Drug development is a complicated, costly and time-consuming process, where data science and machine learning can bring decisive advantages.
For pharmaceutical business, it takes more than ten years and often more than a billion euros to proceed from the first step, target identification, via drug discovery, development and testing, to the final step: drug approval.
During this process about 89 % of the candidate molecules fail to become a drug (Kubiniy 2003). Among the main reasons for these failures are the lack of efficacy, poor pharmacokinetics, and emerging adverse effects (Fig. 1).
In this series of two blog posts, we will explore how certain failures in drug development process can be tackled thanks to Machine Learning models predicting certain properties of the candidate compounds. These properties are related to the structure of the compounds, which is descriptive in the set of features, named "molecular fingerprint".
Models using molecular fingerprints to predict compound properties are known as QSAR/QSPR (quantitative structure activity/property relationships) models. Usage of QSAR/QSPR models significantly reduces both time and cost of drug development, as they lower the number of compounds needed to pass in vitro and in vivo tests.
To illustrate our point, after a general introduction, we will take the particular example of drug interactions occurring within the brain.
Fig. 1 The main reasons for failure in drug development (Kubinyi 2003)
Drug efficacy: know your transporters!
Let us go deeper into understanding some of the underlying mechanisms behind drug failures.
First, drug efficacy largely depends on drug ability to be properly distributed in the human body, which is described by drug pharmacokinetics. Drugs can pass across body barriers by passive diffusion or by active transport, performed by membrane proteins. These proteins can transport drugs from the bloodstream to the target organ (influx transporters) or in the opposite direction (efflux transporters).
Certain barriers in the body, like those in kidneys, liver and brain, have a lot of protein transporters (Fig. 2). Passive diffusion across these barriers is limited, so drug movement largely depends on active transport.
Therefore, knowing the transporters of a certain drug helps predict efficacy of its distribution in the body, particularly in the above-mentioned organs. For example, if a drug is a substrate of efflux transporters, it can be pumped out of the target organ, which makes such drug inefficient (Giacomini 2010). Initial screening for transporter substrates would eliminate ineffective drugs on the first step of drug development.
If these properties are only detected at late stages, time and resources will have already been spent on development, to no avail.
Fig. 2 Protein transporters in the intestinal epithelia (a), liver cells (b), kidney epithelia (c), and brain endothelial cells (d) (Giacomini 2010).
The importance of drug-drug interactions
Another common factor impacting drug efficacy, distribution and occurring adverse side reactions is drug-drug interactions (later on: DDIs).
DDIs may take place at different stages of the drug action pathway, starting with drug distribution in the body. As we have already seen, the distribution largely depends on specialized proteins, drug transporters. These proteins are one of the major sources of DDIs.
Transporter-mediated DDIs happen when two types of drugs are administered together: the first one is a substrate of transporter, the second one is a modulator for the same transporter. The second drug affects distribution of the first drug by modifying the transporter activity. This may lead to limited efficacy of the substrate drug, because it will not be able to reach the desired target due to a blocked transporter. Also, accumulation of the substrate in the non-targeted organs may be the source of drug adverse effects.
Two famous examples of DDIs caused by substrate/modulator interaction occur at the blood-brain barrier (BBB) and are mediated by the ubiquitous efflux transporter P-gp.
A first example is the administration of anticancer drug etoposide, which is a substrate of P-gp, together with antibiotic cyclosporine, which is a P-gp inhibitor. This is known to cause severe nausea as a neurological side-effect, because etoposide accumulates in the brain and is not pumped outside by P-gp (Fig. 3).
A second example is the co-administration of P-gp substrate loperamide, which is a common opiate antidiarrheal drug, with quinidine, a P-gp inhibitor. With suppressed efflux P-gp, loperamide has no obstacles getting in the brain, which causes euphoria and respiratory depression. There are a lot of other known examples of DDIs with similar mechanisms which can be found in the literature (König et al. 2013).
Fig. 3 The mechanism of P-gp mediated interaction between etoposide and cyclosporine. Cyclosporine inhibits P-gp activity, leading to accumulation of etoposide in the brain. (https://basicmedicalkey.com/antineoplastic-and-immunomodulating-drugs-2/)
As a summary, knowing the exact relevant transporters of a drug is essential to predict possible failures in the drug development process and make it cheaper. And an efficient prediction model will also help narrow down the list of drugs, against which the drug candidate should be tested for possible DDIs.
