Predicting human health from biofluid-based metabolomics using machine learning

Biofluid-based metabolomics has the potential to provide highly accurate, minimally invasive diagnostics. Metabolomics studies using mass spectrometry typically reduce the high-dimensional

data to only a small number of statistically significant features, that are often chemically identified—where each feature corresponds to a mass-to-charge ratio, retention time, and

intensity. This practice may remove a substantial amount of predictive signal. To test the utility of the complete feature set, we train machine learning models for health state-prediction

in 35 human metabolomics studies, representing 148 individual data sets. Models trained with all features outperform those using only significant features and frequently provide high

predictive performance across nine health state categories, despite disparate experimental and disease contexts. Using only non-significant features it is still often possible to train

models and achieve high predictive performance, suggesting useful predictive signal. This work highlights the potential for health state diagnostics using all metabolomics features with

data-driven analysis.

While fundamental to personalized healthcare, it is often challenging to diagnose an individual’s health state (a general term encompassing disease and non-disease phenotypes like age) due

to low test sensitivity, specificity or the requirement of invasive procedures. Body-fluid sampling (e.g. blood or urine) offers a minimally invasive approach to identify health conditions

throughout the body. The traditional concept of biofluid-based diagnostics relies on health-state biomarkers. Biomarkers cover a broad spectrum of measurements1, but typically refer to a

small number of select and specific molecules or biopolymers, capable of differentiating healthy from diseased states. Currently, many biomarker-containing tests are used in routine lab

monitoring (e.g. complete blood count, ‘basic’ and ‘comprehensive’ metabolic panels, lipid panels, etc.) providing coarse health-state categorization. Tests for many diseases exist and

display a range of sensitivity and specificity, examples include: apolipoprotein E along with other measurements for Alzheimer’s disease2, the prostate-specific antigen test for prostate

cancer3, alpha fetoprotein (AFP) for liver cancer4, as well as a recent use of the SOMAscan5 for diagnosing tuberculosis6.

Metabolomics rapidly supplies information on thousands of molecules, and provides a method for biofluid-based diagnostics7,8,9. To date, serum, plasma, urine and cerebrospinal fluid (CSF)

metabolomics has been applied to many health states, ranging from cancers10,11,12,13 and infectious diseases14,15 to chronic obstructive pulmonary disease (COPD)16, smoking17 and Alzheimer’s

disease18,19. Metabolomics studies are regularly performed using analytical instrumentation like liquid or gas chromatography mass spectrometry (LC–MS and GC–MS respectively) as well as

nuclear magnetic resonance (NMR)20. Frequently, the goal is to determine the chemical identity of the features that are significantly altered between health states. For LC- and GC–MS

studies, a feature is defined by a mass-to-charge ratio (mz), retention time (rt) and intensity. While a chemical name cannot be assigned to the majority of features, analysis of those that

are identified allows for biological interpretation by differential analysis and biochemical pathway mapping18,21. Select chemically identified features are often used for differential

diagnostics or health state association. For instance, certain amino acids have been associated with diabetes22 as well as urinary formate, alanine, and hippurate with blood pressure23.

For diagnostic modeling purposes, full metabolomics data sets are generally not used for training, validation, and testing. To deal with the large number of features, a host of feature

selection methods and classification techniques are employed. Univariate statistical tests (Student’s t-test or Mann–Whitney U-test, MW-U) are routinely used to isolate statistically

significant features—usually identified using false discovery rate (FDR) adjusted P-values