Tile-Based Random Forest Analysis for Analyte Discovery in Balanced and Unbalanced GC × GC-TOFMS Data Sets

Area of Science:

• Metabolomics
• Machine Learning in Bioinformatics
• Analytical Chemistry

Background:

• Nontargeted metabolomics studies often face challenges with complex biological matrices and unbalanced datasets.
• Traditional Fisher ratio analysis (FRA) is effective but can be limited with varying sample sizes per class.
• Supervised analysis methods are crucial for identifying class-discriminating features in complex datasets.

Purpose of the Study:

• To introduce and evaluate a novel nontargeted, tile-based supervised analysis method combining Random Forest (RF) with a four-grid tiling scheme.
• To assess the performance of the tile-based RF analysis on unbalanced datasets compared to standard FRA.
• To identify class-distinguishing analytes in stool samples from omnivore subjects under different storage conditions.

Main Methods:

• Developed a tile-based RF analysis method integrating a four-grid tiling scheme with RF for tile hit importance estimation.
• Applied the tile-based RF approach to a two-class comparison of liquid (Liq) and lyophilized (Lyo) stool samples.
• Generated hit lists using both balanced (8 vs 8) and unbalanced (8 vs 9) datasets and compared them to FRA results.

Main Results:

• The tile-based RF analysis successfully identified similar class-distinguishing analytes (p < 0.01) as FRA.
• RF analysis yielded a more focused hit list with stringent concentration ratios ([OLiq]/[OLyo] > 2 or < 0.5), unlike the broader FRA list.
• The RF approach demonstrated robustness with both balanced and unbalanced datasets, highlighting its stringent feature selection.

Conclusions:

• Tile-based RF analysis is a promising machine learning method for identifying class-distinguishing analytes in complex biological matrices.
• This approach is advantageous for datasets with high between-class and within-class variance, particularly when dealing with unbalanced sample sizes.
• The RF method provides a more selective feature identification compared to standard FRA, focusing on analytes with significant concentration differences.