Tree-based SIMCA for dealing with heterogeneous and sparse data.

Area of Science:

• Cheminformatics and Bioinformatics
• Machine Learning in Omics Data Analysis

Background:

• One Class Classification (OCC) methods like SIMCA face challenges with omics data, including high dimensionality, sparsity, and non-linearities.
• These issues often lead to inflated decision boundaries and false positives in identifying target classes.
• Tree-based techniques offer inherent robustness to these data characteristics.

Purpose of the Study:

• To explore and evaluate tree-based variants of SIMCA for omics data analysis.
• To compare the performance of these novel methods against traditional OCC strategies.
• To enhance the interpretability of omics data models.

Main Methods:

• Developed non-linear SIMCA variants using sample proximities from Unsupervised Random Forest (URF-SIMCA) and Isolation Forest (IF-SIMCA).
• Compared URF-SIMCA and IF-SIMCA against traditional SIMCA, one-class SVM, and Isolation Forest.
• Utilized five clinical omics datasets and the wine dataset for empirical validation.

Main Results:

• URF-SIMCA demonstrated superior performance across the tested omics datasets.
• Pseudo-sampling principles enabled feature interpretation for class separation.
• Feature trajectories in score and orthogonal distance spaces enhanced model interpretability.

Conclusions:

• URF-SIMCA provides an accessible extension to SIMCA, effectively reducing target class variance for improved separation.
• While feature interpretation is slightly reduced, the pseudo-sampling principle offers a viable solution.
• The method enhances modeling performance in omics data analysis.