Machine Learning Uncovers Food- and Excipient-Drug Interactions

Abstract

Inactive ingredients and generally recognized as safe compounds are regarded by the US Food and Drug Administration (FDA) as benign for human consumption within specified dose ranges, but a growing body of research has revealed that many inactive ingredients might have unknown biological effects at these concentrations and might alter treatment outcomes. To speed up such discoveries, we apply state-of-the-art machine learning to delineate currently unknown biological effects of inactive ingredients-focusing on P-glycoprotein (P-gp) and uridine diphosphate-glucuronosyltransferase-2B7 (UGT2B7), two proteins that impact the pharmacokinetics of approximately 20% of FDA-approved drugs. Our platform identifies vitamin A palmitate and abietic acid as inhibitors of P-gp and UGT2B7, respectively; in silico, in vitro, ex vivo, and in vivo validations support these interactions. Our predictive framework can elucidate biological effects of commonly consumed chemical matter with implications on food- and excipient-drug interactions and functional drug formulation development.

Keywords: data science; drug delivery; excipient-drug interactions; food-drug interactions; inactive ingredients; machine learning; pharmacokinetics; pharmacology; virtual screening.

Figure 1.. Inactive Ingredients and GRAS Compounds Resemble FDA-Approved Drugs and Exert Known or Potentially Novel Bioactivities

(A) Schematic visualizing the general workflow of the study and the utilized datasets. Briefly, CAS numbers for generally recognized as safe (GRAS) and inactive ingredient (IIG) compounds were extracted and curated from the FDA website (https://www.fda.gov) and translated into SMILES structural representations using the CACTUS NIH webserver (https://cactus.nci.nih.gov). These chemical representations were then harnessed to calculate physicochemical properties (http://rdkit.org) and compare the property distributions with approved drugs (https://www.drugbank.ca). Biological activity data were extracted from ChEMBL22 (http://ebi.ac.uk/chembl) to identify previously reported activities for GRAS/IIG compounds and build machine learning models (https://scikit-learn.org) to predict additional biological activities of GRAS/IIG compounds. (B) Distribution of molecular weight (MW), calculated logP, and the fraction of rotational bonds (rot bonds) among GRAS (light blue) and IIG (dark blue) compared to FDA-approved drugs in the DrugBank database (DRUGS, orange). Summary statistics represented through boxplots show considerable overlap in the three distinct distributions. (C) Visualization of chemical space spanned by GRAS (light blue) and IIG (dark blue) compared to approved drugs stored within the DrugBank 5.0 database (orange). Projection based on t-Distributed Stochastic Neighbor Embedding (t-SNE) using Morgan fingerprints (r = 4, 2,048 bits; RDKit) is shown. (D) Pharmacology network of GRAS and IIG. Compounds are shown as light blue (GRAS) or dark blue (IIG) nodes; protein targets (ChEMBL22) are shown in red.A compound and a target are connected either when the compound has been previously measured to interact with the protein (black edge) or when machine learning models predicted that the compound is likely to interact with the protein (Z score > 4; gray edge). (E and F) Distribution of number of previously reported (left, E) and computationally predicted (right, F) activities on the level of different protein families (inner pie charts). Top seven families are labeled. Outer pie charts visualize the number of reported or predicted activities per protein. Proteins for which more than 10 GRAS or IIG compounds have been reported or predicted to modulate their activity have been annotated.

Figure 2.. Gum Rosin and Abietic Acid Inhibit UGT2B7 Activity

(A) Chemical structures of abietic acid (1) and training data compound isolongifolic acid (2). (B) In vitro validation shows that gum rosin (black circles) and abietic acid (orange squares) inhibit UGT2B7 activity in microsomes. (C) The effect of abietic acid (orange) on UGT activity was confirmed in complex tissue liver lysates, where it slowed the conversion of a proprietary UGT substrate (Biovision K692; gray). (D) Computational docking indicates that abietic acid has the potential to interact with UGT2B7 at the interface of the substrate- (gold) and the co-factor-binding (cyan) domains.

Figure 3.. Vitamin A Palmitate Modulates P-gp Activity

(A) P-gp is involved in the transport of 8% of all approved drugs, spanning a broad range of different indications (DrugBank 5.0). Complete bar corresponds to 170 approved drugs. Only sub-classifications with at least three drugs are visualized on the right. (B) Structure of vitamin A palmitate (3). (C) Vitamin A palmitate inhibits P-gp activity in HepG cells with an IC₅₀ of 2.9 ± 3.6 µM. (Biovision K507) Data plotted as mean and standard deviation, curve fitted in Prism using the standard three parameter equation for “inhibitor vs. response”. (D) Schematic of ex vivo tissue permeability experiment in (E). (E) Vitamin A palmitate increases the permeability of the four known P-gp substrates irinotecan, ranitidine, colchicine, and loperamide across porcine intestinal tissue. p ≤ 0.001; two-tailed t test with Holm-Sidak correction. (F) Schematic of in vivo experiment in (G). (G) Vitamin A palmitate shows a mild increase of systemic warfarin, a known P-gp substrate, after oral delivery in mice. p = 0.04; one-tailed t test. (H) Computational docking suggests that vitamin A palmitate can bind the ATPase site of P-gp (blue mesh) with a stabilizing hydrogen bond formed with ARG1047 (dashed yellow line; see black arrow).