AI in drug discovery: ADMET property prediction
Discover how artificial intelligence can help you select lead compounds that are both to be effective and do not have adverse effects on patients
The search for new lead compounds
For every new drug brought to the market, most estimates suggest that researchers employed over 100 screens looking for drug leads, threshing out candidates from tens of thousands of compounds. Some lead compounds later fail during the subsequent development phase for reasons often unforeseen by researchers. This makes the search for lead compounds immensely time-consuming and costly, taking by some estimates around 5 years and over $200 million, not including the time and cost associated with drug development.
It is estimated that close to 50% of drug candidates fail because of low efficacy, and that around 40% of drug candidates fail due to the presence of toxicity. This leaves a major challenge for researchers and scientists to have to decide on lead compounds that are both to be shown as effective and do not have adverse effects on the human body.
Example (a table) of drugs that have been withdrawn from the market
What are ADMET properties?
In pharmacokinetics and pharmacology, ADMET stands as an abbreviation for absorption, distribution, metabolism, excretion, and toxicity, describing the reaction of a drug medication within an organism, human, or animal. These are the key elements that determine the safety, uptake, elimination, metabolic behavior, and effectiveness of drugs. Specifically, ADMET answers the questions on:
- How much of the drug is absorbed and how quickly? (absorption)
- Where is the drug distributed within the body, and what is the rate and extent of the distribution? (distribution)
- How quickly is the drug metabolized, what is the mechanism of action, and what metabolite is formed? (metabolism)
- How is the drug excreted and how quickly? (excretion)
- Does it have a toxic outcome on the body and/or organs? (toxicity)
ADMET properties have been defined to help scientists evaluate their drug candidates and to help them select compounds that are likely to yield the desired effects in patients.
The challenges in traditional computation of ADMET properties
Building ADMET profiles involves exploring a heterogeneous set of properties. For instance, a researcher or biotech might try to figure out if their lead compound has a risk of damaging the liver (also known as the hepatotoxicity of a compound), whether it might cross the blood-brain barrier, or whether it can inhibit important metabolite enzymes such as cytochrome P450 enzymes, which can cause important adverse drug reactions. However, the number of possible causes for a compound to have adverse effects is large, and as such, the number of properties to check in a wet-lab setting can easily become very costly and time-consuming. Because of this, the in-vitro screening of compounds is often limited to only a few ADMET properties, of the few, most promising drug candidates. Much of the validation of the safety of compounds is done through expensive and time-consuming animal or human clinical trials of the drug candidates.
The promise of predicting ADMET properties using machine learning
The increased availability of ADMET dataset and the growing performance of machine learning algorithms make it possible to predict many of a compound’s properties in silico, rather than in vitro or in vivo. Furthermore, rather than limiting the analysis of ADMET properties to a handful of promising leads, it’s possible to screen thousands, if not millions of compounds quickly and cost-effectively through computer algorithms, making it possible to screen compounds at any stage of the drug discovery process, and not just during lead identification. This process makes it possible to better balance the best drug candidate among many that finds a good balance between its effectiveness and safety, rather than simply evaluating the safety of a single compound that is likely effective, and facing potential failure if the candidate proves to be unsafe.
In case of drug repurposing, scanning a library of compounds for toxicity at the very early steps may avoid wasting time on studying poor drug candidates. Early ADMET investigations may also be valuable in other contexts. When working on drug cocktails, metabolism screenings become an important assessment: if one drug from the cocktail is being metabolized by the CYP1A2 enzyme, for example, then the other drugs should avoid binding to this enzyme. Once again, having insights on metabolism at the beginning of the drug discovery process can enhance it.
The challenges of using machine learning to predict ADMET properties
Nevertheless, using computational methods such as machine learning to predict ADMET properties comes with a whole set of challenges. Some of them include:
- Need of data: machine learning algorithms need data of compounds that have been previously screened in vitro (or in vivo) for the properties that one is looking at. On top of that, the compounds that have been screened need to be diverse enough to allow algorithms to generalize to the complete chemical search space - a task that is often challenging.
- Need of machine learning expertise: even if you do have relevant data, the process of training a machine learning algorithm isn’t trivial. In order to ensure that an algorithm makes accurate predictions, machine learning practitioners need to be careful in how they split their data between training and evaluation sets (for example, by evaluating how well algorithms do on compounds that have chemical scaffolds that the algorithm has never seen), handling of imbalanced datasets (where compounds with favorable outcomes are much rarer than compounds with unfavorable features), selection of algorithms, and much more.
- Need of biological expertise: not all properties are bad for all drugs. For example, drugs targeting diseases that affect the central nervous system most often need to penetrate the Blood Brain Barrier (BBB), while this property is mostly considered as a negative for other diseases. As such, the exercise is not to simply collect as many ADMET endpoints as possible, but to assign them a relevant meaning and corresponding weight in the decision-making process.
During drug discovery projects, Kantify is looking at more than 80 ADMET points of compounds to evaluate their safety and effectiveness as drug candidates. Using state-of-the-art machine learning accompanied by many in-house innovations, Kantify has developed an in-silico approach to help:
- Select drug candidates that are both to be effective and not have adverse effects on the patients (ex: heart toxicity, liver toxicity, carcinogenicity, etc.) ;
- Eliminate the risk of the drug being withdrawn from the market;
- Avoid an occurrence of unwanted drug-drug interaction that may lead to unexpected side effects in patients (ex: inhibition of cytochrome P450 enzymes);
- Gather key information on what happens to a drug when it enters a human body (ex: half-life of drug, clearance of drug, etc.)
- Reduce the number of animal trials needed to extract information about the drug compounds.
Let’s get in touch!
At Kantify, we believe the future of life sciences is collaborative. Therefore, we are constantly working on developing state-of-the-art artificial intelligence algorithms to solve some of the most complex problems in modern-day drug discovery and development. If you are interested in collaboration or have a specific project in mind, leave us a message below and our team will get back to you shortly!