predictions (mean AUC=0.58) and strong fluctuations between folds (AUC range: 0.43-0.79).

Conclusions

Unsupervised dimensionality reduction and integration of transcriptome and metabolome data through MOFA identified latent factors that capture clinically and biologically interesting drivers of variance such as enriched immune response pathways, supporting the idea of inflammation-based stratification of patients. The ML model, however, failed to discriminate treatment responders from non-responders. We expect model performance to improve as the number of patients participating in EMBER-MDD increases. Overall, MOFA appears promising for condensing multi-omics data into latent factors, but further research is needed to translate these into meaningful predictors of treatment response.

References

  1. Major Depressive Disorder Working Group of the Psychiatric Genomics Consortium, 2025. Trans-ancestry genome-wide study of depression identifies 697 associations implicating cell types and pharmacotherapies. Cell 188 (3), 640–652.