Optimal Bayesian Filtering for Biomarker Discovery: Performance and Robustness.

Area of Science:

• Computational biology
• Biostatistics
• Bioinformatics

Background:

• Optimal Bayesian feature filtering (OBF) is a computationally efficient algorithm for identifying potential biomarkers.
• OBF assumes Gaussian distributions and independence, which may not hold true in real biological data.
• Understanding OBF's performance under violated assumptions is crucial for reliable biomarker discovery.

Purpose of the Study:

• To evaluate the performance and robustness of Optimal Bayesian feature filtering (OBF) in biomarker discovery.
• To assess OBF's behavior when its underlying modeling assumptions (Gaussianity, independence, sample size balance) are violated.
• To provide practical guidelines for setting OBF input parameters to ensure robust results.

Main Methods:

• A comprehensive simulation study was conducted to test OBF under various conditions.
• Simulations challenged OBF's assumptions regarding input parameters, independence, sample size balance, and Gaussianity.
• Performance was evaluated across a range of non-Gaussian distributions and different prior settings.

Main Results:

• OBF demonstrates robustness even when modeling assumptions are not strictly met.
• The study identified specific advantages and disadvantages associated with different priors and optimization criteria.
• OBF successfully identified known biomarkers in acute myeloid leukemia (AML) and colon cancer datasets, outperforming the moderated t-test.

Conclusions:

• Optimal Bayesian feature filtering (OBF) is a reliable tool for biomarker discovery, even with imperfect data.
• The findings offer practical guidance for optimizing OBF parameter selection for enhanced performance.
• OBF shows promise as an alternative to traditional methods like the moderated t-test for identifying significant biological markers.