Anomaly detection in mixed high-dimensional molecular data | JoVE Visualize

Area of Science:

Bioinformatics
Computational Biology
Data Science

Background:

Mixed molecular data, combining continuous (e.g., omics) and categorical (e.g., genotype, diagnosis) features, is crucial for biological insights.
High-dimensional molecular data is susceptible to errors, often manifesting as single incorrect measurements (anomalies) within a sample, not entire outlier samples.
Existing anomaly detection methods typically identify entire samples as outliers, failing to address subtle, single-feature errors common in mixed data.

Purpose of the Study:

To introduce ADMIRE (Anomaly Detection using MIxed gRaphical modEls), a novel method for detecting and correcting single-dimensional anomalies in mixed high-dimensional data.
To specifically address errors in both continuous and categorical features within molecular datasets.
To improve the accuracy and reliability of analyses using mixed molecular data.

Main Methods:

ADMIRE utilizes mixed graphical models to learn the joint distribution of continuous and categorical features.
Anomalies are identified by quantifying the discrepancy between measured feature values and model-based estimations.
Detected anomalies are corrected using imputation techniques.

Main Results:

ADMIRE successfully detects and corrects single-dimensional anomalies in both continuous and categorical features of mixed molecular data.
In simulation experiments, ADMIRE demonstrated superior performance compared to state-of-the-art anomaly detection methods like Local Outlier Factor, stray, and Isolation Forest.
The method was validated on a real-world metabolic dataset, showcasing its practical applicability.

Conclusions:

ADMIRE offers a robust solution for identifying and rectifying subtle data errors in complex, mixed molecular datasets.
The developed Python package 'adadmire' provides accessible implementation for researchers.
This advancement is expected to enhance the quality and interpretability of findings from high-dimensional omics and clinical data analyses.