Appraisal of Gene Expression-Based Classifiers for Neuropsychiatric Disorders: A Meta-Regression.

Area of Science:

• Neuroscience
• Genomics
• Bioinformatics

Background:

• Gene-expression biomarkers show promise for neuropsychiatric disorders.
• Consensus is lacking on how study design impacts biomarker classifier performance.

Purpose of the Study:

• To identify study characteristics and methodologies influencing the accuracy of transcriptomics-based classification in neuropsychiatric disorders.
• To provide insights for improving the design and evaluation of gene-expression biomarker studies.

Main Methods:

• Conducted a literature review and meta-regression of studies using transcriptomics for neuropsychiatric disorder classification.
• Extracted study characteristics including sample size, validation approach, and classification model.
• Performed univariate and multivariate mixed-effect meta-regression analyses to assess associations with classification accuracy.

Main Results:

• Machine Learning (ML) models were most common (55%), followed by Logistic Regression (25%) and Deep Learning (DL) (20%). Support Vector Machines (SVM) were the most frequent model (17%).
• Withheld test samples were the primary validation approach (56%).
• Significant associations were found between accuracy and study bias risk, model type, class ratio, and validation approach.

Conclusions:

• Study design and methodological choices significantly influence the accuracy of gene-expression-based classifiers for neuropsychiatric disorders.
• Prudent methodologies are crucial for training and evaluating models to avoid biased accuracy estimates.
• Findings can guide future study designs to develop reliable, minimally invasive gene-expression biomarkers for improved neuropsychiatric diagnostics and patient outcomes.