NormAE: Deep Adversarial Learning Model to Remove Batch Effects in Liquid Chromatography Mass Spectrometry-Based...

Area of Science:

• Metabolomics
• Computational Biology
• Data Science

Background:

• Untargeted metabolomics using liquid chromatography-mass spectrometry (LC-MS) is prone to nonlinear batch effects.
• These effects obscure biological signals, reduce reproducibility, and complicate data calibration.
• Existing methods struggle to effectively address these complex batch variations.

Purpose of the Study:

• To develop a novel deep learning model for robust batch effect correction in untargeted metabolomics.
• To enhance the accuracy and reproducibility of LC-MS data analysis.
• To improve the identification of biologically relevant metabolites and biomarkers.

Main Methods:

• Proposed a deep learning model named Normalization Autoencoder (NormAE).
• NormAE utilizes nonlinear autoencoders (AEs) combined with adversarial learning for batch effect removal.
• An integrated classifier and ranker provide adversarial regularization during AE training, enabling effective latent representation extraction and data reconstruction.

Main Results:

• NormAE significantly reduced batch effects in two real-world metabolomics datasets, evidenced by tighter clustering of quality control samples in PCA plots (average distances decreased substantially).
• The model achieved high average correlation coefficients (up to 0.997) for the datasets after calibration.
• NormAE markedly enhanced biomarker discovery, increasing the median number of differential peaks observed.

Conclusions:

• Normalization Autoencoder (NormAE) is a highly effective deep learning approach for correcting nonlinear batch effects in untargeted metabolomics.
• NormAE outperforms commonly used batch effect removal methods in terms of calibration accuracy and data quality.
• The method shows significant promise for improving the reliability and discovery power of metabolomics studies.