Getting into the brain
Now let us look at drug interactions related to one of the most important physiological barrier in the human body: the blood-brain barrier (later on : BBB).
BBB is formed by the tight junctions between endothelial cells of the blood vessels in the brain. These cells have a very low rate of transcytosis, which means that only a small number of chemicals can pass through them by diffusion.
Membranes of the endothelial cells also have a large number of transporter proteins, performing selective barrier function (Fig. 2d). These transporters play a crucial role in the absorption, distribution and elimination of different chemicals in the brain. Efflux transporters are pumping out the molecules that occasionally diffuse into the cell, while influx transporters select vital molecules from the blood and pump them inside the cell.
The transport of molecules through the blood-brain barrier is an important feature to be considered during the drug development process, in two ways. If a drug is targeted to the brain it should pass BBB: permeability to this particular drug is a facilitating feature. On the other side, if a drug is not targeted to the brain, it would be better if it does not pass the barrier, as the drug may cause neurological adverse effects (König et al. 2013)(Girardin 2006).
BBB? DDI? ML!
These fundaments in drug domain expertise were necessary to pave the way for our favourite actors, the ones that will take the main roles in our second act: machine learning and prediction models.
Passive diffusion through BBB is comprehensively described, and machine learning models predicting permeability for chemicals have been developed years ago (Saber et al. 2020)(Miao et al. 2019). The models that predict passive diffusion through BBB are widely used to screen for permeability among the drug candidates.
However, these models do not account for active transport through the barrier. Only a few models for active transport exist, and most of them are dedicated to the efflux transporter P-gp, its substrates and inhibitors (Kadioglu and Efferth 2019)(Esposito et al. 2020). At the same time, no models are available for the active transport directed inside the brain, performed by influx transporters. Two influx transporters are present in the BBB: OATP1A2 and OATP2B1. Both transporters have the ability to transport drugs into the brain, and their substrate specificity is wide.
A model predicting substrates of one or both of these channels can improve screening for BBB permeability, as more substances able to pass BBB would be identified.
Well, as you might guess, this will precisely be the matter of our second blog post on the topic. Stay tuned!
* Follow us on LinkedIn for next blog updates:
* Interested in our skills? Let's discuss your projects together:
* Our public Github repository:
Esposito C, Wang S, Lange UEW, et al (2020) Combining Machine Learning and Molecular Dynamics to Predict P-Glycoprotein Substrates. J Chem Inf Model. doi: 10.1021/acs.jcim.0c00525
Giacomini KM, Huang SM, Tweedie DJ, et al (2010) Membrane transporters in drug development. Nat Rev Drug Discov 9:215–236. doi: 10.1038/nrd3028
Girardin F (2006) Membrane transporter proteins: A challenge for CNS drug development. Dialogues Clin Neurosci 8:311–321
Kadioglu O, Efferth T (2019) A Machine Learning-Based Prediction Platform for P-Glycoprotein Modulators and Its Validation by Molecular Docking. Cells 8:. doi: 10.3390/cells8101286
König J, Müller F, Fromm MF (2013) Transporters and drug-drug interactions: Important determinants of drug disposition and effects. Pharmacol Rev 65:944–966. doi: 10.1124/pr.113.007518
Kubinyi H (2003) Drug research: Myths, hype and reality. Nat Rev Drug Discov 2:665–668. doi: 10.1038/nrd1156
Miao R, Xia LY, Chen HH, et al (2019) Improved Classification of Blood-Brain-Barrier Drugs Using Deep Learning. Sci Rep 9:1–11. doi: 10.1038/s41598-019-44773-4
Saber R, Mhanna R, Rihana S (2020) A machine learning model for the prediction of drug permeability across the Blood-Brain Barrier: a comparative approach. 1–15. doi: 10.21203/rs.3.rs-29117/v1
Sanchez-Covarrubias L, Slosky L, Thompson B, et al (2014) Transporters at CNS Barrier Sites: Obstacles or Opportunities for Drug Delivery? Curr Pharm Des 20:1422–1449. doi: 10.2174/13816128113199990463