Factors affecting the accuracy of a class prediction model in gene expression data

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Class prediction models show variable performance in clinical gene expression datasets.
• Previous cancer-focused studies highlight dataset and classification function influence on model accuracy.
• Limited research exists on how gene expression data characteristics impact classifier performance.

Purpose of the Study:

• To empirically identify data characteristics affecting predictive accuracy of classification models.
• To investigate these factors specifically in non-cancerous gene expression datasets.

Main Methods:

• Downloaded datasets from 25 studies meeting inclusion criteria.
• Selected nine classification functions (discriminant analyses, Bayes classifiers, tree-based, regularization/shrinkage, nearest neighbors).
• Built and evaluated nine class prediction models per dataset, recording study and data characteristics (disease, sample size, gene expression features).

Main Results:

• The number of differentially expressed genes and average fold change significantly impacted classification model accuracy.
• These factors individually explained up to 72% and 57% of prediction accuracy variation.
• Multivariable analysis identified these two factors plus within-class correlation as key drivers, explaining 91.5% of between-study variation.

Conclusions:

• Number of differentially expressed genes, fold change, and within-class correlation significantly influence classification model accuracy in non-cancerous datasets.
• These data characteristics are crucial for understanding and improving model performance.
• Findings extend knowledge beyond cancer research regarding factors affecting predictive modeling in genomics.