Empirically establishing drug exposure records directly from untargeted metabolomics data

Abstract

Despite extensive efforts, extracting information on medication exposure from clinical records remains challenging. To complement this approach, we developed the tandem mass spectrometry (MS/MS) based GNPS Drug Library. This resource integrates MS/MS data for drugs and their metabolites/analogs with controlled vocabularies on exposure sources, pharmacologic classes, therapeutic indications, and mechanisms of action. It enables direct analysis of drug exposure and metabolism from untargeted metabolomics data independent of clinical records. Our library facilitates stratification of individuals in clinical studies based on the empirically detected medications, exemplified by drug-dependent microbiota-derived N-acyl lipid changes in a cohort with human immunodeficiency virus. The GNPS Drug Library holds potential for broader applications in drug discovery and precision medicine.

Figure 1.. The GNPS Drug Library and connected pharmacologic metadata.

a, The GNPS Drug Library comprises four key resources: Drug MS/MS reference spectra, drug metabolite MS/MS reference spectra, propagated drug analogs derived from public metabolomics datasets, and pharmacologic metadata connected to each reference spectrum. b, FastMASST analog search of drug spectra against public metabolomics studies yielded propagated drug-analogous MS/MS spectra, which were filtered by removing analogs for drugs with endogenous and food sources (source filter), removing mass offsets unexplained by common metabolic pathways (mass offset filter), removing analogs with GNPS library matches (library match filter), removing analogs connected to multiple drugs with dissimilar structures after spectra clustering (drug similarity filter), and removing analogs with unrealistic drug exposure indications (dataset testing). c, Illustration of each filter employed in curating FastMASST analog match results. d, Frequency of mass offsets in the propagated drug analog library. The mass offsets were grouped by unit mass and stacked based on the number of analog spectra. The most frequently observed mass offsets are colored while the rests are black. e, An example of structural modification sites predicted by ModiFinder. Purple color highlights modified spectra and substructures, while the green color highlights unmodified ones. f, Overview of the ontology-based drug metadata, highlighting common pharmaceutical classes and specific drugs in the neurology/psychiatry category. Width of the bars and lines reflects the number of unique drug structures in each class. g, The top 20 most detected pharmacologic classes in fecal samples from the American Gut Project, a cohort of the general population from the United States (US), Europe, and Australia (1,993 individuals). h, Detected therapeutic drug class patterns by age and sex (1,845 individuals with age and sex information; age 46 ± 18 years [range 3-93], with 53% being female). Detection of cardiovascular drugs increased with age, while analgesics, antihistamines, and antibiotics were detected across all ages.^, Analgesics were more frequently detected in females, consistent with the literature,^, and drugs for erectile dysfunction were detected only in males. NSAID, non-steroidal anti-inflammatory drugs; ACE, angiotensin converting enzyme; SSRI, selective serotonin reuptake inhibitor; PPI, proton pump inhibitor; DHFR, dihydrofolate reductase; HSV, herpes simplex virus; SNRI, serotonin and norepinephrine reuptake inhibitor.

Figure 2.. Drug exposures in the HIV Neurobehavioral Research Center (HNRC) cohort with connections to microbial metabolism and endogenous metabolites.

From the HNRC cohort, 322 fecal samples were analyzed with 222 samples from people with HIV and 100 samples from people without HIV. a, Peak area visualization of drugs detected with metabolites and analogs. Each column represents one sample and each row represents one drug annotation. Drug annotations were grouped based on the parent drugs and separated by gap spaces. Drug annotations were denoted based on their types (as drug, drug metabolites, or drug analogs) and the pharmacologic classes of the parent drugs. All annotated ion/adduct forms of the parent drugs were visualized, leading to multiple rows of parent annotations for some drugs. Asterisks on the drug name mark parent drug annotations confirmed with commercial standards based on retention time and MS/MS spectral matches. Raw peak areas were log-transformed. b, Retention time and MS/MS spectra mirror matches for drug analogs observed in both the fecal samples and the drug microbial incubations. Purple traces represent the fecal samples, while red traces represent the drug microbial incubation. Blue traces represent mixtures of the fecal samples and the microbial incubations at 1:1 volume ratio. The atomic changes of the drug analogs were based on [M+H]⁺ ion of the parent drug. c, Hierarchical clustering of the samples from people with HIV (n = 222) based on detected antiretroviral drugs (ARV). Each row represents one detected ARV, with peak areas summed for the drug, metabolite, and analog detections followed by log-transformation (visualized with the same color scale as panel a). ARVs detected in <10% of samples are not shown. Each column represents one sample, clustered into four groups by hierarchical clustering with Ward’s linkage and Euclidean distance. d, Sample-to-sample peak areas of N-acyl lipids in people with HIV, separated by the clusters derived from ARV detections shown in panel c. For each compound, the peak area in each sample was standardized to the maximum value observed across all samples. A non-parametric Kruskal-Wallis test followed by pairwise Wilcoxon test and Benjamini-Hochberg correction for multiple comparisons were performed. P-values < 0.05 were noted in the figure. Boxplots showcase the median value, first (lower) and third (upper) quartiles, and whiskers indicate the error range as 1.5 times the interquartile range.