Metabolomic data contain missing values for a number of reasons, including instrumental limits of detection and improper data alignment between samples. While many supervised and unsupervised data analysis techniques have been used to analyze metabolomic data, the topic of how to process data and handle missing data has largely been ignored. These are well-studied issues in the field of microarrays that have an important parallel in metabolomics datasets.

We surveyed a number of techniques from the microarray literature and determined their applicability to metabolomics data. We used a publicly available GC-MS dataset of Medicago truncatula (ten replicates for each sample type of roots, stems, and leaves). We first ran the data through SpectConnect, a metabolomic data analysis tool, requiring 100% metabolite presence in all replicates in order to obtain a complete dataset with no missing values. We then processed the data in three different ways: no data pre-processing, log₂ data transformation, and log₂ transformation with data whitening (zero mean and unit standard deviation).

For each processed data version, we then deleted values at random 1) with a uniform prior distribution and varied total missingness and 2) with a nonuniform prior distribution such that 95% of the missing values are expected to come from some fraction of the smallest GC peaks, also with varied total missingness. Numerous such deletion datasets were created for each processed dataset so as to obtain a satisfactory sample size for performance analysis. We imputed new values for the missing values of each deletion dataset using six prominent techniques from the literature. Error with respect to the complete datasets was calculated based on a normalized root mean square error, and error values for each processing method/imputation method combination were averaged for all iterations at each missingness value.

For the uniform prior distribution, imputation with zeros yielded the highest error for every processing method. The impact of data transformation varied as a function of total data missingness, though general trends were observed that will be described. Depending on data transformation and total missingness, different algorithms may be optimal in imputing missing data. As expected, errors under the nonuniform prior distribution were much higher than those under the uniform prior distribution as the missing data was targeted towards those peaks with the lowest abundance. Trends specific to these circumstances will also be described.

This work has shown that parallels can be drawn between the natures of microarray data and metabolomics data and those techniques used to process and impute missing microarray data can – and perhaps should – be applied to metabolomics data.