MegaTrans – human transporter machine learning models

Collaborations Pharmaceuticals, Inc., Fuquay Varina NC

Investigators

Abstract

Summary Being able to predict interactions with important human transporters would be of value to new drug design to avoid compounds that interact with them and cause undesirable side effects. Conversely, some drug transporters can be used for targeting molecules to specific organs and this may have considerable utility. Understanding the interactions of novel drugs, natural products and environmental toxicants and their interactions with an array of such transporters is, therefore, important for several industries, as well as from a regulatory perspective (e.g. FDA, EPA and EMA). Being able to predict such interactions in a fast and reliable manner effectively requires using computational approaches and learning from in vitro data, the latter a resource that is rapidly growing. Over the past 20 years, we have been at the forefront of applying different machine learning approaches to modeling drug transporters and, in many cases, developing datasets for transporters for which there was scant available data. We now propose doing this for several transporters that may be important for drug discovery. In Phase I we focused on OATP1B1 (SLCO1B1), which is an uptake transporter largely restricted to the sinusoidal aspect of hepatocytes where it mediates transport of a variety of structurally unrelated compounds, including members of several clinically important drug families (incl. statins, sartans and angiotensin converting enzyme (ACE) inhibitors). We tested 476 drugs against one substrate in vitro. We then curated these data and built machine learning models using multiple machine learning methods as well as model evaluation metrics. This enabled us to develop models for integration in a web-based software tool called MegaTrans® that enables the user to input their own compound structures and generate predictions for interactions with transporter/s of interest, as well as visualize the similarity to the training set of each model using several different visualization methods. In addition, during Phase I we also performed preliminary data curation, model building and validation for two equilibrative nucleoside transporters (ENTs), ENT1 and ENT2, that are present at the blood testes barrier (BTB), where they can facilitate drug disposition (e.g. for antivirals, thereby potentially eliminating a sanctuary site for viruses detectable in semen). We generated Bayesian and pharmacophore models and used these to predict numerous compounds that were then tested in vitro against ENTs. We used these ENT models to predict (i) the antivirals used in treating COVID-19, remdesivir and molnupiravir, inhibit ENT activity, and that (ii) remdesivir is an ENT substrate, as well as validating these predictions. In Phase II we plan on building on the foundation of Phase I and propose greatly expanding the ENT1 and ENT2 models through in vitro testing (at the University of Arizona) of >2000 approved drugs, natural products, and environmental toxicants as inhibitors of ENT transport. We will use these data to build and validate machine learning models using several algorithms, at Collaborations Pharmaceuticals, Inc. We will also test these models using external validation with additional molecules from vendor libraries and drug collections that are not in the model. In this process we will also build out the capabilities of MegaTransÒ to use 3D pharmacophore descriptors to incorporate molecular shape features and allow 3D searches. The return on investment of such a commercial tool would be that it could assist in the design and selection of more favorable compounds by avoiding transporters of interest (or, conversely, allow the targeting of specific transporters to increase uptake into organs). It could also identify compounds that are already approved that might present a drug-interaction risk. Predicting such behavior seen in vivo is ideal and will lead to the prioritization of compounds to test in vitro for potential drug-drug interactions. In summary, we propose generating large training sets for ENT1 and ENT2 transporters that we will use to generate an array of validated machine learning models of interest to drug discovery (with specific interest for those generating antivirals). MegaTransÒ will be a commercial product available for licensing by pharmaceutical, consumer product, agrochemical and regulatory groups, as well as fee-for-service consulting provided by Collaborations Pharmaceuticals, Inc.