KADAIF: an anomaly detection method for complex microbiome data.

Area of Science:

• Microbiology
• Bioinformatics
• Computational Biology

Background:

• The gut microbiome significantly impacts human health and disease, generating large datasets requiring robust preprocessing.
• Anomaly detection is crucial for identifying erroneous samples in microbiome data to prevent misleading statistical results.
• Microbiome data's unique characteristics (compositionality, sparsity, high dimensionality) challenge conventional anomaly detection methods.

Purpose of the Study:

• To develop a microbiome-specific anomaly detection method tailored to the unique properties of microbiome data.
• To address the limitations of existing anomaly detection techniques in handling high-dimensional, sparse, and compositional biological data.

Main Methods:

• Introduction of KADAIF (K-Anonymity Detection Anomaly Identification Framework), a generalization of the Isolation Forest (IF) approach.
• KADAIF builds an ensemble of trees, partitioning data using feature subsets and dimensionality reduction to capture species interactions and sparsity.
• The method isolates anomalous samples closer to the root of the trees based on average depth.

Main Results:

• KADAIF demonstrates superior performance over alternative methods in simulated anomaly detection scenarios across diverse datasets.
• The method effectively detects anomalies in other high-dimensional, sparse biological data types, outperforming standard Isolation Forest.
• KADAIF successfully identified disease onset in longitudinal microbiome data and partitioned cases versus controls.

Conclusions:

• KADAIF offers a powerful new tool for microbiome data preprocessing and downstream analysis.
• The method has significant potential to improve the accuracy and reliability of findings in precision medicine studies.
• Availability of KADAIF implementation on GitHub facilitates its adoption and further research.