NS-kNN: a modified k-nearest neighbors approach for imputing metabolomics data

Area of Science:

• Metabolomics
• Bioinformatics
• Data Science

Background:

• Missing values are a common challenge in metabolomics data analysis, potentially biasing results.
• The standard k-nearest neighbors (kNN) imputation method assumes random distribution of missing data, which is often not true in metabolomics.
• Values below the limit of detection contribute to non-random missingness (missing not at random - MNAR).

Purpose of the Study:

• To investigate the impact of MNAR values on imputation performance in metabolomics.
• To introduce a novel algorithm, No-Skip kNN (NS-kNN), designed to handle MNAR values.
• To compare the accuracy of NS-kNN against existing kNN-based imputation methods.

Main Methods:

• Development of a new model for generating synthetic MNAR data.
• Implementation and testing of the NS-kNN algorithm.
• Comparison of imputation errors between kNN, NS-kNN, and KNN-TN on diverse experimental datasets.

Main Results:

• NS-kNN generally outperforms standard kNN when 20-30% of missing values are MNAR.
• NS-kNN demonstrates lower imputation errors than KNN-TN on realistic datasets with at least 50% MNAR values.
• The performance improvement is significant with increasing levels of MNAR data.

Conclusions:

• Addressing the nonuniform distribution of missing values is crucial for accurate metabolomics data imputation.
• NS-kNN provides a more accurate imputation strategy for metabolomics data compared to existing kNN-based methods, especially when MNAR values are prevalent.