Machine learning approaches to predict lupus disease activity from gene expression data.

Area of Science:

• Immunology
• Computational Biology
• Genomics

Background:

• Systemic lupus erythematosus (SLE) exhibits significant patient and cohort heterogeneity, complicating gene expression analysis for disease activity prediction.
• Microarray platform differences further challenge the integration of diverse SLE gene expression datasets.
• Accurate prediction of SLE disease activity is crucial for effective patient management.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting SLE disease activity using integrated gene expression data.
• To compare the performance of using raw gene expression versus gene modules for SLE classification.
• To assess the robustness of prediction models against technical variations across datasets.

Main Methods:

• Machine learning algorithms, including random forest, were applied to integrate gene expression data from three SLE cohorts.
• Weighted Gene Co-expression Network Analysis (WGCNA) was used to generate informative gene modules from purified leukocyte populations.
• Classifiers were evaluated using 10-fold cross-validation and independent training/testing splits to assess performance and robustness.

Main Results:

• A random forest classifier achieved 83% accuracy with 10-fold cross-validation but was sensitive to technical variations.
• Utilizing gene modules demonstrated greater robustness, yielding approximately 70% classification accuracy irrespective of data splitting strategies.
• Technical variations significantly impacted the performance of classifiers trained on raw gene expression data.

Conclusions:

• Gene modules derived from WGCNA provide a more robust approach for predicting SLE disease activity compared to raw gene expression data.
• Machine learning models, particularly those using gene modules, show promise for estimating SLE disease activity, potentially serving as a standalone measure.
• Further algorithm and parameter optimization may enhance the accuracy and clinical utility of these predictive models.