Re-evaluation of publicly available gene-expression databases using machine-learning yields a maximum prognostic...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene expression signatures, patterns of gene activity, are vital for cancer classification, prognosis, and treatment.
• High-throughput technology and machine learning have improved prognostic predictions for various cancer phenotypes.
• The prognostic power of computational methods for analyzing gene expression signatures remains under-evaluated, leading to ongoing debate.

Purpose of the Study:

• To unify and evaluate the prognostic power of different approaches for constructing gene expression signatures.
• To assess the predictive accuracy of gene expression signatures using machine learning models.
• To determine the maximum achievable prognostic power and identify limitations.

Main Methods:

• Re-evaluation of publicly available gene expression data from 8 databases.
• Application of 9 distinct machine learning models to analyze signatures.
• Calculation of the concordance index to measure prognostic discrimination accuracy.

Main Results:

• Gene expression signatures are confirmed as useful tools for predicting patient prognosis.
• A maximum prognostic power was identified across approximately 10,000 signatures.
• Signatures can discriminate patient prognoses with a maximum accuracy of 80%, with over 50% of potential information remaining unutilized.

Conclusions:

• Gene expression signatures offer valuable prognostic insights in cancer.
• The current prognostic power of gene expression signatures is capped, indicating room for improvement.
• Accurate cancer prognosis necessitates integrating molecular data with clinical, histological, and other complementary factors.