(B-61) Machine Learning for the Design of Functional Formulations

Introduction:: Inactive ingredients are added to medications to improve the physical properties of the dosage form and are largely considered biologically inert. Recently, an increasing number of studies have reported previously unknown biological effects of inactive ingredients, for example leading to irritations in highly sensitized patients. [1] Understanding such effects not only stands to improve medication safety by avoiding adverse effects, but also provides an opportunity to design functional formulations where the inactive ingredients improve drug transport and metabolism to create safer and more effective therapeutics. However, experimental testing of such effects can be laborious and expensive. Instead, we propose to design and deploy machine learning algorithms used for drug discovery to uncover previously unknown biological effects of inactive ingredients. We show that such machine learning models can be highly predictive and streamline experimental testing [2]. To validate our predictions, we predict and experimentally validate that novel liquid co-formulations with vitamin A palmitate can inhibit drug transport and improve oral bioavailability of medications in vivo [2]. Similarly, we design novel acetaminophen pills that include rutin or eugenol, inactive ingredients that we predicted and validated to prevent liver toxicity of acetaminophen overdose in mice [3].


Materials and Methods:: Inactive ingredients were extracted from the FDA website and converted using NIH CACTUS. Training data of bioactive molecules was curated from ChEMBL22. Activities were logarithmized and values < 3 or >12 were excluded. Molecules were encoded using Morgan fingerprints (radius 4, 2048 bits) and physicochemical properties (RDkit). We created an external test set of inactive ingredient activities for model selection and tested Random Forest, Support-Vector Machines, and kNN models. For large-scale prioritization of predictions, we normalized predicted activities of inactive ingredients based on the average expected prediction of a random set of compounds extracted from ChemDB that we subsampled to approximate the molecular weight distribution of the inactive ingredient compound libraries through Probability Proportional to Size (PPS) Sampling [2]. For P-gp in vitro experiments, we used a proprietary, fluorogenic P-gp substrate (Biovision K507) in HepG2 cells. For P-gp in vivo experiments, a suspension of 500 mg/kg vitamin A palmitate in 10% DMSO PBS or 10% DMSO PBS buffer control were administered orally to five female Balb/c mice 15 minutes prior to oral treatment with warfarin 20 mg/kg [2]. Blood was sampled 30 minutes after administration and warfarin serum concentrations were determined using UPLC-MS/MS. For in vivo acetaminophen experiments, 6-8 weeks old, female Balb/c mice (Charles River) were administered rutin or eugenol prepared in PEG400 or vehicle control via oral gavage and acetaminophen was administered IP at 350 mg/kg. ALT and AST plasma levels of treated mice were determined after 24h and livers were dissected for histology [3].


Results, Conclusions, and Discussions:: We used more than 1.3 million bioactivity datapoints for 1776 protein targets to create Random Forest, Support-Vector Machine, and kNN models that predict potential biological activity of inactive ingredients from their molecular structure and physicochemical properties. We also curated a set of 256 known and previously reported bioactivity values of inactive ingredients against these 1,776 modeled protein targets as a test set for model selection. Our final random forest model had a mean absolute error (MAE) of 0.95 on this test set, outperforming other machine learning approaches such as support-vector machines and k-nearest neighbor models (MAE > 1.0; p < 0.001 two-tailed paired t test). Using the random forest model, we predicted biological activities for all inactive ingredients and found that multiple enzymes (343 predictions), kinases (343 predictions), and family A GPCRs (280 predictions) were predicted to be modulated by inactive ingredients – indicating a large potential for these ingredients to have relevant bioactivities (Figure 1A).

We investigated whether our model could identify inactive ingredients as inhibitors of P-glycoprotein to modulate drug uptake and distribution. One of the top predictions was for vitamin A palmitate, a common supplement and food ingredient. We confirmed this prediction in a cell-based in vitro assay and in an ex vivo Franz diffusion cell assay. Finally, we validated that vitamin A palmitate increases the oral absorption of warfarin in mice by ca. 31% (Figure 1B) [2].

As an additional case study, we set out to investigate whether our models could help us to reformulate acetaminophen to prevent hepatoxicity. Our model suggested that the inactive ingredients rutin and eugenol could be effective cytochrome inhibitors and thereby prevent the creation of liver toxic acetaminophen metabolites. We validated that rutin and eugenol both are hepatoprotective to acetaminophen overdose in mice (Figure 1C) and prototyped pills that release protective ingredients before acetaminophen to create safer medications [3].

Taken together, our machine learning workflows rapidly and effectively predict bioactivities for inactive ingredients. These predictions have important implications in avoiding side effects but can also be used to design functional formulations to make therapeutics safer and more effective.


Acknowledgements (Optional): :

References (Optional): : [1] Reker et al. Science Translational Medicine 11:483 (2019) 10.1126/scitranslmed.aau6753
[2] Reker et al. Cell Reports 30:11 (2020) 10.1016/j.celrep.2020.02.094
[3] Shi et al. BioRxiv (2022) 10.1101/2022.09.05.506668