An interpretable machine learning pipeline based on transcriptomics predicts phenotypes of lupus patients.

Area of Science:

• Computational biology
• Immunology
• Genomics

Background:

• Machine learning (ML) can identify patient subgroups from gene expression data.
• Current ML approaches often lack systems biology context for phenotype prediction.

Purpose of the Study:

• Develop an interpretable ML approach using blood transcriptomics to predict phenotypes in systemic lupus erythematosus (SLE).
• Integrate systems biology context by evaluating predefined gene sets for predictive power.

Main Methods:

• Utilized a sequential grouped feature importance algorithm.
• Assessed performance of gene sets (immune, metabolic pathways, cell types) in predicting SLE disease activity and organ involvement.
• Focused on blood transcriptomics for phenotype prediction.

Main Results:

• Identified gene sets related to interferon, tumor necrosis factor, mitoribosome, and T cell activation as top predictors of SLE phenotype.
• Achieved excellent performance in predicting disease activity and organ involvement.
• Demonstrated the utility of interpretable ML in understanding disease heterogeneity.

Conclusions:

• The developed ML approach provides insights into molecular pathways associated with SLE manifestations.
• This interpretable ML strategy can be applied to phenotype prediction in other complex diseases and tissues.
• Highlights the importance of systems biology context in ML-driven biomedical research.