MetImputBERT: a pretrained BERT framework for missing value imputation in NMR metabolomics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Metabolomics

Background:

• Missing values in nuclear magnetic resonance (NMR) metabolomics data present a significant challenge for accurate clinical interpretation and downstream analysis.
• Existing imputation methods often struggle to adequately handle the complexity and scale of missing data in large metabolomics datasets.

Purpose of the Study:

• To introduce MetImputBERT, a novel imputation method leveraging a pretrained BERT framework for handling missing values in NMR metabolomics data.
• To evaluate the performance of MetImputBERT against established imputation techniques on independent datasets.

Main Methods:

• Developed MetImputBERT, a method based on a pretrained BERT framework that simulates missing values using masked language model techniques.
• Utilized a masked language model approach to predict and reconstruct missing data points, minimizing reconstruction error during training.
• Pretrained MetImputBERT on an extensive UK Biobank metabolomics dataset (over 230,000 individuals) for robust parameter learning.

Main Results:

• MetImputBERT demonstrated superior imputation performance compared to K-nearest neighbors, multiple imputation by chained equations, and singular value decomposition on two independent test sets.
• The method successfully imputed missing values by inferring reconstructed estimates after loading pretrained parameters.

Conclusions:

• MetImputBERT offers a powerful and accurate solution for imputing missing values in NMR metabolomics data, significantly improving data quality.
• The open-source Python tool facilitates easy adoption for researchers, enabling efficient imputation without requiring additional model training.